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         bch2_assert_btree_nodes_not_locked();
  50
  51         wait_on_bit_lock_io(&b->flags, BTREE_NODE_write_in_flight,
  52                             TASK_UNINTERRUPTIBLE);
  53 }
  54
  55 void __bch2_btree_node_wait_on_read(struct btree *b)
  56 {
  57         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
  58                        TASK_UNINTERRUPTIBLE);
  59 }
  60
  61 void __bch2_btree_node_wait_on_write(struct btree *b)
  62 {
  63         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
  64                        TASK_UNINTERRUPTIBLE);
  65 }
  66
  67 void bch2_btree_node_wait_on_read(struct btree *b)
  68 {
  69         bch2_assert_btree_nodes_not_locked();
  70
  71         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
  72                        TASK_UNINTERRUPTIBLE);
  73 }
  74
  75 void bch2_btree_node_wait_on_write(struct btree *b)
  76 {
  77         bch2_assert_btree_nodes_not_locked();
  78
  79         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
  80                        TASK_UNINTERRUPTIBLE);
  81 }
  82
  83 static void verify_no_dups(struct btree *b,
  84                            struct bkey_packed *start,
  85                            struct bkey_packed *end)
  86 {
  87 #ifdef CONFIG_BCACHEFS_DEBUG
  88         struct bkey_packed *k, *p;
  89
  90         if (start == end)
  91                 return;
  92
  93         for (p = start, k = bkey_p_next(start);
  94              k != end;
  95              p = k, k = bkey_p_next(k)) {
  96                 struct bkey l = bkey_unpack_key(b, p);
  97                 struct bkey r = bkey_unpack_key(b, k);
  98
  99                 BUG_ON(bpos_ge(l.p, bkey_start_pos(&r)));
 100         }
 101 #endif
 102 }
 103
 104 static void set_needs_whiteout(struct bset *i, int v)
 105 {
 106         struct bkey_packed *k;
 107
 108         for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
 109                 k->needs_whiteout = v;
 110 }
 111
 112 static void btree_bounce_free(struct bch_fs *c, size_t size,
 113                               bool used_mempool, void *p)
 114 {
 115         if (used_mempool)
 116                 mempool_free(p, &c->btree_bounce_pool);
 117         else
 118                 kvfree(p);
 119 }
 120
 121 static void *btree_bounce_alloc(struct bch_fs *c, size_t size,
 122                                 bool *used_mempool)
 123 {
 124         unsigned flags = memalloc_nofs_save();
 125         void *p;
 126
 127         BUG_ON(size > c->opts.btree_node_size);
 128
 129         *used_mempool = false;
 130         p = kvmalloc(size, __GFP_NOWARN|GFP_NOWAIT);
 131         if (!p) {
 132                 *used_mempool = true;
 133                 p = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
 134         }
 135         memalloc_nofs_restore(flags);
 136         return p;
 137 }
 138
 139 static void sort_bkey_ptrs(const struct btree *bt,
 140                            struct bkey_packed **ptrs, unsigned nr)
 141 {
 142         unsigned n = nr, a = nr / 2, b, c, d;
 143
 144         if (!a)
 145                 return;
 146
 147         /* Heap sort: see lib/sort.c: */
 148         while (1) {
 149                 if (a)
 150                         a--;
 151                 else if (--n)
 152                         swap(ptrs[0], ptrs[n]);
 153                 else
 154                         break;
 155
 156                 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
 157                         b = bch2_bkey_cmp_packed(bt,
 158                                             ptrs[c],
 159                                             ptrs[d]) >= 0 ? c : d;
 160                 if (d == n)
 161                         b = c;
 162
 163                 while (b != a &&
 164                        bch2_bkey_cmp_packed(bt,
 165                                        ptrs[a],
 166                                        ptrs[b]) >= 0)
 167                         b = (b - 1) / 2;
 168                 c = b;
 169                 while (b != a) {
 170                         b = (b - 1) / 2;
 171                         swap(ptrs[b], ptrs[c]);
 172                 }
 173         }
 174 }
 175
 176 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
 177 {
 178         struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
 179         bool used_mempool = false;
 180         size_t bytes = b->whiteout_u64s * sizeof(u64);
 181
 182         if (!b->whiteout_u64s)
 183                 return;
 184
 185         new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
 186
 187         ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
 188
 189         for (k = unwritten_whiteouts_start(b);
 190              k != unwritten_whiteouts_end(b);
 191              k = bkey_p_next(k))
 192                 *--ptrs = k;
 193
 194         sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
 195
 196         k = new_whiteouts;
 197
 198         while (ptrs != ptrs_end) {
 199                 bkey_p_copy(k, *ptrs);
 200                 k = bkey_p_next(k);
 201                 ptrs++;
 202         }
 203
 204         verify_no_dups(b, new_whiteouts,
 205                        (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
 206
 207         memcpy_u64s(unwritten_whiteouts_start(b),
 208                     new_whiteouts, b->whiteout_u64s);
 209
 210         btree_bounce_free(c, bytes, used_mempool, new_whiteouts);
 211 }
 212
 213 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
 214                                 bool compacting, enum compact_mode mode)
 215 {
 216         if (!bset_dead_u64s(b, t))
 217                 return false;
 218
 219         switch (mode) {
 220         case COMPACT_LAZY:
 221                 return should_compact_bset_lazy(b, t) ||
 222                         (compacting && !bset_written(b, bset(b, t)));
 223         case COMPACT_ALL:
 224                 return true;
 225         default:
 226                 BUG();
 227         }
 228 }
 229
 230 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
 231 {
 232         bool ret = false;
 233
 234         for_each_bset(b, t) {
 235                 struct bset *i = bset(b, t);
 236                 struct bkey_packed *k, *n, *out, *start, *end;
 237                 struct btree_node_entry *src = NULL, *dst = NULL;
 238
 239                 if (t != b->set && !bset_written(b, i)) {
 240                         src = container_of(i, struct btree_node_entry, keys);
 241                         dst = max(write_block(b),
 242                                   (void *) btree_bkey_last(b, t - 1));
 243                 }
 244
 245                 if (src != dst)
 246                         ret = true;
 247
 248                 if (!should_compact_bset(b, t, ret, mode)) {
 249                         if (src != dst) {
 250                                 memmove(dst, src, sizeof(*src) +
 251                                         le16_to_cpu(src->keys.u64s) *
 252                                         sizeof(u64));
 253                                 i = &dst->keys;
 254                                 set_btree_bset(b, t, i);
 255                         }
 256                         continue;
 257                 }
 258
 259                 start   = btree_bkey_first(b, t);
 260                 end     = btree_bkey_last(b, t);
 261
 262                 if (src != dst) {
 263                         memmove(dst, src, sizeof(*src));
 264                         i = &dst->keys;
 265                         set_btree_bset(b, t, i);
 266                 }
 267
 268                 out = i->start;
 269
 270                 for (k = start; k != end; k = n) {
 271                         n = bkey_p_next(k);
 272
 273                         if (!bkey_deleted(k)) {
 274                                 bkey_p_copy(out, k);
 275                                 out = bkey_p_next(out);
 276                         } else {
 277                                 BUG_ON(k->needs_whiteout);
 278                         }
 279                 }
 280
 281                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 282                 set_btree_bset_end(b, t);
 283                 bch2_bset_set_no_aux_tree(b, t);
 284                 ret = true;
 285         }
 286
 287         bch2_verify_btree_nr_keys(b);
 288
 289         bch2_btree_build_aux_trees(b);
 290
 291         return ret;
 292 }
 293
 294 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 295                             enum compact_mode mode)
 296 {
 297         return bch2_drop_whiteouts(b, mode);
 298 }
 299
 300 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 301                             unsigned start_idx,
 302                             unsigned end_idx)
 303 {
 304         struct btree_node *out;
 305         struct sort_iter_stack sort_iter;
 306         struct bset_tree *t;
 307         struct bset *start_bset = bset(b, &b->set[start_idx]);
 308         bool used_mempool = false;
 309         u64 start_time, seq = 0;
 310         unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
 311         bool sorting_entire_node = start_idx == 0 &&
 312                 end_idx == b->nsets;
 313
 314         sort_iter_stack_init(&sort_iter, b);
 315
 316         for (t = b->set + start_idx;
 317              t < b->set + end_idx;
 318              t++) {
 319                 u64s += le16_to_cpu(bset(b, t)->u64s);
 320                 sort_iter_add(&sort_iter.iter,
 321                               btree_bkey_first(b, t),
 322                               btree_bkey_last(b, t));
 323         }
 324
 325         bytes = sorting_entire_node
 326                 ? btree_buf_bytes(b)
 327                 : __vstruct_bytes(struct btree_node, u64s);
 328
 329         out = btree_bounce_alloc(c, bytes, &used_mempool);
 330
 331         start_time = local_clock();
 332
 333         u64s = bch2_sort_keys(out->keys.start, &sort_iter.iter);
 334
 335         out->keys.u64s = cpu_to_le16(u64s);
 336
 337         BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
 338
 339         if (sorting_entire_node)
 340                 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 341                                        start_time);
 342
 343         /* Make sure we preserve bset journal_seq: */
 344         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 345                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 346         start_bset->journal_seq = cpu_to_le64(seq);
 347
 348         if (sorting_entire_node) {
 349                 u64s = le16_to_cpu(out->keys.u64s);
 350
 351                 BUG_ON(bytes != btree_buf_bytes(b));
 352
 353                 /*
 354                  * Our temporary buffer is the same size as the btree node's
 355                  * buffer, we can just swap buffers instead of doing a big
 356                  * memcpy()
 357                  */
 358                 *out = *b->data;
 359                 out->keys.u64s = cpu_to_le16(u64s);
 360                 swap(out, b->data);
 361                 set_btree_bset(b, b->set, &b->data->keys);
 362         } else {
 363                 start_bset->u64s = out->keys.u64s;
 364                 memcpy_u64s(start_bset->start,
 365                             out->keys.start,
 366                             le16_to_cpu(out->keys.u64s));
 367         }
 368
 369         for (i = start_idx + 1; i < end_idx; i++)
 370                 b->nr.bset_u64s[start_idx] +=
 371                         b->nr.bset_u64s[i];
 372
 373         b->nsets -= shift;
 374
 375         for (i = start_idx + 1; i < b->nsets; i++) {
 376                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 377                 b->set[i]               = b->set[i + shift];
 378         }
 379
 380         for (i = b->nsets; i < MAX_BSETS; i++)
 381                 b->nr.bset_u64s[i] = 0;
 382
 383         set_btree_bset_end(b, &b->set[start_idx]);
 384         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 385
 386         btree_bounce_free(c, bytes, used_mempool, out);
 387
 388         bch2_verify_btree_nr_keys(b);
 389 }
 390
 391 void bch2_btree_sort_into(struct bch_fs *c,
 392                          struct btree *dst,
 393                          struct btree *src)
 394 {
 395         struct btree_nr_keys nr;
 396         struct btree_node_iter src_iter;
 397         u64 start_time = local_clock();
 398
 399         BUG_ON(dst->nsets != 1);
 400
 401         bch2_bset_set_no_aux_tree(dst, dst->set);
 402
 403         bch2_btree_node_iter_init_from_start(&src_iter, src);
 404
 405         nr = bch2_sort_repack(btree_bset_first(dst),
 406                         src, &src_iter,
 407                         &dst->format,
 408                         true);
 409
 410         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 411                                start_time);
 412
 413         set_btree_bset_end(dst, dst->set);
 414
 415         dst->nr.live_u64s       += nr.live_u64s;
 416         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 417         dst->nr.packed_keys     += nr.packed_keys;
 418         dst->nr.unpacked_keys   += nr.unpacked_keys;
 419
 420         bch2_verify_btree_nr_keys(dst);
 421 }
 422
 423 /*
 424  * We're about to add another bset to the btree node, so if there's currently
 425  * too many bsets - sort some of them together:
 426  */
 427 static bool btree_node_compact(struct bch_fs *c, struct btree *b)
 428 {
 429         unsigned unwritten_idx;
 430         bool ret = false;
 431
 432         for (unwritten_idx = 0;
 433              unwritten_idx < b->nsets;
 434              unwritten_idx++)
 435                 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
 436                         break;
 437
 438         if (b->nsets - unwritten_idx > 1) {
 439                 btree_node_sort(c, b, unwritten_idx, b->nsets);
 440                 ret = true;
 441         }
 442
 443         if (unwritten_idx > 1) {
 444                 btree_node_sort(c, b, 0, unwritten_idx);
 445                 ret = true;
 446         }
 447
 448         return ret;
 449 }
 450
 451 void bch2_btree_build_aux_trees(struct btree *b)
 452 {
 453         for_each_bset(b, t)
 454                 bch2_bset_build_aux_tree(b, t,
 455                                 !bset_written(b, bset(b, t)) &&
 456                                 t == bset_tree_last(b));
 457 }
 458
 459 /*
 460  * If we have MAX_BSETS (3) bsets, should we sort them all down to just one?
