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 have_good_copy = false;
1183         bool dump_bset_maps = false;
1184         bool have_retry = false;
1185         int ret = 0, write = READ;
1186         unsigned i, written, written2;
1187         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1188                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1189
1190         for (i = 0; i < ra->nr; i++) {
1191                 if (ra->err[i])
1192                         continue;
1193
1194                 if (!have_good_copy) {
1195                         memcpy(b->data, ra->buf[i], btree_bytes(c));
1196                         have_good_copy = true;
1197                         written = btree_node_sectors_written(c, b->data);
1198                 }
1199
1200                 /* Try to get the right btree node: */
1201                 if (have_good_copy &&
1202                     seq &&
1203                     b->data->keys.seq != seq &&
1204                     ((struct btree_node *) ra->buf[i])->keys.seq == seq) {
1205                         memcpy(b->data, ra->buf[i], btree_bytes(c));
1206                         written = btree_node_sectors_written(c, b->data);
1207                 }
1208
1209                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1210                 if (btree_err_on(written2 != written, BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1211                                  "btree node sectors written mismatch: %u != %u",
1212                                  written, written2) ||
1213                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1214                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1215                                  "found bset signature after last bset") ||
1216                     btree_err_on(memcmp(b->data, ra->buf[i], written << 9),
1217                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1218                                  "btree node replicas content mismatch"))
1219                         dump_bset_maps = true;
1220
1221                 if (written2 > written) {
1222                         written = written2;
1223                         memcpy(b->data, ra->buf[i], btree_bytes(c));
1224                 }
1225         }
1226 fsck_err:
1227         if (dump_bset_maps) {
1228                 for (i = 0; i < ra->nr; i++) {
1229                         char buf[200];
1230                         struct printbuf out = PBUF(buf);
1231                         struct btree_node *bn = ra->buf[i];
1232                         struct btree_node_entry *bne = NULL;
1233                         unsigned offset = 0, sectors;
1234                         bool gap = false;
1235
1236                         if (ra->err[i])
1237                                 continue;
1238
1239                         while (offset < c->opts.btree_node_size) {
1240                                 if (!offset) {
1241                                         sectors = vstruct_sectors(bn, c->block_bits);
1242                                 } else {
1243                                         bne = ra->buf[i] + (offset << 9);
1244                                         if (bne->keys.seq != bn->keys.seq)
1245                                                 break;
1246                                         sectors = vstruct_sectors(bne, c->block_bits);
1247                                 }
1248
1249                                 pr_buf(&out, " %u-%u", offset, offset + sectors);
1250                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1251                                                         le64_to_cpu(bne->keys.journal_seq), false))
1252                                         pr_buf(&out, "*");
1253                                 offset += sectors;
1254                         }
1255
1256                         while (offset < c->opts.btree_node_size) {
1257                                 bne = ra->buf[i] + (offset << 9);
1258                                 if (bne->keys.seq == bn->keys.seq) {
1259                                         if (!gap)
1260                                                 pr_buf(&out, " GAP");
1261                                         gap = true;
1262
1263                                         sectors = vstruct_sectors(bne, c->block_bits);
1264                                         pr_buf(&out, " %u-%u", offset, offset + sectors);
1265                                         if (bch2_journal_seq_is_blacklisted(c,
1266                                                         le64_to_cpu(bne->keys.journal_seq), false))
1267                                                 pr_buf(&out, "*");
1268                                 }
1269                                 offset++;
1270                         }
1271
1272                         bch_err(c, "replica %u:%s", i, buf);
1273                 }
1274         }
1275
1276         if (have_good_copy)
1277                 bch2_btree_node_read_done(c, NULL, b, false);
1278         else
1279                 set_btree_node_read_error(b);
1280
1281         for (i = 0; i < ra->nr; i++) {
1282                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1283                 bio_put(ra->bio[i]);
1284         }
1285
1286         closure_debug_destroy(&ra->cl);
