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 static void verify_no_dups(struct btree *b,
  26                            struct bkey_packed *start,
  27                            struct bkey_packed *end)
  28 {
  29 #ifdef CONFIG_BCACHEFS_DEBUG
  30         struct bkey_packed *k, *p;
  31
  32         if (start == end)
  33                 return;
  34
  35         for (p = start, k = bkey_next(start);
  36              k != end;
  37              p = k, k = bkey_next(k)) {
  38                 struct bkey l = bkey_unpack_key(b, p);
  39                 struct bkey r = bkey_unpack_key(b, k);
  40
  41                 BUG_ON(bpos_cmp(l.p, bkey_start_pos(&r)) >= 0);
  42         }
  43 #endif
  44 }
  45
  46 static void set_needs_whiteout(struct bset *i, int v)
  47 {
  48         struct bkey_packed *k;
  49
  50         for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
  51                 k->needs_whiteout = v;
  52 }
  53
  54 static void btree_bounce_free(struct bch_fs *c, size_t size,
  55                               bool used_mempool, void *p)
  56 {
  57         if (used_mempool)
  58                 mempool_free(p, &c->btree_bounce_pool);
  59         else
  60                 vpfree(p, size);
  61 }
  62
  63 static void *btree_bounce_alloc(struct bch_fs *c, size_t size,
  64                                 bool *used_mempool)
  65 {
  66         unsigned flags = memalloc_nofs_save();
  67         void *p;
  68
  69         BUG_ON(size > btree_bytes(c));
  70
  71         *used_mempool = false;
  72         p = vpmalloc(size, __GFP_NOWARN|GFP_NOWAIT);
  73         if (!p) {
  74                 *used_mempool = true;
  75                 p = mempool_alloc(&c->btree_bounce_pool, GFP_NOIO);
  76         }
  77         memalloc_nofs_restore(flags);
  78         return p;
  79 }
  80
  81 static void sort_bkey_ptrs(const struct btree *bt,
  82                            struct bkey_packed **ptrs, unsigned nr)
  83 {
  84         unsigned n = nr, a = nr / 2, b, c, d;
  85
  86         if (!a)
  87                 return;
  88
  89         /* Heap sort: see lib/sort.c: */
  90         while (1) {
  91                 if (a)
  92                         a--;
  93                 else if (--n)
  94                         swap(ptrs[0], ptrs[n]);
  95                 else
  96                         break;
  97
  98                 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
  99                         b = bch2_bkey_cmp_packed(bt,
 100                                             ptrs[c],
 101                                             ptrs[d]) >= 0 ? c : d;
 102                 if (d == n)
 103                         b = c;
 104
 105                 while (b != a &&
 106                        bch2_bkey_cmp_packed(bt,
 107                                        ptrs[a],
 108                                        ptrs[b]) >= 0)
 109                         b = (b - 1) / 2;
 110                 c = b;
 111                 while (b != a) {
 112                         b = (b - 1) / 2;
 113                         swap(ptrs[b], ptrs[c]);
 114                 }
 115         }
 116 }
 117
 118 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
 119 {
 120         struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
 121         bool used_mempool = false;
 122         size_t bytes = b->whiteout_u64s * sizeof(u64);
 123
 124         if (!b->whiteout_u64s)
 125                 return;
 126
 127         new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
 128
 129         ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
 130
 131         for (k = unwritten_whiteouts_start(c, b);
 132              k != unwritten_whiteouts_end(c, b);
 133              k = bkey_next(k))
 134                 *--ptrs = k;
 135
 136         sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
 137
 138         k = new_whiteouts;
 139
 140         while (ptrs != ptrs_end) {
 141                 bkey_copy(k, *ptrs);
 142                 k = bkey_next(k);
 143                 ptrs++;
 144         }
 145
 146         verify_no_dups(b, new_whiteouts,
 147                        (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
 148
 149         memcpy_u64s(unwritten_whiteouts_start(c, b),
 150                     new_whiteouts, b->whiteout_u64s);
 151
 152         btree_bounce_free(c, bytes, used_mempool, new_whiteouts);
 153 }
 154
 155 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
 156                                 bool compacting, enum compact_mode mode)
 157 {
 158         if (!bset_dead_u64s(b, t))
 159                 return false;
 160
 161         switch (mode) {
 162         case COMPACT_LAZY:
 163                 return should_compact_bset_lazy(b, t) ||
 164                         (compacting && !bset_written(b, bset(b, t)));
 165         case COMPACT_ALL:
 166                 return true;
 167         default:
 168                 BUG();
 169         }
 170 }
 171
 172 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
 173 {
 174         struct bset_tree *t;
 175         bool ret = false;
 176
 177         for_each_bset(b, t) {
 178                 struct bset *i = bset(b, t);
 179                 struct bkey_packed *k, *n, *out, *start, *end;
 180                 struct btree_node_entry *src = NULL, *dst = NULL;
 181
 182                 if (t != b->set && !bset_written(b, i)) {
 183                         src = container_of(i, struct btree_node_entry, keys);
 184                         dst = max(write_block(b),
 185                                   (void *) btree_bkey_last(b, t - 1));
 186                 }
 187
 188                 if (src != dst)
 189                         ret = true;
 190
 191                 if (!should_compact_bset(b, t, ret, mode)) {
 192                         if (src != dst) {
 193                                 memmove(dst, src, sizeof(*src) +
 194                                         le16_to_cpu(src->keys.u64s) *
 195                                         sizeof(u64));
 196                                 i = &dst->keys;
 197                                 set_btree_bset(b, t, i);
 198                         }
 199                         continue;
 200                 }
 201
 202                 start   = btree_bkey_first(b, t);
 203                 end     = btree_bkey_last(b, t);
 204
 205                 if (src != dst) {
 206                         memmove(dst, src, sizeof(*src));
 207                         i = &dst->keys;
 208                         set_btree_bset(b, t, i);
 209                 }
 210
 211                 out = i->start;
 212
 213                 for (k = start; k != end; k = n) {
 214                         n = bkey_next(k);
 215
 216                         if (!bkey_deleted(k)) {
 217                                 bkey_copy(out, k);
 218                                 out = bkey_next(out);
 219                         } else {
 220                                 BUG_ON(k->needs_whiteout);
 221                         }
 222                 }
 223
 224                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 225                 set_btree_bset_end(b, t);
 226                 bch2_bset_set_no_aux_tree(b, t);
 227                 ret = true;
 228         }
 229
 230         bch2_verify_btree_nr_keys(b);
 231
 232         bch2_btree_build_aux_trees(b);
 233
 234         return ret;
 235 }
 236
 237 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 238                             enum compact_mode mode)
 239 {
 240         return bch2_drop_whiteouts(b, mode);
 241 }
 242
 243 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 244                             unsigned start_idx,
 245                             unsigned end_idx,
 246                             bool filter_whiteouts)
 247 {
 248         struct btree_node *out;
 249         struct sort_iter sort_iter;
 250         struct bset_tree *t;
 251         struct bset *start_bset = bset(b, &b->set[start_idx]);
 252         bool used_mempool = false;
 253         u64 start_time, seq = 0;
 254         unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
 255         bool sorting_entire_node = start_idx == 0 &&
 256                 end_idx == b->nsets;
 257
 258         sort_iter_init(&sort_iter, b);
 259
 260         for (t = b->set + start_idx;
 261              t < b->set + end_idx;
 262              t++) {
 263                 u64s += le16_to_cpu(bset(b, t)->u64s);
 264                 sort_iter_add(&sort_iter,
 265                               btree_bkey_first(b, t),
 266                               btree_bkey_last(b, t));
 267         }
 268
 269         bytes = sorting_entire_node
 270                 ? btree_bytes(c)
 271                 : __vstruct_bytes(struct btree_node, u64s);
 272
 273         out = btree_bounce_alloc(c, bytes, &used_mempool);
 274
 275         start_time = local_clock();
 276
 277         u64s = bch2_sort_keys(out->keys.start, &sort_iter, filter_whiteouts);
 278
 279         out->keys.u64s = cpu_to_le16(u64s);
 280
 281         BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
 282
 283         if (sorting_entire_node)
 284                 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 285                                        start_time);
 286
 287         /* Make sure we preserve bset journal_seq: */
 288         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 289                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 290         start_bset->journal_seq = cpu_to_le64(seq);
 291
 292         if (sorting_entire_node) {
 293                 unsigned u64s = le16_to_cpu(out->keys.u64s);
 294
 295                 BUG_ON(bytes != btree_bytes(c));
 296
 297                 /*
 298                  * Our temporary buffer is the same size as the btree node's
 299                  * buffer, we can just swap buffers instead of doing a big
 300                  * memcpy()
 301                  */
 302                 *out = *b->data;
 303                 out->keys.u64s = cpu_to_le16(u64s);
 304                 swap(out, b->data);
 305                 set_btree_bset(b, b->set, &b->data->keys);
 306         } else {
 307                 start_bset->u64s = out->keys.u64s;
 308                 memcpy_u64s(start_bset->start,
 309                             out->keys.start,
 310                             le16_to_cpu(out->keys.u64s));
 311         }
 312
 313         for (i = start_idx + 1; i < end_idx; i++)
 314                 b->nr.bset_u64s[start_idx] +=
 315                         b->nr.bset_u64s[i];
 316
 317         b->nsets -= shift;
 318
 319         for (i = start_idx + 1; i < b->nsets; i++) {