 461  *
 462  * The first bset is going to be of similar order to the size of the node, the
 463  * last bset is bounded by btree_write_set_buffer(), which is set to keep the
 464  * memmove on insert from being too expensive: the middle bset should, ideally,
 465  * be the geometric mean of the first and the last.
 466  *
 467  * Returns true if the middle bset is greater than that geometric mean:
 468  */
 469 static inline bool should_compact_all(struct bch_fs *c, struct btree *b)
 470 {
 471         unsigned mid_u64s_bits =
 472                 (ilog2(btree_max_u64s(c)) + BTREE_WRITE_SET_U64s_BITS) / 2;
 473
 474         return bset_u64s(&b->set[1]) > 1U << mid_u64s_bits;
 475 }
 476
 477 /*
 478  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 479  * inserted into
 480  *
 481  * Safe to call if there already is an unwritten bset - will only add a new bset
 482  * if @b doesn't already have one.
 483  *
 484  * Returns true if we sorted (i.e. invalidated iterators
 485  */
 486 void bch2_btree_init_next(struct btree_trans *trans, struct btree *b)
 487 {
 488         struct bch_fs *c = trans->c;
 489         struct btree_node_entry *bne;
 490         bool reinit_iter = false;
 491
 492         EBUG_ON(!six_lock_counts(&b->c.lock).n[SIX_LOCK_write]);
 493         BUG_ON(bset_written(b, bset(b, &b->set[1])));
 494         BUG_ON(btree_node_just_written(b));
 495
 496         if (b->nsets == MAX_BSETS &&
 497             !btree_node_write_in_flight(b) &&
 498             should_compact_all(c, b)) {
 499                 bch2_btree_node_write(c, b, SIX_LOCK_write,
 500                                       BTREE_WRITE_init_next_bset);
 501                 reinit_iter = true;
 502         }
 503
 504         if (b->nsets == MAX_BSETS &&
 505             btree_node_compact(c, b))
 506                 reinit_iter = true;
 507
 508         BUG_ON(b->nsets >= MAX_BSETS);
 509
 510         bne = want_new_bset(c, b);
 511         if (bne)
 512                 bch2_bset_init_next(b, bne);
 513
 514         bch2_btree_build_aux_trees(b);
 515
 516         if (reinit_iter)
 517                 bch2_trans_node_reinit_iter(trans, b);
 518 }
 519
 520 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 521                           struct bch_dev *ca,
 522                           struct btree *b, struct bset *i, struct bkey_packed *k,
 523                           unsigned offset, int write)
 524 {
 525         prt_printf(out, bch2_log_msg(c, "%s"),
 526                    write == READ
 527                    ? "error validating btree node "
 528                    : "corrupt btree node before write ");
 529         if (ca)
 530                 prt_printf(out, "on %s ", ca->name);
 531         prt_printf(out, "at btree ");
 532         bch2_btree_pos_to_text(out, c, b);
 533
 534         printbuf_indent_add(out, 2);
 535
 536         prt_printf(out, "\nnode offset %u/%u",
 537                    b->written, btree_ptr_sectors_written(&b->key));
 538         if (i)
 539                 prt_printf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 540         if (k)
 541                 prt_printf(out, " bset byte offset %lu",
 542                            (unsigned long)(void *)k -
 543                            ((unsigned long)(void *)i & ~511UL));
 544         prt_str(out, ": ");
 545 }
 546
 547 __printf(10, 11)
 548 static int __btree_err(int ret,
 549                        struct bch_fs *c,
 550                        struct bch_dev *ca,
 551                        struct btree *b,
 552                        struct bset *i,
 553                        struct bkey_packed *k,
 554                        int write,
 555                        bool have_retry,
 556                        enum bch_sb_error_id err_type,
 557                        const char *fmt, ...)
 558 {
 559         struct printbuf out = PRINTBUF;
 560         bool silent = c->curr_recovery_pass == BCH_RECOVERY_PASS_scan_for_btree_nodes;
 561         va_list args;
 562
 563         btree_err_msg(&out, c, ca, b, i, k, b->written, write);
 564
 565         va_start(args, fmt);
 566         prt_vprintf(&out, fmt, args);
 567         va_end(args);
 568
 569         if (write == WRITE) {
 570                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 571                 ret = c->opts.errors == BCH_ON_ERROR_continue
 572                         ? 0
 573                         : -BCH_ERR_fsck_errors_not_fixed;
 574                 goto out;
 575         }
 576
 577         if (!have_retry && ret == -BCH_ERR_btree_node_read_err_want_retry)
 578                 ret = -BCH_ERR_btree_node_read_err_fixable;
 579         if (!have_retry && ret == -BCH_ERR_btree_node_read_err_must_retry)
 580                 ret = -BCH_ERR_btree_node_read_err_bad_node;
 581
 582         if (!silent && ret != -BCH_ERR_btree_node_read_err_fixable)
 583                 bch2_sb_error_count(c, err_type);
 584
 585         switch (ret) {
 586         case -BCH_ERR_btree_node_read_err_fixable:
 587                 ret = !silent
 588                         ? bch2_fsck_err(c, FSCK_CAN_FIX, err_type, "%s", out.buf)
 589                         : -BCH_ERR_fsck_fix;
 590                 if (ret != -BCH_ERR_fsck_fix &&
 591                     ret != -BCH_ERR_fsck_ignore)
 592                         goto fsck_err;
 593                 ret = -BCH_ERR_fsck_fix;
 594                 break;
 595         case -BCH_ERR_btree_node_read_err_want_retry:
 596         case -BCH_ERR_btree_node_read_err_must_retry:
 597                 if (!silent)
 598                         bch2_print_string_as_lines(KERN_ERR, out.buf);
 599                 break;
 600         case -BCH_ERR_btree_node_read_err_bad_node:
 601                 if (!silent)
 602                         bch2_print_string_as_lines(KERN_ERR, out.buf);
 603                 ret = bch2_topology_error(c);
 604                 break;
 605         case -BCH_ERR_btree_node_read_err_incompatible:
 606                 if (!silent)
 607                         bch2_print_string_as_lines(KERN_ERR, out.buf);
 608                 ret = -BCH_ERR_fsck_errors_not_fixed;
 609                 break;
 610         default:
 611                 BUG();
 612         }
 613 out:
 614 fsck_err:
 615         printbuf_exit(&out);
 616         return ret;
 617 }
 618
 619 #define btree_err(type, c, ca, b, i, k, _err_type, msg, ...)            \
 620 ({                                                                      \
 621         int _ret = __btree_err(type, c, ca, b, i, k, write, have_retry, \
 622                                BCH_FSCK_ERR_##_err_type,                \
 623                                msg, ##__VA_ARGS__);                     \
 624                                                                         \
 625         if (_ret != -BCH_ERR_fsck_fix) {                                \
 626                 ret = _ret;                                             \
 627                 goto fsck_err;                                          \
 628         }                                                               \
 629                                                                         \
 630         *saw_error = true;                                              \
 631 })
 632
 633 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 634
 635 /*
 636  * When btree topology repair changes the start or end of a node, that might
 637  * mean we have to drop keys that are no longer inside the node:
 638  */
 639 __cold
 640 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
 641 {
 642         for_each_bset(b, t) {
 643                 struct bset *i = bset(b, t);
 644                 struct bkey_packed *k;
 645
 646                 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
 647                         if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
 648                                 break;
 649
 650                 if (k != i->start) {
 651                         unsigned shift = (u64 *) k - (u64 *) i->start;
 652
 653                         memmove_u64s_down(i->start, k,
 654                                           (u64 *) vstruct_end(i) - (u64 *) k);
 655                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
 656                         set_btree_bset_end(b, t);
 657                 }
 658
 659                 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
 660                         if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
 661                                 break;
 662
 663                 if (k != vstruct_last(i)) {
 664                         i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
 665                         set_btree_bset_end(b, t);
 666                 }
 667         }
 668
 669         /*
 670          * Always rebuild search trees: eytzinger search tree nodes directly
 671          * depend on the values of min/max key:
 672          */
 673         bch2_bset_set_no_aux_tree(b, b->set);
 674         bch2_btree_build_aux_trees(b);
 675         b->nr = bch2_btree_node_count_keys(b);
 676
 677         struct bkey_s_c k;
 678         struct bkey unpacked;
 679         struct btree_node_iter iter;
 680         for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
 681                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 682                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 683         }
 684 }
 685
 686 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 687                          struct btree *b, struct bset *i,
 688                          unsigned offset, unsigned sectors,
 689                          int write, bool have_retry, bool *saw_error)
 690 {
 691         unsigned version = le16_to_cpu(i->version);
 692         struct printbuf buf1 = PRINTBUF;
 693         struct printbuf buf2 = PRINTBUF;
 694         int ret = 0;
 695
 696         btree_err_on(!bch2_version_compatible(version),
 697                      -BCH_ERR_btree_node_read_err_incompatible,
 698                      c, ca, b, i, NULL,
 699                      btree_node_unsupported_version,
 700                      "unsupported bset version %u.%u",
 701                      BCH_VERSION_MAJOR(version),
 702                      BCH_VERSION_MINOR(version));
 703
 704         if (btree_err_on(version < c->sb.version_min,
 705                          -BCH_ERR_btree_node_read_err_fixable,
 706                          c, NULL, b, i, NULL,
 707                          btree_node_bset_older_than_sb_min,
 708                          "bset version %u older than superblock version_min %u",
 709                          version, c->sb.version_min)) {
 710                 mutex_lock(&c->sb_lock);
 711                 c->disk_sb.sb->version_min = cpu_to_le16(version);
 712                 bch2_write_super(c);
 713                 mutex_unlock(&c->sb_lock);
 714         }
 715
 716         if (btree_err_on(BCH_VERSION_MAJOR(version) >
 717                          BCH_VERSION_MAJOR(c->sb.version),
 718                          -BCH_ERR_btree_node_read_err_fixable,
 719                          c, NULL, b, i, NULL,
 720                          btree_node_bset_newer_than_sb,
 721                          "bset version %u newer than superblock version %u",
 722                          version, c->sb.version)) {
 723                 mutex_lock(&c->sb_lock);
 724                 c->disk_sb.sb->version = cpu_to_le16(version);
 725                 bch2_write_super(c);
 726                 mutex_unlock(&c->sb_lock);
 727         }
 728
 729         btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
 730                      -BCH_ERR_btree_node_read_err_incompatible,
 731                      c, ca, b, i, NULL,
 732                      btree_node_unsupported_version,