1287         kfree(ra);
1288
1289         clear_btree_node_read_in_flight(b);
1290         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1291 }
1292
1293 static void btree_node_read_all_replicas_endio(struct bio *bio)
1294 {
1295         struct btree_read_bio *rb =
1296                 container_of(bio, struct btree_read_bio, bio);
1297         struct bch_fs *c        = rb->c;
1298         struct btree_node_read_all *ra = rb->ra;
1299
1300         if (rb->have_ioref) {
1301                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1302                 bch2_latency_acct(ca, rb->start_time, READ);
1303         }
1304
1305         ra->err[rb->idx] = bio->bi_status;
1306         closure_put(&ra->cl);
1307 }
1308
1309 /*
1310  * XXX This allocates multiple times from the same mempools, and can deadlock
1311  * under sufficient memory pressure (but is only a debug path)
1312  */
1313 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1314 {
1315         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1316         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1317         const union bch_extent_entry *entry;
1318         struct extent_ptr_decoded pick;
1319         struct btree_node_read_all *ra;
1320         unsigned i;
1321
1322         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1323         if (!ra)
1324                 return -ENOMEM;
1325
1326         closure_init(&ra->cl, NULL);
1327         ra->c   = c;
1328         ra->b   = b;
1329         ra->nr  = bch2_bkey_nr_ptrs(k);
1330
1331         for (i = 0; i < ra->nr; i++) {
1332                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1333                 ra->bio[i] = bio_alloc_bioset(GFP_NOFS, buf_pages(ra->buf[i],
1334                                                                   btree_bytes(c)),
1335                                               &c->btree_bio);
1336         }
1337
1338         i = 0;
1339         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1340                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1341                 struct btree_read_bio *rb =
1342                         container_of(ra->bio[i], struct btree_read_bio, bio);
1343                 rb->c                   = c;
1344                 rb->b                   = b;
1345                 rb->ra                  = ra;
1346                 rb->start_time          = local_clock();
1347                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1348                 rb->idx                 = i;
1349                 rb->pick                = pick;
1350                 rb->bio.bi_opf          = REQ_OP_READ|REQ_SYNC|REQ_META;
1351                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1352                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1353                 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1354
1355                 if (rb->have_ioref) {
1356                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1357                                      bio_sectors(&rb->bio));
1358                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1359
1360                         closure_get(&ra->cl);
1361                         submit_bio(&rb->bio);
1362                 } else {
1363                         ra->err[i] = BLK_STS_REMOVED;
1364                 }
1365
1366                 i++;
1367         }
1368
1369         if (sync) {
1370                 closure_sync(&ra->cl);
1371                 btree_node_read_all_replicas_done(&ra->cl);
1372         } else {
1373                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1374                             c->io_complete_wq);
1375         }
1376
1377         return 0;
1378 }
1379
1380 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1381                           bool sync)
1382 {
1383         struct extent_ptr_decoded pick;
1384         struct btree_read_bio *rb;
1385         struct bch_dev *ca;
1386         struct bio *bio;
1387         char buf[200];
1388         int ret;
1389
1390         btree_pos_to_text(&PBUF(buf), c, b);
1391         trace_btree_read(c, b);
1392
1393         set_btree_node_read_in_flight(b);
1394
1395         if (bch2_verify_all_btree_replicas &&
1396             !btree_node_read_all_replicas(c, b, sync))
1397                 return;
1398
1399         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1400                                          NULL, &pick);
1401         if (bch2_fs_fatal_err_on(ret <= 0, c,
1402                         "btree node read error: no device to read from\n"
1403                         " at %s", buf)) {
1404                 set_btree_node_read_error(b);
1405                 return;
1406         }
1407
1408         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1409
1410         bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1411                                                    btree_bytes(c)),
1412                                &c->btree_bio);
1413         rb = container_of(bio, struct btree_read_bio, bio);
1414         rb->c                   = c;
1415         rb->b                   = b;
1416         rb->ra                  = NULL;
1417         rb->start_time          = local_clock();
1418         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1419         rb->pick                = pick;