 320                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 321                 b->set[i]               = b->set[i + shift];
 322         }
 323
 324         for (i = b->nsets; i < MAX_BSETS; i++)
 325                 b->nr.bset_u64s[i] = 0;
 326
 327         set_btree_bset_end(b, &b->set[start_idx]);
 328         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 329
 330         btree_bounce_free(c, bytes, used_mempool, out);
 331
 332         bch2_verify_btree_nr_keys(b);
 333 }
 334
 335 void bch2_btree_sort_into(struct bch_fs *c,
 336                          struct btree *dst,
 337                          struct btree *src)
 338 {
 339         struct btree_nr_keys nr;
 340         struct btree_node_iter src_iter;
 341         u64 start_time = local_clock();
 342
 343         BUG_ON(dst->nsets != 1);
 344
 345         bch2_bset_set_no_aux_tree(dst, dst->set);
 346
 347         bch2_btree_node_iter_init_from_start(&src_iter, src);
 348
 349         if (btree_node_is_extents(src))
 350                 nr = bch2_sort_repack_merge(c, btree_bset_first(dst),
 351                                 src, &src_iter,
 352                                 &dst->format,
 353                                 true);
 354         else
 355                 nr = bch2_sort_repack(btree_bset_first(dst),
 356                                 src, &src_iter,
 357                                 &dst->format,
 358                                 true);
 359
 360         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 361                                start_time);
 362
 363         set_btree_bset_end(dst, dst->set);
 364
 365         dst->nr.live_u64s       += nr.live_u64s;
 366         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 367         dst->nr.packed_keys     += nr.packed_keys;
 368         dst->nr.unpacked_keys   += nr.unpacked_keys;
 369
 370         bch2_verify_btree_nr_keys(dst);
 371 }
 372
 373 #define SORT_CRIT       (4096 / sizeof(u64))
 374
 375 /*
 376  * We're about to add another bset to the btree node, so if there's currently
 377  * too many bsets - sort some of them together:
 378  */
 379 static bool btree_node_compact(struct bch_fs *c, struct btree *b)
 380 {
 381         unsigned unwritten_idx;
 382         bool ret = false;
 383
 384         for (unwritten_idx = 0;
 385              unwritten_idx < b->nsets;
 386              unwritten_idx++)
 387                 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
 388                         break;
 389
 390         if (b->nsets - unwritten_idx > 1) {
 391                 btree_node_sort(c, b, unwritten_idx,
 392                                 b->nsets, false);
 393                 ret = true;
 394         }
 395
 396         if (unwritten_idx > 1) {
 397                 btree_node_sort(c, b, 0, unwritten_idx, false);
 398                 ret = true;
 399         }
 400
 401         return ret;
 402 }
 403
 404 void bch2_btree_build_aux_trees(struct btree *b)
 405 {
 406         struct bset_tree *t;
 407
 408         for_each_bset(b, t)
 409                 bch2_bset_build_aux_tree(b, t,
 410                                 !bset_written(b, bset(b, t)) &&
 411                                 t == bset_tree_last(b));
 412 }
 413
 414 /*
 415  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 416  * inserted into
 417  *
 418  * Safe to call if there already is an unwritten bset - will only add a new bset
 419  * if @b doesn't already have one.
 420  *
 421  * Returns true if we sorted (i.e. invalidated iterators
 422  */
 423 void bch2_btree_init_next(struct bch_fs *c, struct btree *b,
 424                           struct btree_iter *iter)
 425 {
 426         struct btree_node_entry *bne;
 427         bool reinit_iter = false;
 428
 429         EBUG_ON(!(b->c.lock.state.seq & 1));
 430         EBUG_ON(iter && iter->l[b->c.level].b != b);
 431         BUG_ON(bset_written(b, bset(b, &b->set[1])));
 432
 433         if (b->nsets == MAX_BSETS) {
 434                 unsigned log_u64s[] = {
 435                         ilog2(bset_u64s(&b->set[0])),
 436                         ilog2(bset_u64s(&b->set[1])),
 437                         ilog2(bset_u64s(&b->set[2])),
 438                 };
 439
 440                 if (log_u64s[1] >= (log_u64s[0] + log_u64s[2]) / 2) {
 441                         bch2_btree_node_write(c, b, SIX_LOCK_write);
 442                         reinit_iter = true;
 443                 }
 444         }
 445
 446         if (b->nsets == MAX_BSETS &&
 447             btree_node_compact(c, b))
 448                 reinit_iter = true;
 449
 450         BUG_ON(b->nsets >= MAX_BSETS);
 451
 452         bne = want_new_bset(c, b);
 453         if (bne)
 454                 bch2_bset_init_next(c, b, bne);
 455
 456         bch2_btree_build_aux_trees(b);
 457
 458         if (iter && reinit_iter)
 459                 bch2_btree_iter_reinit_node(iter, b);
 460 }
 461
 462 static void btree_pos_to_text(struct printbuf *out, struct bch_fs *c,
 463                           struct btree *b)
 464 {
 465         pr_buf(out, "%s level %u/%u\n  ",
 466                bch2_btree_ids[b->c.btree_id],
 467                b->c.level,
 468                c->btree_roots[b->c.btree_id].level);
 469         bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 470 }
 471
 472 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 473                           struct bch_dev *ca,
 474                           struct btree *b, struct bset *i,
 475                           unsigned offset, int write)
 476 {
 477         pr_buf(out, "error validating btree node ");
 478         if (write)
 479                 pr_buf(out, "before write ");
 480         if (ca)
 481                 pr_buf(out, "on %s ", ca->name);
 482         pr_buf(out, "at btree ");
 483         btree_pos_to_text(out, c, b);
 484
 485         pr_buf(out, "\n  node offset %u", b->written);
 486         if (i)
 487                 pr_buf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 488 }
 489
 490 enum btree_err_type {
 491         BTREE_ERR_FIXABLE,
 492         BTREE_ERR_WANT_RETRY,
 493         BTREE_ERR_MUST_RETRY,
 494         BTREE_ERR_FATAL,
 495 };
 496
 497 enum btree_validate_ret {
 498         BTREE_RETRY_READ = 64,
 499 };
 500
 501 #define btree_err(type, c, ca, b, i, msg, ...)                          \
 502 ({                                                                      \
 503         __label__ out;                                                  \
 504         char _buf[300];                                                 \
 505         char *_buf2 = _buf;                                             \
 506         struct printbuf out = PBUF(_buf);                               \
 507                                                                         \
 508         _buf2 = kmalloc(4096, GFP_ATOMIC);                              \
 509         if (_buf2)                                                      \
 510                 out = _PBUF(_buf2, 4986);                               \
 511                                                                         \
 512         btree_err_msg(&out, c, ca, b, i, b->written, write);            \
 513         pr_buf(&out, ": " msg, ##__VA_ARGS__);                          \
 514                                                                         \
 515         if (type == BTREE_ERR_FIXABLE &&                                \
 516             write == READ &&                                            \
 517             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {             \
 518                 mustfix_fsck_err(c, "%s", _buf2);                       \
 519                 goto out;                                               \
 520         }                                                               \
 521                                                                         \
 522         switch (write) {                                                \
 523         case READ:                                                      \
 524                 bch_err(c, "%s", _buf2);                                \
 525                                                                         \
 526                 switch (type) {                                         \
 527                 case BTREE_ERR_FIXABLE:                                 \
 528                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 529                         goto fsck_err;                                  \
 530                 case BTREE_ERR_WANT_RETRY:                              \
 531                         if (have_retry) {                               \
 532                                 ret = BTREE_RETRY_READ;                 \
 533                                 goto fsck_err;                          \
 534                         }                                               \
 535                         break;                                          \
 536                 case BTREE_ERR_MUST_RETRY:                              \
 537                         ret = BTREE_RETRY_READ;                         \
 538                         goto fsck_err;                                  \
 539                 case BTREE_ERR_FATAL:                                   \
 540                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 541                         goto fsck_err;                                  \
 542                 }                                                       \
 543                 break;                                                  \
 544         case WRITE:                                                     \
 545                 bch_err(c, "corrupt metadata before write: %s", _buf2); \
 546                                                                         \
 547                 if (bch2_fs_inconsistent(c)) {                          \
 548                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 549                         goto fsck_err;                                  \
 550                 }                                                       \
 551                 break;                                                  \
 552         }                                                               \
 553 out:                                                                    \
 554         if (_buf2 != _buf)                                              \
 555                 kfree(_buf2);                                           \
 556         true;                                                           \
 557 })