 733                      "BSET_SEPARATE_WHITEOUTS no longer supported");
 734
 735         if (btree_err_on(offset + sectors > btree_sectors(c),
 736                          -BCH_ERR_btree_node_read_err_fixable,
 737                          c, ca, b, i, NULL,
 738                          bset_past_end_of_btree_node,
 739                          "bset past end of btree node")) {
 740                 i->u64s = 0;
 741                 ret = 0;
 742                 goto out;
 743         }
 744
 745         btree_err_on(offset && !i->u64s,
 746                      -BCH_ERR_btree_node_read_err_fixable,
 747                      c, ca, b, i, NULL,
 748                      bset_empty,
 749                      "empty bset");
 750
 751         btree_err_on(BSET_OFFSET(i) && BSET_OFFSET(i) != offset,
 752                      -BCH_ERR_btree_node_read_err_want_retry,
 753                      c, ca, b, i, NULL,
 754                      bset_wrong_sector_offset,
 755                      "bset at wrong sector offset");
 756
 757         if (!offset) {
 758                 struct btree_node *bn =
 759                         container_of(i, struct btree_node, keys);
 760                 /* These indicate that we read the wrong btree node: */
 761
 762                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 763                         struct bch_btree_ptr_v2 *bp =
 764                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 765
 766                         /* XXX endianness */
 767                         btree_err_on(bp->seq != bn->keys.seq,
 768                                      -BCH_ERR_btree_node_read_err_must_retry,
 769                                      c, ca, b, NULL, NULL,
 770                                      bset_bad_seq,
 771                                      "incorrect sequence number (wrong btree node)");
 772                 }
 773
 774                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 775                              -BCH_ERR_btree_node_read_err_must_retry,
 776                              c, ca, b, i, NULL,
 777                              btree_node_bad_btree,
 778                              "incorrect btree id");
 779
 780                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 781                              -BCH_ERR_btree_node_read_err_must_retry,
 782                              c, ca, b, i, NULL,
 783                              btree_node_bad_level,
 784                              "incorrect level");
 785
 786                 if (!write)
 787                         compat_btree_node(b->c.level, b->c.btree_id, version,
 788                                           BSET_BIG_ENDIAN(i), write, bn);
 789
 790                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 791                         struct bch_btree_ptr_v2 *bp =
 792                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 793
 794                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 795                                 b->data->min_key = bp->min_key;
 796                                 b->data->max_key = b->key.k.p;
 797                         }
 798
 799                         btree_err_on(!bpos_eq(b->data->min_key, bp->min_key),
 800                                      -BCH_ERR_btree_node_read_err_must_retry,
 801                                      c, ca, b, NULL, NULL,
 802                                      btree_node_bad_min_key,
 803                                      "incorrect min_key: got %s should be %s",
 804                                      (printbuf_reset(&buf1),
 805                                       bch2_bpos_to_text(&buf1, bn->min_key), buf1.buf),
 806                                      (printbuf_reset(&buf2),
 807                                       bch2_bpos_to_text(&buf2, bp->min_key), buf2.buf));
 808                 }
 809
 810                 btree_err_on(!bpos_eq(bn->max_key, b->key.k.p),
 811                              -BCH_ERR_btree_node_read_err_must_retry,
 812                              c, ca, b, i, NULL,
 813                              btree_node_bad_max_key,
 814                              "incorrect max key %s",
 815                              (printbuf_reset(&buf1),
 816                               bch2_bpos_to_text(&buf1, bn->max_key), buf1.buf));
 817
 818                 if (write)
 819                         compat_btree_node(b->c.level, b->c.btree_id, version,
 820                                           BSET_BIG_ENDIAN(i), write, bn);
 821
 822                 btree_err_on(bch2_bkey_format_invalid(c, &bn->format, write, &buf1),
 823                              -BCH_ERR_btree_node_read_err_bad_node,
 824                              c, ca, b, i, NULL,
 825                              btree_node_bad_format,
 826                              "invalid bkey format: %s\n  %s", buf1.buf,
 827                              (printbuf_reset(&buf2),
 828                               bch2_bkey_format_to_text(&buf2, &bn->format), buf2.buf));
 829                 printbuf_reset(&buf1);
 830
 831                 compat_bformat(b->c.level, b->c.btree_id, version,
 832                                BSET_BIG_ENDIAN(i), write,
 833                                &bn->format);
 834         }
 835 out:
 836 fsck_err:
 837         printbuf_exit(&buf2);
 838         printbuf_exit(&buf1);
 839         return ret;
 840 }
 841
 842 static int bset_key_invalid(struct bch_fs *c, struct btree *b,
 843                             struct bkey_s_c k,
 844                             bool updated_range, int rw,
 845                             struct printbuf *err)
 846 {
 847         return __bch2_bkey_invalid(c, k, btree_node_type(b), READ, err) ?:
 848                 (!updated_range ? bch2_bkey_in_btree_node(c, b, k, err) : 0) ?:
 849                 (rw == WRITE ? bch2_bkey_val_invalid(c, k, READ, err) : 0);
 850 }
 851
 852 static bool bkey_packed_valid(struct bch_fs *c, struct btree *b,
 853                          struct bset *i, struct bkey_packed *k)
 854 {
 855         if (bkey_p_next(k) > vstruct_last(i))
 856                 return false;
 857
 858         if (k->format > KEY_FORMAT_CURRENT)
 859                 return false;
 860
 861         if (!bkeyp_u64s_valid(&b->format, k))
 862                 return false;
 863
 864         struct printbuf buf = PRINTBUF;
 865         struct bkey tmp;
 866         struct bkey_s u = __bkey_disassemble(b, k, &tmp);
 867         bool ret = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b), READ, &buf);
 868         printbuf_exit(&buf);
 869         return ret;
 870 }
 871
 872 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 873                          struct bset *i, int write,
 874                          bool have_retry, bool *saw_error)
 875 {
 876         unsigned version = le16_to_cpu(i->version);
 877         struct bkey_packed *k, *prev = NULL;
 878         struct printbuf buf = PRINTBUF;
 879         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 880                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 881         int ret = 0;
 882
 883         for (k = i->start;
 884              k != vstruct_last(i);) {
 885                 struct bkey_s u;
 886                 struct bkey tmp;
 887                 unsigned next_good_key;
 888
 889                 if (btree_err_on(bkey_p_next(k) > vstruct_last(i),
 890                                  -BCH_ERR_btree_node_read_err_fixable,
 891                                  c, NULL, b, i, k,
 892                                  btree_node_bkey_past_bset_end,
 893                                  "key extends past end of bset")) {
 894                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 895                         break;
 896                 }
 897
 898                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 899                                  -BCH_ERR_btree_node_read_err_fixable,
 900                                  c, NULL, b, i, k,
 901                                  btree_node_bkey_bad_format,
 902                                  "invalid bkey format %u", k->format))
 903                         goto drop_this_key;
 904
 905                 if (btree_err_on(!bkeyp_u64s_valid(&b->format, k),
 906                                  -BCH_ERR_btree_node_read_err_fixable,
 907                                  c, NULL, b, i, k,
 908                                  btree_node_bkey_bad_u64s,
 909                                  "bad k->u64s %u (min %u max %zu)", k->u64s,
 910                                  bkeyp_key_u64s(&b->format, k),
 911                                  U8_MAX - BKEY_U64s + bkeyp_key_u64s(&b->format, k)))
 912                         goto drop_this_key;
 913
 914                 if (!write)
 915                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 916                                     BSET_BIG_ENDIAN(i), write,
 917                                     &b->format, k);
 918
 919                 u = __bkey_disassemble(b, k, &tmp);
 920
 921                 printbuf_reset(&buf);
 922                 if (bset_key_invalid(c, b, u.s_c, updated_range, write, &buf)) {
 923                         printbuf_reset(&buf);
 924                         bset_key_invalid(c, b, u.s_c, updated_range, write, &buf);
 925                         prt_printf(&buf, "\n  ");
 926                         bch2_bkey_val_to_text(&buf, c, u.s_c);
 927
 928                         btree_err(-BCH_ERR_btree_node_read_err_fixable,
 929                                   c, NULL, b, i, k,
 930                                   btree_node_bad_bkey,
 931                                   "invalid bkey: %s", buf.buf);
 932                         goto drop_this_key;
 933                 }
 934
 935                 if (write)
 936                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 937                                     BSET_BIG_ENDIAN(i), write,
 938                                     &b->format, k);
 939
 940                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 941                         struct bkey up = bkey_unpack_key(b, prev);
 942
 943                         printbuf_reset(&buf);
 944                         prt_printf(&buf, "keys out of order: ");
 945                         bch2_bkey_to_text(&buf, &up);
 946                         prt_printf(&buf, " > ");
 947                         bch2_bkey_to_text(&buf, u.k);
 948
 949                         if (btree_err(-BCH_ERR_btree_node_read_err_fixable,
 950                                       c, NULL, b, i, k,
 951                                       btree_node_bkey_out_of_order,
 952                                       "%s", buf.buf))
 953                                 goto drop_this_key;
 954                 }
 955
 956                 prev = k;
 957                 k = bkey_p_next(k);
 958                 continue;
 959 drop_this_key:
 960                 next_good_key = k->u64s;
 961
 962                 if (!next_good_key ||
 963                     (BSET_BIG_ENDIAN(i) == CPU_BIG_ENDIAN &&
 964                      version >= bcachefs_metadata_version_snapshot)) {
 965                         /*
 966                          * only do scanning if bch2_bkey_compat() has nothing to
 967                          * do
 968                          */
 969
 970                         if (!bkey_packed_valid(c, b, i, (void *) ((u64 *) k + next_good_key))) {
 971                                 for (next_good_key = 1;
 972                                      next_good_key < (u64 *) vstruct_last(i) - (u64 *) k;
 973                                      next_good_key++)
