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