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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "bkey_methods.h"
   5 #include "bkey_sort.h"
   6 #include "btree_cache.h"
   7 #include "btree_io.h"
   8 #include "btree_iter.h"
   9 #include "btree_locking.h"
  10 #include "btree_update.h"
  11 #include "btree_update_interior.h"
  12 #include "buckets.h"
  13 #include "checksum.h"
  14 #include "debug.h"
  15 #include "error.h"
  16 #include "extents.h"
  17 #include "io.h"
  18 #include "journal_reclaim.h"
  19 #include "journal_seq_blacklist.h"
  20 #include "super-io.h"
  21
  22 #include <linux/sched/mm.h>
  23 #include <trace/events/bcachefs.h>
  24
  25 void bch2_btree_node_io_unlock(struct btree *b)
  26 {
  27         EBUG_ON(!btree_node_write_in_flight(b));
  28
  29         clear_btree_node_write_in_flight_inner(b);
  30         clear_btree_node_write_in_flight(b);
  31         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
  32 }
  33
  34 void bch2_btree_node_io_lock(struct btree *b)
  35 {
  36         BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
  37
  38         wait_on_bit_lock_io(&b->flags, BTREE_NODE_write_in_flight,
  39                             TASK_UNINTERRUPTIBLE);
  40 }
  41
  42 void __bch2_btree_node_wait_on_read(struct btree *b)
  43 {
  44         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
  45                        TASK_UNINTERRUPTIBLE);
  46 }
  47
  48 void __bch2_btree_node_wait_on_write(struct btree *b)
  49 {
  50         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
  51                        TASK_UNINTERRUPTIBLE);
  52 }
  53
  54 void bch2_btree_node_wait_on_read(struct btree *b)
  55 {
  56         BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
  57
  58         wait_on_bit_io(&b->flags, BTREE_NODE_read_in_flight,
  59                        TASK_UNINTERRUPTIBLE);
  60 }
  61
  62 void bch2_btree_node_wait_on_write(struct btree *b)
  63 {
  64         BUG_ON(lock_class_is_held(&bch2_btree_node_lock_key));
  65
  66         wait_on_bit_io(&b->flags, BTREE_NODE_write_in_flight,
  67                        TASK_UNINTERRUPTIBLE);
  68 }
  69
  70 static void verify_no_dups(struct btree *b,
  71                            struct bkey_packed *start,
  72                            struct bkey_packed *end)
  73 {
  74 #ifdef CONFIG_BCACHEFS_DEBUG
  75         struct bkey_packed *k, *p;
  76
  77         if (start == end)
  78                 return;
  79
  80         for (p = start, k = bkey_next(start);
  81              k != end;
  82              p = k, k = bkey_next(k)) {
  83                 struct bkey l = bkey_unpack_key(b, p);
  84                 struct bkey r = bkey_unpack_key(b, k);
  85
  86                 BUG_ON(bpos_cmp(l.p, bkey_start_pos(&r)) >= 0);
  87         }
  88 #endif
  89 }
  90
  91 static void set_needs_whiteout(struct bset *i, int v)
  92 {
  93         struct bkey_packed *k;
  94
  95         for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
  96                 k->needs_whiteout = v;
  97 }
  98
  99 static void btree_bounce_free(struct bch_fs *c, size_t size,
 100                               bool used_mempool, void *p)
 101 {
 102         if (used_mempool)
 103                 mempool_free(p, &c->btree_bounce_pool);
 104         else
 105                 vpfree(p, size);
 106 }
 107
 108 static void *btree_bounce_alloc(struct bch_fs *c, size_t size,
 109                                 bool *used_mempool)
 110 {
 111         unsigned flags = memalloc_nofs_save();
 112         void *p;
 113
 114         BUG_ON(size > btree_bytes(c));
 115
 116         *used_mempool = false;
 117         p = vpmalloc(size, __GFP_NOWARN|GFP_NOWAIT);
 118         if (!p) {
 119                 *used_mempool = true;
 120                 p = mempool_alloc(&c->btree_bounce_pool, GFP_NOIO);
 121         }
 122         memalloc_nofs_restore(flags);
 123         return p;
 124 }
 125
 126 static void sort_bkey_ptrs(const struct btree *bt,
 127                            struct bkey_packed **ptrs, unsigned nr)
 128 {
 129         unsigned n = nr, a = nr / 2, b, c, d;
 130
 131         if (!a)
 132                 return;
 133
 134         /* Heap sort: see lib/sort.c: */
 135         while (1) {
 136                 if (a)
 137                         a--;
 138                 else if (--n)
 139                         swap(ptrs[0], ptrs[n]);
 140                 else
 141                         break;
 142
 143                 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
 144                         b = bch2_bkey_cmp_packed(bt,
 145                                             ptrs[c],
 146                                             ptrs[d]) >= 0 ? c : d;
 147                 if (d == n)
 148                         b = c;
 149
 150                 while (b != a &&
 151                        bch2_bkey_cmp_packed(bt,
 152                                        ptrs[a],
 153                                        ptrs[b]) >= 0)
 154                         b = (b - 1) / 2;
 155                 c = b;
 156                 while (b != a) {
 157                         b = (b - 1) / 2;
 158                         swap(ptrs[b], ptrs[c]);
 159                 }
 160         }
 161 }
 162
 163 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
 164 {
 165         struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
 166         bool used_mempool = false;
 167         size_t bytes = b->whiteout_u64s * sizeof(u64);
 168
 169         if (!b->whiteout_u64s)
 170                 return;
 171
 172         new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
 173
 174         ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
 175
 176         for (k = unwritten_whiteouts_start(c, b);
 177              k != unwritten_whiteouts_end(c, b);
 178              k = bkey_next(k))
 179                 *--ptrs = k;
 180
 181         sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
 182
 183         k = new_whiteouts;
 184
 185         while (ptrs != ptrs_end) {
 186                 bkey_copy(k, *ptrs);
 187                 k = bkey_next(k);
 188                 ptrs++;
 189         }
 190
 191         verify_no_dups(b, new_whiteouts,
 192                        (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
 193
 194         memcpy_u64s(unwritten_whiteouts_start(c, b),
 195                     new_whiteouts, b->whiteout_u64s);
 196
 197         btree_bounce_free(c, bytes, used_mempool, new_whiteouts);
 198 }
 199
 200 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
 201                                 bool compacting, enum compact_mode mode)
 202 {
 203         if (!bset_dead_u64s(b, t))
 204                 return false;
 205
 206         switch (mode) {
 207         case COMPACT_LAZY:
 208                 return should_compact_bset_lazy(b, t) ||
 209                         (compacting && !bset_written(b, bset(b, t)));
 210         case COMPACT_ALL:
 211                 return true;
 212         default:
 213                 BUG();
 214         }
 215 }
 216
 217 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
 218 {
 219         struct bset_tree *t;
 220         bool ret = false;
 221
 222         for_each_bset(b, t) {
 223                 struct bset *i = bset(b, t);
 224                 struct bkey_packed *k, *n, *out, *start, *end;
 225                 struct btree_node_entry *src = NULL, *dst = NULL;
 226
 227                 if (t != b->set && !bset_written(b, i)) {
 228                         src = container_of(i, struct btree_node_entry, keys);
 229                         dst = max(write_block(b),
 230                                   (void *) btree_bkey_last(b, t - 1));
 231                 }
 232
 233                 if (src != dst)
 234                         ret = true;
 235
 236                 if (!should_compact_bset(b, t, ret, mode)) {
 237                         if (src != dst) {
 238                                 memmove(dst, src, sizeof(*src) +
 239                                         le16_to_cpu(src->keys.u64s) *
 240                                         sizeof(u64));