 558
 559 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 560
 561 /*
 562  * When btree topology repair changes the start or end of a node, that might
 563  * mean we have to drop keys that are no longer inside the node:
 564  */
 565 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
 566 {
 567         struct bset_tree *t;
 568
 569         for_each_bset(b, t) {
 570                 struct bset *i = bset(b, t);
 571                 struct bkey_packed *k;
 572
 573                 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
 574                         if (bkey_cmp_left_packed(b, k, &b->data->min_key) < 0)
 575                                 break;
 576
 577                 if (k != i->start) {
 578                         unsigned shift = (u64 *) k - (u64 *) i->start;
 579
 580                         memmove_u64s_down(i->start, k,
 581                                           (u64 *) vstruct_end(i) - (u64 *) k);
 582                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
 583                         set_btree_bset_end(b, t);
 584                         bch2_bset_set_no_aux_tree(b, t);
 585                 }
 586
 587                 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
 588                         if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
 589                                 break;
 590
 591                 if (k != vstruct_last(i)) {
 592                         i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
 593                         set_btree_bset_end(b, t);
 594                         bch2_bset_set_no_aux_tree(b, t);
 595                 }
 596         }
 597
 598         bch2_btree_build_aux_trees(b);
 599 }
 600
 601 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 602                          struct btree *b, struct bset *i,
 603                          unsigned sectors, int write, bool have_retry)
 604 {
 605         unsigned version = le16_to_cpu(i->version);
 606         const char *err;
 607         char buf1[100];
 608         char buf2[100];
 609         int ret = 0;
 610
 611         btree_err_on((version != BCH_BSET_VERSION_OLD &&
 612                       version < bcachefs_metadata_version_min) ||
 613                      version >= bcachefs_metadata_version_max,
 614                      BTREE_ERR_FATAL, c, ca, b, i,
 615                      "unsupported bset version");
 616
 617         if (btree_err_on(version < c->sb.version_min,
 618                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 619                          "bset version %u older than superblock version_min %u",
 620                          version, c->sb.version_min)) {
 621                 mutex_lock(&c->sb_lock);
 622                 c->disk_sb.sb->version_min = cpu_to_le16(version);
 623                 bch2_write_super(c);
 624                 mutex_unlock(&c->sb_lock);
 625         }
 626
 627         if (btree_err_on(version > c->sb.version,
 628                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 629                          "bset version %u newer than superblock version %u",
 630                          version, c->sb.version)) {
 631                 mutex_lock(&c->sb_lock);
 632                 c->disk_sb.sb->version = cpu_to_le16(version);
 633                 bch2_write_super(c);
 634                 mutex_unlock(&c->sb_lock);
 635         }
 636
 637         btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
 638                      BTREE_ERR_FATAL, c, ca, b, i,
 639                      "BSET_SEPARATE_WHITEOUTS no longer supported");
 640
 641         if (btree_err_on(b->written + sectors > c->opts.btree_node_size,
 642                          BTREE_ERR_FIXABLE, c, ca, b, i,
 643                          "bset past end of btree node")) {
 644                 i->u64s = 0;
 645                 return 0;
 646         }
 647
 648         btree_err_on(b->written && !i->u64s,
 649                      BTREE_ERR_FIXABLE, c, ca, b, i,
 650                      "empty bset");
 651
 652         if (!b->written) {
 653                 struct btree_node *bn =
 654                         container_of(i, struct btree_node, keys);
 655                 /* These indicate that we read the wrong btree node: */
 656
 657                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 658                         struct bch_btree_ptr_v2 *bp =
 659                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 660
 661                         /* XXX endianness */
 662                         btree_err_on(bp->seq != bn->keys.seq,
 663                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 664                                      "incorrect sequence number (wrong btree node)");
 665                 }
 666
 667                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 668                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 669                              "incorrect btree id");
 670
 671                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 672                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 673                              "incorrect level");
 674
 675                 if (!write)
 676                         compat_btree_node(b->c.level, b->c.btree_id, version,
 677                                           BSET_BIG_ENDIAN(i), write, bn);
 678
 679                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 680                         struct bch_btree_ptr_v2 *bp =
 681                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 682
 683                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 684                                 b->data->min_key = bp->min_key;
 685                                 b->data->max_key = b->key.k.p;
 686                         }
 687
 688                         btree_err_on(bpos_cmp(b->data->min_key, bp->min_key),
 689                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 690                                      "incorrect min_key: got %s should be %s",
 691                                      (bch2_bpos_to_text(&PBUF(buf1), bn->min_key), buf1),
 692                                      (bch2_bpos_to_text(&PBUF(buf2), bp->min_key), buf2));
 693                 }
 694
 695                 btree_err_on(bpos_cmp(bn->max_key, b->key.k.p),
 696                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 697                              "incorrect max key %s",
 698                              (bch2_bpos_to_text(&PBUF(buf1), bn->max_key), buf1));
 699
 700                 if (write)
 701                         compat_btree_node(b->c.level, b->c.btree_id, version,
 702                                           BSET_BIG_ENDIAN(i), write, bn);
 703
 704                 err = bch2_bkey_format_validate(&bn->format);
 705                 btree_err_on(err,
 706                              BTREE_ERR_FATAL, c, ca, b, i,
 707                              "invalid bkey format: %s", err);
 708
 709                 compat_bformat(b->c.level, b->c.btree_id, version,
 710                                BSET_BIG_ENDIAN(i), write,
 711                                &bn->format);
 712         }
 713 fsck_err:
 714         return ret;
 715 }
 716
 717 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 718                          struct bset *i, unsigned *whiteout_u64s,
 719                          int write, bool have_retry)
 720 {
 721         unsigned version = le16_to_cpu(i->version);
 722         struct bkey_packed *k, *prev = NULL;
 723         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 724                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 725         int ret = 0;
 726
 727         for (k = i->start;
 728              k != vstruct_last(i);) {
 729                 struct bkey_s u;
 730                 struct bkey tmp;
 731                 const char *invalid;
 732
 733                 if (btree_err_on(bkey_next(k) > vstruct_last(i),
 734                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 735                                  "key extends past end of bset")) {
 736                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 737                         break;
 738                 }
 739
 740                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 741                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 742                                  "invalid bkey format %u", k->format)) {
 743                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 744                         memmove_u64s_down(k, bkey_next(k),
 745                                           (u64 *) vstruct_end(i) - (u64 *) k);
 746                         continue;
 747                 }
 748
 749                 /* XXX: validate k->u64s */
 750                 if (!write)
 751                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 752                                     BSET_BIG_ENDIAN(i), write,
 753                                     &b->format, k);
 754
 755                 u = __bkey_disassemble(b, k, &tmp);
 756
 757                 invalid = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b)) ?:
 758                         (!updated_range ?  bch2_bkey_in_btree_node(b, u.s_c) : NULL) ?:
 759                         (write ? bch2_bkey_val_invalid(c, u.s_c) : NULL);
 760                 if (invalid) {
 761                         char buf[160];
 762
 763                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 764                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 765                                   "invalid bkey: %s\n%s", invalid, buf);
 766
 767                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 768                         memmove_u64s_down(k, bkey_next(k),
 769                                           (u64 *) vstruct_end(i) - (u64 *) k);
 770                         continue;
 771                 }
 772
 773                 if (write)
 774                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 775                                     BSET_BIG_ENDIAN(i), write,
 776                                     &b->format, k);
 777
 778                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 779                         char buf1[80];
 780                         char buf2[80];
 781                         struct bkey up = bkey_unpack_key(b, prev);
 782
 783                         bch2_bkey_to_text(&PBUF(buf1), &up);
 784                         bch2_bkey_to_text(&PBUF(buf2), u.k);
 785
 786                         bch2_dump_bset(c, b, i, 0);