1420         INIT_WORK(&rb->work, btree_node_read_work);
1421         bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1422         bio->bi_iter.bi_sector  = pick.ptr.offset;
1423         bio->bi_end_io          = btree_node_read_endio;
1424         bch2_bio_map(bio, b->data, btree_bytes(c));
1425
1426         if (rb->have_ioref) {
1427                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1428                              bio_sectors(bio));
1429                 bio_set_dev(bio, ca->disk_sb.bdev);
1430
1431                 if (sync) {
1432                         submit_bio_wait(bio);
1433
1434                         btree_node_read_work(&rb->work);
1435                 } else {
1436                         submit_bio(bio);
1437                 }
1438         } else {
1439                 bio->bi_status = BLK_STS_REMOVED;
1440
1441                 if (sync)
1442                         btree_node_read_work(&rb->work);
1443                 else
1444                         queue_work(c->io_complete_wq, &rb->work);
1445         }
1446 }
1447
1448 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1449                         const struct bkey_i *k, unsigned level)
1450 {
1451         struct closure cl;
1452         struct btree *b;
1453         int ret;
1454
1455         closure_init_stack(&cl);
1456
1457         do {
1458                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1459                 closure_sync(&cl);
1460         } while (ret);
1461
1462         b = bch2_btree_node_mem_alloc(c);
1463         bch2_btree_cache_cannibalize_unlock(c);
1464
1465         BUG_ON(IS_ERR(b));
1466
1467         bkey_copy(&b->key, k);
1468         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1469
1470         bch2_btree_node_read(c, b, true);
1471
1472         if (btree_node_read_error(b)) {
1473                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1474
1475                 mutex_lock(&c->btree_cache.lock);
1476                 list_move(&b->list, &c->btree_cache.freeable);
1477                 mutex_unlock(&c->btree_cache.lock);
1478
1479                 ret = -EIO;
1480                 goto err;
1481         }
1482
1483         bch2_btree_set_root_for_read(c, b);
1484 err:
1485         six_unlock_write(&b->c.lock);
1486         six_unlock_intent(&b->c.lock);
1487
1488         return ret;
1489 }
1490
1491 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1492                               struct btree_write *w)
1493 {
1494         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1495
1496         do {
1497                 old = new = v;
1498                 if (!(old & 1))
1499                         break;
1500
1501                 new &= ~1UL;
1502         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1503
1504         if (old & 1)
1505                 closure_put(&((struct btree_update *) new)->cl);
1506
1507         bch2_journal_pin_drop(&c->journal, &w->journal);
1508 }
1509
1510 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1511 {
1512         struct btree_write *w = btree_prev_write(b);
1513
1514         bch2_btree_complete_write(c, b, w);
1515         btree_node_io_unlock(b);
1516 }
1517
1518 static void bch2_btree_node_write_error(struct bch_fs *c,
1519                                         struct btree_write_bio *wbio)
1520 {
1521         struct btree *b         = wbio->wbio.bio.bi_private;
1522         struct bkey_buf k;
1523         struct bch_extent_ptr *ptr;
1524         struct btree_trans trans;
1525         struct btree_iter *iter;
1526         int ret;
1527
1528         bch2_bkey_buf_init(&k);
1529         bch2_trans_init(&trans, c, 0, 0);
1530
1531         iter = bch2_trans_get_node_iter(&trans, b->c.btree_id, b->key.k.p,
1532                                         BTREE_MAX_DEPTH, b->c.level, 0);
1533 retry:
1534         ret = bch2_btree_iter_traverse(iter);
1535         if (ret)
1536                 goto err;
1537
1538         /* has node been freed? */
1539         if (iter->l[b->c.level].b != b) {
1540                 /* node has been freed: */
1541                 BUG_ON(!btree_node_dying(b));
1542                 goto out;
1543         }
1544
1545         BUG_ON(!btree_node_hashed(b));
1546
1547         bch2_bkey_buf_copy(&k, c, &b->key);
1548
1549         bch2_bkey_drop_ptrs(bkey_i_to_s(k.k), ptr,
1550                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1551
1552         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(k.k)))
1553                 goto err;
1554
1555         ret = bch2_btree_node_update_key(c, iter, b, k.k);
1556         if (ret == -EINTR)
1557                 goto retry;
1558         if (ret)
1559                 goto err;
1560 out:
1561         bch2_trans_iter_put(&trans, iter);
1562         bch2_trans_exit(&trans);
1563         bch2_bkey_buf_exit(&k, c);
1564         bio_put(&wbio->wbio.bio);
1565         btree_node_write_done(c, b);
1566         return;
1567 err:
1568         set_btree_node_noevict(b);
1569         bch2_fs_fatal_error(c, "fatal error writing btree node");
1570         goto out;
1571 }
1572
1573 void bch2_btree_write_error_work(struct work_struct *work)
1574 {
1575         struct bch_fs *c = container_of(work, struct bch_fs,
1576                                         btree_write_error_work);
1577         struct bio *bio;
