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         struct bkey_s_c k;
 569         struct bkey unpacked;
 570         struct btree_node_iter iter;
 571
 572         for_each_bset(b, t) {
 573                 struct bset *i = bset(b, t);
 574                 struct bkey_packed *k;
 575
 576                 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
 577                         if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
 578                                 break;
 579
 580                 if (k != i->start) {
 581                         unsigned shift = (u64 *) k - (u64 *) i->start;
 582
 583                         memmove_u64s_down(i->start, k,
 584                                           (u64 *) vstruct_end(i) - (u64 *) k);
 585                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
 586                         set_btree_bset_end(b, t);
 587                         bch2_bset_set_no_aux_tree(b, t);
 588                 }
 589
 590                 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
 591                         if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
 592                                 break;
 593
 594                 if (k != vstruct_last(i)) {
 595                         i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
 596                         set_btree_bset_end(b, t);
 597                         bch2_bset_set_no_aux_tree(b, t);
 598                 }
 599         }
 600
 601         bch2_btree_build_aux_trees(b);
 602
 603         for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
 604                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 605                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 606         }
 607 }
 608
 609 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 610                          struct btree *b, struct bset *i,
 611                          unsigned sectors, int write, bool have_retry)
 612 {
 613         unsigned version = le16_to_cpu(i->version);
 614         const char *err;
 615         char buf1[100];
 616         char buf2[100];
 617         int ret = 0;
 618
 619         btree_err_on((version != BCH_BSET_VERSION_OLD &&
 620                       version < bcachefs_metadata_version_min) ||
 621                      version >= bcachefs_metadata_version_max,
 622                      BTREE_ERR_FATAL, c, ca, b, i,
 623                      "unsupported bset version");
 624
 625         if (btree_err_on(version < c->sb.version_min,
 626                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 627                          "bset version %u older than superblock version_min %u",
 628                          version, c->sb.version_min)) {
 629                 mutex_lock(&c->sb_lock);
 630                 c->disk_sb.sb->version_min = cpu_to_le16(version);
 631                 bch2_write_super(c);
 632                 mutex_unlock(&c->sb_lock);
 633         }
 634
 635         if (btree_err_on(version > c->sb.version,
 636                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 637                          "bset version %u newer than superblock version %u",
 638                          version, c->sb.version)) {
 639                 mutex_lock(&c->sb_lock);
 640                 c->disk_sb.sb->version = cpu_to_le16(version);
 641                 bch2_write_super(c);
 642                 mutex_unlock(&c->sb_lock);
 643         }
 644
 645         btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
 646                      BTREE_ERR_FATAL, c, ca, b, i,
 647                      "BSET_SEPARATE_WHITEOUTS no longer supported");
 648
 649         if (btree_err_on(b->written + sectors > c->opts.btree_node_size,
 650                          BTREE_ERR_FIXABLE, c, ca, b, i,
 651                          "bset past end of btree node")) {
 652                 i->u64s = 0;
 653                 return 0;
 654         }
 655
 656         btree_err_on(b->written && !i->u64s,
 657                      BTREE_ERR_FIXABLE, c, ca, b, i,
 658                      "empty bset");
 659
 660         if (!b->written) {
 661                 struct btree_node *bn =
 662                         container_of(i, struct btree_node, keys);
 663                 /* These indicate that we read the wrong btree node: */
 664
 665                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 666                         struct bch_btree_ptr_v2 *bp =
 667                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 668
 669                         /* XXX endianness */
 670                         btree_err_on(bp->seq != bn->keys.seq,
 671                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 672                                      "incorrect sequence number (wrong btree node)");
 673                 }
 674
 675                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 676                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 677                              "incorrect btree id");
 678
 679                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 680                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 681                              "incorrect level");
 682
 683                 if (!write)
 684                         compat_btree_node(b->c.level, b->c.btree_id, version,
 685                                           BSET_BIG_ENDIAN(i), write, bn);
 686
 687                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 688                         struct bch_btree_ptr_v2 *bp =
 689                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 690
 691                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 692                                 b->data->min_key = bp->min_key;
 693                                 b->data->max_key = b->key.k.p;
 694                         }
 695
 696                         btree_err_on(bpos_cmp(b->data->min_key, bp->min_key),
 697                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 698                                      "incorrect min_key: got %s should be %s",
 699                                      (bch2_bpos_to_text(&PBUF(buf1), bn->min_key), buf1),
 700                                      (bch2_bpos_to_text(&PBUF(buf2), bp->min_key), buf2));
 701                 }
 702
 703                 btree_err_on(bpos_cmp(bn->max_key, b->key.k.p),
 704                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 705                              "incorrect max key %s",
 706                              (bch2_bpos_to_text(&PBUF(buf1), bn->max_key), buf1));
 707
 708                 if (write)
 709                         compat_btree_node(b->c.level, b->c.btree_id, version,
 710                                           BSET_BIG_ENDIAN(i), write, bn);
 711
 712                 err = bch2_bkey_format_validate(&bn->format);
 713                 btree_err_on(err,
 714                              BTREE_ERR_FATAL, c, ca, b, i,
 715                              "invalid bkey format: %s", err);
 716
 717                 compat_bformat(b->c.level, b->c.btree_id, version,
 718                                BSET_BIG_ENDIAN(i), write,
 719                                &bn->format);
 720         }
 721 fsck_err:
 722         return ret;
 723 }
 724
 725 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 726                          struct bset *i, unsigned *whiteout_u64s,
 727                          int write, bool have_retry)
 728 {
 729         unsigned version = le16_to_cpu(i->version);
 730         struct bkey_packed *k, *prev = NULL;
 731         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 732                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 733         int ret = 0;
 734
 735         for (k = i->start;
 736              k != vstruct_last(i);) {
 737                 struct bkey_s u;