 241                                 i = &dst->keys;
 242                                 set_btree_bset(b, t, i);
 243                         }
 244                         continue;
 245                 }
 246
 247                 start   = btree_bkey_first(b, t);
 248                 end     = btree_bkey_last(b, t);
 249
 250                 if (src != dst) {
 251                         memmove(dst, src, sizeof(*src));
 252                         i = &dst->keys;
 253                         set_btree_bset(b, t, i);
 254                 }
 255
 256                 out = i->start;
 257
 258                 for (k = start; k != end; k = n) {
 259                         n = bkey_next(k);
 260
 261                         if (!bkey_deleted(k)) {
 262                                 bkey_copy(out, k);
 263                                 out = bkey_next(out);
 264                         } else {
 265                                 BUG_ON(k->needs_whiteout);
 266                         }
 267                 }
 268
 269                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 270                 set_btree_bset_end(b, t);
 271                 bch2_bset_set_no_aux_tree(b, t);
 272                 ret = true;
 273         }
 274
 275         bch2_verify_btree_nr_keys(b);
 276
 277         bch2_btree_build_aux_trees(b);
 278
 279         return ret;
 280 }
 281
 282 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 283                             enum compact_mode mode)
 284 {
 285         return bch2_drop_whiteouts(b, mode);
 286 }
 287
 288 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 289                             unsigned start_idx,
 290                             unsigned end_idx,
 291                             bool filter_whiteouts)
 292 {
 293         struct btree_node *out;
 294         struct sort_iter sort_iter;
 295         struct bset_tree *t;
 296         struct bset *start_bset = bset(b, &b->set[start_idx]);
 297         bool used_mempool = false;
 298         u64 start_time, seq = 0;
 299         unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
 300         bool sorting_entire_node = start_idx == 0 &&
 301                 end_idx == b->nsets;
 302
 303         sort_iter_init(&sort_iter, b);
 304
 305         for (t = b->set + start_idx;
 306              t < b->set + end_idx;
 307              t++) {
 308                 u64s += le16_to_cpu(bset(b, t)->u64s);
 309                 sort_iter_add(&sort_iter,
 310                               btree_bkey_first(b, t),
 311                               btree_bkey_last(b, t));
 312         }
 313
 314         bytes = sorting_entire_node
 315                 ? btree_bytes(c)
 316                 : __vstruct_bytes(struct btree_node, u64s);
 317
 318         out = btree_bounce_alloc(c, bytes, &used_mempool);
 319
 320         start_time = local_clock();
 321
 322         u64s = bch2_sort_keys(out->keys.start, &sort_iter, filter_whiteouts);
 323
 324         out->keys.u64s = cpu_to_le16(u64s);
 325
 326         BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
 327
 328         if (sorting_entire_node)
 329                 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 330                                        start_time);
 331
 332         /* Make sure we preserve bset journal_seq: */
 333         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 334                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 335         start_bset->journal_seq = cpu_to_le64(seq);
 336
 337         if (sorting_entire_node) {
 338                 unsigned u64s = le16_to_cpu(out->keys.u64s);
 339
 340                 BUG_ON(bytes != btree_bytes(c));
 341
 342                 /*
 343                  * Our temporary buffer is the same size as the btree node's
 344                  * buffer, we can just swap buffers instead of doing a big
 345                  * memcpy()
 346                  */
 347                 *out = *b->data;
 348                 out->keys.u64s = cpu_to_le16(u64s);
 349                 swap(out, b->data);
 350                 set_btree_bset(b, b->set, &b->data->keys);
 351         } else {
 352                 start_bset->u64s = out->keys.u64s;
 353                 memcpy_u64s(start_bset->start,
 354                             out->keys.start,
 355                             le16_to_cpu(out->keys.u64s));
 356         }
 357
 358         for (i = start_idx + 1; i < end_idx; i++)
 359                 b->nr.bset_u64s[start_idx] +=
 360                         b->nr.bset_u64s[i];
 361
 362         b->nsets -= shift;
 363
 364         for (i = start_idx + 1; i < b->nsets; i++) {
 365                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 366                 b->set[i]               = b->set[i + shift];
 367         }
 368
 369         for (i = b->nsets; i < MAX_BSETS; i++)
 370                 b->nr.bset_u64s[i] = 0;
 371
 372         set_btree_bset_end(b, &b->set[start_idx]);
 373         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 374
 375         btree_bounce_free(c, bytes, used_mempool, out);
 376
 377         bch2_verify_btree_nr_keys(b);
 378 }
 379
 380 void bch2_btree_sort_into(struct bch_fs *c,
 381                          struct btree *dst,
 382                          struct btree *src)
 383 {
 384         struct btree_nr_keys nr;
 385         struct btree_node_iter src_iter;
 386         u64 start_time = local_clock();
 387
 388         BUG_ON(dst->nsets != 1);
 389
 390         bch2_bset_set_no_aux_tree(dst, dst->set);
 391
 392         bch2_btree_node_iter_init_from_start(&src_iter, src);
 393
 394         if (btree_node_is_extents(src))
 395                 nr = bch2_sort_repack_merge(c, btree_bset_first(dst),
 396                                 src, &src_iter,
 397                                 &dst->format,
 398                                 true);
 399         else
 400                 nr = bch2_sort_repack(btree_bset_first(dst),
 401                                 src, &src_iter,
 402                                 &dst->format,
 403                                 true);
 404
 405         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 406                                start_time);
 407
 408         set_btree_bset_end(dst, dst->set);
 409
 410         dst->nr.live_u64s       += nr.live_u64s;
 411         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 412         dst->nr.packed_keys     += nr.packed_keys;
 413         dst->nr.unpacked_keys   += nr.unpacked_keys;
 414
 415         bch2_verify_btree_nr_keys(dst);
 416 }
 417
 418 #define SORT_CRIT       (4096 / sizeof(u64))
 419
 420 /*
 421  * We're about to add another bset to the btree node, so if there's currently
 422  * too many bsets - sort some of them together:
 423  */
 424 static bool btree_node_compact(struct bch_fs *c, struct btree *b)
 425 {
 426         unsigned unwritten_idx;
 427         bool ret = false;
 428
 429         for (unwritten_idx = 0;
 430              unwritten_idx < b->nsets;
 431              unwritten_idx++)
 432                 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
 433                         break;
 434
 435         if (b->nsets - unwritten_idx > 1) {
 436                 btree_node_sort(c, b, unwritten_idx,
 437                                 b->nsets, false);
 438                 ret = true;
 439         }
 440
 441         if (unwritten_idx > 1) {
 442                 btree_node_sort(c, b, 0, unwritten_idx, false);
 443                 ret = true;
 444         }
 445
 446         return ret;
 447 }
 448
 449 void bch2_btree_build_aux_trees(struct btree *b)
 450 {
 451         struct bset_tree *t;
 452
 453         for_each_bset(b, t)
 454                 bch2_bset_build_aux_tree(b, t,
 455                                 !bset_written(b, bset(b, t)) &&
 456                                 t == bset_tree_last(b));
 457 }
 458
 459 /*
 460  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 461  * inserted into
 462  *
 463  * Safe to call if there already is an unwritten bset - will only add a new bset
 464  * if @b doesn't already have one.