1578
1579         while (1) {
1580                 spin_lock_irq(&c->btree_write_error_lock);
1581                 bio = bio_list_pop(&c->btree_write_error_list);
1582                 spin_unlock_irq(&c->btree_write_error_lock);
1583
1584                 if (!bio)
1585                         break;
1586
1587                 bch2_btree_node_write_error(c,
1588                         container_of(bio, struct btree_write_bio, wbio.bio));
1589         }
1590 }
1591
1592 static void btree_node_write_work(struct work_struct *work)
1593 {
1594         struct btree_write_bio *wbio =
1595                 container_of(work, struct btree_write_bio, work);
1596         struct bch_fs *c        = wbio->wbio.c;
1597         struct btree *b         = wbio->wbio.bio.bi_private;
1598
1599         btree_bounce_free(c,
1600                 wbio->bytes,
1601                 wbio->wbio.used_mempool,
1602                 wbio->data);
1603
1604         if (wbio->wbio.failed.nr) {
1605                 unsigned long flags;
1606
1607                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1608                 bio_list_add(&c->btree_write_error_list, &wbio->wbio.bio);
1609                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1610
1611                 queue_work(c->btree_error_wq, &c->btree_write_error_work);
1612                 return;
1613         }
1614
1615         bio_put(&wbio->wbio.bio);
1616         btree_node_write_done(c, b);
1617 }
1618
1619 static void btree_node_write_endio(struct bio *bio)
1620 {
1621         struct bch_write_bio *wbio      = to_wbio(bio);
1622         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1623         struct bch_write_bio *orig      = parent ?: wbio;
1624         struct bch_fs *c                = wbio->c;
1625         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1626         unsigned long flags;
1627
1628         if (wbio->have_ioref)
1629                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1630
1631         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1632                                bch2_blk_status_to_str(bio->bi_status)) ||
1633             bch2_meta_write_fault("btree")) {
1634                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1635                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1636                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1637         }
1638
1639         if (wbio->have_ioref)
1640                 percpu_ref_put(&ca->io_ref);
1641
1642         if (parent) {
1643                 bio_put(bio);
1644                 bio_endio(&parent->bio);
1645         } else {
1646                 struct btree_write_bio *wb =
1647                         container_of(orig, struct btree_write_bio, wbio);
1648
1649                 INIT_WORK(&wb->work, btree_node_write_work);
1650                 queue_work(c->io_complete_wq, &wb->work);
1651         }
1652 }
1653
1654 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1655                                    struct bset *i, unsigned sectors)
1656 {
1657         unsigned whiteout_u64s = 0;
1658         int ret;
1659
1660         if (bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), BKEY_TYPE_btree))
1661                 return -1;
1662
1663         ret = validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false) ?:
1664                 validate_bset(c, NULL, b, i, sectors, WRITE, false);
1665         if (ret) {
1666                 bch2_inconsistent_error(c);
1667                 dump_stack();
1668         }
1669
1670         return ret;
1671 }
1672
1673 static void btree_write_submit(struct work_struct *work)
1674 {
1675         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1676
1677         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree, &wbio->key);
1678 }
1679
1680 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b)
1681 {
1682         struct btree_write_bio *wbio;
1683         struct bset_tree *t;
1684         struct bset *i;
1685         struct btree_node *bn = NULL;
1686         struct btree_node_entry *bne = NULL;
1687         struct bch_extent_ptr *ptr;
1688         struct sort_iter sort_iter;
1689         struct nonce nonce;
1690         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1691         u64 seq = 0;
1692         bool used_mempool;
1693         unsigned long old, new;
1694         bool validate_before_checksum = false;
1695         void *data;
1696
1697         if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
1698                 return;
1699
1700         /*
1701          * We may only have a read lock on the btree node - the dirty bit is our
1702          * "lock" against racing with other threads that may be trying to start
1703          * a write, we do a write iff we clear the dirty bit. Since setting the
1704          * dirty bit requires a write lock, we can't race with other threads
1705          * redirtying it:
1706          */
1707         do {
1708                 old = new = READ_ONCE(b->flags);
1709
1710                 if (!(old & (1 << BTREE_NODE_dirty)))
1711                         return;
1712
1713                 if (!btree_node_may_write(b))
1714                         return;
1715
1716                 if (old & (1 << BTREE_NODE_never_write))
1717                         return;
1718