 738                 struct bkey tmp;
 739                 const char *invalid;
 740
 741                 if (btree_err_on(bkey_next(k) > vstruct_last(i),
 742                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 743                                  "key extends past end of bset")) {
 744                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 745                         break;
 746                 }
 747
 748                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 749                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 750                                  "invalid bkey format %u", k->format)) {
 751                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 752                         memmove_u64s_down(k, bkey_next(k),
 753                                           (u64 *) vstruct_end(i) - (u64 *) k);
 754                         continue;
 755                 }
 756
 757                 /* XXX: validate k->u64s */
 758                 if (!write)
 759                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 760                                     BSET_BIG_ENDIAN(i), write,
 761                                     &b->format, k);
 762
 763                 u = __bkey_disassemble(b, k, &tmp);
 764
 765                 invalid = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b)) ?:
 766                         (!updated_range ?  bch2_bkey_in_btree_node(b, u.s_c) : NULL) ?:
 767                         (write ? bch2_bkey_val_invalid(c, u.s_c) : NULL);
 768                 if (invalid) {
 769                         char buf[160];
 770
 771                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 772                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 773                                   "invalid bkey: %s\n%s", invalid, buf);
 774
 775                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 776                         memmove_u64s_down(k, bkey_next(k),
 777                                           (u64 *) vstruct_end(i) - (u64 *) k);
 778                         continue;
 779                 }
 780
 781                 if (write)
 782                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 783                                     BSET_BIG_ENDIAN(i), write,
 784                                     &b->format, k);
 785
 786                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 787                         char buf1[80];
 788                         char buf2[80];
 789                         struct bkey up = bkey_unpack_key(b, prev);
 790
 791                         bch2_bkey_to_text(&PBUF(buf1), &up);
 792                         bch2_bkey_to_text(&PBUF(buf2), u.k);
 793
 794                         bch2_dump_bset(c, b, i, 0);
 795
 796                         if (btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 797                                       "keys out of order: %s > %s",
 798                                       buf1, buf2)) {
 799                                 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 800                                 memmove_u64s_down(k, bkey_next(k),
 801                                                   (u64 *) vstruct_end(i) - (u64 *) k);
 802                                 continue;
 803                         }
 804                 }
 805
 806                 prev = k;
 807                 k = bkey_next(k);
 808         }
 809 fsck_err:
 810         return ret;
 811 }
 812
 813 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 814                               struct btree *b, bool have_retry)
 815 {
 816         struct btree_node_entry *bne;
 817         struct sort_iter *iter;
 818         struct btree_node *sorted;
 819         struct bkey_packed *k;
 820         struct bch_extent_ptr *ptr;
 821         struct bset *i;
 822         bool used_mempool, blacklisted;
 823         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 824                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 825         unsigned u64s;
 826         unsigned nonblacklisted_written = 0;
 827         int ret, retry_read = 0, write = READ;
 828
 829         b->version_ondisk = U16_MAX;
 830
 831         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
 832         sort_iter_init(iter, b);
 833         iter->size = (btree_blocks(c) + 1) * 2;
 834
 835         if (bch2_meta_read_fault("btree"))
 836                 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 837                           "dynamic fault");
 838
 839         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 840                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 841                      "bad magic");
 842
 843         btree_err_on(!b->data->keys.seq,
 844                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 845                      "bad btree header");
 846
 847         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 848                 struct bch_btree_ptr_v2 *bp =
 849                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 850
 851                 btree_err_on(b->data->keys.seq != bp->seq,
 852                              BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 853                              "got wrong btree node (seq %llx want %llx)",
 854                              b->data->keys.seq, bp->seq);
 855         }
 856
 857         while (b->written < c->opts.btree_node_size) {
 858                 unsigned sectors, whiteout_u64s = 0;
 859                 struct nonce nonce;
 860                 struct bch_csum csum;
 861                 bool first = !b->written;
 862
 863                 if (!b->written) {
 864                         i = &b->data->keys;
 865
 866                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 867                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 868                                      "unknown checksum type %llu",
 869                                      BSET_CSUM_TYPE(i));
 870
 871                         nonce = btree_nonce(i, b->written << 9);
 872                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 873
 874                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 875                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 876                                      "invalid checksum");
 877
 878                         bset_encrypt(c, i, b->written << 9);
 879
 880                         btree_err_on(btree_node_is_extents(b) &&
 881                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
 882                                      BTREE_ERR_FATAL, c, NULL, b, NULL,
 883                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
 884
 885                         sectors = vstruct_sectors(b->data, c->block_bits);
 886                 } else {
 887                         bne = write_block(b);
 888                         i = &bne->keys;
 889
 890                         if (i->seq != b->data->keys.seq)
 891                                 break;
 892
 893                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 894                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 895                                      "unknown checksum type %llu",
 896                                      BSET_CSUM_TYPE(i));
 897
 898                         nonce = btree_nonce(i, b->written << 9);
 899                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
 900
 901                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
 902                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 903                                      "invalid checksum");
 904
 905                         bset_encrypt(c, i, b->written << 9);
 906
 907                         sectors = vstruct_sectors(bne, c->block_bits);