 465  *
 466  * Returns true if we sorted (i.e. invalidated iterators
 467  */
 468 void bch2_btree_init_next(struct btree_trans *trans, struct btree *b)
 469 {
 470         struct bch_fs *c = trans->c;
 471         struct btree_node_entry *bne;
 472         bool reinit_iter = false;
 473
 474         EBUG_ON(!(b->c.lock.state.seq & 1));
 475         BUG_ON(bset_written(b, bset(b, &b->set[1])));
 476
 477         if (b->nsets == MAX_BSETS &&
 478             !btree_node_write_in_flight(b)) {
 479                 unsigned log_u64s[] = {
 480                         ilog2(bset_u64s(&b->set[0])),
 481                         ilog2(bset_u64s(&b->set[1])),
 482                         ilog2(bset_u64s(&b->set[2])),
 483                 };
 484
 485                 if (log_u64s[1] >= (log_u64s[0] + log_u64s[2]) / 2) {
 486                         bch2_btree_node_write(c, b, SIX_LOCK_write);
 487                         reinit_iter = true;
 488                 }
 489         }
 490
 491         if (b->nsets == MAX_BSETS &&
 492             btree_node_compact(c, b))
 493                 reinit_iter = true;
 494
 495         BUG_ON(b->nsets >= MAX_BSETS);
 496
 497         bne = want_new_bset(c, b);
 498         if (bne)
 499                 bch2_bset_init_next(c, b, bne);
 500
 501         bch2_btree_build_aux_trees(b);
 502
 503         if (reinit_iter)
 504                 bch2_trans_node_reinit_iter(trans, b);
 505 }
 506
 507 static void btree_pos_to_text(struct printbuf *out, struct bch_fs *c,
 508                           struct btree *b)
 509 {
 510         pr_buf(out, "%s level %u/%u\n  ",
 511                bch2_btree_ids[b->c.btree_id],
 512                b->c.level,
 513                c->btree_roots[b->c.btree_id].level);
 514         bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 515 }
 516
 517 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 518                           struct bch_dev *ca,
 519                           struct btree *b, struct bset *i,
 520                           unsigned offset, int write)
 521 {
 522         pr_buf(out, "error validating btree node ");
 523         if (write)
 524                 pr_buf(out, "before write ");
 525         if (ca)
 526                 pr_buf(out, "on %s ", ca->name);
 527         pr_buf(out, "at btree ");
 528         btree_pos_to_text(out, c, b);
 529
 530         pr_buf(out, "\n  node offset %u", b->written);
 531         if (i)
 532                 pr_buf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 533 }
 534
 535 enum btree_err_type {
 536         BTREE_ERR_FIXABLE,
 537         BTREE_ERR_WANT_RETRY,
 538         BTREE_ERR_MUST_RETRY,
 539         BTREE_ERR_FATAL,
 540 };
 541
 542 enum btree_validate_ret {
 543         BTREE_RETRY_READ = 64,
 544 };
 545
 546 #define btree_err(type, c, ca, b, i, msg, ...)                          \
 547 ({                                                                      \
 548         __label__ out;                                                  \
 549         char _buf[300];                                                 \
 550         char *_buf2 = _buf;                                             \
 551         struct printbuf out = PBUF(_buf);                               \
 552                                                                         \
 553         _buf2 = kmalloc(4096, GFP_ATOMIC);                              \
 554         if (_buf2)                                                      \
 555                 out = _PBUF(_buf2, 4986);                               \
 556                                                                         \
 557         btree_err_msg(&out, c, ca, b, i, b->written, write);            \
 558         pr_buf(&out, ": " msg, ##__VA_ARGS__);                          \
 559                                                                         \
 560         if (type == BTREE_ERR_FIXABLE &&                                \
 561             write == READ &&                                            \
 562             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {             \
 563                 mustfix_fsck_err(c, "%s", _buf2);                       \
 564                 goto out;                                               \
 565         }                                                               \
 566                                                                         \
 567         switch (write) {                                                \
 568         case READ:                                                      \
 569                 bch_err(c, "%s", _buf2);                                \
 570                                                                         \
 571                 switch (type) {                                         \
 572                 case BTREE_ERR_FIXABLE:                                 \
 573                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 574                         goto fsck_err;                                  \
 575                 case BTREE_ERR_WANT_RETRY:                              \
 576                         if (have_retry) {                               \
 577                                 ret = BTREE_RETRY_READ;                 \
 578                                 goto fsck_err;                          \
 579                         }                                               \
 580                         break;                                          \
 581                 case BTREE_ERR_MUST_RETRY:                              \
 582                         ret = BTREE_RETRY_READ;                         \
 583                         goto fsck_err;                                  \
 584                 case BTREE_ERR_FATAL:                                   \
 585                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 586                         goto fsck_err;                                  \
 587                 }                                                       \
 588                 break;                                                  \
 589         case WRITE:                                                     \
 590                 bch_err(c, "corrupt metadata before write: %s", _buf2); \
 591                                                                         \
 592                 if (bch2_fs_inconsistent(c)) {                          \
 593                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 594                         goto fsck_err;                                  \
 595                 }                                                       \
 596                 break;                                                  \
 597         }                                                               \
 598 out:                                                                    \
 599         if (_buf2 != _buf)                                              \
 600                 kfree(_buf2);                                           \
 601         true;                                                           \
 602 })
 603
 604 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 605
 606 /*
 607  * When btree topology repair changes the start or end of a node, that might
 608  * mean we have to drop keys that are no longer inside the node:
 609  */
 610 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
 611 {
 612         struct bset_tree *t;
 613         struct bkey_s_c k;
 614         struct bkey unpacked;
 615         struct btree_node_iter iter;
 616
 617         for_each_bset(b, t) {
 618                 struct bset *i = bset(b, t);
 619                 struct bkey_packed *k;
 620
 621                 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
 622                         if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
 623                                 break;
 624
 625                 if (k != i->start) {
 626                         unsigned shift = (u64 *) k - (u64 *) i->start;
 627
 628                         memmove_u64s_down(i->start, k,
 629                                           (u64 *) vstruct_end(i) - (u64 *) k);
 630                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
 631                         set_btree_bset_end(b, t);
 632                         bch2_bset_set_no_aux_tree(b, t);
 633                 }
 634
 635                 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
 636                         if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
 637                                 break;
 638
 639                 if (k != vstruct_last(i)) {
 640                         i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
 641                         set_btree_bset_end(b, t);
 642                         bch2_bset_set_no_aux_tree(b, t);
 643                 }
 644         }
 645
 646         bch2_btree_build_aux_trees(b);
 647
 648         for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
 649                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 650                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 651         }
 652 }
 653
 654 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 655                          struct btree *b, struct bset *i,
 656                          unsigned offset, unsigned sectors,
 657                          int write, bool have_retry)
 658 {
 659         unsigned version = le16_to_cpu(i->version);
 660         const char *err;
 661         char buf1[100];
 662         char buf2[100];
 663         int ret = 0;
 664
 665         btree_err_on((version != BCH_BSET_VERSION_OLD &&
 666                       version < bcachefs_metadata_version_min) ||
 667                      version >= bcachefs_metadata_version_max,
 668                      BTREE_ERR_FATAL, c, ca, b, i,
 669                      "unsupported bset version");
 670
 671         if (btree_err_on(version < c->sb.version_min,
 672                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 673                          "bset version %u older than superblock version_min %u",
 674                          version, c->sb.version_min)) {
 675                 mutex_lock(&c->sb_lock);
 676                 c->disk_sb.sb->version_min = cpu_to_le16(version);
 677                 bch2_write_super(c);
 678                 mutex_unlock(&c->sb_lock);
 679         }
 680
 681         if (btree_err_on(version > c->sb.version,
 682                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 683                          "bset version %u newer than superblock version %u",
 684                          version, c->sb.version)) {
 685                 mutex_lock(&c->sb_lock);
 686                 c->disk_sb.sb->version = cpu_to_le16(version);
 687                 bch2_write_super(c);
 688                 mutex_unlock(&c->sb_lock);
 689         }
 690
 691         btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
 692                      BTREE_ERR_FATAL, c, ca, b, i,
 693                      "BSET_SEPARATE_WHITEOUTS no longer supported");
 694
 695         if (btree_err_on(offset + sectors > c->opts.btree_node_size,
 696                          BTREE_ERR_FIXABLE, c, ca, b, i,
 697                          "bset past end of btree node")) {
 698                 i->u64s = 0;
 699                 return 0;
 700         }
 701
 702         btree_err_on(offset && !i->u64s,
 703                      BTREE_ERR_FIXABLE, c, ca, b, i,
 704                      "empty bset");
 705
 706         btree_err_on(BSET_OFFSET(i) &&
 707                      BSET_OFFSET(i) != offset,
 708                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 709                      "bset at wrong sector offset");
 710
 711         if (!offset) {
 712                 struct btree_node *bn =
 713                         container_of(i, struct btree_node, keys);
 714                 /* These indicate that we read the wrong btree node: */
 715
 716                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 717                         struct bch_btree_ptr_v2 *bp =
 718                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 719
 720                         /* XXX endianness */
 721                         btree_err_on(bp->seq != bn->keys.seq,