1719                 if (old & (1 << BTREE_NODE_write_in_flight)) {
1720                         /*
1721                          * XXX waiting on btree writes with btree locks held -
1722                          * this can deadlock, and we hit the write error path
1723                          */
1724                         btree_node_wait_on_io(b);
1725                         continue;
1726                 }
1727
1728                 new &= ~(1 << BTREE_NODE_dirty);
1729                 new &= ~(1 << BTREE_NODE_need_write);
1730                 new |=  (1 << BTREE_NODE_write_in_flight);
1731                 new |=  (1 << BTREE_NODE_just_written);
1732                 new ^=  (1 << BTREE_NODE_write_idx);
1733         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1734
1735         atomic_dec(&c->btree_cache.dirty);
1736
1737         BUG_ON(btree_node_fake(b));
1738         BUG_ON((b->will_make_reachable != 0) != !b->written);
1739
1740         BUG_ON(b->written >= c->opts.btree_node_size);
1741         BUG_ON(b->written & (c->opts.block_size - 1));
1742         BUG_ON(bset_written(b, btree_bset_last(b)));
1743         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1744         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1745
1746         bch2_sort_whiteouts(c, b);
1747
1748         sort_iter_init(&sort_iter, b);
1749
1750         bytes = !b->written
1751                 ? sizeof(struct btree_node)
1752                 : sizeof(struct btree_node_entry);
1753
1754         bytes += b->whiteout_u64s * sizeof(u64);
1755
1756         for_each_bset(b, t) {
1757                 i = bset(b, t);
1758
1759                 if (bset_written(b, i))
1760                         continue;
1761
1762                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1763                 sort_iter_add(&sort_iter,
1764                               btree_bkey_first(b, t),
1765                               btree_bkey_last(b, t));
1766                 seq = max(seq, le64_to_cpu(i->journal_seq));
1767         }
1768
1769         BUG_ON(b->written && !seq);
1770
1771         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1772         bytes += 8;
1773
1774         /* buffer must be a multiple of the block size */
1775         bytes = round_up(bytes, block_bytes(c));
1776
1777         data = btree_bounce_alloc(c, bytes, &used_mempool);
1778
1779         if (!b->written) {
1780                 bn = data;
1781                 *bn = *b->data;
1782                 i = &bn->keys;
1783         } else {
1784                 bne = data;
1785                 bne->keys = b->data->keys;
1786                 i = &bne->keys;
1787         }
1788
1789         i->journal_seq  = cpu_to_le64(seq);
1790         i->u64s         = 0;
1791
1792         sort_iter_add(&sort_iter,
1793                       unwritten_whiteouts_start(c, b),
1794                       unwritten_whiteouts_end(c, b));
1795         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1796
1797         b->whiteout_u64s = 0;
1798
1799         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1800         le16_add_cpu(&i->u64s, u64s);
1801
1802         set_needs_whiteout(i, false);
1803
1804         /* do we have data to write? */
1805         if (b->written && !i->u64s)
1806                 goto nowrite;
1807
1808         bytes_to_write = vstruct_end(i) - data;
1809         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1810
1811         memset(data + bytes_to_write, 0,
1812                (sectors_to_write << 9) - bytes_to_write);
1813
1814         BUG_ON(b->written + sectors_to_write > c->opts.btree_node_size);
1815         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1816         BUG_ON(i->seq != b->data->keys.seq);
1817
1818         i->version = c->sb.version < bcachefs_metadata_version_new_versioning
1819                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1820                 : cpu_to_le16(c->sb.version);
1821         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1822
1823         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1824                 validate_before_checksum = true;
1825
1826         /* validate_bset will be modifying: */
1827         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
1828                 validate_before_checksum = true;
1829
1830         /* if we're going to be encrypting, check metadata validity first: */
1831         if (validate_before_checksum &&
1832             validate_bset_for_write(c, b, i, sectors_to_write))
1833                 goto err;
1834
1835         bset_encrypt(c, i, b->written << 9);
1836
1837         nonce = btree_nonce(i, b->written << 9);
1838
1839         if (bn)
1840                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1841         else
1842                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1843
1844         /* if we're not encrypting, check metadata after checksumming: */
1845         if (!validate_before_checksum &&
1846             validate_bset_for_write(c, b, i, sectors_to_write))
1847                 goto err;
1848
1849         /*
1850          * We handle btree write errors by immediately halting the journal -
1851          * after we've done that, we can't issue any subsequent btree writes
1852          * because they might have pointers to new nodes that failed to write.