 908                 }
 909
 910                 b->version_ondisk = min(b->version_ondisk,
 911                                         le16_to_cpu(i->version));
 912
 913                 ret = validate_bset(c, ca, b, i, sectors,
 914                                     READ, have_retry);
 915                 if (ret)
 916                         goto fsck_err;
 917
 918                 if (!b->written)
 919                         btree_node_set_format(b, b->data->format);
 920
 921                 ret = validate_bset_keys(c, b, i, &whiteout_u64s,
 922                                     READ, have_retry);
 923                 if (ret)
 924                         goto fsck_err;
 925
 926                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 927
 928                 b->written += sectors;
 929
 930                 blacklisted = bch2_journal_seq_is_blacklisted(c,
 931                                         le64_to_cpu(i->journal_seq),
 932                                         true);
 933
 934                 btree_err_on(blacklisted && first,
 935                              BTREE_ERR_FIXABLE, c, ca, b, i,
 936                              "first btree node bset has blacklisted journal seq");
 937                 if (blacklisted && !first)
 938                         continue;
 939
 940                 sort_iter_add(iter, i->start,
 941                               vstruct_idx(i, whiteout_u64s));
 942
 943                 sort_iter_add(iter,
 944                               vstruct_idx(i, whiteout_u64s),
 945                               vstruct_last(i));
 946
 947                 nonblacklisted_written = b->written;
 948         }
 949
 950         for (bne = write_block(b);
 951              bset_byte_offset(b, bne) < btree_bytes(c);
 952              bne = (void *) bne + block_bytes(c))
 953                 btree_err_on(bne->keys.seq == b->data->keys.seq &&
 954                              !bch2_journal_seq_is_blacklisted(c,
 955                                         le64_to_cpu(bne->keys.journal_seq),
 956                                         true),
 957                              BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
 958                              "found bset signature after last bset");
 959
 960         /*
 961          * Blacklisted bsets are those that were written after the most recent
 962          * (flush) journal write. Since there wasn't a flush, they may not have
 963          * made it to all devices - which means we shouldn't write new bsets
 964          * after them, as that could leave a gap and then reads from that device
 965          * wouldn't find all the bsets in that btree node - which means it's
 966          * important that we start writing new bsets after the most recent _non_
 967          * blacklisted bset:
 968          */
 969         b->written = nonblacklisted_written;
 970
 971         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
 972         sorted->keys.u64s = 0;
 973
 974         set_btree_bset(b, b->set, &b->data->keys);
 975
 976         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
 977
 978         u64s = le16_to_cpu(sorted->keys.u64s);
 979         *sorted = *b->data;
 980         sorted->keys.u64s = cpu_to_le16(u64s);
 981         swap(sorted, b->data);
 982         set_btree_bset(b, b->set, &b->data->keys);
 983         b->nsets = 1;
 984
 985         BUG_ON(b->nr.live_u64s != u64s);
 986
 987         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
 988
 989         if (updated_range)
 990                 bch2_btree_node_drop_keys_outside_node(b);
 991
 992         i = &b->data->keys;
 993         for (k = i->start; k != vstruct_last(i);) {
 994                 struct bkey tmp;
 995                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
 996                 const char *invalid = bch2_bkey_val_invalid(c, u.s_c);
 997
 998                 if (invalid ||
 999                     (bch2_inject_invalid_keys &&
1000                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1001                         char buf[160];
1002
1003                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
1004                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
1005                                   "invalid bkey %s: %s", buf, invalid);
1006
1007                         btree_keys_account_key_drop(&b->nr, 0, k);
1008
1009                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1010                         memmove_u64s_down(k, bkey_next(k),
1011                                           (u64 *) vstruct_end(i) - (u64 *) k);
1012                         set_btree_bset_end(b, b->set);
1013                         continue;
1014                 }
1015
1016                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1017                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1018
1019                         bp.v->mem_ptr = 0;
1020                 }
1021
1022                 k = bkey_next(k);
1023         }
1024
1025         bch2_bset_build_aux_tree(b, b->set, false);
1026
1027         set_needs_whiteout(btree_bset_first(b), true);
1028
1029         btree_node_reset_sib_u64s(b);
1030
1031         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1032                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1033
1034                 if (ca->mi.state != BCH_MEMBER_STATE_rw)
1035                         set_btree_node_need_rewrite(b);
1036         }
1037 out:
1038         mempool_free(iter, &c->fill_iter);
1039         return retry_read;
1040 fsck_err:
1041         if (ret == BTREE_RETRY_READ) {
1042                 retry_read = 1;
1043         } else {
1044                 bch2_inconsistent_error(c);
1045                 set_btree_node_read_error(b);
1046         }
1047         goto out;
1048 }
1049
1050 static void btree_node_read_work(struct work_struct *work)
1051 {
1052         struct btree_read_bio *rb =
1053                 container_of(work, struct btree_read_bio, work);
1054         struct bch_fs *c        = rb->c;
1055         struct btree *b         = rb->b;
1056         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1057         struct bio *bio         = &rb->bio;
1058         struct bch_io_failures failed = { .nr = 0 };
1059         char buf[200];
1060         struct printbuf out;
1061         bool saw_error = false;
1062         bool can_retry;
1063
1064         goto start;
1065         while (1) {
1066                 bch_info(c, "retrying read");
1067                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1068                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1069                 bio_reset(bio);
1070                 bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1071                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1072                 bio->bi_iter.bi_size    = btree_bytes(c);
1073
1074                 if (rb->have_ioref) {
1075                         bio_set_dev(bio, ca->disk_sb.bdev);
1076                         submit_bio_wait(bio);
1077                 } else {
1078                         bio->bi_status = BLK_STS_REMOVED;
1079                 }
1080 start:
1081                 out = PBUF(buf);
1082                 btree_pos_to_text(&out, c, b);
1083                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1084                                    bch2_blk_status_to_str(bio->bi_status), buf);
1085                 if (rb->have_ioref)
1086                         percpu_ref_put(&ca->io_ref);
1087                 rb->have_ioref = false;
1088
1089                 bch2_mark_io_failure(&failed, &rb->pick);
1090
1091                 can_retry = bch2_bkey_pick_read_device(c,
1092                                 bkey_i_to_s_c(&b->key),
1093                                 &failed, &rb->pick) > 0;
1094
1095                 if (!bio->bi_status &&
