fs/bcachefs/btree_io.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_methods.h"
   5 #include "bkey_sort.h"
   6 #include "btree_cache.h"
   7 #include "btree_io.h"
   8 #include "btree_iter.h"
   9 #include "btree_locking.h"
  10 #include "btree_update.h"
  11 #include "btree_update_interior.h"
  12 #include "buckets.h"
  13 #include "checksum.h"
  14 #include "debug.h"
  15 #include "error.h"
  16 #include "extents.h"
  17 #include "io_write.h"
  18 #include "journal_reclaim.h"
  19 #include "journal_seq_blacklist.h"
  20 #include "recovery.h"
  21 #include "super-io.h"
  22 #include "trace.h"
  23
  24 #include <linux/sched/mm.h>
  25
  26 static void bch2_btree_node_header_to_text(struct printbuf *out, struct btree_node *bn)
  27 {
  28         prt_printf(out, "btree=%s l=%u seq %llux\n",
  29                    bch2_btree_id_str(BTREE_NODE_ID(bn)),
  30                    (unsigned) BTREE_NODE_LEVEL(bn), bn->keys.seq);
  31         prt_str(out, "min: ");
  32         bch2_bpos_to_text(out, bn->min_key);
  33         prt_newline(out);
  34         prt_str(out, "max: ");
  35         bch2_bpos_to_text(out, bn->max_key);
  36 }
  37
  38 void bch2_btree_node_io_unlock(struct btree *b)
  39 {
  40         EBUG_ON(!btree_node_write_in_flight(b));
  41
  42         clear_btree_node_write_in_flight_inner(b);
  43         clear_btree_node_write_in_flight(b);
  44         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
  45 }
  46
  47 void bch2_btree_node_io_lock(struct btree *b)
  48 {
  49         wait_on_bit_lock_io(&b->flags, BTREE_NODE_write_in_flight,
  50                             TASK_UNINTERRUPTIBLE);
  51 }
  52
  53 void __bch2_btree_node_wait_on_read(struct btree *b)
  54 {
  55         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
  56                        TASK_UNINTERRUPTIBLE);
  57 }
  58
  59 void __bch2_btree_node_wait_on_write(struct btree *b)
  60 {
  61         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
  62                        TASK_UNINTERRUPTIBLE);
  63 }
  64
  65 void bch2_btree_node_wait_on_read(struct btree *b)
  66 {
  67         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
  68                        TASK_UNINTERRUPTIBLE);
  69 }
  70
  71 void bch2_btree_node_wait_on_write(struct btree *b)
  72 {
  73         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
  74                        TASK_UNINTERRUPTIBLE);
  75 }
  76
  77 static void verify_no_dups(struct btree *b,
  78                            struct bkey_packed *start,
  79                            struct bkey_packed *end)
  80 {
  81 #ifdef CONFIG_BCACHEFS_DEBUG
  82         struct bkey_packed *k, *p;
  83
  84         if (start == end)
  85                 return;
  86
  87         for (p = start, k = bkey_p_next(start);
  88              k != end;
  89              p = k, k = bkey_p_next(k)) {
  90                 struct bkey l = bkey_unpack_key(b, p);
  91                 struct bkey r = bkey_unpack_key(b, k);
  92
  93                 BUG_ON(bpos_ge(l.p, bkey_start_pos(&r)));
  94         }
  95 #endif
  96 }
  97
  98 static void set_needs_whiteout(struct bset *i, int v)
  99 {
 100         struct bkey_packed *k;
 101
 102         for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
 103                 k->needs_whiteout = v;
 104 }
 105
 106 static void btree_bounce_free(struct bch_fs *c, size_t size,
 107                               bool used_mempool, void *p)
 108 {
 109         if (used_mempool)
 110                 mempool_free(p, &c->btree_bounce_pool);
 111         else
 112                 kvfree(p);
 113 }
 114
 115 static void *btree_bounce_alloc(struct bch_fs *c, size_t size,
 116                                 bool *used_mempool)
 117 {
 118         unsigned flags = memalloc_nofs_save();
 119         void *p;
 120
 121         BUG_ON(size > c->opts.btree_node_size);
 122
 123         *used_mempool = false;
 124         p = kvmalloc(size, __GFP_NOWARN|GFP_NOWAIT);
 125         if (!p) {
 126                 *used_mempool = true;
 127                 p = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
 128         }
 129         memalloc_nofs_restore(flags);
 130         return p;
 131 }
 132
 133 static void sort_bkey_ptrs(const struct btree *bt,
 134                            struct bkey_packed **ptrs, unsigned nr)
 135 {
 136         unsigned n = nr, a = nr / 2, b, c, d;
 137
 138         if (!a)
 139                 return;
 140
 141         /* Heap sort: see lib/sort.c: */
 142         while (1) {
 143                 if (a)
 144                         a--;
 145                 else if (--n)
 146                         swap(ptrs[0], ptrs[n]);
 147                 else
 148                         break;
 149
 150                 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
 151                         b = bch2_bkey_cmp_packed(bt,
 152                                             ptrs[c],
 153                                             ptrs[d]) >= 0 ? c : d;
 154                 if (d == n)
 155                         b = c;
 156
 157                 while (b != a &&
 158                        bch2_bkey_cmp_packed(bt,
 159                                        ptrs[a],
 160                                        ptrs[b]) >= 0)
 161                         b = (b - 1) / 2;
 162                 c = b;
 163                 while (b != a) {
 164                         b = (b - 1) / 2;
 165                         swap(ptrs[b], ptrs[c]);
 166                 }
 167         }
 168 }
 169
 170 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
 171 {
 172         struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
 173         bool used_mempool = false;
 174         size_t bytes = b->whiteout_u64s * sizeof(u64);
 175
 176         if (!b->whiteout_u64s)
 177                 return;
 178
 179         new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
 180
 181         ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
 182
 183         for (k = unwritten_whiteouts_start(b);
 184              k != unwritten_whiteouts_end(b);
 185              k = bkey_p_next(k))
 186                 *--ptrs = k;
 187
 188         sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
 189
 190         k = new_whiteouts;
 191
 192         while (ptrs != ptrs_end) {
 193                 bkey_p_copy(k, *ptrs);
 194                 k = bkey_p_next(k);
 195                 ptrs++;
 196         }
 197
 198         verify_no_dups(b, new_whiteouts,
 199                        (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
 200
 201         memcpy_u64s(unwritten_whiteouts_start(b),
 202                     new_whiteouts, b->whiteout_u64s);
 203
 204         btree_bounce_free(c, bytes, used_mempool, new_whiteouts);
 205 }
 206
 207 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
 208                                 bool compacting, enum compact_mode mode)
 209 {
 210         if (!bset_dead_u64s(b, t))
 211                 return false;
 212
 213         switch (mode) {
 214         case COMPACT_LAZY:
 215                 return should_compact_bset_lazy(b, t) ||
 216                         (compacting && !bset_written(b, bset(b, t)));
 217         case COMPACT_ALL:
 218                 return true;
 219         default:
 220                 BUG();
 221         }
 222 }
 223
 224 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
 225 {
 226         bool ret = false;
 227
 228         for_each_bset(b, t) {
 229                 struct bset *i = bset(b, t);
 230                 struct bkey_packed *k, *n, *out, *start, *end;
 231                 struct btree_node_entry *src = NULL, *dst = NULL;
 232
 233                 if (t != b->set && !bset_written(b, i)) {
 234                         src = container_of(i, struct btree_node_entry, keys);
 235                         dst = max(write_block(b),
 236                                   (void *) btree_bkey_last(b, t - 1));
 237                 }
 238
 239                 if (src != dst)
 240                         ret = true;
 241
 242                 if (!should_compact_bset(b, t, ret, mode)) {
 243                         if (src != dst) {
 244                                 memmove(dst, src, sizeof(*src) +
 245                                         le16_to_cpu(src->keys.u64s) *
 246                                         sizeof(u64));
 247                                 i = &dst->keys;
 248                                 set_btree_bset(b, t, i);
 249                         }
 250                         continue;
 251                 }
 252
 253                 start   = btree_bkey_first(b, t);
 254                 end     = btree_bkey_last(b, t);
 255
 256                 if (src != dst) {
 257                         memmove(dst, src, sizeof(*src));
 258                         i = &dst->keys;
 259                         set_btree_bset(b, t, i);
 260                 }
 261
 262                 out = i->start;
 263
 264                 for (k = start; k != end; k = n) {
 265                         n = bkey_p_next(k);
 266
 267                         if (!bkey_deleted(k)) {
 268                                 bkey_p_copy(out, k);
 269                                 out = bkey_p_next(out);
 270                         } else {
 271                                 BUG_ON(k->needs_whiteout);
 272                         }
 273                 }
 274
 275                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 276                 set_btree_bset_end(b, t);
 277                 bch2_bset_set_no_aux_tree(b, t);
 278                 ret = true;
 279         }
 280
 281         bch2_verify_btree_nr_keys(b);
 282
 283         bch2_btree_build_aux_trees(b);
 284
 285         return ret;
 286 }
 287
 288 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 289                             enum compact_mode mode)
 290 {
 291         return bch2_drop_whiteouts(b, mode);
 292 }
 293
 294 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 295                             unsigned start_idx,
 296                             unsigned end_idx)
 297 {
 298         struct btree_node *out;
 299         struct sort_iter_stack sort_iter;
 300         struct bset_tree *t;
 301         struct bset *start_bset = bset(b, &b->set[start_idx]);
 302         bool used_mempool = false;
 303         u64 start_time, seq = 0;
 304         unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
 305         bool sorting_entire_node = start_idx == 0 &&
 306                 end_idx == b->nsets;
 307
 308         sort_iter_stack_init(&sort_iter, b);
 309
 310         for (t = b->set + start_idx;
 311              t < b->set + end_idx;
 312              t++) {
 313                 u64s += le16_to_cpu(bset(b, t)->u64s);
 314                 sort_iter_add(&sort_iter.iter,
 315                               btree_bkey_first(b, t),
 316                               btree_bkey_last(b, t));
 317         }
 318
 319         bytes = sorting_entire_node
 320                 ? btree_buf_bytes(b)
 321                 : __vstruct_bytes(struct btree_node, u64s);
 322
 323         out = btree_bounce_alloc(c, bytes, &used_mempool);
 324
 325         start_time = local_clock();
 326
 327         u64s = bch2_sort_keys(out->keys.start, &sort_iter.iter);
 328
 329         out->keys.u64s = cpu_to_le16(u64s);
 330
 331         BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
 332
 333         if (sorting_entire_node)
 334                 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 335                                        start_time);
 336
 337         /* Make sure we preserve bset journal_seq: */
 338         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 339                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 340         start_bset->journal_seq = cpu_to_le64(seq);
 341
 342         if (sorting_entire_node) {
 343                 u64s = le16_to_cpu(out->keys.u64s);
 344
 345                 BUG_ON(bytes != btree_buf_bytes(b));
 346
 347                 /*
 348                  * Our temporary buffer is the same size as the btree node's