 722                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 723                                      "incorrect sequence number (wrong btree node)");
 724                 }
 725
 726                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 727                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 728                              "incorrect btree id");
 729
 730                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 731                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 732                              "incorrect level");
 733
 734                 if (!write)
 735                         compat_btree_node(b->c.level, b->c.btree_id, version,
 736                                           BSET_BIG_ENDIAN(i), write, bn);
 737
 738                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 739                         struct bch_btree_ptr_v2 *bp =
 740                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 741
 742                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 743                                 b->data->min_key = bp->min_key;
 744                                 b->data->max_key = b->key.k.p;
 745                         }
 746
 747                         btree_err_on(bpos_cmp(b->data->min_key, bp->min_key),
 748                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 749                                      "incorrect min_key: got %s should be %s",
 750                                      (bch2_bpos_to_text(&PBUF(buf1), bn->min_key), buf1),
 751                                      (bch2_bpos_to_text(&PBUF(buf2), bp->min_key), buf2));
 752                 }
 753
 754                 btree_err_on(bpos_cmp(bn->max_key, b->key.k.p),
 755                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 756                              "incorrect max key %s",
 757                              (bch2_bpos_to_text(&PBUF(buf1), bn->max_key), buf1));
 758
 759                 if (write)
 760                         compat_btree_node(b->c.level, b->c.btree_id, version,
 761                                           BSET_BIG_ENDIAN(i), write, bn);
 762
 763                 err = bch2_bkey_format_validate(&bn->format);
 764                 btree_err_on(err,
 765                              BTREE_ERR_FATAL, c, ca, b, i,
 766                              "invalid bkey format: %s", err);
 767
 768                 compat_bformat(b->c.level, b->c.btree_id, version,
 769                                BSET_BIG_ENDIAN(i), write,
 770                                &bn->format);
 771         }
 772 fsck_err:
 773         return ret;
 774 }
 775
 776 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 777                          struct bset *i, unsigned *whiteout_u64s,
 778                          int write, bool have_retry)
 779 {
 780         unsigned version = le16_to_cpu(i->version);
 781         struct bkey_packed *k, *prev = NULL;
 782         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 783                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 784         int ret = 0;
 785
 786         for (k = i->start;
 787              k != vstruct_last(i);) {
 788                 struct bkey_s u;
 789                 struct bkey tmp;
 790                 const char *invalid;
 791
 792                 if (btree_err_on(bkey_next(k) > vstruct_last(i),
 793                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 794                                  "key extends past end of bset")) {
 795                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 796                         break;
 797                 }
 798
 799                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 800                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 801                                  "invalid bkey format %u", k->format)) {
 802                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 803                         memmove_u64s_down(k, bkey_next(k),
 804                                           (u64 *) vstruct_end(i) - (u64 *) k);
 805                         continue;
 806                 }
 807
 808                 /* XXX: validate k->u64s */
 809                 if (!write)
 810                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 811                                     BSET_BIG_ENDIAN(i), write,
 812                                     &b->format, k);
 813
 814                 u = __bkey_disassemble(b, k, &tmp);
 815
 816                 invalid = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b)) ?:
 817                         (!updated_range ?  bch2_bkey_in_btree_node(b, u.s_c) : NULL) ?:
 818                         (write ? bch2_bkey_val_invalid(c, u.s_c) : NULL);
 819                 if (invalid) {
 820                         char buf[160];
 821
 822                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 823                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 824                                   "invalid bkey: %s\n%s", invalid, buf);
 825
 826                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 827                         memmove_u64s_down(k, bkey_next(k),
 828                                           (u64 *) vstruct_end(i) - (u64 *) k);
 829                         continue;
 830                 }
 831
 832                 if (write)
 833                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 834                                     BSET_BIG_ENDIAN(i), write,
 835                                     &b->format, k);
 836
 837                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 838                         char buf1[80];
 839                         char buf2[80];
 840                         struct bkey up = bkey_unpack_key(b, prev);
 841
 842                         bch2_bkey_to_text(&PBUF(buf1), &up);
 843                         bch2_bkey_to_text(&PBUF(buf2), u.k);
 844
 845                         bch2_dump_bset(c, b, i, 0);
 846
 847                         if (btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 848                                       "keys out of order: %s > %s",
 849                                       buf1, buf2)) {
 850                                 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 851                                 memmove_u64s_down(k, bkey_next(k),
 852                                                   (u64 *) vstruct_end(i) - (u64 *) k);
 853                                 continue;
 854                         }
 855                 }
 856
 857                 prev = k;
 858                 k = bkey_next(k);
 859         }
 860 fsck_err:
 861         return ret;
 862 }
 863
 864 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 865                               struct btree *b, bool have_retry)
 866 {
 867         struct btree_node_entry *bne;
 868         struct sort_iter *iter;
 869         struct btree_node *sorted;
 870         struct bkey_packed *k;
 871         struct bch_extent_ptr *ptr;
 872         struct bset *i;
 873         bool used_mempool, blacklisted;
 874         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 875                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 876         unsigned u64s;
 877         unsigned blacklisted_written, nonblacklisted_written = 0;
 878         unsigned ptr_written = btree_ptr_sectors_written(&b->key);
 879         int ret, retry_read = 0, write = READ;
 880
 881         b->version_ondisk = U16_MAX;
 882
 883         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
 884         sort_iter_init(iter, b);
 885         iter->size = (btree_blocks(c) + 1) * 2;
 886
 887         if (bch2_meta_read_fault("btree"))
 888                 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 889                           "dynamic fault");
 890
 891         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 892                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 893                      "bad magic");
 894
 895         btree_err_on(!b->data->keys.seq,
 896                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 897                      "bad btree header");
 898
 899         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 900                 struct bch_btree_ptr_v2 *bp =
 901                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 902
 903                 btree_err_on(b->data->keys.seq != bp->seq,
 904                              BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 905                              "got wrong btree node (seq %llx want %llx)",
 906                              b->data->keys.seq, bp->seq);
 907         }
 908
 909         while (b->written < (ptr_written ?: c->opts.btree_node_size)) {
 910                 unsigned sectors, whiteout_u64s = 0;
 911                 struct nonce nonce;
 912                 struct bch_csum csum;
 913                 bool first = !b->written;
 914
 915                 if (!b->written) {
 916                         i = &b->data->keys;
 917
 918                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 919                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 920                                      "unknown checksum type %llu",
 921                                      BSET_CSUM_TYPE(i));
 922
 923                         nonce = btree_nonce(i, b->written << 9);
 924                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 925
 926                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 927                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 928                                      "invalid checksum");
 929
 930                         bset_encrypt(c, i, b->written << 9);
 931
 932                         btree_err_on(btree_node_is_extents(b) &&
 933                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
 934                                      BTREE_ERR_FATAL, c, NULL, b, NULL,
 935                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
 936
 937                         sectors = vstruct_sectors(b->data, c->block_bits);
 938                 } else {
 939                         bne = write_block(b);
 940                         i = &bne->keys;
 941
 942                         if (i->seq != b->data->keys.seq)
 943                                 break;
 944
 945                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 946                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 947                                      "unknown checksum type %llu",
 948                                      BSET_CSUM_TYPE(i));
 949
 950                         nonce = btree_nonce(i, b->written << 9);
 951                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
 952
 953                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
 954                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 955                                      "invalid checksum");
 956
 957                         bset_encrypt(c, i, b->written << 9);
 958
 959                         sectors = vstruct_sectors(bne, c->block_bits);
 960                 }
 961
 962                 b->version_ondisk = min(b->version_ondisk,
 963                                         le16_to_cpu(i->version));
 964
 965                 ret = validate_bset(c, ca, b, i, b->written, sectors,
 966                                     READ, have_retry);
 967                 if (ret)
 968                         goto fsck_err;
 969
 970                 if (!b->written)
 971                         btree_node_set_format(b, b->data->format);
 972
 973                 ret = validate_bset_keys(c, b, i, &whiteout_u64s,
 974                                     READ, have_retry);
 975                 if (ret)
 976                         goto fsck_err;
 977
 978                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 979
 980                 b->written += sectors;
 981
 982                 blacklisted = bch2_journal_seq_is_blacklisted(c,
 983                                         le64_to_cpu(i->journal_seq),
 984                                         true);
 985
 986                 btree_err_on(blacklisted && first,
 987                              BTREE_ERR_FIXABLE, c, ca, b, i,
 988                              "first btree node bset has blacklisted journal seq");