1096                     !bch2_btree_node_read_done(c, ca, b, can_retry))
1097                         break;
1098
1099                 saw_error = true;
1100
1101                 if (!can_retry) {
1102                         set_btree_node_read_error(b);
1103                         break;
1104                 }
1105         }
1106
1107         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1108                                rb->start_time);
1109         bio_put(&rb->bio);
1110
1111         if (saw_error && !btree_node_read_error(b))
1112                 bch2_btree_node_rewrite_async(c, b);
1113
1114         clear_btree_node_read_in_flight(b);
1115         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1116 }
1117
1118 static void btree_node_read_endio(struct bio *bio)
1119 {
1120         struct btree_read_bio *rb =
1121                 container_of(bio, struct btree_read_bio, bio);
1122         struct bch_fs *c        = rb->c;
1123
1124         if (rb->have_ioref) {
1125                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1126                 bch2_latency_acct(ca, rb->start_time, READ);
1127         }
1128
1129         queue_work(c->io_complete_wq, &rb->work);
1130 }
1131
1132 struct btree_node_read_all {
1133         struct closure          cl;
1134         struct bch_fs           *c;
1135         struct btree            *b;
1136         unsigned                nr;
1137         void                    *buf[BCH_REPLICAS_MAX];
1138         struct bio              *bio[BCH_REPLICAS_MAX];
1139         int                     err[BCH_REPLICAS_MAX];
1140 };
1141
1142 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1143 {
1144         struct btree_node *bn = data;
1145         struct btree_node_entry *bne;
1146         unsigned offset = 0;
1147
1148         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1149                 return 0;
1150
1151         while (offset < c->opts.btree_node_size) {
1152                 if (!offset) {
1153                         offset += vstruct_sectors(bn, c->block_bits);
1154                 } else {
1155                         bne = data + (offset << 9);
1156                         if (bne->keys.seq != bn->keys.seq)
1157                                 break;
1158                         offset += vstruct_sectors(bne, c->block_bits);
1159                 }
1160         }
1161
1162         return offset;
1163 }
1164
1165 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1166 {
1167         struct btree_node *bn = data;
1168         struct btree_node_entry *bne;
1169
1170         if (!offset)
1171                 return false;
1172
1173         while (offset < c->opts.btree_node_size) {
1174                 bne = data + (offset << 9);
1175                 if (bne->keys.seq == bn->keys.seq)
1176                         return true;
1177                 offset++;
1178         }
1179
1180         return false;
1181         return offset;
1182 }
1183
1184 static void btree_node_read_all_replicas_done(struct closure *cl)
1185 {
1186         struct btree_node_read_all *ra =
1187                 container_of(cl, struct btree_node_read_all, cl);
1188         struct bch_fs *c = ra->c;
1189         struct btree *b = ra->b;
1190         bool dump_bset_maps = false;
1191         bool have_retry = false;
1192         int ret = 0, best = -1, write = READ;
1193         unsigned i, written, written2;
1194         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1195                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1196
1197         for (i = 0; i < ra->nr; i++) {
1198                 struct btree_node *bn = ra->buf[i];
1199
1200                 if (ra->err[i])
1201                         continue;
1202
1203                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1204                     (seq && seq != bn->keys.seq))
1205                         continue;
1206
1207                 if (best < 0) {
1208                         best = i;
1209                         written = btree_node_sectors_written(c, bn);
1210                         continue;
1211                 }
1212
1213                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1214                 if (btree_err_on(written2 != written, BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1215                                  "btree node sectors written mismatch: %u != %u",
1216                                  written, written2) ||
1217                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1218                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1219                                  "found bset signature after last bset") ||
1220                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1221                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1222                                  "btree node replicas content mismatch"))
1223                         dump_bset_maps = true;
1224
1225                 if (written2 > written) {
1226                         written = written2;
1227                         best = i;
1228                 }
1229         }
1230 fsck_err:
1231         if (dump_bset_maps) {
1232                 for (i = 0; i < ra->nr; i++) {
1233                         char buf[200];
1234                         struct printbuf out = PBUF(buf);
1235                         struct btree_node *bn = ra->buf[i];
1236                         struct btree_node_entry *bne = NULL;
1237                         unsigned offset = 0, sectors;
1238                         bool gap = false;
1239
1240                         if (ra->err[i])
1241                                 continue;
1242
1243                         while (offset < c->opts.btree_node_size) {
1244                                 if (!offset) {
1245                                         sectors = vstruct_sectors(bn, c->block_bits);
1246                                 } else {
1247                                         bne = ra->buf[i] + (offset << 9);
1248                                         if (bne->keys.seq != bn->keys.seq)
1249                                                 break;
1250                                         sectors = vstruct_sectors(bne, c->block_bits);
1251                                 }
1252
1253                                 pr_buf(&out, " %u-%u", offset, offset + sectors);
1254                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1255                                                         le64_to_cpu(bne->keys.journal_seq), false))
1256                                         pr_buf(&out, "*");
1257                                 offset += sectors;
1258                         }
1259
1260                         while (offset < c->opts.btree_node_size) {
1261                                 bne = ra->buf[i] + (offset << 9);
1262                                 if (bne->keys.seq == bn->keys.seq) {
1263                                         if (!gap)
1264                                                 pr_buf(&out, " GAP");
1265                                         gap = true;
1266
1267                                         sectors = vstruct_sectors(bne, c->block_bits);
1268                                         pr_buf(&out, " %u-%u", offset, offset + sectors);
1269                                         if (bch2_journal_seq_is_blacklisted(c,
1270                                                         le64_to_cpu(bne->keys.journal_seq), false))
1271                                                 pr_buf(&out, "*");
1272                                 }
1273                                 offset++;
1274                         }
1275
1276                         bch_err(c, "replica %u:%s", i, buf);
1277                 }
1278         }
1279
1280         if (best >= 0) {
1281                 memcpy(b->data, ra->buf[best], btree_bytes(c));