 787
 788                         if (btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 789                                       "keys out of order: %s > %s",
 790                                       buf1, buf2)) {
 791                                 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 792                                 memmove_u64s_down(k, bkey_next(k),
 793                                                   (u64 *) vstruct_end(i) - (u64 *) k);
 794                                 continue;
 795                         }
 796                 }
 797
 798                 prev = k;
 799                 k = bkey_next(k);
 800         }
 801 fsck_err:
 802         return ret;
 803 }
 804
 805 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 806                               struct btree *b, bool have_retry)
 807 {
 808         struct btree_node_entry *bne;
 809         struct sort_iter *iter;
 810         struct btree_node *sorted;
 811         struct bkey_packed *k;
 812         struct bch_extent_ptr *ptr;
 813         struct bset *i;
 814         bool used_mempool, blacklisted;
 815         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 816                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 817         unsigned u64s;
 818         unsigned nonblacklisted_written = 0;
 819         int ret, retry_read = 0, write = READ;
 820
 821         b->version_ondisk = U16_MAX;
 822
 823         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
 824         sort_iter_init(iter, b);
 825         iter->size = (btree_blocks(c) + 1) * 2;
 826
 827         if (bch2_meta_read_fault("btree"))
 828                 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 829                           "dynamic fault");
 830
 831         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 832                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 833                      "bad magic");
 834
 835         btree_err_on(!b->data->keys.seq,
 836                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 837                      "bad btree header");
 838
 839         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 840                 struct bch_btree_ptr_v2 *bp =
 841                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 842
 843                 btree_err_on(b->data->keys.seq != bp->seq,
 844                              BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 845                              "got wrong btree node (seq %llx want %llx)",
 846                              b->data->keys.seq, bp->seq);
 847         }
 848
 849         while (b->written < c->opts.btree_node_size) {
 850                 unsigned sectors, whiteout_u64s = 0;
 851                 struct nonce nonce;
 852                 struct bch_csum csum;
 853                 bool first = !b->written;
 854
 855                 if (!b->written) {
 856                         i = &b->data->keys;
 857
 858                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 859                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 860                                      "unknown checksum type %llu",
 861                                      BSET_CSUM_TYPE(i));
 862
 863                         nonce = btree_nonce(i, b->written << 9);
 864                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 865
 866                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 867                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 868                                      "invalid checksum");
 869
 870                         bset_encrypt(c, i, b->written << 9);
 871
 872                         btree_err_on(btree_node_is_extents(b) &&
 873                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
 874                                      BTREE_ERR_FATAL, c, NULL, b, NULL,
 875                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
 876
 877                         sectors = vstruct_sectors(b->data, c->block_bits);
 878                 } else {
 879                         bne = write_block(b);
 880                         i = &bne->keys;
 881
 882                         if (i->seq != b->data->keys.seq)
 883                                 break;
 884
 885                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 886                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 887                                      "unknown checksum type %llu",
 888                                      BSET_CSUM_TYPE(i));
 889
 890                         nonce = btree_nonce(i, b->written << 9);
 891                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
 892
 893                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
 894                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 895                                      "invalid checksum");
 896
 897                         bset_encrypt(c, i, b->written << 9);
 898
 899                         sectors = vstruct_sectors(bne, c->block_bits);
 900                 }
 901
 902                 b->version_ondisk = min(b->version_ondisk,
 903                                         le16_to_cpu(i->version));
 904
 905                 ret = validate_bset(c, ca, b, i, sectors,
 906                                     READ, have_retry);
 907                 if (ret)
 908                         goto fsck_err;
 909
 910                 if (!b->written)
 911                         btree_node_set_format(b, b->data->format);
 912
 913                 ret = validate_bset_keys(c, b, i, &whiteout_u64s,
 914                                     READ, have_retry);
 915                 if (ret)
 916                         goto fsck_err;
 917
 918                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 919
 920                 b->written += sectors;
 921
 922                 blacklisted = bch2_journal_seq_is_blacklisted(c,
 923                                         le64_to_cpu(i->journal_seq),
 924                                         true);
 925
 926                 btree_err_on(blacklisted && first,
 927                              BTREE_ERR_FIXABLE, c, ca, b, i,
 928                              "first btree node bset has blacklisted journal seq");
 929                 if (blacklisted && !first)
 930                         continue;
 931
 932                 sort_iter_add(iter, i->start,
 933                               vstruct_idx(i, whiteout_u64s));
 934
 935                 sort_iter_add(iter,
 936                               vstruct_idx(i, whiteout_u64s),
 937                               vstruct_last(i));
 938
 939                 nonblacklisted_written = b->written;
 940         }
 941
 942         for (bne = write_block(b);
 943              bset_byte_offset(b, bne) < btree_bytes(c);
 944              bne = (void *) bne + block_bytes(c))
 945                 btree_err_on(bne->keys.seq == b->data->keys.seq &&
 946                              !bch2_journal_seq_is_blacklisted(c,
 947                                         le64_to_cpu(bne->keys.journal_seq),
 948                                         true),
 949                              BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
 950                              "found bset signature after last bset");
 951
 952         /*
 953          * Blacklisted bsets are those that were written after the most recent
 954          * (flush) journal write. Since there wasn't a flush, they may not have
 955          * made it to all devices - which means we shouldn't write new bsets
 956          * after them, as that could leave a gap and then reads from that device
 957          * wouldn't find all the bsets in that btree node - which means it's
 958          * important that we start writing new bsets after the most recent _non_
 959          * blacklisted bset:
 960          */
 961         b->written = nonblacklisted_written;
 962
 963         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
 964         sorted->keys.u64s = 0;
 965
 966         set_btree_bset(b, b->set, &b->data->keys);
 967
 968         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
 969
 970         u64s = le16_to_cpu(sorted->keys.u64s);
 971         *sorted = *b->data;
 972         sorted->keys.u64s = cpu_to_le16(u64s);
 973         swap(sorted, b->data);
 974         set_btree_bset(b, b->set, &b->data->keys);
 975         b->nsets = 1;
 976
 977         BUG_ON(b->nr.live_u64s != u64s);
 978
 979         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
 980
 981         if (updated_range)
 982                 bch2_btree_node_drop_keys_outside_node(b);
 983
 984         i = &b->data->keys;
 985         for (k = i->start; k != vstruct_last(i);) {
 986                 struct bkey tmp;
 987                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
 988                 const char *invalid = bch2_bkey_val_invalid(c, u.s_c);
 989
 990                 if (invalid ||
 991                     (bch2_inject_invalid_keys &&
 992                      !bversion_cmp(u.k->version, MAX_VERSION))) {
 993                         char buf[160];
 994
 995                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 996                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 997                                   "invalid bkey %s: %s", buf, invalid);
 998
 999                         btree_keys_account_key_drop(&b->nr, 0, k);
1000
1001                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1002                         memmove_u64s_down(k, bkey_next(k),
1003                                           (u64 *) vstruct_end(i) - (u64 *) k);
1004                         set_btree_bset_end(b, b->set);
1005                         continue;
1006                 }
1007
1008                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1009                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1010
1011                         bp.v->mem_ptr = 0;
1012                 }
1013
1014                 k = bkey_next(k);
1015         }
1016
1017         bch2_bset_build_aux_tree(b, b->set, false);
1018
1019         set_needs_whiteout(btree_bset_first(b), true);
1020
1021         btree_node_reset_sib_u64s(b);
1022
1023         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1024                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1025
1026                 if (ca->mi.state != BCH_MEMBER_STATE_rw)
1027                         set_btree_node_need_rewrite(b);
1028         }
1029 out:
1030         mempool_free(iter, &c->fill_iter);