 989
 990                 btree_err_on(blacklisted && ptr_written,
 991                              BTREE_ERR_FIXABLE, c, ca, b, i,
 992                              "found blacklisted bset in btree node with sectors_written");
 993                 if (blacklisted && !first)
 994                         continue;
 995
 996                 sort_iter_add(iter, i->start,
 997                               vstruct_idx(i, whiteout_u64s));
 998
 999                 sort_iter_add(iter,
1000                               vstruct_idx(i, whiteout_u64s),
1001                               vstruct_last(i));
1002
1003                 nonblacklisted_written = b->written;
1004         }
1005
1006         if (ptr_written) {
1007                 btree_err_on(b->written < ptr_written,
1008                              BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1009                              "btree node data missing: expected %u sectors, found %u",
1010                              ptr_written, b->written);
1011         } else {
1012                 for (bne = write_block(b);
1013                      bset_byte_offset(b, bne) < btree_bytes(c);
1014                      bne = (void *) bne + block_bytes(c))
1015                         btree_err_on(bne->keys.seq == b->data->keys.seq &&
1016                                      !bch2_journal_seq_is_blacklisted(c,
1017                                                                       le64_to_cpu(bne->keys.journal_seq),
1018                                                                       true),
1019                                      BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1020                                      "found bset signature after last bset");
1021
1022                 /*
1023                  * Blacklisted bsets are those that were written after the most recent
1024                  * (flush) journal write. Since there wasn't a flush, they may not have
1025                  * made it to all devices - which means we shouldn't write new bsets
1026                  * after them, as that could leave a gap and then reads from that device
1027                  * wouldn't find all the bsets in that btree node - which means it's
1028                  * important that we start writing new bsets after the most recent _non_
1029                  * blacklisted bset:
1030                  */
1031                 blacklisted_written = b->written;
1032                 b->written = nonblacklisted_written;
1033         }
1034
1035         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1036         sorted->keys.u64s = 0;
1037
1038         set_btree_bset(b, b->set, &b->data->keys);
1039
1040         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1041
1042         u64s = le16_to_cpu(sorted->keys.u64s);
1043         *sorted = *b->data;
1044         sorted->keys.u64s = cpu_to_le16(u64s);
1045         swap(sorted, b->data);
1046         set_btree_bset(b, b->set, &b->data->keys);
1047         b->nsets = 1;
1048
1049         BUG_ON(b->nr.live_u64s != u64s);
1050
1051         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1052
1053         if (updated_range)
1054                 bch2_btree_node_drop_keys_outside_node(b);
1055
1056         i = &b->data->keys;
1057         for (k = i->start; k != vstruct_last(i);) {
1058                 struct bkey tmp;
1059                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1060                 const char *invalid = bch2_bkey_val_invalid(c, u.s_c);
1061
1062                 if (invalid ||
1063                     (bch2_inject_invalid_keys &&
1064                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1065                         char buf[160];
1066
1067                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
1068                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
1069                                   "invalid bkey %s: %s", buf, invalid);
1070
1071                         btree_keys_account_key_drop(&b->nr, 0, k);
1072
1073                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1074                         memmove_u64s_down(k, bkey_next(k),
1075                                           (u64 *) vstruct_end(i) - (u64 *) k);
1076                         set_btree_bset_end(b, b->set);
1077                         continue;
1078                 }
1079
1080                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1081                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1082
1083                         bp.v->mem_ptr = 0;
1084                 }
1085
1086                 k = bkey_next(k);
1087         }
1088
1089         bch2_bset_build_aux_tree(b, b->set, false);
1090
1091         set_needs_whiteout(btree_bset_first(b), true);
1092
1093         btree_node_reset_sib_u64s(b);
1094
1095         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1096                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1097
1098                 if (ca->mi.state != BCH_MEMBER_STATE_rw)
1099                         set_btree_node_need_rewrite(b);
1100         }
1101
1102         if (!ptr_written)
1103                 set_btree_node_need_rewrite(b);
1104 out:
1105         mempool_free(iter, &c->fill_iter);
1106         return retry_read;
1107 fsck_err:
1108         if (ret == BTREE_RETRY_READ) {
1109                 retry_read = 1;
1110         } else {
1111                 bch2_inconsistent_error(c);
1112                 set_btree_node_read_error(b);
1113         }
1114         goto out;
1115 }
1116
1117 static void btree_node_read_work(struct work_struct *work)
1118 {
1119         struct btree_read_bio *rb =
1120                 container_of(work, struct btree_read_bio, work);
1121         struct bch_fs *c        = rb->c;
1122         struct btree *b         = rb->b;
1123         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1124         struct bio *bio         = &rb->bio;
1125         struct bch_io_failures failed = { .nr = 0 };
1126         char buf[200];
1127         struct printbuf out;
1128         bool saw_error = false;
1129         bool can_retry;
1130
1131         goto start;
1132         while (1) {
1133                 bch_info(c, "retrying read");
1134                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1135                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1136                 bio_reset(bio);
1137                 bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1138                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1139                 bio->bi_iter.bi_size    = btree_bytes(c);
1140
1141                 if (rb->have_ioref) {
1142                         bio_set_dev(bio, ca->disk_sb.bdev);
1143                         submit_bio_wait(bio);
1144                 } else {
1145                         bio->bi_status = BLK_STS_REMOVED;
1146                 }
1147 start:
1148                 out = PBUF(buf);
1149                 btree_pos_to_text(&out, c, b);
1150                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1151                                    bch2_blk_status_to_str(bio->bi_status), buf);
1152                 if (rb->have_ioref)
1153                         percpu_ref_put(&ca->io_ref);
1154                 rb->have_ioref = false;
1155
1156                 bch2_mark_io_failure(&failed, &rb->pick);
1157
1158                 can_retry = bch2_bkey_pick_read_device(c,
1159                                 bkey_i_to_s_c(&b->key),
1160                                 &failed, &rb->pick) > 0;
1161
1162                 if (!bio->bi_status &&
1163                     !bch2_btree_node_read_done(c, ca, b, can_retry))
1164                         break;
1165
1166                 saw_error = true;
1167
1168                 if (!can_retry) {
1169                         set_btree_node_read_error(b);
1170                         break;
1171                 }
1172         }
1173
1174         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1175                                rb->start_time);
1176         bio_put(&rb->bio);
1177
1178         if (saw_error && !btree_node_read_error(b))
1179                 bch2_btree_node_rewrite_async(c, b);
1180
1181         clear_btree_node_read_in_flight(b);
1182         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1183 }
1184
1185 static void btree_node_read_endio(struct bio *bio)
1186 {
1187         struct btree_read_bio *rb =
1188                 container_of(bio, struct btree_read_bio, bio);
1189         struct bch_fs *c        = rb->c;
1190
1191         if (rb->have_ioref) {
1192                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1193                 bch2_latency_acct(ca, rb->start_time, READ);
1194         }
1195
1196         queue_work(c->io_complete_wq, &rb->work);
1197 }
1198
1199 struct btree_node_read_all {
1200         struct closure          cl;
1201         struct bch_fs           *c;
1202         struct btree            *b;
1203         unsigned                nr;
1204         void                    *buf[BCH_REPLICAS_MAX];
1205         struct bio              *bio[BCH_REPLICAS_MAX];
1206         int                     err[BCH_REPLICAS_MAX];
1207 };
1208
1209 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1210 {
1211         struct btree_node *bn = data;
1212         struct btree_node_entry *bne;
1213         unsigned offset = 0;
1214
1215         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1216                 return 0;
1217
1218         while (offset < c->opts.btree_node_size) {
1219                 if (!offset) {
1220                         offset += vstruct_sectors(bn, c->block_bits);
1221                 } else {
1222                         bne = data + (offset << 9);
1223                         if (bne->keys.seq != bn->keys.seq)
1224                                 break;
1225                         offset += vstruct_sectors(bne, c->block_bits);
1226                 }
1227         }
1228
1229         return offset;
1230 }
1231
1232 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1233 {
1234         struct btree_node *bn = data;
1235         struct btree_node_entry *bne;
1236
1237         if (!offset)
1238                 return false;
1239
1240         while (offset < c->opts.btree_node_size) {
1241                 bne = data + (offset << 9);
1242                 if (bne->keys.seq == bn->keys.seq)
1243                         return true;
1244                 offset++;
1245         }
1246
1247         return false;
1248         return offset;
1249 }
1250
1251 static void btree_node_read_all_replicas_done(struct closure *cl)
1252 {
1253         struct btree_node_read_all *ra =
1254                 container_of(cl, struct btree_node_read_all, cl);
1255         struct bch_fs *c = ra->c;
1256         struct btree *b = ra->b;
1257         bool dump_bset_maps = false;
1258         bool have_retry = false;
1259         int ret = 0, best = -1, write = READ;
1260         unsigned i, written = 0, written2 = 0;
1261         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1262                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1263
1264         for (i = 0; i < ra->nr; i++) {
1265                 struct btree_node *bn = ra->buf[i];
1266
1267                 if (ra->err[i])
1268                         continue;
1269
1270                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1271                     (seq && seq != bn->keys.seq))
1272                         continue;
1273
1274                 if (best < 0) {
1275                         best = i;
1276                         written = btree_node_sectors_written(c, bn);
1277                         continue;
1278                 }
1279
1280                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1281                 if (btree_err_on(written2 != written, BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1282                                  "btree node sectors written mismatch: %u != %u",
1283                                  written, written2) ||
1284                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1285                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1286                                  "found bset signature after last bset") ||
1287                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1288                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1289                                  "btree node replicas content mismatch"))