1282                 ret = bch2_btree_node_read_done(c, NULL, b, false);
1283         } else {
1284                 ret = -1;
1285         }
1286
1287         if (ret)
1288                 set_btree_node_read_error(b);
1289
1290         for (i = 0; i < ra->nr; i++) {
1291                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1292                 bio_put(ra->bio[i]);
1293         }
1294
1295         closure_debug_destroy(&ra->cl);
1296         kfree(ra);
1297
1298         clear_btree_node_read_in_flight(b);
1299         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1300 }
1301
1302 static void btree_node_read_all_replicas_endio(struct bio *bio)
1303 {
1304         struct btree_read_bio *rb =
1305                 container_of(bio, struct btree_read_bio, bio);
1306         struct bch_fs *c        = rb->c;
1307         struct btree_node_read_all *ra = rb->ra;
1308
1309         if (rb->have_ioref) {
1310                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1311                 bch2_latency_acct(ca, rb->start_time, READ);
1312         }
1313
1314         ra->err[rb->idx] = bio->bi_status;
1315         closure_put(&ra->cl);
1316 }
1317
1318 /*
1319  * XXX This allocates multiple times from the same mempools, and can deadlock
1320  * under sufficient memory pressure (but is only a debug path)
1321  */
1322 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1323 {
1324         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1325         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1326         const union bch_extent_entry *entry;
1327         struct extent_ptr_decoded pick;
1328         struct btree_node_read_all *ra;
1329         unsigned i;
1330
1331         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1332         if (!ra)
1333                 return -ENOMEM;
1334
1335         closure_init(&ra->cl, NULL);
1336         ra->c   = c;
1337         ra->b   = b;
1338         ra->nr  = bch2_bkey_nr_ptrs(k);
1339
1340         for (i = 0; i < ra->nr; i++) {
1341                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1342                 ra->bio[i] = bio_alloc_bioset(GFP_NOFS, buf_pages(ra->buf[i],
1343                                                                   btree_bytes(c)),
1344                                               &c->btree_bio);
1345         }
1346
1347         i = 0;
1348         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1349                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1350                 struct btree_read_bio *rb =
1351                         container_of(ra->bio[i], struct btree_read_bio, bio);
1352                 rb->c                   = c;
1353                 rb->b                   = b;
1354                 rb->ra                  = ra;
1355                 rb->start_time          = local_clock();
1356                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1357                 rb->idx                 = i;
1358                 rb->pick                = pick;
1359                 rb->bio.bi_opf          = REQ_OP_READ|REQ_SYNC|REQ_META;
1360                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1361                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1362                 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1363
1364                 if (rb->have_ioref) {
1365                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1366                                      bio_sectors(&rb->bio));
1367                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1368
1369                         closure_get(&ra->cl);
1370                         submit_bio(&rb->bio);
1371                 } else {
1372                         ra->err[i] = BLK_STS_REMOVED;
1373                 }
1374
1375                 i++;
1376         }
1377
1378         if (sync) {
1379                 closure_sync(&ra->cl);
1380                 btree_node_read_all_replicas_done(&ra->cl);
1381         } else {
1382                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1383                             c->io_complete_wq);
1384         }
1385
1386         return 0;
1387 }
1388
1389 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1390                           bool sync)
1391 {
1392         struct extent_ptr_decoded pick;
1393         struct btree_read_bio *rb;
1394         struct bch_dev *ca;
1395         struct bio *bio;
1396         char buf[200];
1397         int ret;
1398
1399         btree_pos_to_text(&PBUF(buf), c, b);
1400         trace_btree_read(c, b);
1401
1402         set_btree_node_read_in_flight(b);
1403
1404         if (bch2_verify_all_btree_replicas &&
1405             !btree_node_read_all_replicas(c, b, sync))
1406                 return;
1407
1408         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1409                                          NULL, &pick);
1410         if (bch2_fs_fatal_err_on(ret <= 0, c,
1411                         "btree node read error: no device to read from\n"
1412                         " at %s", buf)) {
1413                 set_btree_node_read_error(b);
1414                 return;
1415         }
1416
1417         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1418
1419         bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1420                                                    btree_bytes(c)),
1421                                &c->btree_bio);
1422         rb = container_of(bio, struct btree_read_bio, bio);
1423         rb->c                   = c;
1424         rb->b                   = b;
1425         rb->ra                  = NULL;
1426         rb->start_time          = local_clock();
1427         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1428         rb->pick                = pick;
1429         INIT_WORK(&rb->work, btree_node_read_work);
1430         bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1431         bio->bi_iter.bi_sector  = pick.ptr.offset;
1432         bio->bi_end_io          = btree_node_read_endio;
1433         bch2_bio_map(bio, b->data, btree_bytes(c));
1434
1435         if (rb->have_ioref) {
1436                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1437                              bio_sectors(bio));
1438                 bio_set_dev(bio, ca->disk_sb.bdev);
1439
1440                 if (sync) {
1441                         submit_bio_wait(bio);
1442
1443                         btree_node_read_work(&rb->work);
1444                 } else {
1445                         submit_bio(bio);
1446                 }
1447         } else {
1448                 bio->bi_status = BLK_STS_REMOVED;
1449
1450                 if (sync)
1451                         btree_node_read_work(&rb->work);
1452                 else
1453                         queue_work(c->io_complete_wq, &rb->work);
1454         }
1455 }
1456
1457 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1458                         const struct bkey_i *k, unsigned level)
1459 {
1460         struct closure cl;
1461         struct btree *b;
1462         int ret;
1463
1464         closure_init_stack(&cl);
1465
1466         do {
1467                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1468                 closure_sync(&cl);
1469         } while (ret);
1470
1471         b = bch2_btree_node_mem_alloc(c);
1472         bch2_btree_cache_cannibalize_unlock(c);
1473
1474         BUG_ON(IS_ERR(b));
1475
1476         bkey_copy(&b->key, k);
1477         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1478
1479         bch2_btree_node_read(c, b, true);
1480
1481         if (btree_node_read_error(b)) {
1482                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1483
1484                 mutex_lock(&c->btree_cache.lock);
1485                 list_move(&b->list, &c->btree_cache.freeable);
1486                 mutex_unlock(&c->btree_cache.lock);
1487
1488                 ret = -EIO;