 349                  * buffer, we can just swap buffers instead of doing a big
 350                  * memcpy()
 351                  */
 352                 *out = *b->data;
 353                 out->keys.u64s = cpu_to_le16(u64s);
 354                 swap(out, b->data);
 355                 set_btree_bset(b, b->set, &b->data->keys);
 356         } else {
 357                 start_bset->u64s = out->keys.u64s;
 358                 memcpy_u64s(start_bset->start,
 359                             out->keys.start,
 360                             le16_to_cpu(out->keys.u64s));
 361         }
 362
 363         for (i = start_idx + 1; i < end_idx; i++)
 364                 b->nr.bset_u64s[start_idx] +=
 365                         b->nr.bset_u64s[i];
 366
 367         b->nsets -= shift;
 368
 369         for (i = start_idx + 1; i < b->nsets; i++) {
 370                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 371                 b->set[i]               = b->set[i + shift];
 372         }
 373
 374         for (i = b->nsets; i < MAX_BSETS; i++)
 375                 b->nr.bset_u64s[i] = 0;
 376
 377         set_btree_bset_end(b, &b->set[start_idx]);
 378         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 379
 380         btree_bounce_free(c, bytes, used_mempool, out);
 381
 382         bch2_verify_btree_nr_keys(b);
 383 }
 384
 385 void bch2_btree_sort_into(struct bch_fs *c,
 386                          struct btree *dst,
 387                          struct btree *src)
 388 {
 389         struct btree_nr_keys nr;
 390         struct btree_node_iter src_iter;
 391         u64 start_time = local_clock();
 392
 393         BUG_ON(dst->nsets != 1);
 394
 395         bch2_bset_set_no_aux_tree(dst, dst->set);
 396
 397         bch2_btree_node_iter_init_from_start(&src_iter, src);
 398
 399         nr = bch2_sort_repack(btree_bset_first(dst),
 400                         src, &src_iter,
 401                         &dst->format,
 402                         true);
 403
 404         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 405                                start_time);
 406
 407         set_btree_bset_end(dst, dst->set);
 408
 409         dst->nr.live_u64s       += nr.live_u64s;
 410         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 411         dst->nr.packed_keys     += nr.packed_keys;
 412         dst->nr.unpacked_keys   += nr.unpacked_keys;
 413
 414         bch2_verify_btree_nr_keys(dst);
 415 }
 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, b->nsets);
 434                 ret = true;
 435         }
 436
 437         if (unwritten_idx > 1) {
 438                 btree_node_sort(c, b, 0, unwritten_idx);
 439                 ret = true;
 440         }
 441
 442         return ret;
 443 }
 444
 445 void bch2_btree_build_aux_trees(struct btree *b)
 446 {
 447         for_each_bset(b, t)
 448                 bch2_bset_build_aux_tree(b, t,
 449                                 !bset_written(b, bset(b, t)) &&
 450                                 t == bset_tree_last(b));
 451 }
 452
 453 /*
 454  * If we have MAX_BSETS (3) bsets, should we sort them all down to just one?
 455  *
 456  * The first bset is going to be of similar order to the size of the node, the
 457  * last bset is bounded by btree_write_set_buffer(), which is set to keep the
 458  * memmove on insert from being too expensive: the middle bset should, ideally,
 459  * be the geometric mean of the first and the last.
 460  *
 461  * Returns true if the middle bset is greater than that geometric mean:
 462  */
 463 static inline bool should_compact_all(struct bch_fs *c, struct btree *b)
 464 {
 465         unsigned mid_u64s_bits =
 466                 (ilog2(btree_max_u64s(c)) + BTREE_WRITE_SET_U64s_BITS) / 2;
 467
 468         return bset_u64s(&b->set[1]) > 1U << mid_u64s_bits;
 469 }
 470
 471 /*
 472  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 473  * inserted into
 474  *
 475  * Safe to call if there already is an unwritten bset - will only add a new bset
 476  * if @b doesn't already have one.
 477  *
 478  * Returns true if we sorted (i.e. invalidated iterators
 479  */
 480 void bch2_btree_init_next(struct btree_trans *trans, struct btree *b)
 481 {
 482         struct bch_fs *c = trans->c;
 483         struct btree_node_entry *bne;
 484         bool reinit_iter = false;
 485
 486         EBUG_ON(!six_lock_counts(&b->c.lock).n[SIX_LOCK_write]);
 487         BUG_ON(bset_written(b, bset(b, &b->set[1])));
 488         BUG_ON(btree_node_just_written(b));
 489
 490         if (b->nsets == MAX_BSETS &&
 491             !btree_node_write_in_flight(b) &&
 492             should_compact_all(c, b)) {
 493                 bch2_btree_node_write(c, b, SIX_LOCK_write,
 494                                       BTREE_WRITE_init_next_bset);
 495                 reinit_iter = true;
 496         }
 497
 498         if (b->nsets == MAX_BSETS &&
 499             btree_node_compact(c, b))
 500                 reinit_iter = true;
 501
 502         BUG_ON(b->nsets >= MAX_BSETS);
 503
 504         bne = want_new_bset(c, b);
 505         if (bne)
 506                 bch2_bset_init_next(b, bne);
 507
 508         bch2_btree_build_aux_trees(b);
 509
 510         if (reinit_iter)
 511                 bch2_trans_node_reinit_iter(trans, b);
 512 }
 513
 514 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 515                           struct bch_dev *ca,
 516                           struct btree *b, struct bset *i, struct bkey_packed *k,
 517                           unsigned offset, int write)
 518 {
 519         prt_printf(out, bch2_log_msg(c, "%s"),
 520                    write == READ
 521                    ? "error validating btree node "
 522                    : "corrupt btree node before write ");
 523         if (ca)
 524                 prt_printf(out, "on %s ", ca->name);
 525         prt_printf(out, "at btree ");
 526         bch2_btree_pos_to_text(out, c, b);
 527
 528         printbuf_indent_add(out, 2);
 529
 530         prt_printf(out, "\nnode offset %u/%u",
 531                    b->written, btree_ptr_sectors_written(bkey_i_to_s_c(&b->key)));
 532         if (i)
 533                 prt_printf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 534         if (k)
 535                 prt_printf(out, " bset byte offset %lu",
 536                            (unsigned long)(void *)k -
 537                            ((unsigned long)(void *)i & ~511UL));
 538         prt_str(out, ": ");
 539 }
 540
 541 __printf(10, 11)
 542 static int __btree_err(int ret,
 543                        struct bch_fs *c,
 544                        struct bch_dev *ca,
 545                        struct btree *b,
 546                        struct bset *i,
 547                        struct bkey_packed *k,
 548                        int write,
 549                        bool have_retry,
 550                        enum bch_sb_error_id err_type,
 551                        const char *fmt, ...)
 552 {
 553         struct printbuf out = PRINTBUF;
 554         bool silent = c->curr_recovery_pass == BCH_RECOVERY_PASS_scan_for_btree_nodes;
 555         va_list args;
 556
 557         btree_err_msg(&out, c, ca, b, i, k, b->written, write);
 558
 559         va_start(args, fmt);
 560         prt_vprintf(&out, fmt, args);
 561         va_end(args);
 562
 563         if (write == WRITE) {
 564                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 565                 ret = c->opts.errors == BCH_ON_ERROR_continue
 566                         ? 0
 567                         : -BCH_ERR_fsck_errors_not_fixed;
 568                 goto out;
 569         }
 570
 571         if (!have_retry && ret == -BCH_ERR_btree_node_read_err_want_retry)
 572                 ret = -BCH_ERR_btree_node_read_err_fixable;
 573         if (!have_retry && ret == -BCH_ERR_btree_node_read_err_must_retry)
 574                 ret = -BCH_ERR_btree_node_read_err_bad_node;
 575
 576         if (!silent && ret != -BCH_ERR_btree_node_read_err_fixable)
 577                 bch2_sb_error_count(c, err_type);
 578
 579         switch (ret) {
 580         case -BCH_ERR_btree_node_read_err_fixable:
 581                 ret = !silent
 582                         ? __bch2_fsck_err(c, NULL, FSCK_CAN_FIX, err_type, "%s", out.buf)
 583                         : -BCH_ERR_fsck_fix;
 584                 if (ret != -BCH_ERR_fsck_fix &&
 585                     ret != -BCH_ERR_fsck_ignore)
 586                         goto fsck_err;
 587                 ret = -BCH_ERR_fsck_fix;
 588                 break;
 589         case -BCH_ERR_btree_node_read_err_want_retry:
 590         case -BCH_ERR_btree_node_read_err_must_retry:
 591                 if (!silent)
 592                         bch2_print_string_as_lines(KERN_ERR, out.buf);
 593                 break;
 594         case -BCH_ERR_btree_node_read_err_bad_node:
 595                 if (!silent)
 596                         bch2_print_string_as_lines(KERN_ERR, out.buf);
 597                 ret = bch2_topology_error(c);
 598                 break;
 599         case -BCH_ERR_btree_node_read_err_incompatible:
 600                 if (!silent)
 601                         bch2_print_string_as_lines(KERN_ERR, out.buf);
 602                 ret = -BCH_ERR_fsck_errors_not_fixed;
 603                 break;
 604         default:
 605                 BUG();
 606         }
 607 out:
 608 fsck_err:
 609         printbuf_exit(&out);
 610         return ret;
 611 }
 612
 613 #define btree_err(type, c, ca, b, i, k, _err_type, msg, ...)            \
 614 ({                                                                      \
 615         int _ret = __btree_err(type, c, ca, b, i, k, write, have_retry, \
 616                                BCH_FSCK_ERR_##_err_type,                \
 617                                msg, ##__VA_ARGS__);                     \
 618                                                                         \
 619         if (_ret != -BCH_ERR_fsck_fix) {                                \
 620                 ret = _ret;                                             \
 621                 goto fsck_err;                                          \
 622         }                                                               \
 623                                                                         \
 624         *saw_error = true;                                              \
 625 })
 626
 627 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 628
 629 /*
 630  * When btree topology repair changes the start or end of a node, that might
 631  * mean we have to drop keys that are no longer inside the node:
 632  */
 633 __cold
 634 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
 635 {
 636         for_each_bset(b, t) {
 637                 struct bset *i = bset(b, t);
 638                 struct bkey_packed *k;
 639
 640                 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
 641                         if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
 642                                 break;
 643
 644                 if (k != i->start) {
 645                         unsigned shift = (u64 *) k - (u64 *) i->start;
 646
 647                         memmove_u64s_down(i->start, k,
 648                                           (u64 *) vstruct_end(i) - (u64 *) k);
 649                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
 650                         set_btree_bset_end(b, t);
 651                 }
 652
 653                 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
 654                         if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
 655                                 break;
 656
 657                 if (k != vstruct_last(i)) {
 658                         i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
 659                         set_btree_bset_end(b, t);
 660                 }
 661         }
 662
 663         /*
 664          * Always rebuild search trees: eytzinger search tree nodes directly
 665          * depend on the values of min/max key:
 666          */