1031         return retry_read;
1032 fsck_err:
1033         if (ret == BTREE_RETRY_READ) {
1034                 retry_read = 1;
1035         } else {
1036                 bch2_inconsistent_error(c);
1037                 set_btree_node_read_error(b);
1038         }
1039         goto out;
1040 }
1041
1042 static void btree_node_read_work(struct work_struct *work)
1043 {
1044         struct btree_read_bio *rb =
1045                 container_of(work, struct btree_read_bio, work);
1046         struct bch_fs *c        = rb->c;
1047         struct btree *b         = rb->b;
1048         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1049         struct bio *bio         = &rb->bio;
1050         struct bch_io_failures failed = { .nr = 0 };
1051         char buf[200];
1052         struct printbuf out;
1053         bool saw_error = false;
1054         bool can_retry;
1055
1056         goto start;
1057         while (1) {
1058                 bch_info(c, "retrying read");
1059                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1060                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1061                 bio_reset(bio);
1062                 bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1063                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1064                 bio->bi_iter.bi_size    = btree_bytes(c);
1065
1066                 if (rb->have_ioref) {
1067                         bio_set_dev(bio, ca->disk_sb.bdev);
1068                         submit_bio_wait(bio);
1069                 } else {
1070                         bio->bi_status = BLK_STS_REMOVED;
1071                 }
1072 start:
1073                 out = PBUF(buf);
1074                 btree_pos_to_text(&out, c, b);
1075                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1076                                    bch2_blk_status_to_str(bio->bi_status), buf);
1077                 if (rb->have_ioref)
1078                         percpu_ref_put(&ca->io_ref);
1079                 rb->have_ioref = false;
1080
1081                 bch2_mark_io_failure(&failed, &rb->pick);
1082
1083                 can_retry = bch2_bkey_pick_read_device(c,
1084                                 bkey_i_to_s_c(&b->key),
1085                                 &failed, &rb->pick) > 0;
1086
1087                 if (!bio->bi_status &&
1088                     !bch2_btree_node_read_done(c, ca, b, can_retry))
1089                         break;
1090
1091                 saw_error = true;
1092
1093                 if (!can_retry) {
1094                         set_btree_node_read_error(b);
1095                         break;
1096                 }
1097         }
1098
1099         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1100                                rb->start_time);
1101         bio_put(&rb->bio);
1102
1103         if (saw_error && !btree_node_read_error(b))
1104                 bch2_btree_node_rewrite_async(c, b);
1105
1106         clear_btree_node_read_in_flight(b);
1107         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1108 }
1109
1110 static void btree_node_read_endio(struct bio *bio)
1111 {
1112         struct btree_read_bio *rb =
1113                 container_of(bio, struct btree_read_bio, bio);
1114         struct bch_fs *c        = rb->c;
1115
1116         if (rb->have_ioref) {
1117                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1118                 bch2_latency_acct(ca, rb->start_time, READ);
1119         }
1120
1121         queue_work(c->io_complete_wq, &rb->work);
1122 }
1123
1124 struct btree_node_read_all {
1125         struct closure          cl;
1126         struct bch_fs           *c;
1127         struct btree            *b;
1128         unsigned                nr;
1129         void                    *buf[BCH_REPLICAS_MAX];
1130         struct bio              *bio[BCH_REPLICAS_MAX];
1131         int                     err[BCH_REPLICAS_MAX];
1132 };
1133
1134 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1135 {
1136         struct btree_node *bn = data;
1137         struct btree_node_entry *bne;
1138         unsigned offset = 0;
1139
1140         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1141                 return 0;
1142
1143         while (offset < c->opts.btree_node_size) {
1144                 if (!offset) {
1145                         offset += vstruct_sectors(bn, c->block_bits);
1146                 } else {
1147                         bne = data + (offset << 9);
1148                         if (bne->keys.seq != bn->keys.seq)
1149                                 break;
1150                         offset += vstruct_sectors(bne, c->block_bits);
1151                 }
1152         }
1153
1154         return offset;
1155 }
1156
1157 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1158 {
1159         struct btree_node *bn = data;
1160         struct btree_node_entry *bne;
1161
1162         if (!offset)
1163                 return false;
1164
1165         while (offset < c->opts.btree_node_size) {
1166                 bne = data + (offset << 9);
1167                 if (bne->keys.seq == bn->keys.seq)
1168                         return true;
1169                 offset++;
1170         }
1171
1172         return false;
1173         return offset;
1174 }
1175
1176 static void btree_node_read_all_replicas_done(struct closure *cl)
1177 {
1178         struct btree_node_read_all *ra =
1179                 container_of(cl, struct btree_node_read_all, cl);
1180         struct bch_fs *c = ra->c;
1181         struct btree *b = ra->b;
1182         bool dump_bset_maps = false;
1183         bool have_retry = false;
1184         int ret = 0, best = -1, write = READ;
1185         unsigned i, written, written2;
1186         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1187                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1188
1189         for (i = 0; i < ra->nr; i++) {
1190                 struct btree_node *bn = ra->buf[i];
1191
1192                 if (ra->err[i])
1193                         continue;
1194
1195                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1196                     (seq && seq != bn->keys.seq))
1197                         continue;
1198
1199                 if (best < 0) {
1200                         best = i;
1201                         written = btree_node_sectors_written(c, bn);
1202                         continue;
1203                 }
1204
1205                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1206                 if (btree_err_on(written2 != written, BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1207                                  "btree node sectors written mismatch: %u != %u",
1208                                  written, written2) ||
1209                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1210                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1211                                  "found bset signature after last bset") ||
1212                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1213                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1214                                  "btree node replicas content mismatch"))
1215                         dump_bset_maps = true;
1216
1217                 if (written2 > written) {
1218                         written = written2;
1219                         best = i;
1220                 }
1221         }
1222 fsck_err:
1223         if (dump_bset_maps) {
1224                 for (i = 0; i < ra->nr; i++) {
1225                         char buf[200];
1226                         struct printbuf out = PBUF(buf);
1227                         struct btree_node *bn = ra->buf[i];
1228                         struct btree_node_entry *bne = NULL;
1229                         unsigned offset = 0, sectors;
1230                         bool gap = false;
1231
1232                         if (ra->err[i])
1233                                 continue;
1234
1235                         while (offset < c->opts.btree_node_size) {
1236                                 if (!offset) {
1237                                         sectors = vstruct_sectors(bn, c->block_bits);
1238                                 } else {
1239                                         bne = ra->buf[i] + (offset << 9);
1240                                         if (bne->keys.seq != bn->keys.seq)
1241                                                 break;
1242                                         sectors = vstruct_sectors(bne, c->block_bits);
1243                                 }
1244
1245                                 pr_buf(&out, " %u-%u", offset, offset + sectors);
1246                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1247                                                         le64_to_cpu(bne->keys.journal_seq), false))
1248                                         pr_buf(&out, "*");
1249                                 offset += sectors;
1250                         }
1251
1252                         while (offset < c->opts.btree_node_size) {
1253                                 bne = ra->buf[i] + (offset << 9);
1254                                 if (bne->keys.seq == bn->keys.seq) {
1255                                         if (!gap)
1256                                                 pr_buf(&out, " GAP");
1257                                         gap = true;
1258
1259                                         sectors = vstruct_sectors(bne, c->block_bits);
1260                                         pr_buf(&out, " %u-%u", offset, offset + sectors);
1261                                         if (bch2_journal_seq_is_blacklisted(c,
1262                                                         le64_to_cpu(bne->keys.journal_seq), false))
1263                                                 pr_buf(&out, "*");
1264                                 }
1265                                 offset++;
1266                         }
1267
1268                         bch_err(c, "replica %u:%s", i, buf);
1269                 }
1270         }
1271
1272         if (best >= 0) {
1273                 memcpy(b->data, ra->buf[best], btree_bytes(c));
1274                 ret = bch2_btree_node_read_done(c, NULL, b, false);
1275         } else {
1276                 ret = -1;
1277         }
1278
1279         if (ret)
1280                 set_btree_node_read_error(b);
1281
1282         for (i = 0; i < ra->nr; i++) {
1283                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1284                 bio_put(ra->bio[i]);
1285         }
1286
1287         closure_debug_destroy(&ra->cl);