 974                                         if (bkey_packed_valid(c, b, i, (void *) ((u64 *) k + next_good_key)))
 975                                                 goto got_good_key;
 976                         }
 977
 978                         /*
 979                          * didn't find a good key, have to truncate the rest of
 980                          * the bset
 981                          */
 982                         next_good_key = (u64 *) vstruct_last(i) - (u64 *) k;
 983                 }
 984 got_good_key:
 985                 le16_add_cpu(&i->u64s, -next_good_key);
 986                 memmove_u64s_down(k, bkey_p_next(k), (u64 *) vstruct_end(i) - (u64 *) k);
 987         }
 988 fsck_err:
 989         printbuf_exit(&buf);
 990         return ret;
 991 }
 992
 993 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 994                               struct btree *b, bool have_retry, bool *saw_error)
 995 {
 996         struct btree_node_entry *bne;
 997         struct sort_iter *iter;
 998         struct btree_node *sorted;
 999         struct bkey_packed *k;
1000         struct bset *i;
1001         bool used_mempool, blacklisted;
1002         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
1003                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
1004         unsigned u64s;
1005         unsigned ptr_written = btree_ptr_sectors_written(&b->key);
1006         struct printbuf buf = PRINTBUF;
1007         int ret = 0, retry_read = 0, write = READ;
1008         u64 start_time = local_clock();
1009
1010         b->version_ondisk = U16_MAX;
1011         /* We might get called multiple times on read retry: */
1012         b->written = 0;
1013
1014         iter = mempool_alloc(&c->fill_iter, GFP_NOFS);
1015         sort_iter_init(iter, b, (btree_blocks(c) + 1) * 2);
1016
1017         if (bch2_meta_read_fault("btree"))
1018                 btree_err(-BCH_ERR_btree_node_read_err_must_retry,
1019                           c, ca, b, NULL, NULL,
1020                           btree_node_fault_injected,
1021                           "dynamic fault");
1022
1023         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
1024                      -BCH_ERR_btree_node_read_err_must_retry,
1025                      c, ca, b, NULL, NULL,
1026                      btree_node_bad_magic,
1027                      "bad magic: want %llx, got %llx",
1028                      bset_magic(c), le64_to_cpu(b->data->magic));
1029
1030         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
1031                 struct bch_btree_ptr_v2 *bp =
1032                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
1033
1034                 bch2_bpos_to_text(&buf, b->data->min_key);
1035                 prt_str(&buf, "-");
1036                 bch2_bpos_to_text(&buf, b->data->max_key);
1037
1038                 btree_err_on(b->data->keys.seq != bp->seq,
1039                              -BCH_ERR_btree_node_read_err_must_retry,
1040                              c, ca, b, NULL, NULL,
1041                              btree_node_bad_seq,
1042                              "got wrong btree node: got\n%s",
1043                              (printbuf_reset(&buf),
1044                               bch2_btree_node_header_to_text(&buf, b->data),
1045                               buf.buf));
1046         } else {
1047                 btree_err_on(!b->data->keys.seq,
1048                              -BCH_ERR_btree_node_read_err_must_retry,
1049                              c, ca, b, NULL, NULL,
1050                              btree_node_bad_seq,
1051                              "bad btree header: seq 0\n%s",
1052                              (printbuf_reset(&buf),
1053                               bch2_btree_node_header_to_text(&buf, b->data),
1054                               buf.buf));
1055         }
1056
1057         while (b->written < (ptr_written ?: btree_sectors(c))) {
1058                 unsigned sectors;
1059                 struct nonce nonce;
1060                 bool first = !b->written;
1061                 bool csum_bad;
1062
1063                 if (!b->written) {
1064                         i = &b->data->keys;
1065
1066                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1067                                      -BCH_ERR_btree_node_read_err_want_retry,
1068                                      c, ca, b, i, NULL,
1069                                      bset_unknown_csum,
1070                                      "unknown checksum type %llu", BSET_CSUM_TYPE(i));
1071
1072                         nonce = btree_nonce(i, b->written << 9);
1073
1074                         struct bch_csum csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
1075                         csum_bad = bch2_crc_cmp(b->data->csum, csum);
1076                         if (csum_bad)
1077                                 bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
1078
1079                         btree_err_on(csum_bad,
1080                                      -BCH_ERR_btree_node_read_err_want_retry,
1081                                      c, ca, b, i, NULL,
1082                                      bset_bad_csum,
1083                                      "%s",
1084                                      (printbuf_reset(&buf),
1085                                       bch2_csum_err_msg(&buf, BSET_CSUM_TYPE(i), b->data->csum, csum),
1086                                       buf.buf));
1087
1088                         ret = bset_encrypt(c, i, b->written << 9);
1089                         if (bch2_fs_fatal_err_on(ret, c,
1090                                         "decrypting btree node: %s", bch2_err_str(ret)))
1091                                 goto fsck_err;
1092
1093                         btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
1094                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
1095                                      -BCH_ERR_btree_node_read_err_incompatible,
1096                                      c, NULL, b, NULL, NULL,
1097                                      btree_node_unsupported_version,
1098                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
1099
1100                         sectors = vstruct_sectors(b->data, c->block_bits);
1101                 } else {
1102                         bne = write_block(b);
1103                         i = &bne->keys;
1104
1105                         if (i->seq != b->data->keys.seq)
1106                                 break;
1107
1108                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1109                                      -BCH_ERR_btree_node_read_err_want_retry,
1110                                      c, ca, b, i, NULL,
1111                                      bset_unknown_csum,
1112                                      "unknown checksum type %llu", BSET_CSUM_TYPE(i));
1113
1114                         nonce = btree_nonce(i, b->written << 9);
1115                         struct bch_csum csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1116                         csum_bad = bch2_crc_cmp(bne->csum, csum);
1117                         if (ca && csum_bad)
1118                                 bch2_io_error(ca, BCH_MEMBER_ERROR_checksum);
1119
1120                         btree_err_on(csum_bad,
1121                                      -BCH_ERR_btree_node_read_err_want_retry,
1122                                      c, ca, b, i, NULL,
1123                                      bset_bad_csum,
1124                                      "%s",
1125                                      (printbuf_reset(&buf),
1126                                       bch2_csum_err_msg(&buf, BSET_CSUM_TYPE(i), bne->csum, csum),
1127                                       buf.buf));
1128
1129                         ret = bset_encrypt(c, i, b->written << 9);
1130                         if (bch2_fs_fatal_err_on(ret, c,
1131                                         "decrypting btree node: %s", bch2_err_str(ret)))
1132                                 goto fsck_err;
1133
1134                         sectors = vstruct_sectors(bne, c->block_bits);
1135                 }
1136
1137                 b->version_ondisk = min(b->version_ondisk,
1138                                         le16_to_cpu(i->version));
1139
1140                 ret = validate_bset(c, ca, b, i, b->written, sectors,
1141                                     READ, have_retry, saw_error);
1142                 if (ret)
1143                         goto fsck_err;
1144
1145                 if (!b->written)
1146                         btree_node_set_format(b, b->data->format);
1147
1148                 ret = validate_bset_keys(c, b, i, READ, have_retry, saw_error);
1149                 if (ret)
1150                         goto fsck_err;
1151
1152                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1153
1154                 blacklisted = bch2_journal_seq_is_blacklisted(c,
1155                                         le64_to_cpu(i->journal_seq),
1156                                         true);
1157
1158                 btree_err_on(blacklisted && first,
1159                              -BCH_ERR_btree_node_read_err_fixable,
1160                              c, ca, b, i, NULL,
1161                              bset_blacklisted_journal_seq,
1162                              "first btree node bset has blacklisted journal seq (%llu)",
1163                              le64_to_cpu(i->journal_seq));
1164
1165                 btree_err_on(blacklisted && ptr_written,
1166                              -BCH_ERR_btree_node_read_err_fixable,
1167                              c, ca, b, i, NULL,
1168                              first_bset_blacklisted_journal_seq,
1169                              "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
1170                              le64_to_cpu(i->journal_seq),
1171                              b->written, b->written + sectors, ptr_written);
1172
1173                 b->written += sectors;
1174
1175                 if (blacklisted && !first)
1176                         continue;
1177
1178                 sort_iter_add(iter,
1179                               vstruct_idx(i, 0),
1180                               vstruct_last(i));
1181         }
1182
1183         if (ptr_written) {
1184                 btree_err_on(b->written < ptr_written,
1185                              -BCH_ERR_btree_node_read_err_want_retry,
1186                              c, ca, b, NULL, NULL,
1187                              btree_node_data_missing,
1188                              "btree node data missing: expected %u sectors, found %u",
1189                              ptr_written, b->written);
1190         } else {
1191                 for (bne = write_block(b);
1192                      bset_byte_offset(b, bne) < btree_buf_bytes(b);
1193                      bne = (void *) bne + block_bytes(c))
1194                         btree_err_on(bne->keys.seq == b->data->keys.seq &&
1195                                      !bch2_journal_seq_is_blacklisted(c,
1196                                                                       le64_to_cpu(bne->keys.journal_seq),
1197                                                                       true),
1198                                      -BCH_ERR_btree_node_read_err_want_retry,
1199                                      c, ca, b, NULL, NULL,
1200                                      btree_node_bset_after_end,
1201                                      "found bset signature after last bset");
1202         }
1203
1204         sorted = btree_bounce_alloc(c, btree_buf_bytes(b), &used_mempool);
1205         sorted->keys.u64s = 0;
1206
1207         set_btree_bset(b, b->set, &b->data->keys);
