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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_methods.h"
   5 #include "bkey_sort.h"
   6 #include "btree_cache.h"
   7 #include "btree_io.h"
   8 #include "btree_iter.h"
   9 #include "btree_locking.h"
  10 #include "btree_update.h"
  11 #include "btree_update_interior.h"
  12 #include "buckets.h"
  13 #include "checksum.h"
  14 #include "debug.h"
  15 #include "error.h"
  16 #include "extents.h"
  17 #include "io.h"
  18 #include "journal_reclaim.h"
  19 #include "journal_seq_blacklist.h"
  20 #include "super-io.h"
  21
  22 #include <linux/sched/mm.h>
  23 #include <trace/events/bcachefs.h>
  24
  25 void bch2_btree_node_io_unlock(struct btree *b)
  26 {
  27         EBUG_ON(!btree_node_write_in_flight(b));
  28
  29         clear_btree_node_write_in_flight_inner(b);
  30         clear_btree_node_write_in_flight(b);
  31         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
  32 }
  33
  34 void bch2_btree_node_io_lock(struct btree *b)
  35 {
  36         bch2_assert_btree_nodes_not_locked();
  37
  38         wait_on_bit_lock_io(&b->flags, BTREE_NODE_write_in_flight,
  39                             TASK_UNINTERRUPTIBLE);
  40 }
  41
  42 void __bch2_btree_node_wait_on_read(struct btree *b)
  43 {
  44         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
  45                        TASK_UNINTERRUPTIBLE);
  46 }
  47
  48 void __bch2_btree_node_wait_on_write(struct btree *b)
  49 {
  50         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
  51                        TASK_UNINTERRUPTIBLE);
  52 }
  53
  54 void bch2_btree_node_wait_on_read(struct btree *b)
  55 {
  56         bch2_assert_btree_nodes_not_locked();
  57
  58         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
  59                        TASK_UNINTERRUPTIBLE);
  60 }
  61
  62 void bch2_btree_node_wait_on_write(struct btree *b)
  63 {
  64         bch2_assert_btree_nodes_not_locked();
  65
  66         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
  67                        TASK_UNINTERRUPTIBLE);
  68 }
  69
  70 static void verify_no_dups(struct btree *b,
  71                            struct bkey_packed *start,
  72                            struct bkey_packed *end)
  73 {
  74 #ifdef CONFIG_BCACHEFS_DEBUG
  75         struct bkey_packed *k, *p;
  76
  77         if (start == end)
  78                 return;
  79
  80         for (p = start, k = bkey_p_next(start);
  81              k != end;
  82              p = k, k = bkey_p_next(k)) {
  83                 struct bkey l = bkey_unpack_key(b, p);
  84                 struct bkey r = bkey_unpack_key(b, k);
  85
  86                 BUG_ON(bpos_ge(l.p, bkey_start_pos(&r)));
  87         }
  88 #endif
  89 }
  90
  91 static void set_needs_whiteout(struct bset *i, int v)
  92 {
  93         struct bkey_packed *k;
  94
  95         for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
  96                 k->needs_whiteout = v;
  97 }
  98
  99 static void btree_bounce_free(struct bch_fs *c, size_t size,
 100                               bool used_mempool, void *p)
 101 {
 102         if (used_mempool)
 103                 mempool_free(p, &c->btree_bounce_pool);
 104         else
 105                 vpfree(p, size);
 106 }
 107
 108 static void *_btree_bounce_alloc(struct bch_fs *c, size_t size,
 109                                  bool *used_mempool)
 110 {
 111         unsigned flags = memalloc_nofs_save();
 112         void *p;
 113
 114         BUG_ON(size > btree_bytes(c));
 115
 116         *used_mempool = false;
 117         p = _vpmalloc(size, __GFP_NOWARN|GFP_NOWAIT);
 118         if (!p) {
 119                 *used_mempool = true;
 120                 p = mempool_alloc(&c->btree_bounce_pool, GFP_NOIO);
 121         }
 122         memalloc_nofs_restore(flags);
 123         return p;
 124 }
 125 #define btree_bounce_alloc(_c, _size, _used_mempool)            \
 126         alloc_hooks(_btree_bounce_alloc(_c, _size, _used_mempool), void *, NULL)
 127
 128 static void sort_bkey_ptrs(const struct btree *bt,
 129                            struct bkey_packed **ptrs, unsigned nr)
 130 {
 131         unsigned n = nr, a = nr / 2, b, c, d;
 132
 133         if (!a)
 134                 return;
 135
 136         /* Heap sort: see lib/sort.c: */
 137         while (1) {
 138                 if (a)
 139                         a--;
 140                 else if (--n)
 141                         swap(ptrs[0], ptrs[n]);
 142                 else
 143                         break;
 144
 145                 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
 146                         b = bch2_bkey_cmp_packed(bt,
 147                                             ptrs[c],
 148                                             ptrs[d]) >= 0 ? c : d;
 149                 if (d == n)
 150                         b = c;
 151
 152                 while (b != a &&
 153                        bch2_bkey_cmp_packed(bt,
 154                                        ptrs[a],
 155                                        ptrs[b]) >= 0)
 156                         b = (b - 1) / 2;
 157                 c = b;
 158                 while (b != a) {
 159                         b = (b - 1) / 2;
 160                         swap(ptrs[b], ptrs[c]);
 161                 }
 162         }
 163 }
 164
 165 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
 166 {
 167         struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
 168         bool used_mempool = false;
 169         size_t bytes = b->whiteout_u64s * sizeof(u64);
 170
 171         if (!b->whiteout_u64s)
 172                 return;
 173
 174         new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
 175
 176         ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
 177
 178         for (k = unwritten_whiteouts_start(c, b);
 179              k != unwritten_whiteouts_end(c, b);
 180              k = bkey_p_next(k))
 181                 *--ptrs = k;
 182
 183         sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
 184
 185         k = new_whiteouts;
 186
 187         while (ptrs != ptrs_end) {
 188                 bkey_copy(k, *ptrs);
 189                 k = bkey_p_next(k);
 190                 ptrs++;
 191         }
 192
 193         verify_no_dups(b, new_whiteouts,
 194                        (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
 195
 196         memcpy_u64s(unwritten_whiteouts_start(c, b),
 197                     new_whiteouts, b->whiteout_u64s);
 198
 199         btree_bounce_free(c, bytes, used_mempool, new_whiteouts);
 200 }
 201
 202 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
 203                                 bool compacting, enum compact_mode mode)
 204 {
 205         if (!bset_dead_u64s(b, t))
 206                 return false;
 207
 208         switch (mode) {
 209         case COMPACT_LAZY:
 210                 return should_compact_bset_lazy(b, t) ||
 211                         (compacting && !bset_written(b, bset(b, t)));
 212         case COMPACT_ALL:
 213                 return true;
 214         default:
 215                 BUG();
 216         }
 217 }
 218
 219 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
 220 {
 221         struct bset_tree *t;
 222         bool ret = false;
 223
 224         for_each_bset(b, t) {
 225                 struct bset *i = bset(b, t);
 226                 struct bkey_packed *k, *n, *out, *start, *end;
 227                 struct btree_node_entry *src = NULL, *dst = NULL;
 228
 229                 if (t != b->set && !bset_written(b, i)) {
 230                         src = container_of(i, struct btree_node_entry, keys);
 231                         dst = max(write_block(b),
 232                                   (void *) btree_bkey_last(b, t - 1));
 233                 }
 234
 235                 if (src != dst)
 236                         ret = true;
 237
 238                 if (!should_compact_bset(b, t, ret, mode)) {
 239                         if (src != dst) {
 240                                 memmove(dst, src, sizeof(*src) +
 241                                         le16_to_cpu(src->keys.u64s) *
 242                                         sizeof(u64));
 243                                 i = &dst->keys;
 244                                 set_btree_bset(b, t, i);
 245                         }
 246                         continue;
 247                 }
 248
 249                 start   = btree_bkey_first(b, t);
 250                 end     = btree_bkey_last(b, t);
 251
 252                 if (src != dst) {
 253                         memmove(dst, src, sizeof(*src));
 254                         i = &dst->keys;
 255                         set_btree_bset(b, t, i);
 256                 }
 257
 258                 out = i->start;
 259
 260                 for (k = start; k != end; k = n) {
 261                         n = bkey_p_next(k);
 262
 263                         if (!bkey_deleted(k)) {
 264                                 bkey_copy(out, k);
 265                                 out = bkey_p_next(out);
 266                         } else {
 267                                 BUG_ON(k->needs_whiteout);
 268                         }
 269                 }
 270
 271                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 272                 set_btree_bset_end(b, t);
 273                 bch2_bset_set_no_aux_tree(b, t);
 274                 ret = true;
 275         }
 276
 277         bch2_verify_btree_nr_keys(b);
 278
 279         bch2_btree_build_aux_trees(b);
 280
 281         return ret;
 282 }
 283
 284 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 285                             enum compact_mode mode)
 286 {
 287         return bch2_drop_whiteouts(b, mode);
 288 }
 289
 290 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 291                             unsigned start_idx,
 292                             unsigned end_idx,
 293                             bool filter_whiteouts)
 294 {
 295         struct btree_node *out;
 296         struct sort_iter sort_iter;
 297         struct bset_tree *t;
 298         struct bset *start_bset = bset(b, &b->set[start_idx]);