1853          *
1854          * Furthermore, there's no point in doing any more btree writes because
1855          * with the journal stopped, we're never going to update the journal to
1856          * reflect that those writes were done and the data flushed from the
1857          * journal:
1858          *
1859          * Also on journal error, the pending write may have updates that were
1860          * never journalled (interior nodes, see btree_update_nodes_written()) -
1861          * it's critical that we don't do the write in that case otherwise we
1862          * will have updates visible that weren't in the journal:
1863          *
1864          * Make sure to update b->written so bch2_btree_init_next() doesn't
1865          * break:
1866          */
1867         if (bch2_journal_error(&c->journal) ||
1868             c->opts.nochanges)
1869                 goto err;
1870
1871         trace_btree_write(b, bytes_to_write, sectors_to_write);
1872
1873         wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1874                                 buf_pages(data, sectors_to_write << 9),
1875                                 &c->btree_bio),
1876                             struct btree_write_bio, wbio.bio);
1877         wbio_init(&wbio->wbio.bio);
1878         wbio->data                      = data;
1879         wbio->bytes                     = bytes;
1880         wbio->wbio.c                    = c;
1881         wbio->wbio.used_mempool         = used_mempool;
1882         wbio->wbio.bio.bi_opf           = REQ_OP_WRITE|REQ_META;
1883         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
1884         wbio->wbio.bio.bi_private       = b;
1885
1886         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
1887
1888         /*
1889          * If we're appending to a leaf node, we don't technically need FUA -
1890          * this write just needs to be persisted before the next journal write,
1891          * which will be marked FLUSH|FUA.
1892          *
1893          * Similarly if we're writing a new btree root - the pointer is going to
1894          * be in the next journal entry.
1895          *
1896          * But if we're writing a new btree node (that isn't a root) or
1897          * appending to a non leaf btree node, we need either FUA or a flush
1898          * when we write the parent with the new pointer. FUA is cheaper than a
1899          * flush, and writes appending to leaf nodes aren't blocking anything so
1900          * just make all btree node writes FUA to keep things sane.
1901          */
1902
1903         bkey_copy(&wbio->key, &b->key);
1904
1905         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&wbio->key)), ptr)
1906                 ptr->offset += b->written;
1907
1908         b->written += sectors_to_write;
1909
1910         atomic64_inc(&c->btree_writes_nr);
1911         atomic64_add(sectors_to_write, &c->btree_writes_sectors);
1912
1913         INIT_WORK(&wbio->work, btree_write_submit);
1914         queue_work(c->io_complete_wq, &wbio->work);
1915         return;
1916 err:
1917         set_btree_node_noevict(b);
1918         b->written += sectors_to_write;
1919 nowrite:
1920         btree_bounce_free(c, bytes, used_mempool, data);
1921         btree_node_write_done(c, b);
1922 }
1923
1924 /*
1925  * Work that must be done with write lock held:
1926  */
1927 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
1928 {
1929         bool invalidated_iter = false;
1930         struct btree_node_entry *bne;
1931         struct bset_tree *t;
1932
1933         if (!btree_node_just_written(b))
1934                 return false;
1935
1936         BUG_ON(b->whiteout_u64s);
1937
1938         clear_btree_node_just_written(b);
1939
1940         /*
1941          * Note: immediately after write, bset_written() doesn't work - the
1942          * amount of data we had to write after compaction might have been
1943          * smaller than the offset of the last bset.