1208
1209         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1210
1211         u64s = le16_to_cpu(sorted->keys.u64s);
1212         *sorted = *b->data;
1213         sorted->keys.u64s = cpu_to_le16(u64s);
1214         swap(sorted, b->data);
1215         set_btree_bset(b, b->set, &b->data->keys);
1216         b->nsets = 1;
1217
1218         BUG_ON(b->nr.live_u64s != u64s);
1219
1220         btree_bounce_free(c, btree_buf_bytes(b), used_mempool, sorted);
1221
1222         if (updated_range)
1223                 bch2_btree_node_drop_keys_outside_node(b);
1224
1225         i = &b->data->keys;
1226         for (k = i->start; k != vstruct_last(i);) {
1227                 struct bkey tmp;
1228                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1229
1230                 printbuf_reset(&buf);
1231
1232                 if (bch2_bkey_val_invalid(c, u.s_c, READ, &buf) ||
1233                     (bch2_inject_invalid_keys &&
1234                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1235                         printbuf_reset(&buf);
1236
1237                         prt_printf(&buf, "invalid bkey: ");
1238                         bch2_bkey_val_invalid(c, u.s_c, READ, &buf);
1239                         prt_printf(&buf, "\n  ");
1240                         bch2_bkey_val_to_text(&buf, c, u.s_c);
1241
1242                         btree_err(-BCH_ERR_btree_node_read_err_fixable,
1243                                   c, NULL, b, i, k,
1244                                   btree_node_bad_bkey,
1245                                   "%s", buf.buf);
1246
1247                         btree_keys_account_key_drop(&b->nr, 0, k);
1248
1249                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1250                         memmove_u64s_down(k, bkey_p_next(k),
1251                                           (u64 *) vstruct_end(i) - (u64 *) k);
1252                         set_btree_bset_end(b, b->set);
1253                         continue;
1254                 }
1255
1256                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1257                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1258
1259                         bp.v->mem_ptr = 0;
1260                 }
1261
1262                 k = bkey_p_next(k);
1263         }
1264
1265         bch2_bset_build_aux_tree(b, b->set, false);
1266
1267         set_needs_whiteout(btree_bset_first(b), true);
1268
1269         btree_node_reset_sib_u64s(b);
1270
1271         rcu_read_lock();
1272         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1273                 struct bch_dev *ca2 = bch2_dev_rcu(c, ptr->dev);
1274
1275                 if (!ca2 || ca2->mi.state != BCH_MEMBER_STATE_rw)
1276                         set_btree_node_need_rewrite(b);
1277         }
1278         rcu_read_unlock();
1279
1280         if (!ptr_written)
1281                 set_btree_node_need_rewrite(b);
1282 out:
1283         mempool_free(iter, &c->fill_iter);
1284         printbuf_exit(&buf);
1285         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read_done], start_time);
1286         return retry_read;
1287 fsck_err:
1288         if (ret == -BCH_ERR_btree_node_read_err_want_retry ||
1289             ret == -BCH_ERR_btree_node_read_err_must_retry) {
1290                 retry_read = 1;
1291         } else {
1292                 set_btree_node_read_error(b);
1293                 bch2_btree_lost_data(c, b->c.btree_id);
1294         }
1295         goto out;
1296 }
1297
1298 static void btree_node_read_work(struct work_struct *work)
1299 {
1300         struct btree_read_bio *rb =
1301                 container_of(work, struct btree_read_bio, work);
1302         struct bch_fs *c        = rb->c;
1303         struct bch_dev *ca      = rb->have_ioref ? bch2_dev_have_ref(c, rb->pick.ptr.dev) : NULL;
1304         struct btree *b         = rb->b;
1305         struct bio *bio         = &rb->bio;
1306         struct bch_io_failures failed = { .nr = 0 };
1307         struct printbuf buf = PRINTBUF;
1308         bool saw_error = false;
1309         bool retry = false;
1310         bool can_retry;
1311
1312         goto start;
1313         while (1) {
1314                 retry = true;
1315                 bch_info(c, "retrying read");
1316                 ca = bch2_dev_get_ioref(c, rb->pick.ptr.dev, READ);
1317                 rb->have_ioref          = ca != NULL;
1318                 bio_reset(bio, NULL, REQ_OP_READ|REQ_SYNC|REQ_META);
1319                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1320                 bio->bi_iter.bi_size    = btree_buf_bytes(b);
1321
1322                 if (rb->have_ioref) {
1323                         bio_set_dev(bio, ca->disk_sb.bdev);
1324                         submit_bio_wait(bio);
1325                 } else {
1326                         bio->bi_status = BLK_STS_REMOVED;
1327                 }
1328 start:
1329                 printbuf_reset(&buf);
1330                 bch2_btree_pos_to_text(&buf, c, b);
1331                 bch2_dev_io_err_on(ca && bio->bi_status, ca, BCH_MEMBER_ERROR_read,
1332                                    "btree read error %s for %s",
1333                                    bch2_blk_status_to_str(bio->bi_status), buf.buf);
1334                 if (rb->have_ioref)
1335                         percpu_ref_put(&ca->io_ref);
1336                 rb->have_ioref = false;
1337
1338                 bch2_mark_io_failure(&failed, &rb->pick);
1339
1340                 can_retry = bch2_bkey_pick_read_device(c,
1341                                 bkey_i_to_s_c(&b->key),
1342                                 &failed, &rb->pick) > 0;
1343
1344                 if (!bio->bi_status &&
1345                     !bch2_btree_node_read_done(c, ca, b, can_retry, &saw_error)) {
1346                         if (retry)
1347                                 bch_info(c, "retry success");
1348                         break;
1349                 }
1350
1351                 saw_error = true;
1352
1353                 if (!can_retry) {
1354                         set_btree_node_read_error(b);
1355                         bch2_btree_lost_data(c, b->c.btree_id);
1356                         break;
1357                 }
1358         }
1359
1360         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1361                                rb->start_time);
1362         bio_put(&rb->bio);
1363
1364         if (saw_error &&
1365             !btree_node_read_error(b) &&
1366             c->curr_recovery_pass != BCH_RECOVERY_PASS_scan_for_btree_nodes) {
1367                 printbuf_reset(&buf);
1368                 bch2_bpos_to_text(&buf, b->key.k.p);
1369                 bch_err_ratelimited(c, "%s: rewriting btree node at btree=%s level=%u %s due to error",
1370                          __func__, bch2_btree_id_str(b->c.btree_id), b->c.level, buf.buf);
1371
1372                 bch2_btree_node_rewrite_async(c, b);
1373         }
1374
1375         printbuf_exit(&buf);
1376         clear_btree_node_read_in_flight(b);
1377         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1378 }
1379
1380 static void btree_node_read_endio(struct bio *bio)
1381 {
1382         struct btree_read_bio *rb =
1383                 container_of(bio, struct btree_read_bio, bio);
1384         struct bch_fs *c        = rb->c;
1385
1386         if (rb->have_ioref) {
1387                 struct bch_dev *ca = bch2_dev_have_ref(c, rb->pick.ptr.dev);
1388
1389                 bch2_latency_acct(ca, rb->start_time, READ);
1390         }
1391
1392         queue_work(c->btree_read_complete_wq, &rb->work);
1393 }
1394
1395 struct btree_node_read_all {
1396         struct closure          cl;
1397         struct bch_fs           *c;
1398         struct btree            *b;
1399         unsigned                nr;
1400         void                    *buf[BCH_REPLICAS_MAX];
1401         struct bio              *bio[BCH_REPLICAS_MAX];
1402         blk_status_t            err[BCH_REPLICAS_MAX];
1403 };
1404
1405 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1406 {
1407         struct btree_node *bn = data;
1408         struct btree_node_entry *bne;
1409         unsigned offset = 0;
1410
1411         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1412                 return 0;
1413
1414         while (offset < btree_sectors(c)) {
1415                 if (!offset) {
1416                         offset += vstruct_sectors(bn, c->block_bits);
1417                 } else {
1418                         bne = data + (offset << 9);
1419                         if (bne->keys.seq != bn->keys.seq)
1420                                 break;
1421                         offset += vstruct_sectors(bne, c->block_bits);
1422                 }
1423         }
1424
1425         return offset;
1426 }
1427
1428 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1429 {
1430         struct btree_node *bn = data;
1431         struct btree_node_entry *bne;
1432
1433         if (!offset)
1434                 return false;
1435
1436         while (offset < btree_sectors(c)) {
1437                 bne = data + (offset << 9);
1438                 if (bne->keys.seq == bn->keys.seq)
1439                         return true;
1440                 offset++;
1441         }
1442
1443         return false;
1444         return offset;
1445 }
1446
1447 static CLOSURE_CALLBACK(btree_node_read_all_replicas_done)
1448 {
1449         closure_type(ra, struct btree_node_read_all, cl);
1450         struct bch_fs *c = ra->c;
1451         struct btree *b = ra->b;
1452         struct printbuf buf = PRINTBUF;
1453         bool dump_bset_maps = false;
1454         bool have_retry = false;
1455         int ret = 0, best = -1, write = READ;
1456         unsigned i, written = 0, written2 = 0;
1457         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1458                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1459         bool _saw_error = false, *saw_error = &_saw_error;
1460
1461         for (i = 0; i < ra->nr; i++) {
1462                 struct btree_node *bn = ra->buf[i];
1463
1464                 if (ra->err[i])
1465                         continue;
1466
1467                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1468                     (seq && seq != bn->keys.seq))
1469                         continue;
1470
1471                 if (best < 0) {
1472                         best = i;
1473                         written = btree_node_sectors_written(c, bn);
1474                         continue;
1475                 }
1476
1477                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1478                 if (btree_err_on(written2 != written, -BCH_ERR_btree_node_read_err_fixable,
1479                                  c, NULL, b, NULL, NULL,
1480                                  btree_node_replicas_sectors_written_mismatch,
1481                                  "btree node sectors written mismatch: %u != %u",
1482                                  written, written2) ||
1483                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1484                                  -BCH_ERR_btree_node_read_err_fixable,
1485                                  c, NULL, b, NULL, NULL,
1486                                  btree_node_bset_after_end,
1487                                  "found bset signature after last bset") ||
1488                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1489                                  -BCH_ERR_btree_node_read_err_fixable,
1490                                  c, NULL, b, NULL, NULL,
1491                                  btree_node_replicas_data_mismatch,
1492                                  "btree node replicas content mismatch"))
1493                         dump_bset_maps = true;
1494