 299         bool used_mempool = false;
 300         u64 start_time, seq = 0;
 301         unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
 302         bool sorting_entire_node = start_idx == 0 &&
 303                 end_idx == b->nsets;
 304
 305         sort_iter_init(&sort_iter, b);
 306
 307         for (t = b->set + start_idx;
 308              t < b->set + end_idx;
 309              t++) {
 310                 u64s += le16_to_cpu(bset(b, t)->u64s);
 311                 sort_iter_add(&sort_iter,
 312                               btree_bkey_first(b, t),
 313                               btree_bkey_last(b, t));
 314         }
 315
 316         bytes = sorting_entire_node
 317                 ? btree_bytes(c)
 318                 : __vstruct_bytes(struct btree_node, u64s);
 319
 320         out = btree_bounce_alloc(c, bytes, &used_mempool);
 321
 322         start_time = local_clock();
 323
 324         u64s = bch2_sort_keys(out->keys.start, &sort_iter, filter_whiteouts);
 325
 326         out->keys.u64s = cpu_to_le16(u64s);
 327
 328         BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
 329
 330         if (sorting_entire_node)
 331                 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 332                                        start_time);
 333
 334         /* Make sure we preserve bset journal_seq: */
 335         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 336                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 337         start_bset->journal_seq = cpu_to_le64(seq);
 338
 339         if (sorting_entire_node) {
 340                 unsigned u64s = le16_to_cpu(out->keys.u64s);
 341
 342                 BUG_ON(bytes != btree_bytes(c));
 343
 344                 /*
 345                  * Our temporary buffer is the same size as the btree node's
 346                  * buffer, we can just swap buffers instead of doing a big
 347                  * memcpy()
 348                  */
 349                 *out = *b->data;
 350                 out->keys.u64s = cpu_to_le16(u64s);
 351                 swap(out, b->data);
 352                 set_btree_bset(b, b->set, &b->data->keys);
 353         } else {
 354                 start_bset->u64s = out->keys.u64s;
 355                 memcpy_u64s(start_bset->start,
 356                             out->keys.start,
 357                             le16_to_cpu(out->keys.u64s));
 358         }
 359
 360         for (i = start_idx + 1; i < end_idx; i++)
 361                 b->nr.bset_u64s[start_idx] +=
 362                         b->nr.bset_u64s[i];
 363
 364         b->nsets -= shift;
 365
 366         for (i = start_idx + 1; i < b->nsets; i++) {
 367                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 368                 b->set[i]               = b->set[i + shift];
 369         }
 370
 371         for (i = b->nsets; i < MAX_BSETS; i++)
 372                 b->nr.bset_u64s[i] = 0;
 373
 374         set_btree_bset_end(b, &b->set[start_idx]);
 375         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 376
 377         btree_bounce_free(c, bytes, used_mempool, out);
 378
 379         bch2_verify_btree_nr_keys(b);
 380 }
 381
 382 void bch2_btree_sort_into(struct bch_fs *c,
 383                          struct btree *dst,
 384                          struct btree *src)
 385 {
 386         struct btree_nr_keys nr;
 387         struct btree_node_iter src_iter;
 388         u64 start_time = local_clock();
 389
 390         BUG_ON(dst->nsets != 1);
 391
 392         bch2_bset_set_no_aux_tree(dst, dst->set);
 393
 394         bch2_btree_node_iter_init_from_start(&src_iter, src);
 395
 396         nr = bch2_sort_repack(btree_bset_first(dst),
 397                         src, &src_iter,
 398                         &dst->format,
 399                         true);
 400
 401         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 402                                start_time);
 403
 404         set_btree_bset_end(dst, dst->set);
 405
 406         dst->nr.live_u64s       += nr.live_u64s;
 407         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 408         dst->nr.packed_keys     += nr.packed_keys;
 409         dst->nr.unpacked_keys   += nr.unpacked_keys;
 410
 411         bch2_verify_btree_nr_keys(dst);
 412 }
 413
 414 #define SORT_CRIT       (4096 / sizeof(u64))
 415
 416 /*
 417  * We're about to add another bset to the btree node, so if there's currently
 418  * too many bsets - sort some of them together:
 419  */
 420 static bool btree_node_compact(struct bch_fs *c, struct btree *b)
 421 {
 422         unsigned unwritten_idx;
 423         bool ret = false;
 424
 425         for (unwritten_idx = 0;
 426              unwritten_idx < b->nsets;
 427              unwritten_idx++)
 428                 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
 429                         break;
 430
 431         if (b->nsets - unwritten_idx > 1) {
 432                 btree_node_sort(c, b, unwritten_idx,
 433                                 b->nsets, false);
 434                 ret = true;
 435         }
 436
 437         if (unwritten_idx > 1) {
 438                 btree_node_sort(c, b, 0, unwritten_idx, false);
 439                 ret = true;
 440         }
 441
 442         return ret;
 443 }
 444
 445 void bch2_btree_build_aux_trees(struct btree *b)
 446 {
 447         struct bset_tree *t;
 448
 449         for_each_bset(b, t)
 450                 bch2_bset_build_aux_tree(b, t,
 451                                 !bset_written(b, bset(b, t)) &&
 452                                 t == bset_tree_last(b));
 453 }
 454
 455 /*
 456  * If we have MAX_BSETS (3) bsets, should we sort them all down to just one?
 457  *
 458  * The first bset is going to be of similar order to the size of the node, the
 459  * last bset is bounded by btree_write_set_buffer(), which is set to keep the
 460  * memmove on insert from being too expensive: the middle bset should, ideally,
 461  * be the geometric mean of the first and the last.
 462  *
 463  * Returns true if the middle bset is greater than that geometric mean:
 464  */
 465 static inline bool should_compact_all(struct bch_fs *c, struct btree *b)
 466 {
 467         unsigned mid_u64s_bits =
 468                 (ilog2(btree_max_u64s(c)) + BTREE_WRITE_SET_U64s_BITS) / 2;
 469
 470         return bset_u64s(&b->set[1]) > 1U << mid_u64s_bits;
 471 }
 472
 473 /*
 474  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 475  * inserted into
 476  *
 477  * Safe to call if there already is an unwritten bset - will only add a new bset
 478  * if @b doesn't already have one.
 479  *
 480  * Returns true if we sorted (i.e. invalidated iterators
 481  */
 482 void bch2_btree_init_next(struct btree_trans *trans, struct btree *b)
 483 {
 484         struct bch_fs *c = trans->c;
 485         struct btree_node_entry *bne;
 486         bool reinit_iter = false;
 487
 488         EBUG_ON(!(b->c.lock.state.seq & 1));
 489         BUG_ON(bset_written(b, bset(b, &b->set[1])));
 490         BUG_ON(btree_node_just_written(b));
 491
 492         if (b->nsets == MAX_BSETS &&
 493             !btree_node_write_in_flight(b) &&
 494             should_compact_all(c, b)) {
 495                 bch2_btree_node_write(c, b, SIX_LOCK_write,
 496                                       BTREE_WRITE_init_next_bset);
 497                 reinit_iter = true;
 498         }
 499
 500         if (b->nsets == MAX_BSETS &&
 501             btree_node_compact(c, b))
 502                 reinit_iter = true;
 503
 504         BUG_ON(b->nsets >= MAX_BSETS);
 505
 506         bne = want_new_bset(c, b);
 507         if (bne)
 508                 bch2_bset_init_next(c, b, bne);
 509
 510         bch2_btree_build_aux_trees(b);
 511
 512         if (reinit_iter)
 513                 bch2_trans_node_reinit_iter(trans, b);
 514 }
 515
 516 static void btree_pos_to_text(struct printbuf *out, struct bch_fs *c,
 517                           struct btree *b)
 518 {
 519         prt_printf(out, "%s level %u/%u\n  ",
 520                bch2_btree_ids[b->c.btree_id],
 521                b->c.level,
 522                c->btree_roots[b->c.btree_id].level);
 523         bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 524 }
 525
 526 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 527                           struct bch_dev *ca,
 528                           struct btree *b, struct bset *i,
 529                           unsigned offset, int write)
 530 {
 531         prt_printf(out, bch2_log_msg(c, "%s"),
 532                    write == READ
 533                    ? "error validating btree node "
 534                    : "corrupt btree node before write ");
 535         if (ca)
 536                 prt_printf(out, "on %s ", ca->name);
 537         prt_printf(out, "at btree ");
 538         btree_pos_to_text(out, c, b);
 539
 540         prt_printf(out, "\n  node offset %u", b->written);
 541         if (i)
 542                 prt_printf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 543         prt_str(out, ": ");
 544 }
 545
 546 enum btree_err_type {
 547         /*
 548          * We can repair this locally, and we're after the checksum check so
 549          * there's no need to try another replica:
 550          */
 551         BTREE_ERR_FIXABLE,
 552         /*
 553          * We can repair this if we have to, but we should try reading another
 554          * replica if we can:
 555          */
 556         BTREE_ERR_WANT_RETRY,
 557         /*
 558          * Read another replica if we have one, otherwise consider the whole
 559          * node bad:
 560          */
 561         BTREE_ERR_MUST_RETRY,
 562         BTREE_ERR_BAD_NODE,
 563         BTREE_ERR_INCOMPATIBLE,
 564 };
 565
 566 enum btree_validate_ret {
 567         BTREE_RETRY_READ = 64,
 568 };
 569
 570 static int __btree_err(enum btree_err_type type,
 571                        struct bch_fs *c,
 572                        struct bch_dev *ca,
 573                        struct btree *b,
 574                        struct bset *i,
 575                        int write,
 576                        bool have_retry,
 577                        const char *fmt, ...)