 667         bch2_bset_set_no_aux_tree(b, b->set);
 668         bch2_btree_build_aux_trees(b);
 669         b->nr = bch2_btree_node_count_keys(b);
 670
 671         struct bkey_s_c k;
 672         struct bkey unpacked;
 673         struct btree_node_iter iter;
 674         for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
 675                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 676                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 677         }
 678 }
 679
 680 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 681                          struct btree *b, struct bset *i,
 682                          unsigned offset, unsigned sectors,
 683                          int write, bool have_retry, bool *saw_error)
 684 {
 685         unsigned version = le16_to_cpu(i->version);
 686         unsigned ptr_written = btree_ptr_sectors_written(bkey_i_to_s_c(&b->key));
 687         struct printbuf buf1 = PRINTBUF;
 688         struct printbuf buf2 = PRINTBUF;
 689         int ret = 0;
 690
 691         btree_err_on(!bch2_version_compatible(version),
 692                      -BCH_ERR_btree_node_read_err_incompatible,
 693                      c, ca, b, i, NULL,
 694                      btree_node_unsupported_version,
 695                      "unsupported bset version %u.%u",
 696                      BCH_VERSION_MAJOR(version),
 697                      BCH_VERSION_MINOR(version));
 698
 699         if (btree_err_on(version < c->sb.version_min,
 700                          -BCH_ERR_btree_node_read_err_fixable,
 701                          c, NULL, b, i, NULL,
 702                          btree_node_bset_older_than_sb_min,
 703                          "bset version %u older than superblock version_min %u",
 704                          version, c->sb.version_min)) {
 705                 mutex_lock(&c->sb_lock);
 706                 c->disk_sb.sb->version_min = cpu_to_le16(version);
 707                 bch2_write_super(c);
 708                 mutex_unlock(&c->sb_lock);
 709         }
 710
 711         if (btree_err_on(BCH_VERSION_MAJOR(version) >
 712                          BCH_VERSION_MAJOR(c->sb.version),
 713                          -BCH_ERR_btree_node_read_err_fixable,
 714                          c, NULL, b, i, NULL,
 715                          btree_node_bset_newer_than_sb,
 716                          "bset version %u newer than superblock version %u",
 717                          version, c->sb.version)) {
 718                 mutex_lock(&c->sb_lock);
 719                 c->disk_sb.sb->version = cpu_to_le16(version);
 720                 bch2_write_super(c);
 721                 mutex_unlock(&c->sb_lock);
 722         }
 723
 724         btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
 725                      -BCH_ERR_btree_node_read_err_incompatible,
 726                      c, ca, b, i, NULL,
 727                      btree_node_unsupported_version,
 728                      "BSET_SEPARATE_WHITEOUTS no longer supported");
 729
 730         if (!write &&
 731             btree_err_on(offset + sectors > (ptr_written ?: btree_sectors(c)),
 732                          -BCH_ERR_btree_node_read_err_fixable,
 733                          c, ca, b, i, NULL,
 734                          bset_past_end_of_btree_node,
 735                          "bset past end of btree node (offset %u len %u but written %zu)",
 736                          offset, sectors, ptr_written ?: btree_sectors(c)))
 737                 i->u64s = 0;
 738
 739         btree_err_on(offset && !i->u64s,
 740                      -BCH_ERR_btree_node_read_err_fixable,
 741                      c, ca, b, i, NULL,
 742                      bset_empty,
 743                      "empty bset");
 744
 745         btree_err_on(BSET_OFFSET(i) && BSET_OFFSET(i) != offset,
 746                      -BCH_ERR_btree_node_read_err_want_retry,
 747                      c, ca, b, i, NULL,
 748                      bset_wrong_sector_offset,
 749                      "bset at wrong sector offset");
 750
 751         if (!offset) {
 752                 struct btree_node *bn =
 753                         container_of(i, struct btree_node, keys);
 754                 /* These indicate that we read the wrong btree node: */
 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                         /* XXX endianness */
 761                         btree_err_on(bp->seq != bn->keys.seq,
 762                                      -BCH_ERR_btree_node_read_err_must_retry,
 763                                      c, ca, b, NULL, NULL,
 764                                      bset_bad_seq,
 765                                      "incorrect sequence number (wrong btree node)");
 766                 }
 767
 768                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 769                              -BCH_ERR_btree_node_read_err_must_retry,
 770                              c, ca, b, i, NULL,
 771                              btree_node_bad_btree,
 772                              "incorrect btree id");
 773
 774                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 775                              -BCH_ERR_btree_node_read_err_must_retry,
 776                              c, ca, b, i, NULL,
 777                              btree_node_bad_level,
 778                              "incorrect level");
 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                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 785                         struct bch_btree_ptr_v2 *bp =
 786                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 787
 788                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 789                                 b->data->min_key = bp->min_key;
 790                                 b->data->max_key = b->key.k.p;
 791                         }
 792
 793                         btree_err_on(!bpos_eq(b->data->min_key, bp->min_key),
 794                                      -BCH_ERR_btree_node_read_err_must_retry,
 795                                      c, ca, b, NULL, NULL,
 796                                      btree_node_bad_min_key,
 797                                      "incorrect min_key: got %s should be %s",
 798                                      (printbuf_reset(&buf1),
 799                                       bch2_bpos_to_text(&buf1, bn->min_key), buf1.buf),
 800                                      (printbuf_reset(&buf2),
 801                                       bch2_bpos_to_text(&buf2, bp->min_key), buf2.buf));
 802                 }
 803
 804                 btree_err_on(!bpos_eq(bn->max_key, b->key.k.p),
 805                              -BCH_ERR_btree_node_read_err_must_retry,
 806                              c, ca, b, i, NULL,
 807                              btree_node_bad_max_key,
 808                              "incorrect max key %s",
 809                              (printbuf_reset(&buf1),
 810                               bch2_bpos_to_text(&buf1, bn->max_key), buf1.buf));
 811
 812                 if (write)
 813                         compat_btree_node(b->c.level, b->c.btree_id, version,
 814                                           BSET_BIG_ENDIAN(i), write, bn);
 815
 816                 btree_err_on(bch2_bkey_format_invalid(c, &bn->format, write, &buf1),
 817                              -BCH_ERR_btree_node_read_err_bad_node,
 818                              c, ca, b, i, NULL,
 819                              btree_node_bad_format,
 820                              "invalid bkey format: %s\n  %s", buf1.buf,
 821                              (printbuf_reset(&buf2),
 822                               bch2_bkey_format_to_text(&buf2, &bn->format), buf2.buf));
 823                 printbuf_reset(&buf1);
 824
 825                 compat_bformat(b->c.level, b->c.btree_id, version,
 826                                BSET_BIG_ENDIAN(i), write,
 827                                &bn->format);
 828         }
 829 fsck_err:
 830         printbuf_exit(&buf2);
 831         printbuf_exit(&buf1);
 832         return ret;
 833 }
 834
 835 static int bset_key_validate(struct bch_fs *c, struct btree *b,
 836                              struct bkey_s_c k,
 837                              bool updated_range, int rw)
 838 {
 839         return __bch2_bkey_validate(c, k, btree_node_type(b), 0) ?:
 840                 (!updated_range ? bch2_bkey_in_btree_node(c, b, k, 0) : 0) ?:
 841                 (rw == WRITE ? bch2_bkey_val_validate(c, k, 0) : 0);
 842 }
 843
 844 static bool bkey_packed_valid(struct bch_fs *c, struct btree *b,
 845                          struct bset *i, struct bkey_packed *k)
 846 {
 847         if (bkey_p_next(k) > vstruct_last(i))
 848                 return false;
 849
 850         if (k->format > KEY_FORMAT_CURRENT)
 851                 return false;
 852
 853         if (!bkeyp_u64s_valid(&b->format, k))
 854                 return false;
 855
 856         struct bkey tmp;
 857         struct bkey_s u = __bkey_disassemble(b, k, &tmp);
 858         return !__bch2_bkey_validate(c, u.s_c, btree_node_type(b), BCH_VALIDATE_silent);
 859 }
 860
 861 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 862                          struct bset *i, int write,
 863                          bool have_retry, bool *saw_error)
 864 {
 865         unsigned version = le16_to_cpu(i->version);
 866         struct bkey_packed *k, *prev = NULL;
 867         struct printbuf buf = PRINTBUF;
 868         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 869                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 870         int ret = 0;
 871
 872         for (k = i->start;
 873              k != vstruct_last(i);) {
 874                 struct bkey_s u;
 875                 struct bkey tmp;
 876                 unsigned next_good_key;
 877
 878                 if (btree_err_on(bkey_p_next(k) > vstruct_last(i),
 879                                  -BCH_ERR_btree_node_read_err_fixable,
 880                                  c, NULL, b, i, k,
 881                                  btree_node_bkey_past_bset_end,
 882                                  "key extends past end of bset")) {
 883                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 884                         break;
 885                 }
 886
 887                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 888                                  -BCH_ERR_btree_node_read_err_fixable,
 889                                  c, NULL, b, i, k,
 890                                  btree_node_bkey_bad_format,
 891                                  "invalid bkey format %u", k->format))
 892                         goto drop_this_key;
 893
 894                 if (btree_err_on(!bkeyp_u64s_valid(&b->format, k),
 895                                  -BCH_ERR_btree_node_read_err_fixable,
 896                                  c, NULL, b, i, k,
 897                                  btree_node_bkey_bad_u64s,
 898                                  "bad k->u64s %u (min %u max %zu)", k->u64s,
 899                                  bkeyp_key_u64s(&b->format, k),
 900                                  U8_MAX - BKEY_U64s + bkeyp_key_u64s(&b->format, k)))
 901                         goto drop_this_key;
 902
 903                 if (!write)
 904                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 905                                     BSET_BIG_ENDIAN(i), write,
 906                                     &b->format, k);
 907
 908                 u = __bkey_disassemble(b, k, &tmp);
 909
 910                 ret = bset_key_validate(c, b, u.s_c, updated_range, write);
 911                 if (ret == -BCH_ERR_fsck_delete_bkey)
 912                         goto drop_this_key;
 913                 if (ret)
 914                         goto fsck_err;
 915
 916                 if (write)
 917                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 918                                     BSET_BIG_ENDIAN(i), write,
 919                                     &b->format, k);
 920