1495                 if (written2 > written) {
1496                         written = written2;
1497                         best = i;
1498                 }
1499         }
1500 fsck_err:
1501         if (dump_bset_maps) {
1502                 for (i = 0; i < ra->nr; i++) {
1503                         struct btree_node *bn = ra->buf[i];
1504                         struct btree_node_entry *bne = NULL;
1505                         unsigned offset = 0, sectors;
1506                         bool gap = false;
1507
1508                         if (ra->err[i])
1509                                 continue;
1510
1511                         printbuf_reset(&buf);
1512
1513                         while (offset < btree_sectors(c)) {
1514                                 if (!offset) {
1515                                         sectors = vstruct_sectors(bn, c->block_bits);
1516                                 } else {
1517                                         bne = ra->buf[i] + (offset << 9);
1518                                         if (bne->keys.seq != bn->keys.seq)
1519                                                 break;
1520                                         sectors = vstruct_sectors(bne, c->block_bits);
1521                                 }
1522
1523                                 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1524                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1525                                                         le64_to_cpu(bne->keys.journal_seq), false))
1526                                         prt_printf(&buf, "*");
1527                                 offset += sectors;
1528                         }
1529
1530                         while (offset < btree_sectors(c)) {
1531                                 bne = ra->buf[i] + (offset << 9);
1532                                 if (bne->keys.seq == bn->keys.seq) {
1533                                         if (!gap)
1534                                                 prt_printf(&buf, " GAP");
1535                                         gap = true;
1536
1537                                         sectors = vstruct_sectors(bne, c->block_bits);
1538                                         prt_printf(&buf, " %u-%u", offset, offset + sectors);
1539                                         if (bch2_journal_seq_is_blacklisted(c,
1540                                                         le64_to_cpu(bne->keys.journal_seq), false))
1541                                                 prt_printf(&buf, "*");
1542                                 }
1543                                 offset++;
1544                         }
1545
1546                         bch_err(c, "replica %u:%s", i, buf.buf);
1547                 }
1548         }
1549
1550         if (best >= 0) {
1551                 memcpy(b->data, ra->buf[best], btree_buf_bytes(b));
1552                 ret = bch2_btree_node_read_done(c, NULL, b, false, saw_error);
1553         } else {
1554                 ret = -1;
1555         }
1556
1557         if (ret) {
1558                 set_btree_node_read_error(b);
1559                 bch2_btree_lost_data(c, b->c.btree_id);
1560         } else if (*saw_error)
1561                 bch2_btree_node_rewrite_async(c, b);
1562
1563         for (i = 0; i < ra->nr; i++) {
1564                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1565                 bio_put(ra->bio[i]);
1566         }
1567
1568         closure_debug_destroy(&ra->cl);
1569         kfree(ra);
1570         printbuf_exit(&buf);
1571
1572         clear_btree_node_read_in_flight(b);
1573         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1574 }
1575
1576 static void btree_node_read_all_replicas_endio(struct bio *bio)
1577 {
1578         struct btree_read_bio *rb =
1579                 container_of(bio, struct btree_read_bio, bio);
1580         struct bch_fs *c        = rb->c;
1581         struct btree_node_read_all *ra = rb->ra;
1582
1583         if (rb->have_ioref) {
1584                 struct bch_dev *ca = bch2_dev_have_ref(c, rb->pick.ptr.dev);
1585
1586                 bch2_latency_acct(ca, rb->start_time, READ);
1587         }
1588
1589         ra->err[rb->idx] = bio->bi_status;
1590         closure_put(&ra->cl);
1591 }
1592
1593 /*
1594  * XXX This allocates multiple times from the same mempools, and can deadlock
1595  * under sufficient memory pressure (but is only a debug path)
1596  */
1597 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1598 {
1599         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1600         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1601         const union bch_extent_entry *entry;
1602         struct extent_ptr_decoded pick;
1603         struct btree_node_read_all *ra;
1604         unsigned i;
1605
1606         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1607         if (!ra)
1608                 return -BCH_ERR_ENOMEM_btree_node_read_all_replicas;
1609
1610         closure_init(&ra->cl, NULL);
1611         ra->c   = c;
1612         ra->b   = b;
1613         ra->nr  = bch2_bkey_nr_ptrs(k);
1614
1615         for (i = 0; i < ra->nr; i++) {
1616                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1617                 ra->bio[i] = bio_alloc_bioset(NULL,
1618                                               buf_pages(ra->buf[i], btree_buf_bytes(b)),
1619                                               REQ_OP_READ|REQ_SYNC|REQ_META,
1620                                               GFP_NOFS,
1621                                               &c->btree_bio);
1622         }
1623
1624         i = 0;
1625         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1626                 struct bch_dev *ca = bch2_dev_get_ioref(c, pick.ptr.dev, READ);
1627                 struct btree_read_bio *rb =
1628                         container_of(ra->bio[i], struct btree_read_bio, bio);
1629                 rb->c                   = c;
1630                 rb->b                   = b;
1631                 rb->ra                  = ra;
1632                 rb->start_time          = local_clock();
1633                 rb->have_ioref          = ca != NULL;
1634                 rb->idx                 = i;
1635                 rb->pick                = pick;
1636                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1637                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1638                 bch2_bio_map(&rb->bio, ra->buf[i], btree_buf_bytes(b));
1639
1640                 if (rb->have_ioref) {
1641                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1642                                      bio_sectors(&rb->bio));
1643                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1644
1645                         closure_get(&ra->cl);
1646                         submit_bio(&rb->bio);
1647                 } else {
1648                         ra->err[i] = BLK_STS_REMOVED;
1649                 }
1650
1651                 i++;
1652         }
1653
1654         if (sync) {
1655                 closure_sync(&ra->cl);
1656                 btree_node_read_all_replicas_done(&ra->cl.work);
1657         } else {
1658                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1659                             c->btree_read_complete_wq);
1660         }
1661
1662         return 0;
1663 }
1664
1665 void bch2_btree_node_read(struct btree_trans *trans, struct btree *b,
1666                           bool sync)
1667 {
1668         struct bch_fs *c = trans->c;
1669         struct extent_ptr_decoded pick;
1670         struct btree_read_bio *rb;
1671         struct bch_dev *ca;
1672         struct bio *bio;
1673         int ret;
1674
1675         trace_and_count(c, btree_node_read, trans, b);
1676
1677         if (bch2_verify_all_btree_replicas &&
1678             !btree_node_read_all_replicas(c, b, sync))
1679                 return;
1680
1681         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1682                                          NULL, &pick);
1683
1684         if (ret <= 0) {
1685                 struct printbuf buf = PRINTBUF;
1686
1687                 prt_str(&buf, "btree node read error: no device to read from\n at ");
1688                 bch2_btree_pos_to_text(&buf, c, b);
1689                 bch_err_ratelimited(c, "%s", buf.buf);
1690
1691                 if (c->recovery_passes_explicit & BIT_ULL(BCH_RECOVERY_PASS_check_topology) &&
1692                     c->curr_recovery_pass > BCH_RECOVERY_PASS_check_topology)
1693                         bch2_fatal_error(c);
1694
1695                 set_btree_node_read_error(b);
1696                 bch2_btree_lost_data(c, b->c.btree_id);
1697                 clear_btree_node_read_in_flight(b);
1698                 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1699                 printbuf_exit(&buf);
1700                 return;
1701         }
1702
1703         ca = bch2_dev_get_ioref(c, pick.ptr.dev, READ);
1704
1705         bio = bio_alloc_bioset(NULL,
1706                                buf_pages(b->data, btree_buf_bytes(b)),
1707                                REQ_OP_READ|REQ_SYNC|REQ_META,
1708                                GFP_NOFS,
1709                                &c->btree_bio);
1710         rb = container_of(bio, struct btree_read_bio, bio);
1711         rb->c                   = c;
1712         rb->b                   = b;
1713         rb->ra                  = NULL;
1714         rb->start_time          = local_clock();
1715         rb->have_ioref          = ca != NULL;
1716         rb->pick                = pick;
1717         INIT_WORK(&rb->work, btree_node_read_work);
1718         bio->bi_iter.bi_sector  = pick.ptr.offset;
1719         bio->bi_end_io          = btree_node_read_endio;
1720         bch2_bio_map(bio, b->data, btree_buf_bytes(b));
1721
1722         if (rb->have_ioref) {
1723                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1724                              bio_sectors(bio));
1725                 bio_set_dev(bio, ca->disk_sb.bdev);
1726
1727                 if (sync) {
1728                         submit_bio_wait(bio);
1729                         bch2_latency_acct(ca, rb->start_time, READ);
1730                         btree_node_read_work(&rb->work);
1731                 } else {
1732                         submit_bio(bio);
1733                 }
1734         } else {
1735                 bio->bi_status = BLK_STS_REMOVED;
1736
1737                 if (sync)
1738                         btree_node_read_work(&rb->work);
1739                 else
1740                         queue_work(c->btree_read_complete_wq, &rb->work);
1741         }
1742 }
1743
1744 static int __bch2_btree_root_read(struct btree_trans *trans, enum btree_id id,
1745                                   const struct bkey_i *k, unsigned level)
1746 {
1747         struct bch_fs *c = trans->c;
1748         struct closure cl;
1749         struct btree *b;
1750         int ret;
1751
1752         closure_init_stack(&cl);
1753
1754         do {
1755                 ret = bch2_btree_cache_cannibalize_lock(trans, &cl);
1756                 closure_sync(&cl);
1757         } while (ret);
1758
1759         b = bch2_btree_node_mem_alloc(trans, level != 0);
1760         bch2_btree_cache_cannibalize_unlock(trans);
1761
1762         BUG_ON(IS_ERR(b));
1763
1764         bkey_copy(&b->key, k);
1765         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1766
1767         set_btree_node_read_in_flight(b);
1768
1769         bch2_btree_node_read(trans, b, true);
1770
1771         if (btree_node_read_error(b)) {
1772                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1773
1774                 mutex_lock(&c->btree_cache.lock);
1775                 list_move(&b->list, &c->btree_cache.freeable);