1290                         dump_bset_maps = true;
1291
1292                 if (written2 > written) {
1293                         written = written2;
1294                         best = i;
1295                 }
1296         }
1297 fsck_err:
1298         if (dump_bset_maps) {
1299                 for (i = 0; i < ra->nr; i++) {
1300                         char buf[200];
1301                         struct printbuf out = PBUF(buf);
1302                         struct btree_node *bn = ra->buf[i];
1303                         struct btree_node_entry *bne = NULL;
1304                         unsigned offset = 0, sectors;
1305                         bool gap = false;
1306
1307                         if (ra->err[i])
1308                                 continue;
1309
1310                         while (offset < c->opts.btree_node_size) {
1311                                 if (!offset) {
1312                                         sectors = vstruct_sectors(bn, c->block_bits);
1313                                 } else {
1314                                         bne = ra->buf[i] + (offset << 9);
1315                                         if (bne->keys.seq != bn->keys.seq)
1316                                                 break;
1317                                         sectors = vstruct_sectors(bne, c->block_bits);
1318                                 }
1319
1320                                 pr_buf(&out, " %u-%u", offset, offset + sectors);
1321                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1322                                                         le64_to_cpu(bne->keys.journal_seq), false))
1323                                         pr_buf(&out, "*");
1324                                 offset += sectors;
1325                         }
1326
1327                         while (offset < c->opts.btree_node_size) {
1328                                 bne = ra->buf[i] + (offset << 9);
1329                                 if (bne->keys.seq == bn->keys.seq) {
1330                                         if (!gap)
1331                                                 pr_buf(&out, " GAP");
1332                                         gap = true;
1333
1334                                         sectors = vstruct_sectors(bne, c->block_bits);
1335                                         pr_buf(&out, " %u-%u", offset, offset + sectors);
1336                                         if (bch2_journal_seq_is_blacklisted(c,
1337                                                         le64_to_cpu(bne->keys.journal_seq), false))
1338                                                 pr_buf(&out, "*");
1339                                 }
1340                                 offset++;
1341                         }
1342
1343                         bch_err(c, "replica %u:%s", i, buf);
1344                 }
1345         }
1346
1347         if (best >= 0) {
1348                 memcpy(b->data, ra->buf[best], btree_bytes(c));
1349                 ret = bch2_btree_node_read_done(c, NULL, b, false);
1350         } else {
1351                 ret = -1;
1352         }
1353
1354         if (ret)
1355                 set_btree_node_read_error(b);
1356
1357         for (i = 0; i < ra->nr; i++) {
1358                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1359                 bio_put(ra->bio[i]);
1360         }
1361
1362         closure_debug_destroy(&ra->cl);
1363         kfree(ra);
1364
1365         clear_btree_node_read_in_flight(b);
1366         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1367 }
1368
1369 static void btree_node_read_all_replicas_endio(struct bio *bio)
1370 {
1371         struct btree_read_bio *rb =
1372                 container_of(bio, struct btree_read_bio, bio);
1373         struct bch_fs *c        = rb->c;
1374         struct btree_node_read_all *ra = rb->ra;
1375
1376         if (rb->have_ioref) {
1377                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1378                 bch2_latency_acct(ca, rb->start_time, READ);
1379         }
1380
1381         ra->err[rb->idx] = bio->bi_status;
1382         closure_put(&ra->cl);
1383 }
1384
1385 /*
1386  * XXX This allocates multiple times from the same mempools, and can deadlock
1387  * under sufficient memory pressure (but is only a debug path)
1388  */
1389 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1390 {
1391         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1392         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1393         const union bch_extent_entry *entry;
1394         struct extent_ptr_decoded pick;
1395         struct btree_node_read_all *ra;
1396         unsigned i;
1397
1398         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1399         if (!ra)
1400                 return -ENOMEM;
1401
1402         closure_init(&ra->cl, NULL);
1403         ra->c   = c;
1404         ra->b   = b;
1405         ra->nr  = bch2_bkey_nr_ptrs(k);
1406
1407         for (i = 0; i < ra->nr; i++) {
1408                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1409                 ra->bio[i] = bio_alloc_bioset(GFP_NOFS, buf_pages(ra->buf[i],
1410                                                                   btree_bytes(c)),
1411                                               &c->btree_bio);
1412         }
1413
1414         i = 0;
1415         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1416                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1417                 struct btree_read_bio *rb =
1418                         container_of(ra->bio[i], struct btree_read_bio, bio);
1419                 rb->c                   = c;
1420                 rb->b                   = b;
1421                 rb->ra                  = ra;
1422                 rb->start_time          = local_clock();
1423                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1424                 rb->idx                 = i;
1425                 rb->pick                = pick;
1426                 rb->bio.bi_opf          = REQ_OP_READ|REQ_SYNC|REQ_META;
1427                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1428                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1429                 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1430
1431                 if (rb->have_ioref) {
1432                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1433                                      bio_sectors(&rb->bio));
1434                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1435
1436                         closure_get(&ra->cl);
1437                         submit_bio(&rb->bio);
1438                 } else {
1439                         ra->err[i] = BLK_STS_REMOVED;
1440                 }
1441
1442                 i++;
1443         }
1444
1445         if (sync) {
1446                 closure_sync(&ra->cl);
1447                 btree_node_read_all_replicas_done(&ra->cl);
1448         } else {
1449                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1450                             c->io_complete_wq);
1451         }
1452
1453         return 0;
1454 }
1455
1456 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1457                           bool sync)
1458 {
1459         struct extent_ptr_decoded pick;
1460         struct btree_read_bio *rb;
1461         struct bch_dev *ca;
1462         struct bio *bio;
1463         char buf[200];
1464         int ret;
1465
1466         btree_pos_to_text(&PBUF(buf), c, b);
1467         trace_btree_read(c, b);
1468
1469         if (bch2_verify_all_btree_replicas &&
1470             !btree_node_read_all_replicas(c, b, sync))
1471                 return;
1472
1473         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1474                                          NULL, &pick);
1475         if (bch2_fs_fatal_err_on(ret <= 0, c,
1476                         "btree node read error: no device to read from\n"
1477                         " at %s", buf)) {
1478                 set_btree_node_read_error(b);
1479                 return;
1480         }
1481
1482         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1483
1484         bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1485                                                    btree_bytes(c)),
1486                                &c->btree_bio);
1487         rb = container_of(bio, struct btree_read_bio, bio);
1488         rb->c                   = c;
1489         rb->b                   = b;
1490         rb->ra                  = NULL;
1491         rb->start_time          = local_clock();
1492         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1493         rb->pick                = pick;
1494         INIT_WORK(&rb->work, btree_node_read_work);
1495         bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1496         bio->bi_iter.bi_sector  = pick.ptr.offset;
1497         bio->bi_end_io          = btree_node_read_endio;
1498         bch2_bio_map(bio, b->data, btree_bytes(c));
1499
1500         if (rb->have_ioref) {
1501                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1502                              bio_sectors(bio));
1503                 bio_set_dev(bio, ca->disk_sb.bdev);
1504
1505                 if (sync) {
1506                         submit_bio_wait(bio);
1507
1508                         btree_node_read_work(&rb->work);
1509                 } else {
1510                         submit_bio(bio);
1511                 }
1512         } else {
1513                 bio->bi_status = BLK_STS_REMOVED;
1514
1515                 if (sync)
1516                         btree_node_read_work(&rb->work);
1517                 else
1518                         queue_work(c->io_complete_wq, &rb->work);
1519         }
1520 }
1521
1522 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1523                         const struct bkey_i *k, unsigned level)
1524 {
1525         struct closure cl;
1526         struct btree *b;
1527         int ret;
1528
1529         closure_init_stack(&cl);
1530
1531         do {
1532                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1533                 closure_sync(&cl);
1534         } while (ret);
1535
1536         b = bch2_btree_node_mem_alloc(c);
1537         bch2_btree_cache_cannibalize_unlock(c);
1538
1539         BUG_ON(IS_ERR(b));
1540
1541         bkey_copy(&b->key, k);
1542         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1543
1544         set_btree_node_read_in_flight(b);
1545
1546         bch2_btree_node_read(c, b, true);
1547
1548         if (btree_node_read_error(b)) {
1549                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1550
1551                 mutex_lock(&c->btree_cache.lock);
1552                 list_move(&b->list, &c->btree_cache.freeable);
1553                 mutex_unlock(&c->btree_cache.lock);
1554
1555                 ret = -EIO;
1556                 goto err;
1557         }
1558
1559         bch2_btree_set_root_for_read(c, b);
1560 err:
1561         six_unlock_write(&b->c.lock);
1562         six_unlock_intent(&b->c.lock);
1563
1564         return ret;
1565 }
1566
1567 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1568                               struct btree_write *w)
1569 {
1570         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1571
1572         do {
1573                 old = new = v;
1574                 if (!(old & 1))
1575                         break;
1576
1577                 new &= ~1UL;
1578         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1579
1580         if (old & 1)
1581                 closure_put(&((struct btree_update *) new)->cl);
1582
1583         bch2_journal_pin_drop(&c->journal, &w->journal);
1584 }
1585
1586 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1587 {
1588         struct btree_write *w = btree_prev_write(b);
1589         unsigned long old, new, v;
1590
1591         bch2_btree_complete_write(c, b, w);
1592
1593         v = READ_ONCE(b->flags);
1594         do {
1595                 old = new = v;
1596
1597                 if (old & (1U << BTREE_NODE_need_write))
1598                         goto do_write;
1599
1600                 new &= ~(1U << BTREE_NODE_write_in_flight);
1601                 new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1602         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1603
