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