 921                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 922                         struct bkey up = bkey_unpack_key(b, prev);
 923
 924                         printbuf_reset(&buf);
 925                         prt_printf(&buf, "keys out of order: ");
 926                         bch2_bkey_to_text(&buf, &up);
 927                         prt_printf(&buf, " > ");
 928                         bch2_bkey_to_text(&buf, u.k);
 929
 930                         if (btree_err(-BCH_ERR_btree_node_read_err_fixable,
 931                                       c, NULL, b, i, k,
 932                                       btree_node_bkey_out_of_order,
 933                                       "%s", buf.buf))
 934                                 goto drop_this_key;
 935                 }
 936
 937                 prev = k;
 938                 k = bkey_p_next(k);
 939                 continue;
 940 drop_this_key:
 941                 next_good_key = k->u64s;
 942
 943                 if (!next_good_key ||
 944                     (BSET_BIG_ENDIAN(i) == CPU_BIG_ENDIAN &&
 945                      version >= bcachefs_metadata_version_snapshot)) {
 946                         /*
 947                          * only do scanning if bch2_bkey_compat() has nothing to
 948                          * do
 949                          */
 950
 951                         if (!bkey_packed_valid(c, b, i, (void *) ((u64 *) k + next_good_key))) {
 952                                 for (next_good_key = 1;
 953                                      next_good_key < (u64 *) vstruct_last(i) - (u64 *) k;
 954                                      next_good_key++)
 955                                         if (bkey_packed_valid(c, b, i, (void *) ((u64 *) k + next_good_key)))
 956                                                 goto got_good_key;
 957                         }
 958
 959                         /*
 960                          * didn't find a good key, have to truncate the rest of
 961                          * the bset
 962                          */
 963                         next_good_key = (u64 *) vstruct_last(i) - (u64 *) k;
 964                 }
 965 got_good_key:
 966                 le16_add_cpu(&i->u64s, -next_good_key);
 967                 memmove_u64s_down(k, bkey_p_next(k), (u64 *) vstruct_end(i) - (u64 *) k);
 968         }
 969 fsck_err:
 970         printbuf_exit(&buf);
 971         return ret;
 972 }
 973
 974 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 975                               struct btree *b, bool have_retry, bool *saw_error)
 976 {
 977         struct btree_node_entry *bne;
 978         struct sort_iter *iter;
 979         struct btree_node *sorted;
 980         struct bkey_packed *k;
 981         struct bset *i;
 982         bool used_mempool, blacklisted;
 983         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 984                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 985         unsigned u64s;
 986         unsigned ptr_written = btree_ptr_sectors_written(bkey_i_to_s_c(&b->key));
 987         u64 max_journal_seq = 0;
 988         struct printbuf buf = PRINTBUF;
 989         int ret = 0, retry_read = 0, write = READ;
 990         u64 start_time = local_clock();
 991
 992         b->version_ondisk = U16_MAX;
 993         /* We might get called multiple times on read retry: */
 994         b->written = 0;
 995
 996         iter = mempool_alloc(&c->fill_iter, GFP_NOFS);
 997         sort_iter_init(iter, b, (btree_blocks(c) + 1) * 2);
 998
 999         if (bch2_meta_read_fault("btree"))
1000                 btree_err(-BCH_ERR_btree_node_read_err_must_retry,
1001                           c, ca, b, NULL, NULL,
1002                           btree_node_fault_injected,
1003                           "dynamic fault");
1004
1005         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
1006                      -BCH_ERR_btree_node_read_err_must_retry,
1007                      c, ca, b, NULL, NULL,
1008                      btree_node_bad_magic,
1009                      "bad magic: want %llx, got %llx",
1010                      bset_magic(c), le64_to_cpu(b->data->magic));
1011
1012         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
1013                 struct bch_btree_ptr_v2 *bp =
1014                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
1015
1016                 bch2_bpos_to_text(&buf, b->data->min_key);
1017                 prt_str(&buf, "-");
1018                 bch2_bpos_to_text(&buf, b->data->max_key);
1019
1020                 btree_err_on(b->data->keys.seq != bp->seq,
1021                              -BCH_ERR_btree_node_read_err_must_retry,
1022                              c, ca, b, NULL, NULL,
1023                              btree_node_bad_seq,
1024                              "got wrong btree node: got\n%s",
1025                              (printbuf_reset(&buf),
1026                               bch2_btree_node_header_to_text(&buf, b->data),
1027                               buf.buf));
1028         } else {
1029                 btree_err_on(!b->data->keys.seq,
1030                              -BCH_ERR_btree_node_read_err_must_retry,
1031                              c, ca, b, NULL, NULL,
1032                              btree_node_bad_seq,
1033                              "bad btree header: seq 0\n%s",
1034                              (printbuf_reset(&buf),
1035                               bch2_btree_node_header_to_text(&buf, b->data),
1036                               buf.buf));
1037         }
1038
1039         while (b->written < (ptr_written ?: btree_sectors(c))) {
1040                 unsigned sectors;
1041                 struct nonce nonce;
1042                 bool first = !b->written;
1043                 bool csum_bad;
1044
1045                 if (!b->written) {
1046                         i = &b->data->keys;
1047
1048                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1049                                      -BCH_ERR_btree_node_read_err_want_retry,
1050                                      c, ca, b, i, NULL,
1051                                      bset_unknown_csum,
1052                                      "unknown checksum type %llu", BSET_CSUM_TYPE(i));
1053
1054                         nonce = btree_nonce(i, b->written << 9);
1055
1056                         struct bch_csum csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
1057                         csum_bad = bch2_crc_cmp(b->data->csum, csum);
1058                         if (csum_bad)
1059                                 bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
1060
1061                         btree_err_on(csum_bad,
1062                                      -BCH_ERR_btree_node_read_err_want_retry,
1063                                      c, ca, b, i, NULL,
1064                                      bset_bad_csum,
1065                                      "%s",
1066                                      (printbuf_reset(&buf),
1067                                       bch2_csum_err_msg(&buf, BSET_CSUM_TYPE(i), b->data->csum, csum),
1068                                       buf.buf));
1069
1070                         ret = bset_encrypt(c, i, b->written << 9);
1071                         if (bch2_fs_fatal_err_on(ret, c,
1072                                         "decrypting btree node: %s", bch2_err_str(ret)))
1073                                 goto fsck_err;
1074
1075                         btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
1076                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
1077                                      -BCH_ERR_btree_node_read_err_incompatible,
1078                                      c, NULL, b, NULL, NULL,
1079                                      btree_node_unsupported_version,
1080                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
1081
1082                         sectors = vstruct_sectors(b->data, c->block_bits);
1083                 } else {
1084                         bne = write_block(b);
1085                         i = &bne->keys;
1086
1087                         if (i->seq != b->data->keys.seq)
1088                                 break;
1089
1090                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1091                                      -BCH_ERR_btree_node_read_err_want_retry,
1092                                      c, ca, b, i, NULL,
1093                                      bset_unknown_csum,
1094                                      "unknown checksum type %llu", BSET_CSUM_TYPE(i));
1095
1096                         nonce = btree_nonce(i, b->written << 9);
1097                         struct bch_csum csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1098                         csum_bad = bch2_crc_cmp(bne->csum, csum);
1099                         if (ca && csum_bad)
1100                                 bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
1101
1102                         btree_err_on(csum_bad,
1103                                      -BCH_ERR_btree_node_read_err_want_retry,
1104                                      c, ca, b, i, NULL,
1105                                      bset_bad_csum,
1106                                      "%s",
1107                                      (printbuf_reset(&buf),
1108                                       bch2_csum_err_msg(&buf, BSET_CSUM_TYPE(i), bne->csum, csum),
1109                                       buf.buf));
1110
1111                         ret = bset_encrypt(c, i, b->written << 9);
1112                         if (bch2_fs_fatal_err_on(ret, c,
1113                                         "decrypting btree node: %s", bch2_err_str(ret)))
1114                                 goto fsck_err;
1115
1116                         sectors = vstruct_sectors(bne, c->block_bits);
1117                 }
1118
1119                 b->version_ondisk = min(b->version_ondisk,
1120                                         le16_to_cpu(i->version));
1121
1122                 ret = validate_bset(c, ca, b, i, b->written, sectors,
1123                                     READ, have_retry, saw_error);
1124                 if (ret)
1125                         goto fsck_err;
1126
1127                 if (!b->written)
1128                         btree_node_set_format(b, b->data->format);
1129
1130                 ret = validate_bset_keys(c, b, i, READ, have_retry, saw_error);
1131                 if (ret)
1132                         goto fsck_err;
1133
1134                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1135
1136                 blacklisted = bch2_journal_seq_is_blacklisted(c,
1137                                         le64_to_cpu(i->journal_seq),
1138                                         true);
1139
1140                 btree_err_on(blacklisted && first,
1141                              -BCH_ERR_btree_node_read_err_fixable,
1142                              c, ca, b, i, NULL,
1143                              bset_blacklisted_journal_seq,
1144                              "first btree node bset has blacklisted journal seq (%llu)",
1145                              le64_to_cpu(i->journal_seq));
1146
1147                 btree_err_on(blacklisted && ptr_written,
1148                              -BCH_ERR_btree_node_read_err_fixable,
1149                              c, ca, b, i, NULL,
1150                              first_bset_blacklisted_journal_seq,
1151                              "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
1152                              le64_to_cpu(i->journal_seq),
1153                              b->written, b->written + sectors, ptr_written);
1154
1155                 b->written += sectors;
1156
1157                 if (blacklisted && !first)
1158                         continue;
1159
1160                 sort_iter_add(iter,
1161                               vstruct_idx(i, 0),
1162                               vstruct_last(i));
1163
1164                 max_journal_seq = max(max_journal_seq, le64_to_cpu(i->journal_seq));