 578 {
 579         struct printbuf out = PRINTBUF;
 580         va_list args;
 581         int ret = -BCH_ERR_fsck_fix;
 582
 583         btree_err_msg(&out, c, ca, b, i, b->written, write);
 584
 585         va_start(args, fmt);
 586         prt_vprintf(&out, fmt, args);
 587         va_end(args);
 588
 589         if (write == WRITE) {
 590                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 591                 ret = c->opts.errors == BCH_ON_ERROR_continue
 592                         ? 0
 593                         : -BCH_ERR_fsck_errors_not_fixed;
 594                 goto out;
 595         }
 596
 597         if (!have_retry && type == BTREE_ERR_WANT_RETRY)
 598                 type = BTREE_ERR_FIXABLE;
 599         if (!have_retry && type == BTREE_ERR_MUST_RETRY)
 600                 type = BTREE_ERR_BAD_NODE;
 601
 602         switch (type) {
 603         case BTREE_ERR_FIXABLE:
 604                 mustfix_fsck_err(c, "%s", out.buf);
 605                 ret = -BCH_ERR_fsck_fix;
 606                 break;
 607         case BTREE_ERR_WANT_RETRY:
 608         case BTREE_ERR_MUST_RETRY:
 609                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 610                 ret = BTREE_RETRY_READ;
 611                 break;
 612         case BTREE_ERR_BAD_NODE:
 613                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 614                 bch2_topology_error(c);
 615                 ret = -BCH_ERR_need_topology_repair;
 616                 break;
 617         case BTREE_ERR_INCOMPATIBLE:
 618                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 619                 ret = -BCH_ERR_fsck_errors_not_fixed;
 620                 break;
 621         default:
 622                 BUG();
 623         }
 624 out:
 625 fsck_err:
 626         printbuf_exit(&out);
 627         return ret;
 628 }
 629
 630 #define btree_err(type, c, ca, b, i, msg, ...)                          \
 631 ({                                                                      \
 632         int _ret = __btree_err(type, c, ca, b, i, write, have_retry, msg, ##__VA_ARGS__);\
 633                                                                         \
 634         if (_ret != -BCH_ERR_fsck_fix)                                  \
 635                 goto fsck_err;                                          \
 636         *saw_error = true;                                              \
 637 })
 638
 639 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 640
 641 /*
 642  * When btree topology repair changes the start or end of a node, that might
 643  * mean we have to drop keys that are no longer inside the node:
 644  */
 645 __cold
 646 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
 647 {
 648         struct bset_tree *t;
 649         struct bkey_s_c k;
 650         struct bkey unpacked;
 651         struct btree_node_iter iter;
 652
 653         for_each_bset(b, t) {
 654                 struct bset *i = bset(b, t);
 655                 struct bkey_packed *k;
 656
 657                 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
 658                         if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
 659                                 break;
 660
 661                 if (k != i->start) {
 662                         unsigned shift = (u64 *) k - (u64 *) i->start;
 663
 664                         memmove_u64s_down(i->start, k,
 665                                           (u64 *) vstruct_end(i) - (u64 *) k);
 666                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
 667                         set_btree_bset_end(b, t);
 668                 }
 669
 670                 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
 671                         if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
 672                                 break;
 673
 674                 if (k != vstruct_last(i)) {
 675                         i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
 676                         set_btree_bset_end(b, t);
 677                 }
 678         }
 679
 680         /*
 681          * Always rebuild search trees: eytzinger search tree nodes directly
 682          * depend on the values of min/max key:
 683          */
 684         bch2_bset_set_no_aux_tree(b, b->set);
 685         bch2_btree_build_aux_trees(b);
 686
 687         for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
 688                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 689                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 690         }
 691 }
 692
 693 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 694                          struct btree *b, struct bset *i,
 695                          unsigned offset, unsigned sectors,
 696                          int write, bool have_retry, bool *saw_error)
 697 {
 698         unsigned version = le16_to_cpu(i->version);
 699         const char *err;
 700         struct printbuf buf1 = PRINTBUF;
 701         struct printbuf buf2 = PRINTBUF;
 702         int ret = 0;
 703
 704         btree_err_on((version != BCH_BSET_VERSION_OLD &&
 705                       version < bcachefs_metadata_version_min) ||
 706                      version >= bcachefs_metadata_version_max,
 707                      BTREE_ERR_INCOMPATIBLE, c, ca, b, i,
 708                      "unsupported bset version");
 709
 710         if (btree_err_on(version < c->sb.version_min,
 711                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 712                          "bset version %u older than superblock version_min %u",
 713                          version, c->sb.version_min)) {
 714                 mutex_lock(&c->sb_lock);
 715                 c->disk_sb.sb->version_min = cpu_to_le16(version);
 716                 bch2_write_super(c);
 717                 mutex_unlock(&c->sb_lock);
 718         }
 719
 720         if (btree_err_on(version > c->sb.version,
 721                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 722                          "bset version %u newer than superblock version %u",
 723                          version, c->sb.version)) {
 724                 mutex_lock(&c->sb_lock);
 725                 c->disk_sb.sb->version = cpu_to_le16(version);
 726                 bch2_write_super(c);
 727                 mutex_unlock(&c->sb_lock);
 728         }
 729
 730         btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
 731                      BTREE_ERR_INCOMPATIBLE, c, ca, b, i,
 732                      "BSET_SEPARATE_WHITEOUTS no longer supported");
 733
 734         if (btree_err_on(offset + sectors > btree_sectors(c),
 735                          BTREE_ERR_FIXABLE, c, ca, b, i,
 736                          "bset past end of btree node")) {
 737                 i->u64s = 0;
 738                 ret = 0;
 739                 goto out;
 740         }
 741
 742         btree_err_on(offset && !i->u64s,
 743                      BTREE_ERR_FIXABLE, c, ca, b, i,
 744                      "empty bset");
 745
 746         btree_err_on(BSET_OFFSET(i) &&
 747                      BSET_OFFSET(i) != offset,
 748                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 749                      "bset at wrong sector offset");
 750
 751         if (!offset) {
 752                 struct btree_node *bn =
 753                         container_of(i, struct btree_node, keys);
 754                 /* These indicate that we read the wrong btree node: */
 755
 756                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 757                         struct bch_btree_ptr_v2 *bp =
 758                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 759
 760                         /* XXX endianness */
 761                         btree_err_on(bp->seq != bn->keys.seq,
 762                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 763                                      "incorrect sequence number (wrong btree node)");
 764                 }
 765
 766                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 767                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 768                              "incorrect btree id");
 769
 770                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 771                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 772                              "incorrect level");
 773
 774                 if (!write)
 775                         compat_btree_node(b->c.level, b->c.btree_id, version,
 776                                           BSET_BIG_ENDIAN(i), write, bn);
 777
 778                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 779                         struct bch_btree_ptr_v2 *bp =
 780                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 781
 782                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 783                                 b->data->min_key = bp->min_key;
 784                                 b->data->max_key = b->key.k.p;
 785                         }
 786
 787                         btree_err_on(!bpos_eq(b->data->min_key, bp->min_key),
 788                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 789                                      "incorrect min_key: got %s should be %s",
 790                                      (printbuf_reset(&buf1),
 791                                       bch2_bpos_to_text(&buf1, bn->min_key), buf1.buf),
 792                                      (printbuf_reset(&buf2),
 793                                       bch2_bpos_to_text(&buf2, bp->min_key), buf2.buf));
 794                 }
 795
 796                 btree_err_on(!bpos_eq(bn->max_key, b->key.k.p),
 797                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 798                              "incorrect max key %s",
 799                              (printbuf_reset(&buf1),
 800                               bch2_bpos_to_text(&buf1, bn->max_key), buf1.buf));
 801
 802                 if (write)
 803                         compat_btree_node(b->c.level, b->c.btree_id, version,
 804                                           BSET_BIG_ENDIAN(i), write, bn);
 805
 806                 err = bch2_bkey_format_validate(&bn->format);
 807                 btree_err_on(err,
 808                              BTREE_ERR_BAD_NODE, c, ca, b, i,
 809                              "invalid bkey format: %s", err);
 810
 811                 compat_bformat(b->c.level, b->c.btree_id, version,
 812                                BSET_BIG_ENDIAN(i), write,
 813                                &bn->format);
 814         }
 815 out:
 816 fsck_err:
 817         printbuf_exit(&buf2);
 818         printbuf_exit(&buf1);
 819         return ret;
 820 }
 821
 822 static int bset_key_invalid(struct bch_fs *c, struct btree *b,
 823                             struct bkey_s_c k,
 824                             bool updated_range, int rw,
 825                             struct printbuf *err)
 826 {
 827         return __bch2_bkey_invalid(c, k, btree_node_type(b), READ, err) ?:
 828                 (!updated_range ? bch2_bkey_in_btree_node(b, k, err) : 0) ?:
 829                 (rw == WRITE ? bch2_bkey_val_invalid(c, k, READ, err) : 0);
 830 }
 831
 832 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 833                          struct bset *i, int write,
 834                          bool have_retry, bool *saw_error)
 835 {
 836         unsigned version = le16_to_cpu(i->version);
 837         struct bkey_packed *k, *prev = NULL;
 838         struct printbuf buf = PRINTBUF;
 839         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 840                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 841         int ret = 0;
 842
 843         for (k = i->start;
 844              k != vstruct_last(i);) {
 845                 struct bkey_s u;
 846                 struct bkey tmp;
 847
 848                 if (btree_err_on(bkey_p_next(k) > vstruct_last(i),
 849                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 850                                  "key extends past end of bset")) {
 851                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 852                         break;
 853                 }
 854
 855                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 856                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 857                                  "invalid bkey format %u", k->format)) {