1288         kfree(ra);
1289
1290         clear_btree_node_read_in_flight(b);
1291         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1292 }
1293
1294 static void btree_node_read_all_replicas_endio(struct bio *bio)
1295 {
1296         struct btree_read_bio *rb =
1297                 container_of(bio, struct btree_read_bio, bio);
1298         struct bch_fs *c        = rb->c;
1299         struct btree_node_read_all *ra = rb->ra;
1300
1301         if (rb->have_ioref) {
1302                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1303                 bch2_latency_acct(ca, rb->start_time, READ);
1304         }
1305
1306         ra->err[rb->idx] = bio->bi_status;
1307         closure_put(&ra->cl);
1308 }
1309
1310 /*
1311  * XXX This allocates multiple times from the same mempools, and can deadlock
1312  * under sufficient memory pressure (but is only a debug path)
1313  */
1314 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1315 {
1316         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1317         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1318         const union bch_extent_entry *entry;
1319         struct extent_ptr_decoded pick;
1320         struct btree_node_read_all *ra;
1321         unsigned i;
1322
1323         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1324         if (!ra)
1325                 return -ENOMEM;
1326
1327         closure_init(&ra->cl, NULL);
1328         ra->c   = c;
1329         ra->b   = b;
1330         ra->nr  = bch2_bkey_nr_ptrs(k);
1331
1332         for (i = 0; i < ra->nr; i++) {
1333                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1334                 ra->bio[i] = bio_alloc_bioset(GFP_NOFS, buf_pages(ra->buf[i],
1335                                                                   btree_bytes(c)),
1336                                               &c->btree_bio);
1337         }
1338
1339         i = 0;
1340         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1341                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1342                 struct btree_read_bio *rb =
1343                         container_of(ra->bio[i], struct btree_read_bio, bio);
1344                 rb->c                   = c;
1345                 rb->b                   = b;
1346                 rb->ra                  = ra;
1347                 rb->start_time          = local_clock();
1348                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1349                 rb->idx                 = i;
1350                 rb->pick                = pick;
1351                 rb->bio.bi_opf          = REQ_OP_READ|REQ_SYNC|REQ_META;
1352                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1353                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1354                 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1355
1356                 if (rb->have_ioref) {
1357                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1358                                      bio_sectors(&rb->bio));
1359                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1360
1361                         closure_get(&ra->cl);
1362                         submit_bio(&rb->bio);
1363                 } else {
1364                         ra->err[i] = BLK_STS_REMOVED;
1365                 }
1366
1367                 i++;
1368         }
1369
1370         if (sync) {
1371                 closure_sync(&ra->cl);
1372                 btree_node_read_all_replicas_done(&ra->cl);
1373         } else {
1374                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1375                             c->io_complete_wq);
1376         }
1377
1378         return 0;
1379 }
1380
1381 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1382                           bool sync)
1383 {
1384         struct extent_ptr_decoded pick;
1385         struct btree_read_bio *rb;
1386         struct bch_dev *ca;
1387         struct bio *bio;
1388         char buf[200];
1389         int ret;
1390
1391         btree_pos_to_text(&PBUF(buf), c, b);
1392         trace_btree_read(c, b);
1393
1394         set_btree_node_read_in_flight(b);
1395
1396         if (bch2_verify_all_btree_replicas &&
1397             !btree_node_read_all_replicas(c, b, sync))
1398                 return;
1399
1400         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1401                                          NULL, &pick);
1402         if (bch2_fs_fatal_err_on(ret <= 0, c,
1403                         "btree node read error: no device to read from\n"
1404                         " at %s", buf)) {
1405                 set_btree_node_read_error(b);
1406                 return;
1407         }
1408
1409         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1410
1411         bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1412                                                    btree_bytes(c)),
1413                                &c->btree_bio);
1414         rb = container_of(bio, struct btree_read_bio, bio);
1415         rb->c                   = c;
1416         rb->b                   = b;
1417         rb->ra                  = NULL;
1418         rb->start_time          = local_clock();
1419         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1420         rb->pick                = pick;
1421         INIT_WORK(&rb->work, btree_node_read_work);
1422         bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1423         bio->bi_iter.bi_sector  = pick.ptr.offset;
1424         bio->bi_end_io          = btree_node_read_endio;
1425         bch2_bio_map(bio, b->data, btree_bytes(c));
1426
1427         if (rb->have_ioref) {
1428                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1429                              bio_sectors(bio));
1430                 bio_set_dev(bio, ca->disk_sb.bdev);
1431
1432                 if (sync) {
1433                         submit_bio_wait(bio);
1434
1435                         btree_node_read_work(&rb->work);
1436                 } else {
1437                         submit_bio(bio);
1438                 }
1439         } else {
1440                 bio->bi_status = BLK_STS_REMOVED;
1441
1442                 if (sync)
1443                         btree_node_read_work(&rb->work);
1444                 else
1445                         queue_work(c->io_complete_wq, &rb->work);
1446         }
1447 }
1448
1449 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1450                         const struct bkey_i *k, unsigned level)
1451 {
1452         struct closure cl;
1453         struct btree *b;
1454         int ret;
1455
1456         closure_init_stack(&cl);
1457
1458         do {
1459                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1460                 closure_sync(&cl);
1461         } while (ret);
1462
1463         b = bch2_btree_node_mem_alloc(c);
1464         bch2_btree_cache_cannibalize_unlock(c);
1465
1466         BUG_ON(IS_ERR(b));
1467
1468         bkey_copy(&b->key, k);
1469         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1470
1471         bch2_btree_node_read(c, b, true);
1472
1473         if (btree_node_read_error(b)) {
1474                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1475
1476                 mutex_lock(&c->btree_cache.lock);
1477                 list_move(&b->list, &c->btree_cache.freeable);
1478                 mutex_unlock(&c->btree_cache.lock);
1479
1480                 ret = -EIO;
1481                 goto err;
1482         }
1483
1484         bch2_btree_set_root_for_read(c, b);
1485 err:
1486         six_unlock_write(&b->c.lock);
1487         six_unlock_intent(&b->c.lock);
1488
1489         return ret;
1490 }
1491
1492 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1493                               struct btree_write *w)
1494 {
1495         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1496
1497         do {
1498                 old = new = v;
1499                 if (!(old & 1))
1500                         break;
1501
1502                 new &= ~1UL;
1503         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1504
1505         if (old & 1)
1506                 closure_put(&((struct btree_update *) new)->cl);
1507
1508         bch2_journal_pin_drop(&c->journal, &w->journal);
1509 }
1510
1511 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1512 {
1513         struct btree_write *w = btree_prev_write(b);
1514
1515         bch2_btree_complete_write(c, b, w);
1516         btree_node_io_unlock(b);
1517 }
1518
1519 static void bch2_btree_node_write_error(struct bch_fs *c,
1520                                         struct btree_write_bio *wbio)
1521 {
1522         struct btree *b         = wbio->wbio.bio.bi_private;
1523         struct bkey_buf k;
1524         struct bch_extent_ptr *ptr;
1525         struct btree_trans trans;
1526         struct btree_iter *iter;
1527         int ret;
1528
1529         bch2_bkey_buf_init(&k);
1530         bch2_trans_init(&trans, c, 0, 0);
1531
1532         iter = bch2_trans_get_node_iter(&trans, b->c.btree_id, b->key.k.p,
1533                                         BTREE_MAX_DEPTH, b->c.level, 0);
1534 retry:
1535         ret = bch2_btree_iter_traverse(iter);
1536         if (ret)
1537                 goto err;
1538
1539         /* has node been freed? */
1540         if (iter->l[b->c.level].b != b) {
1541                 /* node has been freed: */
1542                 BUG_ON(!btree_node_dying(b));
1543                 goto out;
1544         }
1545
1546         BUG_ON(!btree_node_hashed(b));
1547
1548         bch2_bkey_buf_copy(&k, c, &b->key);
1549
1550         bch2_bkey_drop_ptrs(bkey_i_to_s(k.k), ptr,
1551                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1552
1553         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(k.k)))
1554                 goto err;
1555
1556         ret = bch2_btree_node_update_key(c, iter, b, k.k);
1557         if (ret == -EINTR)
1558                 goto retry;
1559         if (ret)
1560                 goto err;
1561 out:
1562         bch2_trans_iter_put(&trans, iter);
1563         bch2_trans_exit(&trans);
1564         bch2_bkey_buf_exit(&k, c);
1565         bio_put(&wbio->wbio.bio);
1566         btree_node_write_done(c, b);
1567         return;
1568 err:
1569         set_btree_node_noevict(b);
1570         bch2_fs_fatal_error(c, "fatal error writing btree node");
1571         goto out;
1572 }
1573
1574 void bch2_btree_write_error_work(struct work_struct *work)
1575 {
1576         struct bch_fs *c = container_of(work, struct bch_fs,
1577                                         btree_write_error_work);
1578         struct bio *bio;