1489                 goto err;
1490         }
1491
1492         bch2_btree_set_root_for_read(c, b);
1493 err:
1494         six_unlock_write(&b->c.lock);
1495         six_unlock_intent(&b->c.lock);
1496
1497         return ret;
1498 }
1499
1500 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1501                               struct btree_write *w)
1502 {
1503         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1504
1505         do {
1506                 old = new = v;
1507                 if (!(old & 1))
1508                         break;
1509
1510                 new &= ~1UL;
1511         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1512
1513         if (old & 1)
1514                 closure_put(&((struct btree_update *) new)->cl);
1515
1516         bch2_journal_pin_drop(&c->journal, &w->journal);
1517 }
1518
1519 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1520 {
1521         struct btree_write *w = btree_prev_write(b);
1522
1523         bch2_btree_complete_write(c, b, w);
1524         btree_node_io_unlock(b);
1525 }
1526
1527 static void bch2_btree_node_write_error(struct bch_fs *c,
1528                                         struct btree_write_bio *wbio)
1529 {
1530         struct btree *b         = wbio->wbio.bio.bi_private;
1531         struct bkey_buf k;
1532         struct bch_extent_ptr *ptr;
1533         struct btree_trans trans;
1534         struct btree_iter *iter;
1535         int ret;
1536
1537         bch2_bkey_buf_init(&k);
1538         bch2_trans_init(&trans, c, 0, 0);
1539
1540         iter = bch2_trans_get_node_iter(&trans, b->c.btree_id, b->key.k.p,
1541                                         BTREE_MAX_DEPTH, b->c.level, 0);
1542 retry:
1543         ret = bch2_btree_iter_traverse(iter);
1544         if (ret)
1545                 goto err;
1546
1547         /* has node been freed? */
1548         if (iter->l[b->c.level].b != b) {
1549                 /* node has been freed: */
1550                 BUG_ON(!btree_node_dying(b));
1551                 goto out;
1552         }
1553
1554         BUG_ON(!btree_node_hashed(b));
1555
1556         bch2_bkey_buf_copy(&k, c, &b->key);
1557
1558         bch2_bkey_drop_ptrs(bkey_i_to_s(k.k), ptr,
1559                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1560
1561         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(k.k)))
1562                 goto err;
1563
1564         ret = bch2_btree_node_update_key(c, iter, b, k.k);
1565         if (ret == -EINTR)
1566                 goto retry;
1567         if (ret)
1568                 goto err;
1569 out:
1570         bch2_trans_iter_put(&trans, iter);
1571         bch2_trans_exit(&trans);
1572         bch2_bkey_buf_exit(&k, c);
1573         bio_put(&wbio->wbio.bio);
1574         btree_node_write_done(c, b);
1575         return;
1576 err:
1577         set_btree_node_noevict(b);
1578         bch2_fs_fatal_error(c, "fatal error writing btree node");
1579         goto out;
1580 }
1581
1582 void bch2_btree_write_error_work(struct work_struct *work)
1583 {
1584         struct bch_fs *c = container_of(work, struct bch_fs,
1585                                         btree_write_error_work);
1586         struct bio *bio;
1587
1588         while (1) {
1589                 spin_lock_irq(&c->btree_write_error_lock);
1590                 bio = bio_list_pop(&c->btree_write_error_list);
1591                 spin_unlock_irq(&c->btree_write_error_lock);
1592
1593                 if (!bio)
1594                         break;
1595
1596                 bch2_btree_node_write_error(c,
1597                         container_of(bio, struct btree_write_bio, wbio.bio));
1598         }
1599 }
1600
1601 static void btree_node_write_work(struct work_struct *work)
1602 {
1603         struct btree_write_bio *wbio =
1604                 container_of(work, struct btree_write_bio, work);
1605         struct bch_fs *c        = wbio->wbio.c;
1606         struct btree *b         = wbio->wbio.bio.bi_private;
1607
1608         btree_bounce_free(c,
1609                 wbio->bytes,
1610                 wbio->wbio.used_mempool,
1611                 wbio->data);
1612
1613         if (wbio->wbio.failed.nr) {
1614                 unsigned long flags;
1615
1616                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1617                 bio_list_add(&c->btree_write_error_list, &wbio->wbio.bio);
1618                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1619
1620                 queue_work(c->btree_error_wq, &c->btree_write_error_work);
1621                 return;
1622         }
1623
1624         bio_put(&wbio->wbio.bio);
1625         btree_node_write_done(c, b);
1626 }
1627
1628 static void btree_node_write_endio(struct bio *bio)
1629 {
1630         struct bch_write_bio *wbio      = to_wbio(bio);
1631         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1632         struct bch_write_bio *orig      = parent ?: wbio;
1633         struct bch_fs *c                = wbio->c;
1634         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1635         unsigned long flags;
1636
1637         if (wbio->have_ioref)
1638                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1639
1640         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1641                                bch2_blk_status_to_str(bio->bi_status)) ||
1642             bch2_meta_write_fault("btree")) {
1643                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1644                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1645                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1646         }
1647
1648         if (wbio->have_ioref)
1649                 percpu_ref_put(&ca->io_ref);
1650
1651         if (parent) {
1652                 bio_put(bio);
1653                 bio_endio(&parent->bio);
1654         } else {
1655                 struct btree_write_bio *wb =
1656                         container_of(orig, struct btree_write_bio, wbio);
1657
1658                 INIT_WORK(&wb->work, btree_node_write_work);
1659                 queue_work(c->io_complete_wq, &wb->work);
1660         }
1661 }
1662
1663 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1664                                    struct bset *i, unsigned sectors)
1665 {
1666         unsigned whiteout_u64s = 0;
1667         int ret;
1668
1669         if (bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), BKEY_TYPE_btree))
1670                 return -1;
1671
1672         ret = validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false) ?:
1673                 validate_bset(c, NULL, b, i, sectors, WRITE, false);
1674         if (ret) {
1675                 bch2_inconsistent_error(c);
1676                 dump_stack();
1677         }
1678
1679         return ret;
1680 }
1681
1682 static void btree_write_submit(struct work_struct *work)
1683 {
1684         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1685
1686         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree, &wbio->key);
1687 }
1688
1689 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b)
1690 {
1691         struct btree_write_bio *wbio;
1692         struct bset_tree *t;
1693         struct bset *i;
1694         struct btree_node *bn = NULL;
1695         struct btree_node_entry *bne = NULL;
1696         struct bch_extent_ptr *ptr;
1697         struct sort_iter sort_iter;
1698         struct nonce nonce;
1699         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1700         u64 seq = 0;
1701         bool used_mempool;
1702         unsigned long old, new;
1703         bool validate_before_checksum = false;
1704         void *data;
1705
1706         if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
1707                 return;
1708
1709         /*
1710          * We may only have a read lock on the btree node - the dirty bit is our
1711          * "lock" against racing with other threads that may be trying to start