 858                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 859                         memmove_u64s_down(k, bkey_p_next(k),
 860                                           (u64 *) vstruct_end(i) - (u64 *) k);
 861                         continue;
 862                 }
 863
 864                 /* XXX: validate k->u64s */
 865                 if (!write)
 866                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 867                                     BSET_BIG_ENDIAN(i), write,
 868                                     &b->format, k);
 869
 870                 u = __bkey_disassemble(b, k, &tmp);
 871
 872                 printbuf_reset(&buf);
 873                 if (bset_key_invalid(c, b, u.s_c, updated_range, write, &buf)) {
 874                         printbuf_reset(&buf);
 875                         prt_printf(&buf, "invalid bkey:  ");
 876                         bset_key_invalid(c, b, u.s_c, updated_range, write, &buf);
 877                         prt_printf(&buf, "\n  ");
 878                         bch2_bkey_val_to_text(&buf, c, u.s_c);
 879
 880                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf);
 881
 882                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 883                         memmove_u64s_down(k, bkey_p_next(k),
 884                                           (u64 *) vstruct_end(i) - (u64 *) k);
 885                         continue;
 886                 }
 887
 888                 if (write)
 889                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 890                                     BSET_BIG_ENDIAN(i), write,
 891                                     &b->format, k);
 892
 893                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 894                         struct bkey up = bkey_unpack_key(b, prev);
 895
 896                         printbuf_reset(&buf);
 897                         prt_printf(&buf, "keys out of order: ");
 898                         bch2_bkey_to_text(&buf, &up);
 899                         prt_printf(&buf, " > ");
 900                         bch2_bkey_to_text(&buf, u.k);
 901
 902                         bch2_dump_bset(c, b, i, 0);
 903
 904                         if (btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf)) {
 905                                 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 906                                 memmove_u64s_down(k, bkey_p_next(k),
 907                                                   (u64 *) vstruct_end(i) - (u64 *) k);
 908                                 continue;
 909                         }
 910                 }
 911
 912                 prev = k;
 913                 k = bkey_p_next(k);
 914         }
 915 fsck_err:
 916         printbuf_exit(&buf);
 917         return ret;
 918 }
 919
 920 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 921                               struct btree *b, bool have_retry, bool *saw_error)
 922 {
 923         struct btree_node_entry *bne;
 924         struct sort_iter *iter;
 925         struct btree_node *sorted;
 926         struct bkey_packed *k;
 927         struct bch_extent_ptr *ptr;
 928         struct bset *i;
 929         bool used_mempool, blacklisted;
 930         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 931                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 932         unsigned u64s;
 933         unsigned blacklisted_written, nonblacklisted_written = 0;
 934         unsigned ptr_written = btree_ptr_sectors_written(&b->key);
 935         struct printbuf buf = PRINTBUF;
 936         int ret = 0, retry_read = 0, write = READ;
 937
 938         b->version_ondisk = U16_MAX;
 939         /* We might get called multiple times on read retry: */
 940         b->written = 0;
 941
 942         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
 943         sort_iter_init(iter, b);
 944         iter->size = (btree_blocks(c) + 1) * 2;
 945
 946         if (bch2_meta_read_fault("btree"))
 947                 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 948                           "dynamic fault");
 949
 950         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 951                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 952                      "bad magic: want %llx, got %llx",
 953                      bset_magic(c), le64_to_cpu(b->data->magic));
 954
 955         btree_err_on(!b->data->keys.seq,
 956                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 957                      "bad btree header: seq 0");
 958
 959         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 960                 struct bch_btree_ptr_v2 *bp =
 961                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 962
 963                 btree_err_on(b->data->keys.seq != bp->seq,
 964                              BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 965                              "got wrong btree node (seq %llx want %llx)",
 966                              b->data->keys.seq, bp->seq);
 967         }
 968
 969         while (b->written < (ptr_written ?: btree_sectors(c))) {
 970                 unsigned sectors;
 971                 struct nonce nonce;
 972                 struct bch_csum csum;
 973                 bool first = !b->written;
 974
 975                 if (!b->written) {
 976                         i = &b->data->keys;
 977
 978                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 979                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 980                                      "unknown checksum type %llu",
 981                                      BSET_CSUM_TYPE(i));
 982
 983                         nonce = btree_nonce(i, b->written << 9);
 984                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 985
 986                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 987                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 988                                      "invalid checksum");
 989
 990                         ret = bset_encrypt(c, i, b->written << 9);
 991                         if (bch2_fs_fatal_err_on(ret, c,
 992                                         "error decrypting btree node: %i", ret))
 993                                 goto fsck_err;
 994
 995                         btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
 996                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
 997                                      BTREE_ERR_INCOMPATIBLE, c, NULL, b, NULL,
 998                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
 999
1000                         sectors = vstruct_sectors(b->data, c->block_bits);
1001                 } else {
1002                         bne = write_block(b);
1003                         i = &bne->keys;
1004
1005                         if (i->seq != b->data->keys.seq)
1006                                 break;
1007
1008                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1009                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
1010                                      "unknown checksum type %llu",
1011                                      BSET_CSUM_TYPE(i));
1012
1013                         nonce = btree_nonce(i, b->written << 9);
1014                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1015
1016                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
1017                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
1018                                      "invalid checksum");
1019
1020                         ret = bset_encrypt(c, i, b->written << 9);
1021                         if (bch2_fs_fatal_err_on(ret, c,
1022                                         "error decrypting btree node: %i\n", ret))
1023                                 goto fsck_err;
1024
1025                         sectors = vstruct_sectors(bne, c->block_bits);
1026                 }
1027
1028                 b->version_ondisk = min(b->version_ondisk,
1029                                         le16_to_cpu(i->version));
1030
1031                 ret = validate_bset(c, ca, b, i, b->written, sectors,
1032                                     READ, have_retry, saw_error);
1033                 if (ret)
1034                         goto fsck_err;
1035
1036                 if (!b->written)
1037                         btree_node_set_format(b, b->data->format);
1038
1039                 ret = validate_bset_keys(c, b, i, READ, have_retry, saw_error);
1040                 if (ret)
1041                         goto fsck_err;
1042
1043                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1044
1045                 blacklisted = bch2_journal_seq_is_blacklisted(c,
1046                                         le64_to_cpu(i->journal_seq),
1047                                         true);
1048
1049                 btree_err_on(blacklisted && first,
1050                              BTREE_ERR_FIXABLE, c, ca, b, i,
1051                              "first btree node bset has blacklisted journal seq (%llu)",
1052                              le64_to_cpu(i->journal_seq));
1053
1054                 btree_err_on(blacklisted && ptr_written,
1055                              BTREE_ERR_FIXABLE, c, ca, b, i,
1056                              "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
1057                              le64_to_cpu(i->journal_seq),
1058                              b->written, b->written + sectors, ptr_written);
1059
1060                 b->written += sectors;
1061
1062                 if (blacklisted && !first)
1063                         continue;
1064
1065                 sort_iter_add(iter,
1066                               vstruct_idx(i, 0),
1067                               vstruct_last(i));
1068
1069                 nonblacklisted_written = b->written;
1070         }
1071
1072         if (ptr_written) {
1073                 btree_err_on(b->written < ptr_written,
1074                              BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1075                              "btree node data missing: expected %u sectors, found %u",
1076                              ptr_written, b->written);
1077         } else {
1078                 for (bne = write_block(b);
1079                      bset_byte_offset(b, bne) < btree_bytes(c);
1080                      bne = (void *) bne + block_bytes(c))
1081                         btree_err_on(bne->keys.seq == b->data->keys.seq &&
1082                                      !bch2_journal_seq_is_blacklisted(c,
1083                                                                       le64_to_cpu(bne->keys.journal_seq),
1084                                                                       true),
1085                                      BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1086                                      "found bset signature after last bset");
1087
1088                 /*
1089                  * Blacklisted bsets are those that were written after the most recent
1090                  * (flush) journal write. Since there wasn't a flush, they may not have
1091                  * made it to all devices - which means we shouldn't write new bsets
1092                  * after them, as that could leave a gap and then reads from that device
1093                  * wouldn't find all the bsets in that btree node - which means it's
1094                  * important that we start writing new bsets after the most recent _non_
1095                  * blacklisted bset:
1096                  */
1097                 blacklisted_written = b->written;
1098                 b->written = nonblacklisted_written;
1099         }
1100
1101         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1102         sorted->keys.u64s = 0;
1103
1104         set_btree_bset(b, b->set, &b->data->keys);
1105
1106         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1107
1108         u64s = le16_to_cpu(sorted->keys.u64s);
1109         *sorted = *b->data;
1110         sorted->keys.u64s = cpu_to_le16(u64s);
1111         swap(sorted, b->data);
1112         set_btree_bset(b, b->set, &b->data->keys);
1113         b->nsets = 1;
1114
1115         BUG_ON(b->nr.live_u64s != u64s);
1116
