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