1944          *
1945          * However, we know that all bsets have been written here, as long as
1946          * we're still holding the write lock:
1947          */
1948
1949         /*
1950          * XXX: decide if we really want to unconditionally sort down to a
1951          * single bset:
1952          */
1953         if (b->nsets > 1) {
1954                 btree_node_sort(c, b, 0, b->nsets, true);
1955                 invalidated_iter = true;
1956         } else {
1957                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
1958         }
1959
1960         for_each_bset(b, t)
1961                 set_needs_whiteout(bset(b, t), true);
1962
1963         bch2_btree_verify(c, b);
1964
1965         /*
1966          * If later we don't unconditionally sort down to a single bset, we have
1967          * to ensure this is still true:
1968          */
1969         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
1970
1971         bne = want_new_bset(c, b);
1972         if (bne)
1973                 bch2_bset_init_next(c, b, bne);
1974
1975         bch2_btree_build_aux_trees(b);
1976
1977         return invalidated_iter;
1978 }
1979
1980 /*
1981  * Use this one if the node is intent locked:
1982  */
1983 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1984                            enum six_lock_type lock_type_held)
1985 {
1986         if (lock_type_held == SIX_LOCK_intent ||
1987             (lock_type_held == SIX_LOCK_read &&
1988              six_lock_tryupgrade(&b->c.lock))) {
1989                 __bch2_btree_node_write(c, b);
1990
1991                 /* don't cycle lock unnecessarily: */
1992                 if (btree_node_just_written(b) &&
1993                     six_trylock_write(&b->c.lock)) {
1994                         bch2_btree_post_write_cleanup(c, b);
1995                         six_unlock_write(&b->c.lock);
1996                 }
1997
1998                 if (lock_type_held == SIX_LOCK_read)
1999                         six_lock_downgrade(&b->c.lock);
2000         } else {
2001                 __bch2_btree_node_write(c, b);
2002                 if (lock_type_held == SIX_LOCK_write &&
2003                     btree_node_just_written(b))
2004                         bch2_btree_post_write_cleanup(c, b);
2005         }
2006 }
2007
2008 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2009 {
2010         struct bucket_table *tbl;
2011         struct rhash_head *pos;
2012         struct btree *b;
2013         unsigned i;
2014 restart:
2015         rcu_read_lock();
2016         for_each_cached_btree(b, c, tbl, i, pos)
2017                 if (test_bit(flag, &b->flags)) {
2018                         rcu_read_unlock();
2019                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2020                         goto restart;
2021
2022                 }
2023         rcu_read_unlock();
2024 }
2025
2026 void bch2_btree_flush_all_reads(struct bch_fs *c)
2027 {
2028         __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2029 }
2030
2031 void bch2_btree_flush_all_writes(struct bch_fs *c)
2032 {
2033         __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2034 }
2035
2036 void bch2_dirty_btree_nodes_to_text(struct printbuf *out, struct bch_fs *c)
2037 {
2038         struct bucket_table *tbl;
2039         struct rhash_head *pos;
2040         struct btree *b;
2041         unsigned i;
2042
2043         rcu_read_lock();
2044         for_each_cached_btree(b, c, tbl, i, pos) {
2045                 unsigned long flags = READ_ONCE(b->flags);
2046
2047                 if (!(flags & (1 << BTREE_NODE_dirty)))
2048                         continue;
2049
2050                 pr_buf(out, "%p d %u n %u l %u w %u b %u r %u:%lu\n",
2051                        b,
2052                        (flags & (1 << BTREE_NODE_dirty)) != 0,
2053                        (flags & (1 << BTREE_NODE_need_write)) != 0,
2054                        b->c.level,
2055                        b->written,
2056                        !list_empty_careful(&b->write_blocked),
2057                        b->will_make_reachable != 0,
2058                        b->will_make_reachable & 1);
2059         }
2060         rcu_read_unlock();
2061 }