1604         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1605         return;
1606
1607 do_write:
1608         six_lock_read(&b->c.lock, NULL, NULL);
1609         v = READ_ONCE(b->flags);
1610         do {
1611                 old = new = v;
1612
1613                 if ((old & (1U << BTREE_NODE_dirty)) &&
1614                     (old & (1U << BTREE_NODE_need_write)) &&
1615                     !(old & (1U << BTREE_NODE_never_write)) &&
1616                     btree_node_may_write(b)) {
1617                         new &= ~(1U << BTREE_NODE_dirty);
1618                         new &= ~(1U << BTREE_NODE_need_write);
1619                         new |=  (1U << BTREE_NODE_write_in_flight);
1620                         new |=  (1U << BTREE_NODE_write_in_flight_inner);
1621                         new |=  (1U << BTREE_NODE_just_written);
1622                         new ^=  (1U << BTREE_NODE_write_idx);
1623                 } else {
1624                         new &= ~(1U << BTREE_NODE_write_in_flight);
1625                         new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1626                 }
1627         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1628
1629         if (new & (1U << BTREE_NODE_write_in_flight))
1630                 __bch2_btree_node_write(c, b, true);
1631
1632         six_unlock_read(&b->c.lock);
1633 }
1634
1635 static void btree_node_write_work(struct work_struct *work)
1636 {
1637         struct btree_write_bio *wbio =
1638                 container_of(work, struct btree_write_bio, work);
1639         struct bch_fs *c        = wbio->wbio.c;
1640         struct btree *b         = wbio->wbio.bio.bi_private;
1641         struct bch_extent_ptr *ptr;
1642         int ret;
1643
1644         btree_bounce_free(c,
1645                 wbio->data_bytes,
1646                 wbio->wbio.used_mempool,
1647                 wbio->data);
1648
1649         bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1650                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1651
1652         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key)))
1653                 goto err;
1654
1655         if (wbio->wbio.first_btree_write) {
1656                 if (wbio->wbio.failed.nr) {
1657
1658                 }
1659         } else {
1660                 ret = bch2_trans_do(c, NULL, NULL, 0,
1661                         bch2_btree_node_update_key_get_iter(&trans, b, &wbio->key,
1662                                                             !wbio->wbio.failed.nr));
1663                 if (ret)
1664                         goto err;
1665         }
1666 out:
1667         bio_put(&wbio->wbio.bio);
1668         btree_node_write_done(c, b);
1669         return;
1670 err:
1671         set_btree_node_noevict(b);
1672         bch2_fs_fatal_error(c, "fatal error writing btree node");
1673         goto out;
1674 }
1675
1676 static void btree_node_write_endio(struct bio *bio)
1677 {
1678         struct bch_write_bio *wbio      = to_wbio(bio);
1679         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1680         struct bch_write_bio *orig      = parent ?: wbio;
1681         struct btree_write_bio *wb      = container_of(orig, struct btree_write_bio, wbio);
1682         struct bch_fs *c                = wbio->c;
1683         struct btree *b                 = wbio->bio.bi_private;
1684         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1685         unsigned long flags;
1686
1687         if (wbio->have_ioref)
1688                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1689
1690         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1691                                bch2_blk_status_to_str(bio->bi_status)) ||
1692             bch2_meta_write_fault("btree")) {
1693                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1694                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1695                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1696         }
1697
1698         if (wbio->have_ioref)
1699                 percpu_ref_put(&ca->io_ref);
1700
1701         if (parent) {
1702                 bio_put(bio);
1703                 bio_endio(&parent->bio);
1704                 return;
1705         }
1706
1707         clear_btree_node_write_in_flight_inner(b);
1708         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1709         INIT_WORK(&wb->work, btree_node_write_work);
1710         queue_work(c->btree_io_complete_wq, &wb->work);
1711 }
1712
1713 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1714                                    struct bset *i, unsigned sectors)
1715 {
1716         unsigned whiteout_u64s = 0;
1717         int ret;
1718
1719         if (bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), BKEY_TYPE_btree))
1720                 return -1;
1721
1722         ret = validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false) ?:
1723                 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false);
1724         if (ret) {
1725                 bch2_inconsistent_error(c);
1726                 dump_stack();
1727         }
1728
1729         return ret;
1730 }
1731
1732 static void btree_write_submit(struct work_struct *work)
1733 {
1734         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1735         struct bch_extent_ptr *ptr;
1736         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1737
1738         bkey_copy(&tmp.k, &wbio->key);
1739
1740         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1741                 ptr->offset += wbio->sector_offset;
1742
1743         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree, &tmp.k);
1744 }
1745
1746 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, bool already_started)
1747 {
1748         struct btree_write_bio *wbio;
1749         struct bset_tree *t;
1750         struct bset *i;
1751         struct btree_node *bn = NULL;
1752         struct btree_node_entry *bne = NULL;
1753         struct sort_iter sort_iter;
1754         struct nonce nonce;
1755         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1756         u64 seq = 0;
1757         bool used_mempool;
1758         unsigned long old, new;
1759         bool validate_before_checksum = false;
1760         void *data;
1761
1762         if (already_started)
1763                 goto do_write;
1764
1765         if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
1766                 return;
1767
1768         /*
1769          * We may only have a read lock on the btree node - the dirty bit is our
1770          * "lock" against racing with other threads that may be trying to start
1771          * a write, we do a write iff we clear the dirty bit. Since setting the
1772          * dirty bit requires a write lock, we can't race with other threads
1773          * redirtying it:
1774          */
1775         do {
1776                 old = new = READ_ONCE(b->flags);
1777
1778                 if (!(old & (1 << BTREE_NODE_dirty)))
1779                         return;
1780
1781                 if (!btree_node_may_write(b))
1782                         return;
1783
1784                 if (old & (1 << BTREE_NODE_never_write))
1785                         return;
1786
1787                 BUG_ON(old & (1 << BTREE_NODE_write_in_flight));
1788
1789                 new &= ~(1 << BTREE_NODE_dirty);
1790                 new &= ~(1 << BTREE_NODE_need_write);
1791                 new |=  (1 << BTREE_NODE_write_in_flight);
1792                 new |=  (1 << BTREE_NODE_write_in_flight_inner);
1793                 new |=  (1 << BTREE_NODE_just_written);
1794                 new ^=  (1 << BTREE_NODE_write_idx);
1795         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1796
1797         if (new & (1U << BTREE_NODE_need_write))
1798                 return;
1799 do_write:
1800         atomic_dec(&c->btree_cache.dirty);
1801
1802         BUG_ON(btree_node_fake(b));
1803         BUG_ON((b->will_make_reachable != 0) != !b->written);
1804
1805         BUG_ON(b->written >= c->opts.btree_node_size);
1806         BUG_ON(b->written & (c->opts.block_size - 1));
1807         BUG_ON(bset_written(b, btree_bset_last(b)));
1808         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1809         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1810
1811         bch2_sort_whiteouts(c, b);
1812
1813         sort_iter_init(&sort_iter, b);
1814
1815         bytes = !b->written
1816                 ? sizeof(struct btree_node)
1817                 : sizeof(struct btree_node_entry);
1818
1819         bytes += b->whiteout_u64s * sizeof(u64);
1820
1821         for_each_bset(b, t) {
1822                 i = bset(b, t);
1823
1824                 if (bset_written(b, i))
1825                         continue;
1826
1827                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1828                 sort_iter_add(&sort_iter,
1829                               btree_bkey_first(b, t),
1830                               btree_bkey_last(b, t));
1831                 seq = max(seq, le64_to_cpu(i->journal_seq));
1832         }
1833
1834         BUG_ON(b->written && !seq);
1835
1836         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1837         bytes += 8;
1838
1839         /* buffer must be a multiple of the block size */
1840         bytes = round_up(bytes, block_bytes(c));
1841
1842         data = btree_bounce_alloc(c, bytes, &used_mempool);
1843
1844         if (!b->written) {
1845                 bn = data;
1846                 *bn = *b->data;
1847                 i = &bn->keys;
1848         } else {
1849                 bne = data;
1850                 bne->keys = b->data->keys;
1851                 i = &bne->keys;
1852         }
1853
1854         i->journal_seq  = cpu_to_le64(seq);
1855         i->u64s         = 0;
1856
1857         sort_iter_add(&sort_iter,
1858                       unwritten_whiteouts_start(c, b),
1859                       unwritten_whiteouts_end(c, b));
1860         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1861
1862         b->whiteout_u64s = 0;
1863
1864         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1865         le16_add_cpu(&i->u64s, u64s);
1866
1867         set_needs_whiteout(i, false);
1868
1869         /* do we have data to write? */
1870         if (b->written && !i->u64s)
1871                 goto nowrite;
1872
1873         bytes_to_write = vstruct_end(i) - data;
1874         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1875
1876         memset(data + bytes_to_write, 0,
1877                (sectors_to_write << 9) - bytes_to_write);
1878
1879         BUG_ON(b->written + sectors_to_write > c->opts.btree_node_size);
1880         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1881         BUG_ON(i->seq != b->data->keys.seq);
1882
1883         i->version = c->sb.version < bcachefs_metadata_version_new_versioning
1884                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1885                 : cpu_to_le16(c->sb.version);
1886         SET_BSET_OFFSET(i, b->written);
1887         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1888
1889         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1890                 validate_before_checksum = true;
1891
1892         /* validate_bset will be modifying: */
1893         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
1894                 validate_before_checksum = true;
1895
1896         /* if we're going to be encrypting, check metadata validity first: */
1897         if (validate_before_checksum &&
1898             validate_bset_for_write(c, b, i, sectors_to_write))
1899                 goto err;
1900
1901         bset_encrypt(c, i, b->written << 9);
1902
1903         nonce = btree_nonce(i, b->written << 9);
1904
1905         if (bn)
1906                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1907         else
1908                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1909
1910         /* if we're not encrypting, check metadata after checksumming: */
1911         if (!validate_before_checksum &&
1912             validate_bset_for_write(c, b, i, sectors_to_write))
1913                 goto err;
1914
1915         /*
1916          * We handle btree write errors by immediately halting the journal -
1917          * after we've done that, we can't issue any subsequent btree writes
1918          * because they might have pointers to new nodes that failed to write.