1712          * a write, we do a write iff we clear the dirty bit. Since setting the
1713          * dirty bit requires a write lock, we can't race with other threads
1714          * redirtying it:
1715          */
1716         do {
1717                 old = new = READ_ONCE(b->flags);
1718
1719                 if (!(old & (1 << BTREE_NODE_dirty)))
1720                         return;
1721
1722                 if (!btree_node_may_write(b))
1723                         return;
1724
1725                 if (old & (1 << BTREE_NODE_never_write))
1726                         return;
1727
1728                 if (old & (1 << BTREE_NODE_write_in_flight)) {
1729                         /*
1730                          * XXX waiting on btree writes with btree locks held -
1731                          * this can deadlock, and we hit the write error path
1732                          */
1733                         btree_node_wait_on_io(b);
1734                         continue;
1735                 }
1736
1737                 new &= ~(1 << BTREE_NODE_dirty);
1738                 new &= ~(1 << BTREE_NODE_need_write);
1739                 new |=  (1 << BTREE_NODE_write_in_flight);
1740                 new |=  (1 << BTREE_NODE_just_written);
1741                 new ^=  (1 << BTREE_NODE_write_idx);
1742         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1743
1744         atomic_dec(&c->btree_cache.dirty);
1745
1746         BUG_ON(btree_node_fake(b));
1747         BUG_ON((b->will_make_reachable != 0) != !b->written);
1748
1749         BUG_ON(b->written >= c->opts.btree_node_size);
1750         BUG_ON(b->written & (c->opts.block_size - 1));
1751         BUG_ON(bset_written(b, btree_bset_last(b)));
1752         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1753         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1754
1755         bch2_sort_whiteouts(c, b);
1756
1757         sort_iter_init(&sort_iter, b);
1758
1759         bytes = !b->written
1760                 ? sizeof(struct btree_node)
1761                 : sizeof(struct btree_node_entry);
1762
1763         bytes += b->whiteout_u64s * sizeof(u64);
1764
1765         for_each_bset(b, t) {
1766                 i = bset(b, t);
1767
1768                 if (bset_written(b, i))
1769                         continue;
1770
1771                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1772                 sort_iter_add(&sort_iter,
1773                               btree_bkey_first(b, t),
1774                               btree_bkey_last(b, t));
1775                 seq = max(seq, le64_to_cpu(i->journal_seq));
1776         }
1777
1778         BUG_ON(b->written && !seq);
1779
1780         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1781         bytes += 8;
1782
1783         /* buffer must be a multiple of the block size */
1784         bytes = round_up(bytes, block_bytes(c));
1785
1786         data = btree_bounce_alloc(c, bytes, &used_mempool);
1787
1788         if (!b->written) {
1789                 bn = data;
1790                 *bn = *b->data;
1791                 i = &bn->keys;
1792         } else {
1793                 bne = data;
1794                 bne->keys = b->data->keys;
1795                 i = &bne->keys;
1796         }
1797
1798         i->journal_seq  = cpu_to_le64(seq);
1799         i->u64s         = 0;
1800
1801         sort_iter_add(&sort_iter,
1802                       unwritten_whiteouts_start(c, b),
1803                       unwritten_whiteouts_end(c, b));
1804         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1805
1806         b->whiteout_u64s = 0;
1807
1808         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1809         le16_add_cpu(&i->u64s, u64s);
1810
1811         set_needs_whiteout(i, false);
1812
1813         /* do we have data to write? */
1814         if (b->written && !i->u64s)
1815                 goto nowrite;
1816
1817         bytes_to_write = vstruct_end(i) - data;
1818         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1819
1820         memset(data + bytes_to_write, 0,
1821                (sectors_to_write << 9) - bytes_to_write);
1822
1823         BUG_ON(b->written + sectors_to_write > c->opts.btree_node_size);
1824         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1825         BUG_ON(i->seq != b->data->keys.seq);
1826
1827         i->version = c->sb.version < bcachefs_metadata_version_new_versioning
1828                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1829                 : cpu_to_le16(c->sb.version);
1830         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1831
1832         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1833                 validate_before_checksum = true;
1834
1835         /* validate_bset will be modifying: */
1836         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
1837                 validate_before_checksum = true;
1838
1839         /* if we're going to be encrypting, check metadata validity first: */
1840         if (validate_before_checksum &&
1841             validate_bset_for_write(c, b, i, sectors_to_write))
1842                 goto err;
1843
1844         bset_encrypt(c, i, b->written << 9);
1845
1846         nonce = btree_nonce(i, b->written << 9);
1847
1848         if (bn)
1849                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1850         else
1851                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1852
1853         /* if we're not encrypting, check metadata after checksumming: */
1854         if (!validate_before_checksum &&
1855             validate_bset_for_write(c, b, i, sectors_to_write))
1856                 goto err;
1857
1858         /*
1859          * We handle btree write errors by immediately halting the journal -
1860          * after we've done that, we can't issue any subsequent btree writes
1861          * because they might have pointers to new nodes that failed to write.
1862          *
1863          * Furthermore, there's no point in doing any more btree writes because
1864          * with the journal stopped, we're never going to update the journal to
1865          * reflect that those writes were done and the data flushed from the
1866          * journal:
1867          *
1868          * Also on journal error, the pending write may have updates that were
1869          * never journalled (interior nodes, see btree_update_nodes_written()) -
1870          * it's critical that we don't do the write in that case otherwise we
1871          * will have updates visible that weren't in the journal:
1872          *
1873          * Make sure to update b->written so bch2_btree_init_next() doesn't
1874          * break:
1875          */
1876         if (bch2_journal_error(&c->journal) ||
1877             c->opts.nochanges)
1878                 goto err;
1879
1880         trace_btree_write(b, bytes_to_write, sectors_to_write);
1881
1882         wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1883                                 buf_pages(data, sectors_to_write << 9),
1884                                 &c->btree_bio),
1885                             struct btree_write_bio, wbio.bio);
1886         wbio_init(&wbio->wbio.bio);
1887         wbio->data                      = data;
1888         wbio->bytes                     = bytes;
1889         wbio->wbio.c                    = c;
1890         wbio->wbio.used_mempool         = used_mempool;
1891         wbio->wbio.bio.bi_opf           = REQ_OP_WRITE|REQ_META;
1892         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
1893         wbio->wbio.bio.bi_private       = b;
1894
1895         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
1896
1897         /*
1898          * If we're appending to a leaf node, we don't technically need FUA -
1899          * this write just needs to be persisted before the next journal write,
1900          * which will be marked FLUSH|FUA.