1117         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1118
1119         if (updated_range)
1120                 bch2_btree_node_drop_keys_outside_node(b);
1121
1122         i = &b->data->keys;
1123         for (k = i->start; k != vstruct_last(i);) {
1124                 struct bkey tmp;
1125                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1126
1127                 printbuf_reset(&buf);
1128
1129                 if (bch2_bkey_val_invalid(c, u.s_c, READ, &buf) ||
1130                     (bch2_inject_invalid_keys &&
1131                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1132                         printbuf_reset(&buf);
1133
1134                         prt_printf(&buf, "invalid bkey: ");
1135                         bch2_bkey_val_invalid(c, u.s_c, READ, &buf);
1136                         prt_printf(&buf, "\n  ");
1137                         bch2_bkey_val_to_text(&buf, c, u.s_c);
1138
1139                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf);
1140
1141                         btree_keys_account_key_drop(&b->nr, 0, k);
1142
1143                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1144                         memmove_u64s_down(k, bkey_p_next(k),
1145                                           (u64 *) vstruct_end(i) - (u64 *) k);
1146                         set_btree_bset_end(b, b->set);
1147                         continue;
1148                 }
1149
1150                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1151                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1152
1153                         bp.v->mem_ptr = 0;
1154                 }
1155
1156                 k = bkey_p_next(k);
1157         }
1158
1159         bch2_bset_build_aux_tree(b, b->set, false);
1160
1161         set_needs_whiteout(btree_bset_first(b), true);
1162
1163         btree_node_reset_sib_u64s(b);
1164
1165         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1166                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1167
1168                 if (ca->mi.state != BCH_MEMBER_STATE_rw)
1169                         set_btree_node_need_rewrite(b);
1170         }
1171
1172         if (!ptr_written)
1173                 set_btree_node_need_rewrite(b);
1174 out:
1175         mempool_free(iter, &c->fill_iter);
1176         printbuf_exit(&buf);
1177         return retry_read;
1178 fsck_err:
1179         if (ret == BTREE_RETRY_READ)
1180                 retry_read = 1;
1181         else
1182                 set_btree_node_read_error(b);
1183         goto out;
1184 }
1185
1186 static void btree_node_read_work(struct work_struct *work)
1187 {
1188         struct btree_read_bio *rb =
1189                 container_of(work, struct btree_read_bio, work);
1190         struct bch_fs *c        = rb->c;
1191         struct btree *b         = rb->b;
1192         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1193         struct bio *bio         = &rb->bio;
1194         struct bch_io_failures failed = { .nr = 0 };
1195         struct printbuf buf = PRINTBUF;
1196         bool saw_error = false;
1197         bool retry = false;
1198         bool can_retry;
1199
1200         goto start;
1201         while (1) {
1202                 retry = true;
1203                 bch_info(c, "retrying read");
1204                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1205                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1206                 bio_reset(bio, NULL, REQ_OP_READ|REQ_SYNC|REQ_META);
1207                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1208                 bio->bi_iter.bi_size    = btree_bytes(c);
1209
1210                 if (rb->have_ioref) {
1211                         bio_set_dev(bio, ca->disk_sb.bdev);
1212                         submit_bio_wait(bio);
1213                 } else {
1214                         bio->bi_status = BLK_STS_REMOVED;
1215                 }
1216 start:
1217                 printbuf_reset(&buf);
1218                 btree_pos_to_text(&buf, c, b);
1219                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1220                                    bch2_blk_status_to_str(bio->bi_status), buf.buf);
1221                 if (rb->have_ioref)
1222                         percpu_ref_put(&ca->io_ref);
1223                 rb->have_ioref = false;
1224
1225                 bch2_mark_io_failure(&failed, &rb->pick);
1226
1227                 can_retry = bch2_bkey_pick_read_device(c,
1228                                 bkey_i_to_s_c(&b->key),
1229                                 &failed, &rb->pick) > 0;
1230
1231                 if (!bio->bi_status &&
1232                     !bch2_btree_node_read_done(c, ca, b, can_retry, &saw_error)) {
1233                         if (retry)
1234                                 bch_info(c, "retry success");
1235                         break;
1236                 }
1237
1238                 saw_error = true;
1239
1240                 if (!can_retry) {
1241                         set_btree_node_read_error(b);
1242                         break;
1243                 }
1244         }
1245
1246         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1247                                rb->start_time);
1248         bio_put(&rb->bio);
1249         printbuf_exit(&buf);
1250
1251         if (saw_error && !btree_node_read_error(b)) {
1252                 struct printbuf buf = PRINTBUF;
1253
1254                 bch2_bpos_to_text(&buf, b->key.k.p);
1255                 bch_info(c, "%s: rewriting btree node at btree=%s level=%u %s due to error",
1256                          __func__, bch2_btree_ids[b->c.btree_id], b->c.level, buf.buf);
1257                 printbuf_exit(&buf);
1258
1259                 bch2_btree_node_rewrite_async(c, b);
1260         }
1261
1262         clear_btree_node_read_in_flight(b);
1263         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1264 }
1265
1266 static void btree_node_read_endio(struct bio *bio)
1267 {
1268         struct btree_read_bio *rb =
1269                 container_of(bio, struct btree_read_bio, bio);
1270         struct bch_fs *c        = rb->c;
1271
1272         if (rb->have_ioref) {
1273                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1274
1275                 bch2_latency_acct(ca, rb->start_time, READ);
1276         }
1277
1278         queue_work(c->io_complete_wq, &rb->work);
1279 }
1280
1281 struct btree_node_read_all {
1282         struct closure          cl;
1283         struct bch_fs           *c;
1284         struct btree            *b;
1285         unsigned                nr;
1286         void                    *buf[BCH_REPLICAS_MAX];
1287         struct bio              *bio[BCH_REPLICAS_MAX];
1288         int                     err[BCH_REPLICAS_MAX];
1289 };
1290
1291 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1292 {
1293         struct btree_node *bn = data;
1294         struct btree_node_entry *bne;
1295         unsigned offset = 0;
1296
1297         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1298                 return 0;
1299
1300         while (offset < btree_sectors(c)) {
1301                 if (!offset) {
1302                         offset += vstruct_sectors(bn, c->block_bits);
1303                 } else {
1304                         bne = data + (offset << 9);
1305                         if (bne->keys.seq != bn->keys.seq)
1306                                 break;
1307                         offset += vstruct_sectors(bne, c->block_bits);
1308                 }
1309         }
1310
1311         return offset;
1312 }
1313
1314 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1315 {
1316         struct btree_node *bn = data;
1317         struct btree_node_entry *bne;
1318
1319         if (!offset)
1320                 return false;
1321
1322         while (offset < btree_sectors(c)) {
1323                 bne = data + (offset << 9);
1324                 if (bne->keys.seq == bn->keys.seq)
1325                         return true;
1326                 offset++;
1327         }
1328
1329         return false;
1330         return offset;
1331 }
1332
1333 static void btree_node_read_all_replicas_done(struct closure *cl)
1334 {
1335         struct btree_node_read_all *ra =
1336                 container_of(cl, struct btree_node_read_all, cl);
1337         struct bch_fs *c = ra->c;
1338         struct btree *b = ra->b;
1339         struct printbuf buf = PRINTBUF;
1340         bool dump_bset_maps = false;
1341         bool have_retry = false;
1342         int ret = 0, best = -1, write = READ;
1343         unsigned i, written = 0, written2 = 0;
1344         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1345                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1346         bool _saw_error = false, *saw_error = &_saw_error;
1347
1348         for (i = 0; i < ra->nr; i++) {
1349                 struct btree_node *bn = ra->buf[i];
1350
1351                 if (ra->err[i])
1352                         continue;
1353
1354                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1355                     (seq && seq != bn->keys.seq))
1356                         continue;
1357
1358                 if (best < 0) {
1359                         best = i;
1360                         written = btree_node_sectors_written(c, bn);
1361                         continue;
1362                 }
1363
1364                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1365                 if (btree_err_on(written2 != written, BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1366                                  "btree node sectors written mismatch: %u != %u",
1367                                  written, written2) ||
1368                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1369                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1370                                  "found bset signature after last bset") ||
1371                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1372                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1373                                  "btree node replicas content mismatch"))
1374                         dump_bset_maps = true;
1375
1376                 if (written2 > written) {
1377                         written = written2;
1378                         best = i;
1379                 }
1380         }
1381 fsck_err:
1382         if (dump_bset_maps) {
1383                 for (i = 0; i < ra->nr; i++) {
1384                         struct btree_node *bn = ra->buf[i];
1385                         struct btree_node_entry *bne = NULL;
1386                         unsigned offset = 0, sectors;
1387                         bool gap = false;
1388
1389                         if (ra->err[i])
1390                                 continue;
1391
1392                         printbuf_reset(&buf);
1393
1394                         while (offset < btree_sectors(c)) {
1395                                 if (!offset) {
1396                                         sectors = vstruct_sectors(bn, c->block_bits);
1397                                 } else {
1398                                         bne = ra->buf[i] + (offset << 9);
1399                                         if (bne->keys.seq != bn->keys.seq)
1400                                                 break;
1401                                         sectors = vstruct_sectors(bne, c->block_bits);
1402                                 }
1403
1404                                 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1405                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1406                                                         le64_to_cpu(bne->keys.journal_seq), false))
1407                                         prt_printf(&buf, "*");
1408                                 offset += sectors;
1409                         }
1410
1411                         while (offset < btree_sectors(c)) {
1412                                 bne = ra->buf[i] + (offset << 9);
1413                                 if (bne->keys.seq == bn->keys.seq) {
1414                                         if (!gap)
1415                                                 prt_printf(&buf, " GAP");
1416                                         gap = true;
1417
1418                                         sectors = vstruct_sectors(bne, c->block_bits);