1579
1580         while (1) {
1581                 spin_lock_irq(&c->btree_write_error_lock);
1582                 bio = bio_list_pop(&c->btree_write_error_list);
1583                 spin_unlock_irq(&c->btree_write_error_lock);
1584
1585                 if (!bio)
1586                         break;
1587
1588                 bch2_btree_node_write_error(c,
1589                         container_of(bio, struct btree_write_bio, wbio.bio));
1590         }
1591 }
1592
1593 static void btree_node_write_work(struct work_struct *work)
1594 {
1595         struct btree_write_bio *wbio =
1596                 container_of(work, struct btree_write_bio, work);
1597         struct bch_fs *c        = wbio->wbio.c;
1598         struct btree *b         = wbio->wbio.bio.bi_private;
1599
1600         btree_bounce_free(c,
1601                 wbio->bytes,
1602                 wbio->wbio.used_mempool,
1603                 wbio->data);
1604
1605         if (wbio->wbio.failed.nr) {
1606                 unsigned long flags;
1607
1608                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1609                 bio_list_add(&c->btree_write_error_list, &wbio->wbio.bio);
1610                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1611
1612                 queue_work(c->btree_error_wq, &c->btree_write_error_work);
1613                 return;
1614         }
1615
1616         bio_put(&wbio->wbio.bio);
1617         btree_node_write_done(c, b);
1618 }
1619
1620 static void btree_node_write_endio(struct bio *bio)
1621 {
1622         struct bch_write_bio *wbio      = to_wbio(bio);
1623         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1624         struct bch_write_bio *orig      = parent ?: wbio;
1625         struct bch_fs *c                = wbio->c;
1626         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1627         unsigned long flags;
1628
1629         if (wbio->have_ioref)
1630                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1631
1632         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1633                                bch2_blk_status_to_str(bio->bi_status)) ||
1634             bch2_meta_write_fault("btree")) {
1635                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1636                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1637                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1638         }
1639
1640         if (wbio->have_ioref)
1641                 percpu_ref_put(&ca->io_ref);
1642
1643         if (parent) {
1644                 bio_put(bio);
1645                 bio_endio(&parent->bio);
1646         } else {
1647                 struct btree_write_bio *wb =
1648                         container_of(orig, struct btree_write_bio, wbio);
1649
1650                 INIT_WORK(&wb->work, btree_node_write_work);
1651                 queue_work(c->io_complete_wq, &wb->work);
1652         }
1653 }
1654
1655 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1656                                    struct bset *i, unsigned sectors)
1657 {
1658         unsigned whiteout_u64s = 0;
1659         int ret;
1660
1661         if (bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), BKEY_TYPE_btree))
1662                 return -1;
1663
1664         ret = validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false) ?:
1665                 validate_bset(c, NULL, b, i, sectors, WRITE, false);
1666         if (ret) {
1667                 bch2_inconsistent_error(c);
1668                 dump_stack();
1669         }
1670
1671         return ret;
1672 }
1673
1674 static void btree_write_submit(struct work_struct *work)
1675 {
1676         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1677
1678         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree, &wbio->key);
1679 }
1680
1681 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b)
1682 {
1683         struct btree_write_bio *wbio;
1684         struct bset_tree *t;
1685         struct bset *i;
1686         struct btree_node *bn = NULL;
1687         struct btree_node_entry *bne = NULL;
1688         struct bch_extent_ptr *ptr;
1689         struct sort_iter sort_iter;
1690         struct nonce nonce;
1691         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1692         u64 seq = 0;
1693         bool used_mempool;
1694         unsigned long old, new;
1695         bool validate_before_checksum = false;
1696         void *data;
1697
1698         if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
1699                 return;
1700
1701         /*
1702          * We may only have a read lock on the btree node - the dirty bit is our
1703          * "lock" against racing with other threads that may be trying to start
1704          * a write, we do a write iff we clear the dirty bit. Since setting the
1705          * dirty bit requires a write lock, we can't race with other threads
1706          * redirtying it:
1707          */
1708         do {
1709                 old = new = READ_ONCE(b->flags);
1710
1711                 if (!(old & (1 << BTREE_NODE_dirty)))
1712                         return;
1713
1714                 if (!btree_node_may_write(b))
1715                         return;
1716
1717                 if (old & (1 << BTREE_NODE_never_write))
1718                         return;
1719
1720                 if (old & (1 << BTREE_NODE_write_in_flight)) {
1721                         /*
1722                          * XXX waiting on btree writes with btree locks held -
1723                          * this can deadlock, and we hit the write error path
1724                          */
1725                         btree_node_wait_on_io(b);
1726                         continue;
1727                 }
1728
1729                 new &= ~(1 << BTREE_NODE_dirty);
1730                 new &= ~(1 << BTREE_NODE_need_write);
1731                 new |=  (1 << BTREE_NODE_write_in_flight);
1732                 new |=  (1 << BTREE_NODE_just_written);
1733                 new ^=  (1 << BTREE_NODE_write_idx);
1734         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1735
1736         atomic_dec(&c->btree_cache.dirty);
1737
1738         BUG_ON(btree_node_fake(b));
1739         BUG_ON((b->will_make_reachable != 0) != !b->written);
1740
1741         BUG_ON(b->written >= c->opts.btree_node_size);
1742         BUG_ON(b->written & (c->opts.block_size - 1));
1743         BUG_ON(bset_written(b, btree_bset_last(b)));
1744         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1745         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1746
1747         bch2_sort_whiteouts(c, b);
1748
1749         sort_iter_init(&sort_iter, b);
1750
1751         bytes = !b->written
1752                 ? sizeof(struct btree_node)
1753                 : sizeof(struct btree_node_entry);
1754
1755         bytes += b->whiteout_u64s * sizeof(u64);
1756
1757         for_each_bset(b, t) {
1758                 i = bset(b, t);
1759
1760                 if (bset_written(b, i))
1761                         continue;
1762
1763                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1764                 sort_iter_add(&sort_iter,
1765                               btree_bkey_first(b, t),
1766                               btree_bkey_last(b, t));
1767                 seq = max(seq, le64_to_cpu(i->journal_seq));
1768         }
1769
1770         BUG_ON(b->written && !seq);
1771
1772         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1773         bytes += 8;
1774
1775         /* buffer must be a multiple of the block size */
1776         bytes = round_up(bytes, block_bytes(c));
1777
1778         data = btree_bounce_alloc(c, bytes, &used_mempool);
1779
1780         if (!b->written) {
1781                 bn = data;
1782                 *bn = *b->data;
1783                 i = &bn->keys;
1784         } else {
1785                 bne = data;
1786                 bne->keys = b->data->keys;
1787                 i = &bne->keys;
1788         }
1789
1790         i->journal_seq  = cpu_to_le64(seq);
1791         i->u64s         = 0;
1792
1793         sort_iter_add(&sort_iter,
1794                       unwritten_whiteouts_start(c, b),
1795                       unwritten_whiteouts_end(c, b));
1796         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1797
1798         b->whiteout_u64s = 0;
1799
1800         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1801         le16_add_cpu(&i->u64s, u64s);
1802
1803         set_needs_whiteout(i, false);
1804
1805         /* do we have data to write? */
1806         if (b->written && !i->u64s)
1807                 goto nowrite;
1808
1809         bytes_to_write = vstruct_end(i) - data;
1810         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1811
1812         memset(data + bytes_to_write, 0,
1813                (sectors_to_write << 9) - bytes_to_write);
1814
1815         BUG_ON(b->written + sectors_to_write > c->opts.btree_node_size);
1816         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1817         BUG_ON(i->seq != b->data->keys.seq);
1818
1819         i->version = c->sb.version < bcachefs_metadata_version_new_versioning
1820                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1821                 : cpu_to_le16(c->sb.version);
1822         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1823
1824         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1825                 validate_before_checksum = true;
1826
1827         /* validate_bset will be modifying: */
1828         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
1829                 validate_before_checksum = true;
1830
1831         /* if we're going to be encrypting, check metadata validity first: */
1832         if (validate_before_checksum &&
1833             validate_bset_for_write(c, b, i, sectors_to_write))
1834                 goto err;
1835
1836         bset_encrypt(c, i, b->written << 9);
1837
1838         nonce = btree_nonce(i, b->written << 9);
1839
1840         if (bn)
1841                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1842         else
1843                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1844
1845         /* if we're not encrypting, check metadata after checksumming: */
1846         if (!validate_before_checksum &&
1847             validate_bset_for_write(c, b, i, sectors_to_write))
1848                 goto err;
1849
1850         /*
1851          * We handle btree write errors by immediately halting the journal -
1852          * after we've done that, we can't issue any subsequent btree writes
1853          * because they might have pointers to new nodes that failed to write.