1776                 mutex_unlock(&c->btree_cache.lock);
1777
1778                 ret = -BCH_ERR_btree_node_read_error;
1779                 goto err;
1780         }
1781
1782         bch2_btree_set_root_for_read(c, b);
1783 err:
1784         six_unlock_write(&b->c.lock);
1785         six_unlock_intent(&b->c.lock);
1786
1787         return ret;
1788 }
1789
1790 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1791                         const struct bkey_i *k, unsigned level)
1792 {
1793         return bch2_trans_run(c, __bch2_btree_root_read(trans, id, k, level));
1794 }
1795
1796 static void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1797                                       struct btree_write *w)
1798 {
1799         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1800
1801         do {
1802                 old = new = v;
1803                 if (!(old & 1))
1804                         break;
1805
1806                 new &= ~1UL;
1807         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1808
1809         if (old & 1)
1810                 closure_put(&((struct btree_update *) new)->cl);
1811
1812         bch2_journal_pin_drop(&c->journal, &w->journal);
1813 }
1814
1815 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1816 {
1817         struct btree_write *w = btree_prev_write(b);
1818         unsigned long old, new, v;
1819         unsigned type = 0;
1820
1821         bch2_btree_complete_write(c, b, w);
1822
1823         v = READ_ONCE(b->flags);
1824         do {
1825                 old = new = v;
1826
1827                 if ((old & (1U << BTREE_NODE_dirty)) &&
1828                     (old & (1U << BTREE_NODE_need_write)) &&
1829                     !(old & (1U << BTREE_NODE_never_write)) &&
1830                     !(old & (1U << BTREE_NODE_write_blocked)) &&
1831                     !(old & (1U << BTREE_NODE_will_make_reachable))) {
1832                         new &= ~(1U << BTREE_NODE_dirty);
1833                         new &= ~(1U << BTREE_NODE_need_write);
1834                         new |=  (1U << BTREE_NODE_write_in_flight);
1835                         new |=  (1U << BTREE_NODE_write_in_flight_inner);
1836                         new |=  (1U << BTREE_NODE_just_written);
1837                         new ^=  (1U << BTREE_NODE_write_idx);
1838
1839                         type = new & BTREE_WRITE_TYPE_MASK;
1840                         new &= ~BTREE_WRITE_TYPE_MASK;
1841                 } else {
1842                         new &= ~(1U << BTREE_NODE_write_in_flight);
1843                         new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1844                 }
1845         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1846
1847         if (new & (1U << BTREE_NODE_write_in_flight))
1848                 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED|type);
1849         else
1850                 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1851 }
1852
1853 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1854 {
1855         struct btree_trans *trans = bch2_trans_get(c);
1856
1857         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
1858         __btree_node_write_done(c, b);
1859         six_unlock_read(&b->c.lock);
1860
1861         bch2_trans_put(trans);
1862 }
1863
1864 static void btree_node_write_work(struct work_struct *work)
1865 {
1866         struct btree_write_bio *wbio =
1867                 container_of(work, struct btree_write_bio, work);
1868         struct bch_fs *c        = wbio->wbio.c;
1869         struct btree *b         = wbio->wbio.bio.bi_private;
1870         int ret = 0;
1871
1872         btree_bounce_free(c,
1873                 wbio->data_bytes,
1874                 wbio->wbio.used_mempool,
1875                 wbio->data);
1876
1877         bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1878                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1879
1880         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key))) {
1881                 ret = -BCH_ERR_btree_node_write_all_failed;
1882                 goto err;
1883         }
1884
1885         if (wbio->wbio.first_btree_write) {
1886                 if (wbio->wbio.failed.nr) {
1887
1888                 }
1889         } else {
1890                 ret = bch2_trans_do(c, NULL, NULL, 0,
1891                         bch2_btree_node_update_key_get_iter(trans, b, &wbio->key,
1892                                         BCH_WATERMARK_interior_updates|
1893                                         BCH_TRANS_COMMIT_journal_reclaim|
1894                                         BCH_TRANS_COMMIT_no_enospc|
1895                                         BCH_TRANS_COMMIT_no_check_rw,
1896                                         !wbio->wbio.failed.nr));
1897                 if (ret)
1898                         goto err;
1899         }
1900 out:
1901         bio_put(&wbio->wbio.bio);
1902         btree_node_write_done(c, b);
1903         return;
1904 err:
1905         set_btree_node_noevict(b);
1906         bch2_fs_fatal_err_on(!bch2_err_matches(ret, EROFS), c,
1907                              "writing btree node: %s", bch2_err_str(ret));
1908         goto out;
1909 }
1910
1911 static void btree_node_write_endio(struct bio *bio)
1912 {
1913         struct bch_write_bio *wbio      = to_wbio(bio);
1914         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1915         struct bch_write_bio *orig      = parent ?: wbio;
1916         struct btree_write_bio *wb      = container_of(orig, struct btree_write_bio, wbio);
1917         struct bch_fs *c                = wbio->c;
1918         struct btree *b                 = wbio->bio.bi_private;
1919         struct bch_dev *ca              = wbio->have_ioref ? bch2_dev_have_ref(c, wbio->dev) : NULL;
1920         unsigned long flags;
1921
1922         if (wbio->have_ioref)
1923                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1924
1925         if (!ca ||
1926             bch2_dev_io_err_on(bio->bi_status, ca, BCH_MEMBER_ERROR_write,
1927                                "btree write error: %s",
1928                                bch2_blk_status_to_str(bio->bi_status)) ||
1929             bch2_meta_write_fault("btree")) {
1930                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1931                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1932                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1933         }
1934
1935         if (wbio->have_ioref)
1936                 percpu_ref_put(&ca->io_ref);
1937
1938         if (parent) {
1939                 bio_put(bio);
1940                 bio_endio(&parent->bio);
1941                 return;
1942         }
1943
1944         clear_btree_node_write_in_flight_inner(b);
1945         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1946         INIT_WORK(&wb->work, btree_node_write_work);
1947         queue_work(c->btree_io_complete_wq, &wb->work);
1948 }
1949
1950 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1951                                    struct bset *i, unsigned sectors)
1952 {
1953         struct printbuf buf = PRINTBUF;
1954         bool saw_error;
1955         int ret;
1956
1957         ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key),
1958                                 BKEY_TYPE_btree, WRITE, &buf);
1959
1960         if (ret)
1961                 bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf);
1962         printbuf_exit(&buf);
1963         if (ret)
1964                 return ret;
1965
1966         ret = validate_bset_keys(c, b, i, WRITE, false, &saw_error) ?:
1967                 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false, &saw_error);
1968         if (ret) {
1969                 bch2_inconsistent_error(c);
1970                 dump_stack();
1971         }
1972
1973         return ret;
1974 }
1975
1976 static void btree_write_submit(struct work_struct *work)
1977 {
1978         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1979         BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1980
1981         bkey_copy(&tmp.k, &wbio->key);
1982
1983         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1984                 ptr->offset += wbio->sector_offset;
1985
1986         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree,
1987                                   &tmp.k, false);
1988 }
1989
1990 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1991 {
1992         struct btree_write_bio *wbio;
1993         struct bset *i;
1994         struct btree_node *bn = NULL;
1995         struct btree_node_entry *bne = NULL;
1996         struct sort_iter_stack sort_iter;
1997         struct nonce nonce;
1998         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1999         u64 seq = 0;
2000         bool used_mempool;
2001         unsigned long old, new;
2002         bool validate_before_checksum = false;
2003         enum btree_write_type type = flags & BTREE_WRITE_TYPE_MASK;
2004         void *data;
2005         int ret;
2006
2007         if (flags & BTREE_WRITE_ALREADY_STARTED)
2008                 goto do_write;
2009
2010         /*
2011          * We may only have a read lock on the btree node - the dirty bit is our
2012          * "lock" against racing with other threads that may be trying to start
2013          * a write, we do a write iff we clear the dirty bit. Since setting the
2014          * dirty bit requires a write lock, we can't race with other threads
2015          * redirtying it:
2016          */
2017         do {
2018                 old = new = READ_ONCE(b->flags);
2019
2020                 if (!(old & (1 << BTREE_NODE_dirty)))
2021                         return;
2022
2023                 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
2024                     !(old & (1 << BTREE_NODE_need_write)))
2025                         return;
2026
2027                 if (old &
2028                     ((1 << BTREE_NODE_never_write)|
2029                      (1 << BTREE_NODE_write_blocked)))
2030                         return;
2031
2032                 if (b->written &&
2033                     (old & (1 << BTREE_NODE_will_make_reachable)))
2034                         return;
2035
2036                 if (old & (1 << BTREE_NODE_write_in_flight))
2037                         return;
2038
2039                 if (flags & BTREE_WRITE_ONLY_IF_NEED)
2040                         type = new & BTREE_WRITE_TYPE_MASK;
2041                 new &= ~BTREE_WRITE_TYPE_MASK;
2042
2043                 new &= ~(1 << BTREE_NODE_dirty);
2044                 new &= ~(1 << BTREE_NODE_need_write);
2045                 new |=  (1 << BTREE_NODE_write_in_flight);
2046                 new |=  (1 << BTREE_NODE_write_in_flight_inner);
2047                 new |=  (1 << BTREE_NODE_just_written);
2048                 new ^=  (1 << BTREE_NODE_write_idx);
2049         } while (cmpxchg_acquire(&b->flags, old, new) != old);
2050
2051         if (new & (1U << BTREE_NODE_need_write))
2052                 return;
2053 do_write:
2054         BUG_ON((type == BTREE_WRITE_initial) != (b->written == 0));
2055
2056         atomic_dec(&c->btree_cache.dirty);
2057
2058         BUG_ON(btree_node_fake(b));
2059         BUG_ON((b->will_make_reachable != 0) != !b->written);
2060
2061         BUG_ON(b->written >= btree_sectors(c));
2062         BUG_ON(b->written & (block_sectors(c) - 1));
2063         BUG_ON(bset_written(b, btree_bset_last(b)));
2064         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
2065         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
2066
2067         bch2_sort_whiteouts(c, b);
2068
2069         sort_iter_stack_init(&sort_iter, b);
2070
2071         bytes = !b->written