1419                                         prt_printf(&buf, " %u-%u", offset, offset + sectors);
1420                                         if (bch2_journal_seq_is_blacklisted(c,
1421                                                         le64_to_cpu(bne->keys.journal_seq), false))
1422                                                 prt_printf(&buf, "*");
1423                                 }
1424                                 offset++;
1425                         }
1426
1427                         bch_err(c, "replica %u:%s", i, buf.buf);
1428                 }
1429         }
1430
1431         if (best >= 0) {
1432                 memcpy(b->data, ra->buf[best], btree_bytes(c));
1433                 ret = bch2_btree_node_read_done(c, NULL, b, false, saw_error);
1434         } else {
1435                 ret = -1;
1436         }
1437
1438         if (ret)
1439                 set_btree_node_read_error(b);
1440         else if (*saw_error)
1441                 bch2_btree_node_rewrite_async(c, b);
1442
1443         for (i = 0; i < ra->nr; i++) {
1444                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1445                 bio_put(ra->bio[i]);
1446         }
1447
1448         closure_debug_destroy(&ra->cl);
1449         kfree(ra);
1450         printbuf_exit(&buf);
1451
1452         clear_btree_node_read_in_flight(b);
1453         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1454 }
1455
1456 static void btree_node_read_all_replicas_endio(struct bio *bio)
1457 {
1458         struct btree_read_bio *rb =
1459                 container_of(bio, struct btree_read_bio, bio);
1460         struct bch_fs *c        = rb->c;
1461         struct btree_node_read_all *ra = rb->ra;
1462
1463         if (rb->have_ioref) {
1464                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1465
1466                 bch2_latency_acct(ca, rb->start_time, READ);
1467         }
1468
1469         ra->err[rb->idx] = bio->bi_status;
1470         closure_put(&ra->cl);
1471 }
1472
1473 /*
1474  * XXX This allocates multiple times from the same mempools, and can deadlock
1475  * under sufficient memory pressure (but is only a debug path)
1476  */
1477 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1478 {
1479         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1480         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1481         const union bch_extent_entry *entry;
1482         struct extent_ptr_decoded pick;
1483         struct btree_node_read_all *ra;
1484         unsigned i;
1485
1486         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1487         if (!ra)
1488                 return -ENOMEM;
1489
1490         closure_init(&ra->cl, NULL);
1491         ra->c   = c;
1492         ra->b   = b;
1493         ra->nr  = bch2_bkey_nr_ptrs(k);
1494
1495         for (i = 0; i < ra->nr; i++) {
1496                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1497                 ra->bio[i] = bio_alloc_bioset(NULL,
1498                                               buf_pages(ra->buf[i], btree_bytes(c)),
1499                                               REQ_OP_READ|REQ_SYNC|REQ_META,
1500                                               GFP_NOFS,
1501                                               &c->btree_bio);
1502         }
1503
1504         i = 0;
1505         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1506                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1507                 struct btree_read_bio *rb =
1508                         container_of(ra->bio[i], struct btree_read_bio, bio);
1509                 rb->c                   = c;
1510                 rb->b                   = b;
1511                 rb->ra                  = ra;
1512                 rb->start_time          = local_clock();
1513                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1514                 rb->idx                 = i;
1515                 rb->pick                = pick;
1516                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1517                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1518                 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1519
1520                 if (rb->have_ioref) {
1521                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1522                                      bio_sectors(&rb->bio));
1523                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1524
1525                         closure_get(&ra->cl);
1526                         submit_bio(&rb->bio);
1527                 } else {
1528                         ra->err[i] = BLK_STS_REMOVED;
1529                 }
1530
1531                 i++;
1532         }
1533
1534         if (sync) {
1535                 closure_sync(&ra->cl);
1536                 btree_node_read_all_replicas_done(&ra->cl);
1537         } else {
1538                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1539                             c->io_complete_wq);
1540         }
1541
1542         return 0;
1543 }
1544
1545 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1546                           bool sync)
1547 {
1548         struct extent_ptr_decoded pick;
1549         struct btree_read_bio *rb;
1550         struct bch_dev *ca;
1551         struct bio *bio;
1552         int ret;
1553
1554         trace_and_count(c, btree_node_read, c, b);
1555
1556         if (bch2_verify_all_btree_replicas &&
1557             !btree_node_read_all_replicas(c, b, sync))
1558                 return;
1559
1560         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1561                                          NULL, &pick);
1562
1563         if (ret <= 0) {
1564                 struct printbuf buf = PRINTBUF;
1565
1566                 prt_str(&buf, "btree node read error: no device to read from\n at ");
1567                 btree_pos_to_text(&buf, c, b);
1568                 bch_err(c, "%s", buf.buf);
1569
1570                 if (test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags))
1571                         bch2_fatal_error(c);
1572
1573                 set_btree_node_read_error(b);
1574                 clear_btree_node_read_in_flight(b);
1575                 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1576                 printbuf_exit(&buf);
1577                 return;
1578         }
1579
1580         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1581
1582         bio = bio_alloc_bioset(NULL,
1583                                buf_pages(b->data, btree_bytes(c)),
1584                                REQ_OP_READ|REQ_SYNC|REQ_META,
1585                                GFP_NOIO,
1586                                &c->btree_bio);
1587         rb = container_of(bio, struct btree_read_bio, bio);
1588         rb->c                   = c;
1589         rb->b                   = b;
1590         rb->ra                  = NULL;
1591         rb->start_time          = local_clock();
1592         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1593         rb->pick                = pick;
1594         INIT_WORK(&rb->work, btree_node_read_work);
1595         bio->bi_iter.bi_sector  = pick.ptr.offset;
1596         bio->bi_end_io          = btree_node_read_endio;
1597         bch2_bio_map(bio, b->data, btree_bytes(c));
1598
1599         if (rb->have_ioref) {
1600                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1601                              bio_sectors(bio));
1602                 bio_set_dev(bio, ca->disk_sb.bdev);
1603
1604                 if (sync) {
1605                         submit_bio_wait(bio);
1606
1607                         btree_node_read_work(&rb->work);
1608                 } else {
1609                         submit_bio(bio);
1610                 }
1611         } else {
1612                 bio->bi_status = BLK_STS_REMOVED;
1613
1614                 if (sync)
1615                         btree_node_read_work(&rb->work);
1616                 else
1617                         queue_work(c->io_complete_wq, &rb->work);
1618         }
1619 }
1620
1621 static int __bch2_btree_root_read(struct btree_trans *trans, enum btree_id id,
1622                                   const struct bkey_i *k, unsigned level)
1623 {
1624         struct bch_fs *c = trans->c;
1625         struct closure cl;
1626         struct btree *b;
1627         int ret;
1628
1629         closure_init_stack(&cl);
1630
1631         do {
1632                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1633                 closure_sync(&cl);
1634         } while (ret);
1635
1636         b = bch2_btree_node_mem_alloc(trans, level != 0);
1637         bch2_btree_cache_cannibalize_unlock(c);
1638
1639         BUG_ON(IS_ERR(b));
1640
1641         bkey_copy(&b->key, k);
1642         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1643
1644         set_btree_node_read_in_flight(b);
1645
1646         bch2_btree_node_read(c, b, true);
1647
1648         if (btree_node_read_error(b)) {
1649                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1650
1651                 mutex_lock(&c->btree_cache.lock);
1652                 list_move(&b->list, &c->btree_cache.freeable);
1653                 mutex_unlock(&c->btree_cache.lock);
1654
1655                 ret = -EIO;
1656                 goto err;
1657         }
1658
1659         bch2_btree_set_root_for_read(c, b);
1660 err:
1661         six_unlock_write(&b->c.lock);
1662         six_unlock_intent(&b->c.lock);
1663
1664         return ret;
1665 }
1666
1667 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1668                         const struct bkey_i *k, unsigned level)
1669 {
1670         return bch2_trans_run(c, __bch2_btree_root_read(&trans, id, k, level));
1671
1672 }
1673
1674 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1675                               struct btree_write *w)
1676 {
1677         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1678
1679         do {
1680                 old = new = v;
1681                 if (!(old & 1))
1682                         break;
1683
1684                 new &= ~1UL;
1685         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1686
1687         if (old & 1)
1688                 closure_put(&((struct btree_update *) new)->cl);
1689
1690         bch2_journal_pin_drop(&c->journal, &w->journal);
1691 }
1692
1693 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1694 {
1695         struct btree_write *w = btree_prev_write(b);
1696         unsigned long old, new, v;
1697         unsigned type = 0;
1698
1699         bch2_btree_complete_write(c, b, w);
1700
1701         v = READ_ONCE(b->flags);
1702         do {
1703                 old = new = v;
1704
1705                 if ((old & (1U << BTREE_NODE_dirty)) &&
1706                     (old & (1U << BTREE_NODE_need_write)) &&
1707                     !(old & (1U << BTREE_NODE_never_write)) &&
1708                     !(old & (1U << BTREE_NODE_write_blocked)) &&
1709                     !(old & (1U << BTREE_NODE_will_make_reachable))) {
1710                         new &= ~(1U << BTREE_NODE_dirty);
1711                         new &= ~(1U << BTREE_NODE_need_write);
1712                         new |=  (1U << BTREE_NODE_write_in_flight);
1713                         new |=  (1U << BTREE_NODE_write_in_flight_inner);
1714                         new |=  (1U << BTREE_NODE_just_written);
1715                         new ^=  (1U << BTREE_NODE_write_idx);
1716
1717                         type = new & BTREE_WRITE_TYPE_MASK;
1718                         new &= ~BTREE_WRITE_TYPE_MASK;
1719                 } else {
1720                         new &= ~(1U << BTREE_NODE_write_in_flight);
1721                         new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1722                 }
1723         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1724
1725         if (new & (1U << BTREE_NODE_write_in_flight))
1726                 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED|type);
1727         else
1728                 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1729 }
1730
1731 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1732 {
1733         struct btree_trans trans;
1734
1735         bch2_trans_init(&trans, c, 0, 0);
1736
1737         btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
1738         __btree_node_write_done(c, b);
1739         six_unlock_read(&b->c.lock);