1165         }
1166
1167         if (ptr_written) {
1168                 btree_err_on(b->written < ptr_written,
1169                              -BCH_ERR_btree_node_read_err_want_retry,
1170                              c, ca, b, NULL, NULL,
1171                              btree_node_data_missing,
1172                              "btree node data missing: expected %u sectors, found %u",
1173                              ptr_written, b->written);
1174         } else {
1175                 for (bne = write_block(b);
1176                      bset_byte_offset(b, bne) < btree_buf_bytes(b);
1177                      bne = (void *) bne + block_bytes(c))
1178                         btree_err_on(bne->keys.seq == b->data->keys.seq &&
1179                                      !bch2_journal_seq_is_blacklisted(c,
1180                                                                       le64_to_cpu(bne->keys.journal_seq),
1181                                                                       true),
1182                                      -BCH_ERR_btree_node_read_err_want_retry,
1183                                      c, ca, b, NULL, NULL,
1184                                      btree_node_bset_after_end,
1185                                      "found bset signature after last bset");
1186         }
1187
1188         sorted = btree_bounce_alloc(c, btree_buf_bytes(b), &used_mempool);
1189         sorted->keys.u64s = 0;
1190
1191         set_btree_bset(b, b->set, &b->data->keys);
1192
1193         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1194         memset((uint8_t *)(sorted + 1) + b->nr.live_u64s * sizeof(u64), 0,
1195                         btree_buf_bytes(b) -
1196                         sizeof(struct btree_node) -
1197                         b->nr.live_u64s * sizeof(u64));
1198
1199         u64s = le16_to_cpu(sorted->keys.u64s);
1200         *sorted = *b->data;
1201         sorted->keys.u64s = cpu_to_le16(u64s);
1202         swap(sorted, b->data);
1203         set_btree_bset(b, b->set, &b->data->keys);
1204         b->nsets = 1;
1205         b->data->keys.journal_seq = cpu_to_le64(max_journal_seq);
1206
1207         BUG_ON(b->nr.live_u64s != u64s);
1208
1209         btree_bounce_free(c, btree_buf_bytes(b), used_mempool, sorted);
1210
1211         if (updated_range)
1212                 bch2_btree_node_drop_keys_outside_node(b);
1213
1214         i = &b->data->keys;
1215         for (k = i->start; k != vstruct_last(i);) {
1216                 struct bkey tmp;
1217                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1218
1219                 ret = bch2_bkey_val_validate(c, u.s_c, READ);
1220                 if (ret == -BCH_ERR_fsck_delete_bkey ||
1221                     (bch2_inject_invalid_keys &&
1222                      !bversion_cmp(u.k->bversion, MAX_VERSION))) {
1223                         btree_keys_account_key_drop(&b->nr, 0, k);
1224
1225                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1226                         memmove_u64s_down(k, bkey_p_next(k),
1227                                           (u64 *) vstruct_end(i) - (u64 *) k);
1228                         set_btree_bset_end(b, b->set);
1229                         continue;
1230                 }
1231                 if (ret)
1232                         goto fsck_err;
1233
1234                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1235                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1236
1237                         bp.v->mem_ptr = 0;
1238                 }
1239
1240                 k = bkey_p_next(k);
1241         }
1242
1243         bch2_bset_build_aux_tree(b, b->set, false);
1244
1245         set_needs_whiteout(btree_bset_first(b), true);
1246
1247         btree_node_reset_sib_u64s(b);
1248
1249         rcu_read_lock();
1250         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1251                 struct bch_dev *ca2 = bch2_dev_rcu(c, ptr->dev);
1252
1253                 if (!ca2 || ca2->mi.state != BCH_MEMBER_STATE_rw)
1254                         set_btree_node_need_rewrite(b);
1255         }
1256         rcu_read_unlock();
1257
1258         if (!ptr_written)
1259                 set_btree_node_need_rewrite(b);
1260 out:
1261         mempool_free(iter, &c->fill_iter);
1262         printbuf_exit(&buf);
1263         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read_done], start_time);
1264         return retry_read;
1265 fsck_err:
1266         if (ret == -BCH_ERR_btree_node_read_err_want_retry ||
1267             ret == -BCH_ERR_btree_node_read_err_must_retry) {
1268                 retry_read = 1;
1269         } else {
1270                 set_btree_node_read_error(b);
1271                 bch2_btree_lost_data(c, b->c.btree_id);
1272         }
1273         goto out;
1274 }
1275
1276 static void btree_node_read_work(struct work_struct *work)
1277 {
1278         struct btree_read_bio *rb =
1279                 container_of(work, struct btree_read_bio, work);
1280         struct bch_fs *c        = rb->c;
1281         struct bch_dev *ca      = rb->have_ioref ? bch2_dev_have_ref(c, rb->pick.ptr.dev) : NULL;
1282         struct btree *b         = rb->b;
1283         struct bio *bio         = &rb->bio;
1284         struct bch_io_failures failed = { .nr = 0 };
1285         struct printbuf buf = PRINTBUF;
1286         bool saw_error = false;
1287         bool retry = false;
1288         bool can_retry;
1289
1290         goto start;
1291         while (1) {
1292                 retry = true;
1293                 bch_info(c, "retrying read");
1294                 ca = bch2_dev_get_ioref(c, rb->pick.ptr.dev, READ);
1295                 rb->have_ioref          = ca != NULL;
1296                 bio_reset(bio, NULL, REQ_OP_READ|REQ_SYNC|REQ_META);
1297                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1298                 bio->bi_iter.bi_size    = btree_buf_bytes(b);
1299
1300                 if (rb->have_ioref) {
1301                         bio_set_dev(bio, ca->disk_sb.bdev);
1302                         submit_bio_wait(bio);
1303                 } else {
1304                         bio->bi_status = BLK_STS_REMOVED;
1305                 }
1306 start:
1307                 printbuf_reset(&buf);
1308                 bch2_btree_pos_to_text(&buf, c, b);
1309                 bch2_dev_io_err_on(ca && bio->bi_status, ca, BCH_MEMBER_ERROR_read,
1310                                    "btree read error %s for %s",
1311                                    bch2_blk_status_to_str(bio->bi_status), buf.buf);
1312                 if (rb->have_ioref)
1313                         percpu_ref_put(&ca->io_ref);
1314                 rb->have_ioref = false;
1315
1316                 bch2_mark_io_failure(&failed, &rb->pick);
1317
1318                 can_retry = bch2_bkey_pick_read_device(c,
1319                                 bkey_i_to_s_c(&b->key),
1320                                 &failed, &rb->pick) > 0;
1321
1322                 if (!bio->bi_status &&
1323                     !bch2_btree_node_read_done(c, ca, b, can_retry, &saw_error)) {
1324                         if (retry)
1325                                 bch_info(c, "retry success");
1326                         break;
1327                 }
1328
1329                 saw_error = true;
1330
1331                 if (!can_retry) {
1332                         set_btree_node_read_error(b);
1333                         bch2_btree_lost_data(c, b->c.btree_id);
1334                         break;
1335                 }
1336         }
1337
1338         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1339                                rb->start_time);
1340         bio_put(&rb->bio);
1341
1342         if (saw_error &&
1343             !btree_node_read_error(b) &&
1344             c->curr_recovery_pass != BCH_RECOVERY_PASS_scan_for_btree_nodes) {
1345                 printbuf_reset(&buf);
1346                 bch2_bpos_to_text(&buf, b->key.k.p);
1347                 bch_err_ratelimited(c, "%s: rewriting btree node at btree=%s level=%u %s due to error",
1348                          __func__, bch2_btree_id_str(b->c.btree_id), b->c.level, buf.buf);
1349
1350                 bch2_btree_node_rewrite_async(c, b);
1351         }
1352
1353         printbuf_exit(&buf);
1354         clear_btree_node_read_in_flight(b);
1355         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1356 }
1357
1358 static void btree_node_read_endio(struct bio *bio)
1359 {
1360         struct btree_read_bio *rb =
1361                 container_of(bio, struct btree_read_bio, bio);
1362         struct bch_fs *c        = rb->c;
1363
1364         if (rb->have_ioref) {
1365                 struct bch_dev *ca = bch2_dev_have_ref(c, rb->pick.ptr.dev);
1366
1367                 bch2_latency_acct(ca, rb->start_time, READ);
1368         }
1369
1370         queue_work(c->btree_read_complete_wq, &rb->work);
1371 }
1372
1373 struct btree_node_read_all {
1374         struct closure          cl;
1375         struct bch_fs           *c;
1376         struct btree            *b;
1377         unsigned                nr;
1378         void                    *buf[BCH_REPLICAS_MAX];
1379         struct bio              *bio[BCH_REPLICAS_MAX];
1380         blk_status_t            err[BCH_REPLICAS_MAX];
1381 };
1382
1383 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1384 {
1385         struct btree_node *bn = data;
1386         struct btree_node_entry *bne;
1387         unsigned offset = 0;
1388
1389         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1390                 return 0;
1391
1392         while (offset < btree_sectors(c)) {
1393                 if (!offset) {
1394                         offset += vstruct_sectors(bn, c->block_bits);
1395                 } else {
1396                         bne = data + (offset << 9);
1397                         if (bne->keys.seq != bn->keys.seq)
1398                                 break;
1399                         offset += vstruct_sectors(bne, c->block_bits);
1400                 }
1401         }
1402
1403         return offset;
1404 }
1405
1406 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1407 {
1408         struct btree_node *bn = data;
1409         struct btree_node_entry *bne;
1410
1411         if (!offset)
1412                 return false;
1413
1414         while (offset < btree_sectors(c)) {
1415                 bne = data + (offset << 9);
1416                 if (bne->keys.seq == bn->keys.seq)
1417                         return true;
1418                 offset++;
1419         }
1420
1421         return false;
1422         return offset;
1423 }
1424
1425 static CLOSURE_CALLBACK(btree_node_read_all_replicas_done)
1426 {
1427         closure_type(ra, struct btree_node_read_all, cl);
1428         struct bch_fs *c = ra->c;
1429         struct btree *b = ra->b;
1430         struct printbuf buf = PRINTBUF;
1431         bool dump_bset_maps = false;
1432         bool have_retry = false;
1433         int ret = 0, best = -1, write = READ;
1434         unsigned i, written = 0, written2 = 0;
1435         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1436                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1437         bool _saw_error = false, *saw_error = &_saw_error;
1438
1439         for (i = 0; i < ra->nr; i++) {
1440                 struct btree_node *bn = ra->buf[i];
1441
1442                 if (ra->err[i])
1443                         continue;
1444
1445                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1446                     (seq && seq != bn->keys.seq))
1447                         continue;
1448
1449                 if (best < 0) {
1450                         best = i;
1451                         written = btree_node_sectors_written(c, bn);
1452                         continue;
1453                 }
1454
1455                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1456                 if (btree_err_on(written2 != written, -BCH_ERR_btree_node_read_err_fixable,
1457                                  c, NULL, b, NULL, NULL,
1458                                  btree_node_replicas_sectors_written_mismatch,