2072                 ? sizeof(struct btree_node)
2073                 : sizeof(struct btree_node_entry);
2074
2075         bytes += b->whiteout_u64s * sizeof(u64);
2076
2077         for_each_bset(b, t) {
2078                 i = bset(b, t);
2079
2080                 if (bset_written(b, i))
2081                         continue;
2082
2083                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
2084                 sort_iter_add(&sort_iter.iter,
2085                               btree_bkey_first(b, t),
2086                               btree_bkey_last(b, t));
2087                 seq = max(seq, le64_to_cpu(i->journal_seq));
2088         }
2089
2090         BUG_ON(b->written && !seq);
2091
2092         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
2093         bytes += 8;
2094
2095         /* buffer must be a multiple of the block size */
2096         bytes = round_up(bytes, block_bytes(c));
2097
2098         data = btree_bounce_alloc(c, bytes, &used_mempool);
2099
2100         if (!b->written) {
2101                 bn = data;
2102                 *bn = *b->data;
2103                 i = &bn->keys;
2104         } else {
2105                 bne = data;
2106                 bne->keys = b->data->keys;
2107                 i = &bne->keys;
2108         }
2109
2110         i->journal_seq  = cpu_to_le64(seq);
2111         i->u64s         = 0;
2112
2113         sort_iter_add(&sort_iter.iter,
2114                       unwritten_whiteouts_start(b),
2115                       unwritten_whiteouts_end(b));
2116         SET_BSET_SEPARATE_WHITEOUTS(i, false);
2117
2118         u64s = bch2_sort_keys_keep_unwritten_whiteouts(i->start, &sort_iter.iter);
2119         le16_add_cpu(&i->u64s, u64s);
2120
2121         b->whiteout_u64s = 0;
2122
2123         BUG_ON(!b->written && i->u64s != b->data->keys.u64s);
2124
2125         set_needs_whiteout(i, false);
2126
2127         /* do we have data to write? */
2128         if (b->written && !i->u64s)
2129                 goto nowrite;
2130
2131         bytes_to_write = vstruct_end(i) - data;
2132         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
2133
2134         if (!b->written &&
2135             b->key.k.type == KEY_TYPE_btree_ptr_v2)
2136                 BUG_ON(btree_ptr_sectors_written(&b->key) != sectors_to_write);
2137
2138         memset(data + bytes_to_write, 0,
2139                (sectors_to_write << 9) - bytes_to_write);
2140
2141         BUG_ON(b->written + sectors_to_write > btree_sectors(c));
2142         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
2143         BUG_ON(i->seq != b->data->keys.seq);
2144
2145         i->version = cpu_to_le16(c->sb.version);
2146         SET_BSET_OFFSET(i, b->written);
2147         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
2148
2149         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
2150                 validate_before_checksum = true;
2151
2152         /* validate_bset will be modifying: */
2153         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
2154                 validate_before_checksum = true;
2155
2156         /* if we're going to be encrypting, check metadata validity first: */
2157         if (validate_before_checksum &&
2158             validate_bset_for_write(c, b, i, sectors_to_write))
2159                 goto err;
2160
2161         ret = bset_encrypt(c, i, b->written << 9);
2162         if (bch2_fs_fatal_err_on(ret, c,
2163                         "encrypting btree node: %s", bch2_err_str(ret)))
2164                 goto err;
2165
2166         nonce = btree_nonce(i, b->written << 9);
2167
2168         if (bn)
2169                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
2170         else
2171                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
2172
2173         /* if we're not encrypting, check metadata after checksumming: */
2174         if (!validate_before_checksum &&
2175             validate_bset_for_write(c, b, i, sectors_to_write))
2176                 goto err;
2177
2178         /*
2179          * We handle btree write errors by immediately halting the journal -
2180          * after we've done that, we can't issue any subsequent btree writes
2181          * because they might have pointers to new nodes that failed to write.
2182          *
2183          * Furthermore, there's no point in doing any more btree writes because
2184          * with the journal stopped, we're never going to update the journal to
2185          * reflect that those writes were done and the data flushed from the
2186          * journal:
2187          *
2188          * Also on journal error, the pending write may have updates that were
2189          * never journalled (interior nodes, see btree_update_nodes_written()) -
2190          * it's critical that we don't do the write in that case otherwise we
2191          * will have updates visible that weren't in the journal:
2192          *
2193          * Make sure to update b->written so bch2_btree_init_next() doesn't
2194          * break:
2195          */
2196         if (bch2_journal_error(&c->journal) ||
2197             c->opts.nochanges)
2198                 goto err;
2199
2200         trace_and_count(c, btree_node_write, b, bytes_to_write, sectors_to_write);
2201
2202         wbio = container_of(bio_alloc_bioset(NULL,
2203                                 buf_pages(data, sectors_to_write << 9),
2204                                 REQ_OP_WRITE|REQ_META,
2205                                 GFP_NOFS,
2206                                 &c->btree_bio),
2207                             struct btree_write_bio, wbio.bio);
2208         wbio_init(&wbio->wbio.bio);
2209         wbio->data                      = data;
2210         wbio->data_bytes                = bytes;
2211         wbio->sector_offset             = b->written;
2212         wbio->wbio.c                    = c;
2213         wbio->wbio.used_mempool         = used_mempool;
2214         wbio->wbio.first_btree_write    = !b->written;
2215         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
2216         wbio->wbio.bio.bi_private       = b;
2217
2218         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
2219
2220         bkey_copy(&wbio->key, &b->key);
2221
2222         b->written += sectors_to_write;
2223
2224         if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
2225                 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2226                         cpu_to_le16(b->written);
2227
2228         atomic64_inc(&c->btree_write_stats[type].nr);
2229         atomic64_add(bytes_to_write, &c->btree_write_stats[type].bytes);
2230
2231         INIT_WORK(&wbio->work, btree_write_submit);
2232         queue_work(c->btree_write_submit_wq, &wbio->work);
2233         return;
2234 err:
2235         set_btree_node_noevict(b);
2236         b->written += sectors_to_write;
2237 nowrite:
2238         btree_bounce_free(c, bytes, used_mempool, data);
2239         __btree_node_write_done(c, b);
2240 }
2241
2242 /*
2243  * Work that must be done with write lock held:
2244  */
2245 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2246 {
2247         bool invalidated_iter = false;
2248         struct btree_node_entry *bne;
2249
2250         if (!btree_node_just_written(b))
2251                 return false;
2252
2253         BUG_ON(b->whiteout_u64s);
2254
2255         clear_btree_node_just_written(b);
2256
2257         /*
2258          * Note: immediately after write, bset_written() doesn't work - the
2259          * amount of data we had to write after compaction might have been
2260          * smaller than the offset of the last bset.
2261          *
2262          * However, we know that all bsets have been written here, as long as
2263          * we're still holding the write lock:
2264          */
2265
2266         /*
2267          * XXX: decide if we really want to unconditionally sort down to a
2268          * single bset:
2269          */
2270         if (b->nsets > 1) {
2271                 btree_node_sort(c, b, 0, b->nsets);
2272                 invalidated_iter = true;
2273         } else {
2274                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2275         }
2276
2277         for_each_bset(b, t)
2278                 set_needs_whiteout(bset(b, t), true);
2279
2280         bch2_btree_verify(c, b);
2281
2282         /*
2283          * If later we don't unconditionally sort down to a single bset, we have
2284          * to ensure this is still true:
2285          */
2286         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2287
2288         bne = want_new_bset(c, b);
2289         if (bne)
2290                 bch2_bset_init_next(b, bne);
2291
2292         bch2_btree_build_aux_trees(b);
2293
2294         return invalidated_iter;
2295 }
2296
2297 /*
2298  * Use this one if the node is intent locked:
2299  */
2300 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2301                            enum six_lock_type lock_type_held,
2302                            unsigned flags)
2303 {
2304         if (lock_type_held == SIX_LOCK_intent ||
2305             (lock_type_held == SIX_LOCK_read &&
2306              six_lock_tryupgrade(&b->c.lock))) {
2307                 __bch2_btree_node_write(c, b, flags);
2308
2309                 /* don't cycle lock unnecessarily: */
2310                 if (btree_node_just_written(b) &&
2311                     six_trylock_write(&b->c.lock)) {
2312                         bch2_btree_post_write_cleanup(c, b);
2313                         six_unlock_write(&b->c.lock);
2314                 }
2315
2316                 if (lock_type_held == SIX_LOCK_read)
2317                         six_lock_downgrade(&b->c.lock);
2318         } else {
2319                 __bch2_btree_node_write(c, b, flags);
2320                 if (lock_type_held == SIX_LOCK_write &&
2321                     btree_node_just_written(b))
2322                         bch2_btree_post_write_cleanup(c, b);
2323         }
2324 }
2325
2326 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2327 {
2328         struct bucket_table *tbl;
2329         struct rhash_head *pos;
2330         struct btree *b;
2331         unsigned i;
2332         bool ret = false;
2333 restart:
2334         rcu_read_lock();
2335         for_each_cached_btree(b, c, tbl, i, pos)
2336                 if (test_bit(flag, &b->flags)) {
2337                         rcu_read_unlock();
2338                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2339                         ret = true;
2340                         goto restart;
2341                 }
2342         rcu_read_unlock();
2343
2344         return ret;
2345 }
2346
2347 bool bch2_btree_flush_all_reads(struct bch_fs *c)
2348 {
2349         return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2350 }
2351
2352 bool bch2_btree_flush_all_writes(struct bch_fs *c)
2353 {
2354         return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2355 }
2356
2357 static const char * const bch2_btree_write_types[] = {
2358 #define x(t, n) [n] = #t,
2359         BCH_BTREE_WRITE_TYPES()
2360         NULL
2361 };
2362
2363 void bch2_btree_write_stats_to_text(struct printbuf *out, struct bch_fs *c)
2364 {
2365         printbuf_tabstop_push(out, 20);
2366         printbuf_tabstop_push(out, 10);
2367
2368         prt_printf(out, "\tnr\tsize\n");
2369
2370         for (unsigned i = 0; i < BTREE_WRITE_TYPE_NR; i++) {
2371                 u64 nr          = atomic64_read(&c->btree_write_stats[i].nr);
2372                 u64 bytes       = atomic64_read(&c->btree_write_stats[i].bytes);
2373
2374                 prt_printf(out, "%s:\t%llu\t", bch2_btree_write_types[i], nr);
2375                 prt_human_readable_u64(out, nr ? div64_u64(bytes, nr) : 0);
2376                 prt_newline(out);
2377         }
2378 }