1740
1741         bch2_trans_exit(&trans);
1742 }
1743
1744 static void btree_node_write_work(struct work_struct *work)
1745 {
1746         struct btree_write_bio *wbio =
1747                 container_of(work, struct btree_write_bio, work);
1748         struct bch_fs *c        = wbio->wbio.c;
1749         struct btree *b         = wbio->wbio.bio.bi_private;
1750         struct bch_extent_ptr *ptr;
1751         int ret;
1752
1753         btree_bounce_free(c,
1754                 wbio->data_bytes,
1755                 wbio->wbio.used_mempool,
1756                 wbio->data);
1757
1758         bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1759                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1760
1761         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key)))
1762                 goto err;
1763
1764         if (wbio->wbio.first_btree_write) {
1765                 if (wbio->wbio.failed.nr) {
1766
1767                 }
1768         } else {
1769                 ret = bch2_trans_do(c, NULL, NULL, 0,
1770                         bch2_btree_node_update_key_get_iter(&trans, b, &wbio->key,
1771                                                             !wbio->wbio.failed.nr));
1772                 if (ret)
1773                         goto err;
1774         }
1775 out:
1776         bio_put(&wbio->wbio.bio);
1777         btree_node_write_done(c, b);
1778         return;
1779 err:
1780         set_btree_node_noevict(b);
1781         bch2_fs_fatal_error(c, "fatal error writing btree node");
1782         goto out;
1783 }
1784
1785 static void btree_node_write_endio(struct bio *bio)
1786 {
1787         struct bch_write_bio *wbio      = to_wbio(bio);
1788         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1789         struct bch_write_bio *orig      = parent ?: wbio;
1790         struct btree_write_bio *wb      = container_of(orig, struct btree_write_bio, wbio);
1791         struct bch_fs *c                = wbio->c;
1792         struct btree *b                 = wbio->bio.bi_private;
1793         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1794         unsigned long flags;
1795
1796         if (wbio->have_ioref)
1797                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1798
1799         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1800                                bch2_blk_status_to_str(bio->bi_status)) ||
1801             bch2_meta_write_fault("btree")) {
1802                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1803                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1804                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1805         }
1806
1807         if (wbio->have_ioref)
1808                 percpu_ref_put(&ca->io_ref);
1809
1810         if (parent) {
1811                 bio_put(bio);
1812                 bio_endio(&parent->bio);
1813                 return;
1814         }
1815
1816         clear_btree_node_write_in_flight_inner(b);
1817         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1818         INIT_WORK(&wb->work, btree_node_write_work);
1819         queue_work(c->btree_io_complete_wq, &wb->work);
1820 }
1821
1822 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1823                                    struct bset *i, unsigned sectors)
1824 {
1825         struct printbuf buf = PRINTBUF;
1826         bool saw_error;
1827         int ret;
1828
1829         ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key),
1830                                 BKEY_TYPE_btree, WRITE, &buf);
1831
1832         if (ret)
1833                 bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf);
1834         printbuf_exit(&buf);
1835         if (ret)
1836                 return ret;
1837
1838         ret = validate_bset_keys(c, b, i, WRITE, false, &saw_error) ?:
1839                 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false, &saw_error);
1840         if (ret) {
1841                 bch2_inconsistent_error(c);
1842                 dump_stack();
1843         }
1844
1845         return ret;
1846 }
1847
1848 static void btree_write_submit(struct work_struct *work)
1849 {
1850         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1851         struct bch_extent_ptr *ptr;
1852         BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1853
1854         bkey_copy(&tmp.k, &wbio->key);
1855
1856         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1857                 ptr->offset += wbio->sector_offset;
1858
1859         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree,
1860                                   &tmp.k, false);
1861 }
1862
1863 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1864 {
1865         struct btree_write_bio *wbio;
1866         struct bset_tree *t;
1867         struct bset *i;
1868         struct btree_node *bn = NULL;
1869         struct btree_node_entry *bne = NULL;
1870         struct sort_iter sort_iter;
1871         struct nonce nonce;
1872         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1873         u64 seq = 0;
1874         bool used_mempool;
1875         unsigned long old, new;
1876         bool validate_before_checksum = false;
1877         enum btree_write_type type = flags & BTREE_WRITE_TYPE_MASK;
1878         void *data;
1879         int ret;
1880
1881         if (flags & BTREE_WRITE_ALREADY_STARTED)
1882                 goto do_write;
1883
1884         /*
1885          * We may only have a read lock on the btree node - the dirty bit is our
1886          * "lock" against racing with other threads that may be trying to start
1887          * a write, we do a write iff we clear the dirty bit. Since setting the
1888          * dirty bit requires a write lock, we can't race with other threads
1889          * redirtying it:
1890          */
1891         do {
1892                 old = new = READ_ONCE(b->flags);
1893
1894                 if (!(old & (1 << BTREE_NODE_dirty)))
1895                         return;
1896
1897                 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
1898                     !(old & (1 << BTREE_NODE_need_write)))
1899                         return;
1900
1901                 if (old &
1902                     ((1 << BTREE_NODE_never_write)|
1903                      (1 << BTREE_NODE_write_blocked)))
1904                         return;
1905
1906                 if (b->written &&
1907                     (old & (1 << BTREE_NODE_will_make_reachable)))
1908                         return;
1909
1910                 if (old & (1 << BTREE_NODE_write_in_flight))
1911                         return;
1912
1913                 if (flags & BTREE_WRITE_ONLY_IF_NEED)
1914                         type = new & BTREE_WRITE_TYPE_MASK;
1915                 new &= ~BTREE_WRITE_TYPE_MASK;
1916
1917                 new &= ~(1 << BTREE_NODE_dirty);
1918                 new &= ~(1 << BTREE_NODE_need_write);
1919                 new |=  (1 << BTREE_NODE_write_in_flight);
1920                 new |=  (1 << BTREE_NODE_write_in_flight_inner);
1921                 new |=  (1 << BTREE_NODE_just_written);
1922                 new ^=  (1 << BTREE_NODE_write_idx);
1923         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1924
1925         if (new & (1U << BTREE_NODE_need_write))
1926                 return;
1927 do_write:
1928         BUG_ON((type == BTREE_WRITE_initial) != (b->written == 0));
1929
1930         atomic_dec(&c->btree_cache.dirty);
1931
1932         BUG_ON(btree_node_fake(b));
1933         BUG_ON((b->will_make_reachable != 0) != !b->written);
1934
1935         BUG_ON(b->written >= btree_sectors(c));
1936         BUG_ON(b->written & (block_sectors(c) - 1));
1937         BUG_ON(bset_written(b, btree_bset_last(b)));
1938         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1939         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1940
1941         bch2_sort_whiteouts(c, b);
1942
1943         sort_iter_init(&sort_iter, b);
1944
1945         bytes = !b->written
1946                 ? sizeof(struct btree_node)
1947                 : sizeof(struct btree_node_entry);
1948
1949         bytes += b->whiteout_u64s * sizeof(u64);
1950
1951         for_each_bset(b, t) {
1952                 i = bset(b, t);
1953
1954                 if (bset_written(b, i))
1955                         continue;
1956
1957                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1958                 sort_iter_add(&sort_iter,
1959                               btree_bkey_first(b, t),
1960                               btree_bkey_last(b, t));
1961                 seq = max(seq, le64_to_cpu(i->journal_seq));
1962         }
1963
1964         BUG_ON(b->written && !seq);
1965
1966         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1967         bytes += 8;
1968
1969         /* buffer must be a multiple of the block size */
1970         bytes = round_up(bytes, block_bytes(c));
1971
1972         data = btree_bounce_alloc(c, bytes, &used_mempool);
1973
1974         if (!b->written) {
1975                 bn = data;
1976                 *bn = *b->data;
1977                 i = &bn->keys;
1978         } else {
1979                 bne = data;
1980                 bne->keys = b->data->keys;
1981                 i = &bne->keys;
1982         }
1983
1984         i->journal_seq  = cpu_to_le64(seq);
1985         i->u64s         = 0;
1986
1987         sort_iter_add(&sort_iter,
1988                       unwritten_whiteouts_start(c, b),
1989                       unwritten_whiteouts_end(c, b));
1990         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1991
1992         b->whiteout_u64s = 0;
1993
1994         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1995         le16_add_cpu(&i->u64s, u64s);
1996
1997         BUG_ON(!b->written && i->u64s != b->data->keys.u64s);
1998
1999         set_needs_whiteout(i, false);
2000
2001         /* do we have data to write? */
2002         if (b->written && !i->u64s)
2003                 goto nowrite;
2004
2005         bytes_to_write = vstruct_end(i) - data;
2006         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
2007
2008         if (!b->written &&
2009             b->key.k.type == KEY_TYPE_btree_ptr_v2)
2010                 BUG_ON(btree_ptr_sectors_written(&b->key) != sectors_to_write);
2011
2012         memset(data + bytes_to_write, 0,
2013                (sectors_to_write << 9) - bytes_to_write);
2014
2015         BUG_ON(b->written + sectors_to_write > btree_sectors(c));
2016         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
2017         BUG_ON(i->seq != b->data->keys.seq);
2018
2019         i->version = c->sb.version < bcachefs_metadata_version_bkey_renumber
2020                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
2021                 : cpu_to_le16(c->sb.version);
2022         SET_BSET_OFFSET(i, b->written);
2023         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
2024
2025         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
2026                 validate_before_checksum = true;
2027
2028         /* validate_bset will be modifying: */
2029         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
2030                 validate_before_checksum = true;
2031
2032         /* if we're going to be encrypting, check metadata validity first: */
2033         if (validate_before_checksum &&
2034             validate_bset_for_write(c, b, i, sectors_to_write))
2035                 goto err;
2036
2037         ret = bset_encrypt(c, i, b->written << 9);
2038         if (bch2_fs_fatal_err_on(ret, c,
2039                         "error encrypting btree node: %i\n", ret))
2040                 goto err;
2041
2042         nonce = btree_nonce(i, b->written << 9);
2043
2044         if (bn)
2045                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
2046         else
2047                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
2048
2049         /* if we're not encrypting, check metadata after checksumming: */
2050         if (!validate_before_checksum &&
2051             validate_bset_for_write(c, b, i, sectors_to_write))
2052                 goto err;
2053
2054         /*
2055          * We handle btree write errors by immediately halting the journal -
2056          * after we've done that, we can't issue any subsequent btree writes
2057          * because they might have pointers to new nodes that failed to write.