1459                                  "btree node sectors written mismatch: %u != %u",
1460                                  written, written2) ||
1461                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1462                                  -BCH_ERR_btree_node_read_err_fixable,
1463                                  c, NULL, b, NULL, NULL,
1464                                  btree_node_bset_after_end,
1465                                  "found bset signature after last bset") ||
1466                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1467                                  -BCH_ERR_btree_node_read_err_fixable,
1468                                  c, NULL, b, NULL, NULL,
1469                                  btree_node_replicas_data_mismatch,
1470                                  "btree node replicas content mismatch"))
1471                         dump_bset_maps = true;
1472
1473                 if (written2 > written) {
1474                         written = written2;
1475                         best = i;
1476                 }
1477         }
1478 fsck_err:
1479         if (dump_bset_maps) {
1480                 for (i = 0; i < ra->nr; i++) {
1481                         struct btree_node *bn = ra->buf[i];
1482                         struct btree_node_entry *bne = NULL;
1483                         unsigned offset = 0, sectors;
1484                         bool gap = false;
1485
1486                         if (ra->err[i])
1487                                 continue;
1488
1489                         printbuf_reset(&buf);
1490
1491                         while (offset < btree_sectors(c)) {
1492                                 if (!offset) {
1493                                         sectors = vstruct_sectors(bn, c->block_bits);
1494                                 } else {
1495                                         bne = ra->buf[i] + (offset << 9);
1496                                         if (bne->keys.seq != bn->keys.seq)
1497                                                 break;
1498                                         sectors = vstruct_sectors(bne, c->block_bits);
1499                                 }
1500
1501                                 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1502                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1503                                                         le64_to_cpu(bne->keys.journal_seq), false))
1504                                         prt_printf(&buf, "*");
1505                                 offset += sectors;
1506                         }
1507
1508                         while (offset < btree_sectors(c)) {
1509                                 bne = ra->buf[i] + (offset << 9);
1510                                 if (bne->keys.seq == bn->keys.seq) {
1511                                         if (!gap)
1512                                                 prt_printf(&buf, " GAP");
1513                                         gap = true;
1514
1515                                         sectors = vstruct_sectors(bne, c->block_bits);
1516                                         prt_printf(&buf, " %u-%u", offset, offset + sectors);
1517                                         if (bch2_journal_seq_is_blacklisted(c,
1518                                                         le64_to_cpu(bne->keys.journal_seq), false))
1519                                                 prt_printf(&buf, "*");
1520                                 }
1521                                 offset++;
1522                         }
1523
1524                         bch_err(c, "replica %u:%s", i, buf.buf);
1525                 }
1526         }
1527
1528         if (best >= 0) {
1529                 memcpy(b->data, ra->buf[best], btree_buf_bytes(b));
1530                 ret = bch2_btree_node_read_done(c, NULL, b, false, saw_error);
1531         } else {
1532                 ret = -1;
1533         }
1534
1535         if (ret) {
1536                 set_btree_node_read_error(b);
1537                 bch2_btree_lost_data(c, b->c.btree_id);
1538         } else if (*saw_error)
1539                 bch2_btree_node_rewrite_async(c, b);
1540
1541         for (i = 0; i < ra->nr; i++) {
1542                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1543                 bio_put(ra->bio[i]);
1544         }
1545
1546         closure_debug_destroy(&ra->cl);
1547         kfree(ra);
1548         printbuf_exit(&buf);
1549
1550         clear_btree_node_read_in_flight(b);
1551         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1552 }
1553
1554 static void btree_node_read_all_replicas_endio(struct bio *bio)
1555 {
1556         struct btree_read_bio *rb =
1557                 container_of(bio, struct btree_read_bio, bio);
1558         struct bch_fs *c        = rb->c;
1559         struct btree_node_read_all *ra = rb->ra;
1560
1561         if (rb->have_ioref) {
1562                 struct bch_dev *ca = bch2_dev_have_ref(c, rb->pick.ptr.dev);
1563
1564                 bch2_latency_acct(ca, rb->start_time, READ);
1565         }
1566
1567         ra->err[rb->idx] = bio->bi_status;
1568         closure_put(&ra->cl);
1569 }
1570
1571 /*
1572  * XXX This allocates multiple times from the same mempools, and can deadlock
1573  * under sufficient memory pressure (but is only a debug path)
1574  */
1575 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1576 {
1577         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1578         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1579         const union bch_extent_entry *entry;
1580         struct extent_ptr_decoded pick;
1581         struct btree_node_read_all *ra;
1582         unsigned i;
1583
1584         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1585         if (!ra)
1586                 return -BCH_ERR_ENOMEM_btree_node_read_all_replicas;
1587
1588         closure_init(&ra->cl, NULL);
1589         ra->c   = c;
1590         ra->b   = b;
1591         ra->nr  = bch2_bkey_nr_ptrs(k);
1592
1593         for (i = 0; i < ra->nr; i++) {
1594                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1595                 ra->bio[i] = bio_alloc_bioset(NULL,
1596                                               buf_pages(ra->buf[i], btree_buf_bytes(b)),
1597                                               REQ_OP_READ|REQ_SYNC|REQ_META,
1598                                               GFP_NOFS,
1599                                               &c->btree_bio);
1600         }
1601
1602         i = 0;
1603         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1604                 struct bch_dev *ca = bch2_dev_get_ioref(c, pick.ptr.dev, READ);
1605                 struct btree_read_bio *rb =
1606                         container_of(ra->bio[i], struct btree_read_bio, bio);
1607                 rb->c                   = c;
1608                 rb->b                   = b;
1609                 rb->ra                  = ra;
1610                 rb->start_time          = local_clock();
1611                 rb->have_ioref          = ca != NULL;
1612                 rb->idx                 = i;
1613                 rb->pick                = pick;
1614                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1615                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1616                 bch2_bio_map(&rb->bio, ra->buf[i], btree_buf_bytes(b));
1617
1618                 if (rb->have_ioref) {
1619                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1620                                      bio_sectors(&rb->bio));
1621                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1622
1623                         closure_get(&ra->cl);
1624                         submit_bio(&rb->bio);
1625                 } else {
1626                         ra->err[i] = BLK_STS_REMOVED;
1627                 }
1628
1629                 i++;
1630         }
1631
1632         if (sync) {
1633                 closure_sync(&ra->cl);
1634                 btree_node_read_all_replicas_done(&ra->cl.work);
1635         } else {
1636                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1637                             c->btree_read_complete_wq);
1638         }
1639
1640         return 0;
1641 }
1642
1643 void bch2_btree_node_read(struct btree_trans *trans, struct btree *b,
1644                           bool sync)
1645 {
1646         struct bch_fs *c = trans->c;
1647         struct extent_ptr_decoded pick;
1648         struct btree_read_bio *rb;
1649         struct bch_dev *ca;
1650         struct bio *bio;
1651         int ret;
1652
1653         trace_and_count(c, btree_node_read, trans, b);
1654
1655         if (bch2_verify_all_btree_replicas &&
1656             !btree_node_read_all_replicas(c, b, sync))
1657                 return;
1658
1659         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1660                                          NULL, &pick);
1661
1662         if (ret <= 0) {
1663                 struct printbuf buf = PRINTBUF;
1664
1665                 prt_str(&buf, "btree node read error: no device to read from\n at ");
1666                 bch2_btree_pos_to_text(&buf, c, b);
1667                 bch_err_ratelimited(c, "%s", buf.buf);
1668
1669                 if (c->opts.recovery_passes & BIT_ULL(BCH_RECOVERY_PASS_check_topology) &&
1670                     c->curr_recovery_pass > BCH_RECOVERY_PASS_check_topology)
1671                         bch2_fatal_error(c);
1672
1673                 set_btree_node_read_error(b);
1674                 bch2_btree_lost_data(c, b->c.btree_id);
1675                 clear_btree_node_read_in_flight(b);
1676                 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1677                 printbuf_exit(&buf);
1678                 return;
1679         }
1680
1681         ca = bch2_dev_get_ioref(c, pick.ptr.dev, READ);
1682
1683         bio = bio_alloc_bioset(NULL,
1684                                buf_pages(b->data, btree_buf_bytes(b)),
1685                                REQ_OP_READ|REQ_SYNC|REQ_META,
1686                                GFP_NOFS,
1687                                &c->btree_bio);
1688         rb = container_of(bio, struct btree_read_bio, bio);
1689         rb->c                   = c;
1690         rb->b                   = b;
1691         rb->ra                  = NULL;
1692         rb->start_time          = local_clock();
1693         rb->have_ioref          = ca != NULL;
1694         rb->pick                = pick;
1695         INIT_WORK(&rb->work, btree_node_read_work);
1696         bio->bi_iter.bi_sector  = pick.ptr.offset;
1697         bio->bi_end_io          = btree_node_read_endio;
1698         bch2_bio_map(bio, b->data, btree_buf_bytes(b));
1699
1700         if (rb->have_ioref) {
1701                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1702                              bio_sectors(bio));
1703                 bio_set_dev(bio, ca->disk_sb.bdev);
1704
1705                 if (sync) {
1706                         submit_bio_wait(bio);
1707                         bch2_latency_acct(ca, rb->start_time, READ);
1708                         btree_node_read_work(&rb->work);
1709                 } else {
1710                         submit_bio(bio);
1711                 }
1712         } else {
1713                 bio->bi_status = BLK_STS_REMOVED;
1714
1715                 if (sync)
1716                         btree_node_read_work(&rb->work);
1717                 else
1718                         queue_work(c->btree_read_complete_wq, &rb->work);
1719         }
1720 }
1721
1722 static int __bch2_btree_root_read(struct btree_trans *trans, enum btree_id id,
1723                                   const struct bkey_i *k, unsigned level)
1724 {
1725         struct bch_fs *c = trans->c;
1726         struct closure cl;
1727         struct btree *b;
1728         int ret;
1729
1730         closure_init_stack(&cl);
1731
1732         do {
1733                 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
1734                 closure_sync(&cl);