1919          *
1920          * Furthermore, there's no point in doing any more btree writes because
1921          * with the journal stopped, we're never going to update the journal to
1922          * reflect that those writes were done and the data flushed from the
1923          * journal:
1924          *
1925          * Also on journal error, the pending write may have updates that were
1926          * never journalled (interior nodes, see btree_update_nodes_written()) -
1927          * it's critical that we don't do the write in that case otherwise we
1928          * will have updates visible that weren't in the journal:
1929          *
1930          * Make sure to update b->written so bch2_btree_init_next() doesn't
1931          * break:
1932          */
1933         if (bch2_journal_error(&c->journal) ||
1934             c->opts.nochanges)
1935                 goto err;
1936
1937         trace_btree_write(b, bytes_to_write, sectors_to_write);
1938
1939         wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1940                                 buf_pages(data, sectors_to_write << 9),
1941                                 &c->btree_bio),
1942                             struct btree_write_bio, wbio.bio);
1943         wbio_init(&wbio->wbio.bio);
1944         wbio->data                      = data;
1945         wbio->data_bytes                = bytes;
1946         wbio->sector_offset             = b->written;
1947         wbio->wbio.c                    = c;
1948         wbio->wbio.used_mempool         = used_mempool;
1949         wbio->wbio.first_btree_write    = !b->written;
1950         wbio->wbio.bio.bi_opf           = REQ_OP_WRITE|REQ_META;
1951         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
1952         wbio->wbio.bio.bi_private       = b;
1953
1954         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
1955
1956         bkey_copy(&wbio->key, &b->key);
1957
1958         b->written += sectors_to_write;
1959
1960         if (wbio->wbio.first_btree_write &&
1961             b->key.k.type == KEY_TYPE_btree_ptr_v2)
1962                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
1963                         cpu_to_le16(b->written);
1964
1965         if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
1966                 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
1967                         cpu_to_le16(b->written);
1968
1969         atomic64_inc(&c->btree_writes_nr);
1970         atomic64_add(sectors_to_write, &c->btree_writes_sectors);
1971
1972         INIT_WORK(&wbio->work, btree_write_submit);
1973         queue_work(c->io_complete_wq, &wbio->work);
1974         return;
1975 err:
1976         set_btree_node_noevict(b);
1977         if (!b->written &&
1978             b->key.k.type == KEY_TYPE_btree_ptr_v2)
1979                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
1980                         cpu_to_le16(sectors_to_write);
1981         b->written += sectors_to_write;
1982 nowrite:
1983         btree_bounce_free(c, bytes, used_mempool, data);
1984         btree_node_write_done(c, b);
1985 }
1986
1987 /*
1988  * Work that must be done with write lock held:
1989  */
1990 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
1991 {
1992         bool invalidated_iter = false;
1993         struct btree_node_entry *bne;
1994         struct bset_tree *t;
1995
1996         if (!btree_node_just_written(b))
1997                 return false;
1998
1999         BUG_ON(b->whiteout_u64s);
2000
2001         clear_btree_node_just_written(b);
2002
2003         /*
2004          * Note: immediately after write, bset_written() doesn't work - the
2005          * amount of data we had to write after compaction might have been
2006          * smaller than the offset of the last bset.
2007          *
2008          * However, we know that all bsets have been written here, as long as
2009          * we're still holding the write lock:
2010          */
2011
2012         /*
2013          * XXX: decide if we really want to unconditionally sort down to a
2014          * single bset:
2015          */
2016         if (b->nsets > 1) {
2017                 btree_node_sort(c, b, 0, b->nsets, true);
2018                 invalidated_iter = true;
2019         } else {
2020                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2021         }
2022
2023         for_each_bset(b, t)
2024                 set_needs_whiteout(bset(b, t), true);
2025
2026         bch2_btree_verify(c, b);
2027
2028         /*
2029          * If later we don't unconditionally sort down to a single bset, we have
2030          * to ensure this is still true:
2031          */
2032         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2033
2034         bne = want_new_bset(c, b);
2035         if (bne)
2036                 bch2_bset_init_next(c, b, bne);
2037
2038         bch2_btree_build_aux_trees(b);
2039
2040         return invalidated_iter;
2041 }
2042
2043 /*
2044  * Use this one if the node is intent locked:
2045  */
2046 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2047                            enum six_lock_type lock_type_held)
2048 {
2049         if (lock_type_held == SIX_LOCK_intent ||
2050             (lock_type_held == SIX_LOCK_read &&
2051              six_lock_tryupgrade(&b->c.lock))) {
2052                 __bch2_btree_node_write(c, b, false);
2053
2054                 /* don't cycle lock unnecessarily: */
2055                 if (btree_node_just_written(b) &&
2056                     six_trylock_write(&b->c.lock)) {
2057                         bch2_btree_post_write_cleanup(c, b);
2058                         six_unlock_write(&b->c.lock);
2059                 }
2060
2061                 if (lock_type_held == SIX_LOCK_read)
2062                         six_lock_downgrade(&b->c.lock);
2063         } else {
2064                 __bch2_btree_node_write(c, b, false);
2065                 if (lock_type_held == SIX_LOCK_write &&
2066                     btree_node_just_written(b))
2067                         bch2_btree_post_write_cleanup(c, b);
2068         }
2069 }
2070
2071 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2072 {
2073         struct bucket_table *tbl;
2074         struct rhash_head *pos;
2075         struct btree *b;
2076         unsigned i;
2077 restart:
2078         rcu_read_lock();
2079         for_each_cached_btree(b, c, tbl, i, pos)
2080                 if (test_bit(flag, &b->flags)) {
2081                         rcu_read_unlock();
2082                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2083                         goto restart;
2084
2085                 }
2086         rcu_read_unlock();
2087 }
2088
2089 void bch2_btree_flush_all_reads(struct bch_fs *c)
2090 {
2091         __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2092 }
2093
2094 void bch2_btree_flush_all_writes(struct bch_fs *c)
2095 {
2096         __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2097 }
2098
2099 void bch2_dirty_btree_nodes_to_text(struct printbuf *out, struct bch_fs *c)
2100 {
2101         struct bucket_table *tbl;
2102         struct rhash_head *pos;
2103         struct btree *b;
2104         unsigned i;
2105
2106         rcu_read_lock();
2107         for_each_cached_btree(b, c, tbl, i, pos) {
2108                 unsigned long flags = READ_ONCE(b->flags);
2109
2110                 if (!(flags & (1 << BTREE_NODE_dirty)))
2111                         continue;
2112
2113                 pr_buf(out, "%p d %u n %u l %u w %u b %u r %u:%lu\n",
2114                        b,
2115                        (flags & (1 << BTREE_NODE_dirty)) != 0,
2116                        (flags & (1 << BTREE_NODE_need_write)) != 0,
2117                        b->c.level,
2118                        b->written,
2119                        !list_empty_careful(&b->write_blocked),
2120                        b->will_make_reachable != 0,
2121                        b->will_make_reachable & 1);
2122         }
2123         rcu_read_unlock();
2124 }