1854          *
1855          * Furthermore, there's no point in doing any more btree writes because
1856          * with the journal stopped, we're never going to update the journal to
1857          * reflect that those writes were done and the data flushed from the
1858          * journal:
1859          *
1860          * Also on journal error, the pending write may have updates that were
1861          * never journalled (interior nodes, see btree_update_nodes_written()) -
1862          * it's critical that we don't do the write in that case otherwise we
1863          * will have updates visible that weren't in the journal:
1864          *
1865          * Make sure to update b->written so bch2_btree_init_next() doesn't
1866          * break:
1867          */
1868         if (bch2_journal_error(&c->journal) ||
1869             c->opts.nochanges)
1870                 goto err;
1871
1872         trace_btree_write(b, bytes_to_write, sectors_to_write);
1873
1874         wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1875                                 buf_pages(data, sectors_to_write << 9),
1876                                 &c->btree_bio),
1877                             struct btree_write_bio, wbio.bio);
1878         wbio_init(&wbio->wbio.bio);
1879         wbio->data                      = data;
1880         wbio->bytes                     = bytes;
1881         wbio->wbio.c                    = c;
1882         wbio->wbio.used_mempool         = used_mempool;
1883         wbio->wbio.bio.bi_opf           = REQ_OP_WRITE|REQ_META;
1884         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
1885         wbio->wbio.bio.bi_private       = b;
1886
1887         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
1888
1889         /*
1890          * If we're appending to a leaf node, we don't technically need FUA -
1891          * this write just needs to be persisted before the next journal write,
1892          * which will be marked FLUSH|FUA.
1893          *
1894          * Similarly if we're writing a new btree root - the pointer is going to
1895          * be in the next journal entry.
1896          *
1897          * But if we're writing a new btree node (that isn't a root) or
1898          * appending to a non leaf btree node, we need either FUA or a flush
1899          * when we write the parent with the new pointer. FUA is cheaper than a
1900          * flush, and writes appending to leaf nodes aren't blocking anything so
1901          * just make all btree node writes FUA to keep things sane.
1902          */
1903
1904         bkey_copy(&wbio->key, &b->key);
1905
1906         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&wbio->key)), ptr)
1907                 ptr->offset += b->written;
1908
1909         b->written += sectors_to_write;
1910
1911         atomic64_inc(&c->btree_writes_nr);
1912         atomic64_add(sectors_to_write, &c->btree_writes_sectors);
1913
1914         INIT_WORK(&wbio->work, btree_write_submit);
1915         queue_work(c->io_complete_wq, &wbio->work);
1916         return;
1917 err:
1918         set_btree_node_noevict(b);
1919         b->written += sectors_to_write;
1920 nowrite:
1921         btree_bounce_free(c, bytes, used_mempool, data);
1922         btree_node_write_done(c, b);
1923 }
1924
1925 /*
1926  * Work that must be done with write lock held:
1927  */
1928 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
1929 {
1930         bool invalidated_iter = false;
1931         struct btree_node_entry *bne;
1932         struct bset_tree *t;
1933
1934         if (!btree_node_just_written(b))
1935                 return false;
1936
1937         BUG_ON(b->whiteout_u64s);
1938
1939         clear_btree_node_just_written(b);
1940
1941         /*
1942          * Note: immediately after write, bset_written() doesn't work - the
1943          * amount of data we had to write after compaction might have been
1944          * smaller than the offset of the last bset.
1945          *
1946          * However, we know that all bsets have been written here, as long as
1947          * we're still holding the write lock:
1948          */
1949
1950         /*
1951          * XXX: decide if we really want to unconditionally sort down to a
1952          * single bset:
1953          */
1954         if (b->nsets > 1) {
1955                 btree_node_sort(c, b, 0, b->nsets, true);
1956                 invalidated_iter = true;
1957         } else {
1958                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
1959         }
1960
1961         for_each_bset(b, t)
1962                 set_needs_whiteout(bset(b, t), true);
1963
1964         bch2_btree_verify(c, b);
1965
1966         /*
1967          * If later we don't unconditionally sort down to a single bset, we have
1968          * to ensure this is still true:
1969          */
1970         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
1971
1972         bne = want_new_bset(c, b);
1973         if (bne)
1974                 bch2_bset_init_next(c, b, bne);
1975
1976         bch2_btree_build_aux_trees(b);
1977
1978         return invalidated_iter;
1979 }
1980
1981 /*
1982  * Use this one if the node is intent locked:
1983  */
1984 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1985                            enum six_lock_type lock_type_held)
1986 {
1987         if (lock_type_held == SIX_LOCK_intent ||
1988             (lock_type_held == SIX_LOCK_read &&
1989              six_lock_tryupgrade(&b->c.lock))) {
1990                 __bch2_btree_node_write(c, b);
1991
1992                 /* don't cycle lock unnecessarily: */
1993                 if (btree_node_just_written(b) &&
1994                     six_trylock_write(&b->c.lock)) {
1995                         bch2_btree_post_write_cleanup(c, b);
1996                         six_unlock_write(&b->c.lock);
1997                 }
1998
1999                 if (lock_type_held == SIX_LOCK_read)
2000                         six_lock_downgrade(&b->c.lock);
2001         } else {
2002                 __bch2_btree_node_write(c, b);
2003                 if (lock_type_held == SIX_LOCK_write &&
2004                     btree_node_just_written(b))
2005                         bch2_btree_post_write_cleanup(c, b);
2006         }
2007 }
2008
2009 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2010 {
2011         struct bucket_table *tbl;
2012         struct rhash_head *pos;
2013         struct btree *b;
2014         unsigned i;
2015 restart:
2016         rcu_read_lock();
2017         for_each_cached_btree(b, c, tbl, i, pos)
2018                 if (test_bit(flag, &b->flags)) {
2019                         rcu_read_unlock();
2020                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2021                         goto restart;
2022
2023                 }
2024         rcu_read_unlock();
2025 }
2026
2027 void bch2_btree_flush_all_reads(struct bch_fs *c)
2028 {
2029         __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2030 }
2031
2032 void bch2_btree_flush_all_writes(struct bch_fs *c)
2033 {
2034         __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2035 }
2036
2037 void bch2_dirty_btree_nodes_to_text(struct printbuf *out, struct bch_fs *c)
2038 {
2039         struct bucket_table *tbl;
2040         struct rhash_head *pos;
2041         struct btree *b;
2042         unsigned i;
2043
2044         rcu_read_lock();
2045         for_each_cached_btree(b, c, tbl, i, pos) {
2046                 unsigned long flags = READ_ONCE(b->flags);
2047
2048                 if (!(flags & (1 << BTREE_NODE_dirty)))
2049                         continue;
2050
2051                 pr_buf(out, "%p d %u n %u l %u w %u b %u r %u:%lu\n",
2052                        b,
2053                        (flags & (1 << BTREE_NODE_dirty)) != 0,
2054                        (flags & (1 << BTREE_NODE_need_write)) != 0,
2055                        b->c.level,
2056                        b->written,
2057                        !list_empty_careful(&b->write_blocked),
2058                        b->will_make_reachable != 0,
2059                        b->will_make_reachable & 1);
2060         }
2061         rcu_read_unlock();
2062 }