1735         } while (ret);
1736
1737         b = bch2_btree_node_mem_alloc(trans, level != 0);
1738         bch2_btree_cache_cannibalize_unlock(trans);
1739
1740         BUG_ON(IS_ERR(b));
1741
1742         bkey_copy(&b->key, k);
1743         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1744
1745         set_btree_node_read_in_flight(b);
1746
1747         /* we can't pass the trans to read_done() for fsck errors, so it must be unlocked */
1748         bch2_trans_unlock(trans);
1749         bch2_btree_node_read(trans, b, true);
1750
1751         if (btree_node_read_error(b)) {
1752                 mutex_lock(&c->btree_cache.lock);
1753                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1754                 mutex_unlock(&c->btree_cache.lock);
1755
1756                 ret = -BCH_ERR_btree_node_read_error;
1757                 goto err;
1758         }
1759
1760         bch2_btree_set_root_for_read(c, b);
1761 err:
1762         six_unlock_write(&b->c.lock);
1763         six_unlock_intent(&b->c.lock);
1764
1765         return ret;
1766 }
1767
1768 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1769                         const struct bkey_i *k, unsigned level)
1770 {
1771         return bch2_trans_run(c, __bch2_btree_root_read(trans, id, k, level));
1772 }
1773
1774 static void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1775                                       struct btree_write *w)
1776 {
1777         unsigned long old, new;
1778
1779         old = READ_ONCE(b->will_make_reachable);
1780         do {
1781                 new = old;
1782                 if (!(old & 1))
1783                         break;
1784
1785                 new &= ~1UL;
1786         } while (!try_cmpxchg(&b->will_make_reachable, &old, new));
1787
1788         if (old & 1)
1789                 closure_put(&((struct btree_update *) new)->cl);
1790
1791         bch2_journal_pin_drop(&c->journal, &w->journal);
1792 }
1793
1794 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1795 {
1796         struct btree_write *w = btree_prev_write(b);
1797         unsigned long old, new;
1798         unsigned type = 0;
1799
1800         bch2_btree_complete_write(c, b, w);
1801
1802         old = READ_ONCE(b->flags);
1803         do {
1804                 new = old;
1805
1806                 if ((old & (1U << BTREE_NODE_dirty)) &&
1807                     (old & (1U << BTREE_NODE_need_write)) &&
1808                     !(old & (1U << BTREE_NODE_never_write)) &&
1809                     !(old & (1U << BTREE_NODE_write_blocked)) &&
1810                     !(old & (1U << BTREE_NODE_will_make_reachable))) {
1811                         new &= ~(1U << BTREE_NODE_dirty);
1812                         new &= ~(1U << BTREE_NODE_need_write);
1813                         new |=  (1U << BTREE_NODE_write_in_flight);
1814                         new |=  (1U << BTREE_NODE_write_in_flight_inner);
1815                         new |=  (1U << BTREE_NODE_just_written);
1816                         new ^=  (1U << BTREE_NODE_write_idx);
1817
1818                         type = new & BTREE_WRITE_TYPE_MASK;
1819                         new &= ~BTREE_WRITE_TYPE_MASK;
1820                 } else {
1821                         new &= ~(1U << BTREE_NODE_write_in_flight);
1822                         new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1823                 }
1824         } while (!try_cmpxchg(&b->flags, &old, new));
1825
1826         if (new & (1U << BTREE_NODE_write_in_flight))
1827                 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED|type);
1828         else
1829                 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1830 }
1831
1832 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1833 {
1834         struct btree_trans *trans = bch2_trans_get(c);
1835
1836         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
1837
1838         /* we don't need transaction context anymore after we got the lock. */
1839         bch2_trans_put(trans);
1840         __btree_node_write_done(c, b);
1841         six_unlock_read(&b->c.lock);
1842 }
1843
1844 static void btree_node_write_work(struct work_struct *work)
1845 {
1846         struct btree_write_bio *wbio =
1847                 container_of(work, struct btree_write_bio, work);
1848         struct bch_fs *c        = wbio->wbio.c;
1849         struct btree *b         = wbio->wbio.bio.bi_private;
1850         int ret = 0;
1851
1852         btree_bounce_free(c,
1853                 wbio->data_bytes,
1854                 wbio->wbio.used_mempool,
1855                 wbio->data);
1856
1857         bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1858                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1859
1860         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key))) {
1861                 ret = -BCH_ERR_btree_node_write_all_failed;
1862                 goto err;
1863         }
1864
1865         if (wbio->wbio.first_btree_write) {
1866                 if (wbio->wbio.failed.nr) {
1867
1868                 }
1869         } else {
1870                 ret = bch2_trans_do(c,
1871                         bch2_btree_node_update_key_get_iter(trans, b, &wbio->key,
1872                                         BCH_WATERMARK_interior_updates|
1873                                         BCH_TRANS_COMMIT_journal_reclaim|
1874                                         BCH_TRANS_COMMIT_no_enospc|
1875                                         BCH_TRANS_COMMIT_no_check_rw,
1876                                         !wbio->wbio.failed.nr));
1877                 if (ret)
1878                         goto err;
1879         }
1880 out:
1881         bio_put(&wbio->wbio.bio);
1882         btree_node_write_done(c, b);
1883         return;
1884 err:
1885         set_btree_node_noevict(b);
1886         bch2_fs_fatal_err_on(!bch2_err_matches(ret, EROFS), c,
1887                              "writing btree node: %s", bch2_err_str(ret));
1888         goto out;
1889 }
1890
1891 static void btree_node_write_endio(struct bio *bio)
1892 {
1893         struct bch_write_bio *wbio      = to_wbio(bio);
1894         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1895         struct bch_write_bio *orig      = parent ?: wbio;
1896         struct btree_write_bio *wb      = container_of(orig, struct btree_write_bio, wbio);
1897         struct bch_fs *c                = wbio->c;
1898         struct btree *b                 = wbio->bio.bi_private;
1899         struct bch_dev *ca              = wbio->have_ioref ? bch2_dev_have_ref(c, wbio->dev) : NULL;
1900         unsigned long flags;
1901
1902         if (wbio->have_ioref)
1903                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1904
1905         if (!ca ||
1906             bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
1907                                "btree write error: %s",
1908                                bch2_blk_status_to_str(bio->bi_status)) ||
1909             bch2_meta_write_fault("btree")) {
1910                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1911                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1912                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1913         }
1914
1915         if (wbio->have_ioref)
1916                 percpu_ref_put(&ca->io_ref);
1917
1918         if (parent) {
1919                 bio_put(bio);
1920                 bio_endio(&parent->bio);
1921                 return;
1922         }
1923
1924         clear_btree_node_write_in_flight_inner(b);
1925         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1926         INIT_WORK(&wb->work, btree_node_write_work);
1927         queue_work(c->btree_io_complete_wq, &wb->work);
1928 }
1929
1930 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1931                                    struct bset *i, unsigned sectors)
1932 {
1933         bool saw_error;
1934
1935         int ret = bch2_bkey_validate(c, bkey_i_to_s_c(&b->key),
1936                                      BKEY_TYPE_btree, WRITE);
1937         if (ret) {
1938                 bch2_fs_inconsistent(c, "invalid btree node key before write");
1939                 return ret;
1940         }
1941
1942         ret = validate_bset_keys(c, b, i, WRITE, false, &saw_error) ?:
1943                 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false, &saw_error);
1944         if (ret) {
1945                 bch2_inconsistent_error(c);
1946                 dump_stack();
1947         }
1948
1949         return ret;
1950 }
1951
1952 static void btree_write_submit(struct work_struct *work)
1953 {
1954         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1955         BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1956
1957         bkey_copy(&tmp.k, &wbio->key);
1958
1959         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1960                 ptr->offset += wbio->sector_offset;
1961
1962         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree,
1963                                   &tmp.k, false);
1964 }
1965
1966 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1967 {
1968         struct btree_write_bio *wbio;
1969         struct bset *i;
1970         struct btree_node *bn = NULL;
1971         struct btree_node_entry *bne = NULL;
1972         struct sort_iter_stack sort_iter;
1973         struct nonce nonce;
1974         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1975         u64 seq = 0;
1976         bool used_mempool;
1977         unsigned long old, new;
1978         bool validate_before_checksum = false;
1979         enum btree_write_type type = flags & BTREE_WRITE_TYPE_MASK;
1980         void *data;
1981         int ret;
1982
1983         if (flags & BTREE_WRITE_ALREADY_STARTED)
1984                 goto do_write;
1985
1986         /*
1987          * We may only have a read lock on the btree node - the dirty bit is our
1988          * "lock" against racing with other threads that may be trying to start
1989          * a write, we do a write iff we clear the dirty bit. Since setting the
1990          * dirty bit requires a write lock, we can't race with other threads
1991          * redirtying it:
1992          */
1993         old = READ_ONCE(b->flags);
1994         do {
1995                 new = old;
1996
1997                 if (!(old & (1 << BTREE_NODE_dirty)))
1998                         return;
1999
2000                 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
2001                     !(old & (1 << BTREE_NODE_need_write)))
2002                         return;
2003
2004                 if (old &
2005                     ((1 << BTREE_NODE_never_write)|
2006                      (1 << BTREE_NODE_write_blocked)))
2007                         return;
2008
2009                 if (b->written &&
2010                     (old & (1 << BTREE_NODE_will_make_reachable)))
2011                         return;
2012
2013                 if (old & (1 << BTREE_NODE_write_in_flight))
2014                         return;
2015
2016                 if (flags & BTREE_WRITE_ONLY_IF_NEED)
2017                         type = new & BTREE_WRITE_TYPE_MASK;
2018                 new &= ~BTREE_WRITE_TYPE_MASK;
2019
2020                 new &= ~(1 << BTREE_NODE_dirty);
2021                 new &= ~(1 << BTREE_NODE_need_write);
2022                 new |=  (1 << BTREE_NODE_write_in_flight);
2023                 new |=  (1 << BTREE_NODE_write_in_flight_inner);
2024                 new |=  (1 << BTREE_NODE_just_written);
2025                 new ^=  (1 << BTREE_NODE_write_idx);
2026         } while (!try_cmpxchg_acquire(&b->flags, &old, new));
2027
2028         if (new & (1U << BTREE_NODE_need_write))
2029                 return;
2030 do_write:
2031         BUG_ON((type == BTREE_WRITE_initial) != (b->written == 0));
2032
2033         atomic_long_dec(&c->btree_cache.nr_dirty);
2034
2035         BUG_ON(btree_node_fake(b));
2036         BUG_ON((b->will_make_reachable != 0) != !b->written);
2037
2038         BUG_ON(b->written >= btree_sectors(c));
2039         BUG_ON(b->written & (block_sectors(c) - 1));