1901          *
1902          * Similarly if we're writing a new btree root - the pointer is going to
1903          * be in the next journal entry.
1904          *
1905          * But if we're writing a new btree node (that isn't a root) or
1906          * appending to a non leaf btree node, we need either FUA or a flush
1907          * when we write the parent with the new pointer. FUA is cheaper than a
1908          * flush, and writes appending to leaf nodes aren't blocking anything so
1909          * just make all btree node writes FUA to keep things sane.
1910          */
1911
1912         bkey_copy(&wbio->key, &b->key);
1913
1914         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&wbio->key)), ptr)
1915                 ptr->offset += b->written;
1916
1917         b->written += sectors_to_write;
1918
1919         atomic64_inc(&c->btree_writes_nr);
1920         atomic64_add(sectors_to_write, &c->btree_writes_sectors);
1921
1922         INIT_WORK(&wbio->work, btree_write_submit);
1923         queue_work(c->io_complete_wq, &wbio->work);
1924         return;
1925 err:
1926         set_btree_node_noevict(b);
1927         b->written += sectors_to_write;
1928 nowrite:
1929         btree_bounce_free(c, bytes, used_mempool, data);
1930         btree_node_write_done(c, b);
1931 }
1932
1933 /*
1934  * Work that must be done with write lock held:
1935  */
1936 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
1937 {
1938         bool invalidated_iter = false;
1939         struct btree_node_entry *bne;
1940         struct bset_tree *t;
1941
1942         if (!btree_node_just_written(b))
1943                 return false;
1944
1945         BUG_ON(b->whiteout_u64s);
1946
1947         clear_btree_node_just_written(b);
1948
1949         /*
1950          * Note: immediately after write, bset_written() doesn't work - the
1951          * amount of data we had to write after compaction might have been
1952          * smaller than the offset of the last bset.
1953          *
1954          * However, we know that all bsets have been written here, as long as
1955          * we're still holding the write lock:
1956          */
1957
1958         /*
1959          * XXX: decide if we really want to unconditionally sort down to a
1960          * single bset:
1961          */
1962         if (b->nsets > 1) {
1963                 btree_node_sort(c, b, 0, b->nsets, true);
1964                 invalidated_iter = true;
1965         } else {
1966                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
1967         }
1968
1969         for_each_bset(b, t)
1970                 set_needs_whiteout(bset(b, t), true);
1971
1972         bch2_btree_verify(c, b);
1973
1974         /*
1975          * If later we don't unconditionally sort down to a single bset, we have
1976          * to ensure this is still true:
1977          */
1978         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
1979
1980         bne = want_new_bset(c, b);
1981         if (bne)
1982                 bch2_bset_init_next(c, b, bne);
1983
1984         bch2_btree_build_aux_trees(b);
1985
1986         return invalidated_iter;
1987 }
1988
1989 /*
1990  * Use this one if the node is intent locked:
1991  */
1992 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1993                            enum six_lock_type lock_type_held)
1994 {
1995         if (lock_type_held == SIX_LOCK_intent ||
1996             (lock_type_held == SIX_LOCK_read &&
1997              six_lock_tryupgrade(&b->c.lock))) {
1998                 __bch2_btree_node_write(c, b);
1999
2000                 /* don't cycle lock unnecessarily: */
2001                 if (btree_node_just_written(b) &&
2002                     six_trylock_write(&b->c.lock)) {
2003                         bch2_btree_post_write_cleanup(c, b);
2004                         six_unlock_write(&b->c.lock);
2005                 }
2006
2007                 if (lock_type_held == SIX_LOCK_read)
2008                         six_lock_downgrade(&b->c.lock);
2009         } else {
2010                 __bch2_btree_node_write(c, b);
2011                 if (lock_type_held == SIX_LOCK_write &&
2012                     btree_node_just_written(b))
2013                         bch2_btree_post_write_cleanup(c, b);
2014         }
2015 }
2016
2017 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2018 {
2019         struct bucket_table *tbl;
2020         struct rhash_head *pos;
2021         struct btree *b;
2022         unsigned i;
2023 restart:
2024         rcu_read_lock();
2025         for_each_cached_btree(b, c, tbl, i, pos)
2026                 if (test_bit(flag, &b->flags)) {
2027                         rcu_read_unlock();
2028                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2029                         goto restart;
2030
2031                 }
2032         rcu_read_unlock();
2033 }
2034
2035 void bch2_btree_flush_all_reads(struct bch_fs *c)
2036 {
2037         __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2038 }
2039
2040 void bch2_btree_flush_all_writes(struct bch_fs *c)
2041 {
2042         __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2043 }
2044
2045 void bch2_dirty_btree_nodes_to_text(struct printbuf *out, struct bch_fs *c)
2046 {
2047         struct bucket_table *tbl;
2048         struct rhash_head *pos;
2049         struct btree *b;
2050         unsigned i;
2051
2052         rcu_read_lock();
2053         for_each_cached_btree(b, c, tbl, i, pos) {
2054                 unsigned long flags = READ_ONCE(b->flags);
2055
2056                 if (!(flags & (1 << BTREE_NODE_dirty)))
2057                         continue;
2058
2059                 pr_buf(out, "%p d %u n %u l %u w %u b %u r %u:%lu\n",
2060                        b,
2061                        (flags & (1 << BTREE_NODE_dirty)) != 0,
2062                        (flags & (1 << BTREE_NODE_need_write)) != 0,
2063                        b->c.level,
2064                        b->written,
2065                        !list_empty_careful(&b->write_blocked),
2066                        b->will_make_reachable != 0,
2067                        b->will_make_reachable & 1);
2068         }
2069         rcu_read_unlock();
2070 }