2058          *
2059          * Furthermore, there's no point in doing any more btree writes because
2060          * with the journal stopped, we're never going to update the journal to
2061          * reflect that those writes were done and the data flushed from the
2062          * journal:
2063          *
2064          * Also on journal error, the pending write may have updates that were
2065          * never journalled (interior nodes, see btree_update_nodes_written()) -
2066          * it's critical that we don't do the write in that case otherwise we
2067          * will have updates visible that weren't in the journal:
2068          *
2069          * Make sure to update b->written so bch2_btree_init_next() doesn't
2070          * break:
2071          */
2072         if (bch2_journal_error(&c->journal) ||
2073             c->opts.nochanges)
2074                 goto err;
2075
2076         trace_and_count(c, btree_node_write, b, bytes_to_write, sectors_to_write);
2077
2078         wbio = container_of(bio_alloc_bioset(NULL,
2079                                 buf_pages(data, sectors_to_write << 9),
2080                                 REQ_OP_WRITE|REQ_META,
2081                                 GFP_NOIO,
2082                                 &c->btree_bio),
2083                             struct btree_write_bio, wbio.bio);
2084         wbio_init(&wbio->wbio.bio);
2085         wbio->data                      = data;
2086         wbio->data_bytes                = bytes;
2087         wbio->sector_offset             = b->written;
2088         wbio->wbio.c                    = c;
2089         wbio->wbio.used_mempool         = used_mempool;
2090         wbio->wbio.first_btree_write    = !b->written;
2091         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
2092         wbio->wbio.bio.bi_private       = b;
2093
2094         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
2095
2096         bkey_copy(&wbio->key, &b->key);
2097
2098         b->written += sectors_to_write;
2099
2100         if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
2101                 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2102                         cpu_to_le16(b->written);
2103
2104         atomic64_inc(&c->btree_write_stats[type].nr);
2105         atomic64_add(bytes_to_write, &c->btree_write_stats[type].bytes);
2106
2107         INIT_WORK(&wbio->work, btree_write_submit);
2108         queue_work(c->io_complete_wq, &wbio->work);
2109         return;
2110 err:
2111         set_btree_node_noevict(b);
2112         b->written += sectors_to_write;
2113 nowrite:
2114         btree_bounce_free(c, bytes, used_mempool, data);
2115         __btree_node_write_done(c, b);
2116 }
2117
2118 /*
2119  * Work that must be done with write lock held:
2120  */
2121 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2122 {
2123         bool invalidated_iter = false;
2124         struct btree_node_entry *bne;
2125         struct bset_tree *t;
2126
2127         if (!btree_node_just_written(b))
2128                 return false;
2129
2130         BUG_ON(b->whiteout_u64s);
2131
2132         clear_btree_node_just_written(b);
2133
2134         /*
2135          * Note: immediately after write, bset_written() doesn't work - the
2136          * amount of data we had to write after compaction might have been
2137          * smaller than the offset of the last bset.
2138          *
2139          * However, we know that all bsets have been written here, as long as
2140          * we're still holding the write lock:
2141          */
2142
2143         /*
2144          * XXX: decide if we really want to unconditionally sort down to a
2145          * single bset:
2146          */
2147         if (b->nsets > 1) {
2148                 btree_node_sort(c, b, 0, b->nsets, true);
2149                 invalidated_iter = true;
2150         } else {
2151                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2152         }
2153
2154         for_each_bset(b, t)
2155                 set_needs_whiteout(bset(b, t), true);
2156
2157         bch2_btree_verify(c, b);
2158
2159         /*
2160          * If later we don't unconditionally sort down to a single bset, we have
2161          * to ensure this is still true:
2162          */
2163         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2164
2165         bne = want_new_bset(c, b);
2166         if (bne)
2167                 bch2_bset_init_next(c, b, bne);
2168
2169         bch2_btree_build_aux_trees(b);
2170
2171         return invalidated_iter;
2172 }
2173
2174 /*
2175  * Use this one if the node is intent locked:
2176  */
2177 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2178                            enum six_lock_type lock_type_held,
2179                            unsigned flags)
2180 {
2181         if (lock_type_held == SIX_LOCK_intent ||
2182             (lock_type_held == SIX_LOCK_read &&
2183              six_lock_tryupgrade(&b->c.lock))) {
2184                 __bch2_btree_node_write(c, b, flags);
2185
2186                 /* don't cycle lock unnecessarily: */
2187                 if (btree_node_just_written(b) &&
2188                     six_trylock_write(&b->c.lock)) {
2189                         bch2_btree_post_write_cleanup(c, b);
2190                         six_unlock_write(&b->c.lock);
2191                 }
2192
2193                 if (lock_type_held == SIX_LOCK_read)
2194                         six_lock_downgrade(&b->c.lock);
2195         } else {
2196                 __bch2_btree_node_write(c, b, flags);
2197                 if (lock_type_held == SIX_LOCK_write &&
2198                     btree_node_just_written(b))
2199                         bch2_btree_post_write_cleanup(c, b);
2200         }
2201 }
2202
2203 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2204 {
2205         struct bucket_table *tbl;
2206         struct rhash_head *pos;
2207         struct btree *b;
2208         unsigned i;
2209         bool ret = false;
2210 restart:
2211         rcu_read_lock();
2212         for_each_cached_btree(b, c, tbl, i, pos)
2213                 if (test_bit(flag, &b->flags)) {
2214                         rcu_read_unlock();
2215                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2216                         ret = true;
2217                         goto restart;
2218                 }
2219         rcu_read_unlock();
2220
2221         return ret;
2222 }
2223
2224 bool bch2_btree_flush_all_reads(struct bch_fs *c)
2225 {
2226         return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2227 }
2228
2229 bool bch2_btree_flush_all_writes(struct bch_fs *c)
2230 {
2231         return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2232 }
2233
2234 const char * const bch2_btree_write_types[] = {
2235 #define x(t, n) [n] = #t,
2236         BCH_BTREE_WRITE_TYPES()
2237         NULL
2238 };
2239
2240 void bch2_btree_write_stats_to_text(struct printbuf *out, struct bch_fs *c)
2241 {
2242         printbuf_tabstop_push(out, 20);
2243         printbuf_tabstop_push(out, 10);
2244
2245         prt_tab(out);
2246         prt_str(out, "nr");
2247         prt_tab(out);
2248         prt_str(out, "size");
2249         prt_newline(out);
2250
2251         for (unsigned i = 0; i < BTREE_WRITE_TYPE_NR; i++) {
2252                 u64 nr          = atomic64_read(&c->btree_write_stats[i].nr);
2253                 u64 bytes       = atomic64_read(&c->btree_write_stats[i].bytes);
2254
2255                 prt_printf(out, "%s:", bch2_btree_write_types[i]);
2256                 prt_tab(out);
2257                 prt_u64(out, nr);
2258                 prt_tab(out);
2259                 prt_human_readable_u64(out, nr ? div64_u64(bytes, nr) : 0);
2260                 prt_newline(out);
2261         }
2262 }