2040         BUG_ON(bset_written(b, btree_bset_last(b)));
2041         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
2042         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
2043
2044         bch2_sort_whiteouts(c, b);
2045
2046         sort_iter_stack_init(&sort_iter, b);
2047
2048         bytes = !b->written
2049                 ? sizeof(struct btree_node)
2050                 : sizeof(struct btree_node_entry);
2051
2052         bytes += b->whiteout_u64s * sizeof(u64);
2053
2054         for_each_bset(b, t) {
2055                 i = bset(b, t);
2056
2057                 if (bset_written(b, i))
2058                         continue;
2059
2060                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
2061                 sort_iter_add(&sort_iter.iter,
2062                               btree_bkey_first(b, t),
2063                               btree_bkey_last(b, t));
2064                 seq = max(seq, le64_to_cpu(i->journal_seq));
2065         }
2066
2067         BUG_ON(b->written && !seq);
2068
2069         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
2070         bytes += 8;
2071
2072         /* buffer must be a multiple of the block size */
2073         bytes = round_up(bytes, block_bytes(c));
2074
2075         data = btree_bounce_alloc(c, bytes, &used_mempool);
2076
2077         if (!b->written) {
2078                 bn = data;
2079                 *bn = *b->data;
2080                 i = &bn->keys;
2081         } else {
2082                 bne = data;
2083                 bne->keys = b->data->keys;
2084                 i = &bne->keys;
2085         }
2086
2087         i->journal_seq  = cpu_to_le64(seq);
2088         i->u64s         = 0;
2089
2090         sort_iter_add(&sort_iter.iter,
2091                       unwritten_whiteouts_start(b),
2092                       unwritten_whiteouts_end(b));
2093         SET_BSET_SEPARATE_WHITEOUTS(i, false);
2094
2095         u64s = bch2_sort_keys_keep_unwritten_whiteouts(i->start, &sort_iter.iter);
2096         le16_add_cpu(&i->u64s, u64s);
2097
2098         b->whiteout_u64s = 0;
2099
2100         BUG_ON(!b->written && i->u64s != b->data->keys.u64s);
2101
2102         set_needs_whiteout(i, false);
2103
2104         /* do we have data to write? */
2105         if (b->written && !i->u64s)
2106                 goto nowrite;
2107
2108         bytes_to_write = vstruct_end(i) - data;
2109         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
2110
2111         if (!b->written &&
2112             b->key.k.type == KEY_TYPE_btree_ptr_v2)
2113                 BUG_ON(btree_ptr_sectors_written(bkey_i_to_s_c(&b->key)) != sectors_to_write);
2114
2115         memset(data + bytes_to_write, 0,
2116                (sectors_to_write << 9) - bytes_to_write);
2117
2118         BUG_ON(b->written + sectors_to_write > btree_sectors(c));
2119         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
2120         BUG_ON(i->seq != b->data->keys.seq);
2121
2122         i->version = cpu_to_le16(c->sb.version);
2123         SET_BSET_OFFSET(i, b->written);
2124         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
2125
2126         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
2127                 validate_before_checksum = true;
2128
2129         /* validate_bset will be modifying: */
2130         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
2131                 validate_before_checksum = true;
2132
2133         /* if we're going to be encrypting, check metadata validity first: */
2134         if (validate_before_checksum &&
2135             validate_bset_for_write(c, b, i, sectors_to_write))
2136                 goto err;
2137
2138         ret = bset_encrypt(c, i, b->written << 9);
2139         if (bch2_fs_fatal_err_on(ret, c,
2140                         "encrypting btree node: %s", bch2_err_str(ret)))
2141                 goto err;
2142
2143         nonce = btree_nonce(i, b->written << 9);
2144
2145         if (bn)
2146                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
2147         else
2148                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
2149
2150         /* if we're not encrypting, check metadata after checksumming: */
2151         if (!validate_before_checksum &&
2152             validate_bset_for_write(c, b, i, sectors_to_write))
2153                 goto err;
2154
2155         /*
2156          * We handle btree write errors by immediately halting the journal -
2157          * after we've done that, we can't issue any subsequent btree writes
2158          * because they might have pointers to new nodes that failed to write.
2159          *
2160          * Furthermore, there's no point in doing any more btree writes because
2161          * with the journal stopped, we're never going to update the journal to
2162          * reflect that those writes were done and the data flushed from the
2163          * journal:
2164          *
2165          * Also on journal error, the pending write may have updates that were
2166          * never journalled (interior nodes, see btree_update_nodes_written()) -
2167          * it's critical that we don't do the write in that case otherwise we
2168          * will have updates visible that weren't in the journal:
2169          *
2170          * Make sure to update b->written so bch2_btree_init_next() doesn't
2171          * break:
2172          */
2173         if (bch2_journal_error(&c->journal) ||
2174             c->opts.nochanges)
2175                 goto err;
2176
2177         trace_and_count(c, btree_node_write, b, bytes_to_write, sectors_to_write);
2178
2179         wbio = container_of(bio_alloc_bioset(NULL,
2180                                 buf_pages(data, sectors_to_write << 9),
2181                                 REQ_OP_WRITE|REQ_META,
2182                                 GFP_NOFS,
2183                                 &c->btree_bio),
2184                             struct btree_write_bio, wbio.bio);
2185         wbio_init(&wbio->wbio.bio);
2186         wbio->data                      = data;
2187         wbio->data_bytes                = bytes;
2188         wbio->sector_offset             = b->written;
2189         wbio->wbio.c                    = c;
2190         wbio->wbio.used_mempool         = used_mempool;
2191         wbio->wbio.first_btree_write    = !b->written;
2192         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
2193         wbio->wbio.bio.bi_private       = b;
2194
2195         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
2196
2197         bkey_copy(&wbio->key, &b->key);
2198
2199         b->written += sectors_to_write;
2200
2201         if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
2202                 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2203                         cpu_to_le16(b->written);
2204
2205         atomic64_inc(&c->btree_write_stats[type].nr);
2206         atomic64_add(bytes_to_write, &c->btree_write_stats[type].bytes);
2207
2208         INIT_WORK(&wbio->work, btree_write_submit);
2209         queue_work(c->btree_write_submit_wq, &wbio->work);
2210         return;
2211 err:
2212         set_btree_node_noevict(b);
2213         b->written += sectors_to_write;
2214 nowrite:
2215         btree_bounce_free(c, bytes, used_mempool, data);
2216         __btree_node_write_done(c, b);
2217 }
2218
2219 /*
2220  * Work that must be done with write lock held:
2221  */
2222 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2223 {
2224         bool invalidated_iter = false;
2225         struct btree_node_entry *bne;
2226
2227         if (!btree_node_just_written(b))
2228                 return false;
2229
2230         BUG_ON(b->whiteout_u64s);
2231
2232         clear_btree_node_just_written(b);
2233
2234         /*
2235          * Note: immediately after write, bset_written() doesn't work - the
2236          * amount of data we had to write after compaction might have been
2237          * smaller than the offset of the last bset.
2238          *
2239          * However, we know that all bsets have been written here, as long as
2240          * we're still holding the write lock:
2241          */
2242
2243         /*
2244          * XXX: decide if we really want to unconditionally sort down to a
2245          * single bset:
2246          */
2247         if (b->nsets > 1) {
2248                 btree_node_sort(c, b, 0, b->nsets);
2249                 invalidated_iter = true;
2250         } else {
2251                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2252         }
2253
2254         for_each_bset(b, t)
2255                 set_needs_whiteout(bset(b, t), true);
2256
2257         bch2_btree_verify(c, b);
2258
2259         /*
2260          * If later we don't unconditionally sort down to a single bset, we have
2261          * to ensure this is still true:
2262          */
2263         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2264
2265         bne = want_new_bset(c, b);
2266         if (bne)
2267                 bch2_bset_init_next(b, bne);
2268
2269         bch2_btree_build_aux_trees(b);
2270
2271         return invalidated_iter;
2272 }
2273
2274 /*
2275  * Use this one if the node is intent locked:
2276  */
2277 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2278                            enum six_lock_type lock_type_held,
2279                            unsigned flags)
2280 {
2281         if (lock_type_held == SIX_LOCK_intent ||
2282             (lock_type_held == SIX_LOCK_read &&
2283              six_lock_tryupgrade(&b->c.lock))) {
2284                 __bch2_btree_node_write(c, b, flags);
2285
2286                 /* don't cycle lock unnecessarily: */
2287                 if (btree_node_just_written(b) &&
2288                     six_trylock_write(&b->c.lock)) {
2289                         bch2_btree_post_write_cleanup(c, b);
2290                         six_unlock_write(&b->c.lock);
2291                 }
2292
2293                 if (lock_type_held == SIX_LOCK_read)
2294                         six_lock_downgrade(&b->c.lock);
2295         } else {
2296                 __bch2_btree_node_write(c, b, flags);
2297                 if (lock_type_held == SIX_LOCK_write &&
2298                     btree_node_just_written(b))
2299                         bch2_btree_post_write_cleanup(c, b);
2300         }
2301 }
2302
2303 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2304 {
2305         struct bucket_table *tbl;
2306         struct rhash_head *pos;
2307         struct btree *b;
2308         unsigned i;
2309         bool ret = false;
2310 restart:
2311         rcu_read_lock();
2312         for_each_cached_btree(b, c, tbl, i, pos)
2313                 if (test_bit(flag, &b->flags)) {
2314                         rcu_read_unlock();
2315                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2316                         ret = true;
2317                         goto restart;
2318                 }
2319         rcu_read_unlock();
2320
2321         return ret;
2322 }
2323
2324 bool bch2_btree_flush_all_reads(struct bch_fs *c)
2325 {
2326         return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2327 }
2328
2329 bool bch2_btree_flush_all_writes(struct bch_fs *c)
2330 {
2331         return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2332 }
2333
2334 static const char * const bch2_btree_write_types[] = {
2335 #define x(t, n) [n] = #t,
2336         BCH_BTREE_WRITE_TYPES()
2337         NULL
2338 };
2339
2340 void bch2_btree_write_stats_to_text(struct printbuf *out, struct bch_fs *c)
2341 {
2342         printbuf_tabstop_push(out, 20);
2343         printbuf_tabstop_push(out, 10);
2344
2345         prt_printf(out, "\tnr\tsize\n");
2346
2347         for (unsigned i = 0; i < BTREE_WRITE_TYPE_NR; i++) {
2348                 u64 nr          = atomic64_read(&c->btree_write_stats[i].nr);
2349                 u64 bytes       = atomic64_read(&c->btree_write_stats[i].bytes);
2350
2351                 prt_printf(out, "%s:\t%llu\t", bch2_btree_write_types[i], nr);
2352                 prt_human_readable_u64(out, nr ? div64_u64(bytes, nr) : 0);
2353                 prt_newline(out);
2354         }
2355 }