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 bset_key_invalid(struct bch_fs *c, struct btree *b,
 771                             struct bkey_s_c k,
 772                             bool updated_range, int rw,
 773                             struct printbuf *err)
 774 {
 775         return __bch2_bkey_invalid(c, k, btree_node_type(b), READ, err) ?:
 776                 (!updated_range ? bch2_bkey_in_btree_node(b, k, err) : 0) ?:
 777                 (rw == WRITE ? bch2_bkey_val_invalid(c, k, READ, err) : 0);
 778 }
 779
 780 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 781                          struct bset *i, unsigned *whiteout_u64s,
 782                          int write, bool have_retry)
 783 {
 784         unsigned version = le16_to_cpu(i->version);
 785         struct bkey_packed *k, *prev = NULL;
 786         struct printbuf buf = PRINTBUF;
 787         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 788                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 789         int ret = 0;
 790
 791         for (k = i->start;
 792              k != vstruct_last(i);) {
 793                 struct bkey_s u;
 794                 struct bkey tmp;
 795
 796                 if (btree_err_on(bkey_next(k) > vstruct_last(i),
 797                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 798                                  "key extends past end of bset")) {
 799                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 800                         break;
 801                 }
 802
 803                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 804                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 805                                  "invalid bkey format %u", k->format)) {
 806                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 807                         memmove_u64s_down(k, bkey_next(k),
 808                                           (u64 *) vstruct_end(i) - (u64 *) k);
 809                         continue;
 810                 }
 811
 812                 /* XXX: validate k->u64s */
 813                 if (!write)
 814                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 815                                     BSET_BIG_ENDIAN(i), write,
 816                                     &b->format, k);
 817
 818                 u = __bkey_disassemble(b, k, &tmp);
 819
 820                 printbuf_reset(&buf);
 821                 if (bset_key_invalid(c, b, u.s_c, updated_range, write, &buf)) {
 822                         printbuf_reset(&buf);
 823                         pr_buf(&buf, "invalid bkey:\n  ");
 824                         bch2_bkey_val_to_text(&buf, c, u.s_c);
 825                         pr_buf(&buf, "  \n");
 826                         bset_key_invalid(c, b, u.s_c, updated_range, write, &buf);
 827
 828                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf);
 829
 830                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 831                         memmove_u64s_down(k, bkey_next(k),
 832                                           (u64 *) vstruct_end(i) - (u64 *) k);
 833                         continue;
 834                 }
 835
 836                 if (write)
 837                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 838                                     BSET_BIG_ENDIAN(i), write,
 839                                     &b->format, k);
 840
 841                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 842                         struct bkey up = bkey_unpack_key(b, prev);
 843
 844                         printbuf_reset(&buf);
 845                         pr_buf(&buf, "keys out of order: ");
 846                         bch2_bkey_to_text(&buf, &up);
 847                         pr_buf(&buf, " > ");
 848                         bch2_bkey_to_text(&buf, u.k);
 849
 850                         bch2_dump_bset(c, b, i, 0);
 851
 852                         if (btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf)) {
 853                                 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 854                                 memmove_u64s_down(k, bkey_next(k),
 855                                                   (u64 *) vstruct_end(i) - (u64 *) k);
 856                                 continue;
 857                         }
 858                 }
 859
 860                 prev = k;
 861                 k = bkey_next(k);
 862         }
 863 fsck_err:
 864         printbuf_exit(&buf);
 865         return ret;
 866 }
 867
 868 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 869                               struct btree *b, bool have_retry)
 870 {
 871         struct btree_node_entry *bne;
 872         struct sort_iter *iter;
 873         struct btree_node *sorted;
 874         struct bkey_packed *k;
 875         struct bch_extent_ptr *ptr;
 876         struct bset *i;
 877         bool used_mempool, blacklisted;
 878         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 879                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 880         unsigned u64s;
 881         unsigned blacklisted_written, nonblacklisted_written = 0;
 882         unsigned ptr_written = btree_ptr_sectors_written(&b->key);
 883         struct printbuf buf = PRINTBUF;
 884         int ret, retry_read = 0, write = READ;
 885
 886         b->version_ondisk = U16_MAX;
 887
 888         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
 889         sort_iter_init(iter, b);
 890         iter->size = (btree_blocks(c) + 1) * 2;
 891
 892         if (bch2_meta_read_fault("btree"))
 893                 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 894                           "dynamic fault");
 895
 896         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 897                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 898                      "bad magic: want %llx, got %llx",
 899                      bset_magic(c), le64_to_cpu(b->data->magic));
 900
 901         btree_err_on(!b->data->keys.seq,
 902                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 903                      "bad btree header: seq 0");
 904
 905         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 906                 struct bch_btree_ptr_v2 *bp =
 907                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 908
 909                 btree_err_on(b->data->keys.seq != bp->seq,
 910                              BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 911                              "got wrong btree node (seq %llx want %llx)",
 912                              b->data->keys.seq, bp->seq);
 913         }
 914
 915         while (b->written < (ptr_written ?: btree_sectors(c))) {
 916                 unsigned sectors, whiteout_u64s = 0;
 917                 struct nonce nonce;
 918                 struct bch_csum csum;
 919                 bool first = !b->written;
 920
 921                 if (!b->written) {
 922                         i = &b->data->keys;
 923
 924                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 925                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 926                                      "unknown checksum type %llu",
 927                                      BSET_CSUM_TYPE(i));
 928
 929                         nonce = btree_nonce(i, b->written << 9);
 930                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 931
 932                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 933                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 934                                      "invalid checksum");
 935
 936                         ret = bset_encrypt(c, i, b->written << 9);
 937                         if (bch2_fs_fatal_err_on(ret, c,
 938                                         "error decrypting btree node: %i", ret))
 939                                 goto fsck_err;
 940
 941                         btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
 942                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
 943                                      BTREE_ERR_FATAL, c, NULL, b, NULL,
 944                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
 945
 946                         sectors = vstruct_sectors(b->data, c->block_bits);
 947                 } else {
 948                         bne = write_block(b);
 949                         i = &bne->keys;
 950
 951                         if (i->seq != b->data->keys.seq)
 952                                 break;
 953
 954                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 955                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 956                                      "unknown checksum type %llu",
 957                                      BSET_CSUM_TYPE(i));
 958
 959                         nonce = btree_nonce(i, b->written << 9);
 960                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
 961
 962                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
 963                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 964                                      "invalid checksum");
 965
 966                         ret = bset_encrypt(c, i, b->written << 9);
 967                         if (bch2_fs_fatal_err_on(ret, c,
 968                                         "error decrypting btree node: %i\n", ret))
 969                                 goto fsck_err;
 970
 971                         sectors = vstruct_sectors(bne, c->block_bits);
 972                 }
 973
 974                 b->version_ondisk = min(b->version_ondisk,
 975                                         le16_to_cpu(i->version));
 976
 977                 ret = validate_bset(c, ca, b, i, b->written, sectors,
 978                                     READ, have_retry);
 979                 if (ret)
 980                         goto fsck_err;
 981
 982                 if (!b->written)
 983                         btree_node_set_format(b, b->data->format);
 984
 985                 ret = validate_bset_keys(c, b, i, &whiteout_u64s,
 986                                     READ, have_retry);
 987                 if (ret)
 988                         goto fsck_err;
 989
 990                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 991
 992                 blacklisted = bch2_journal_seq_is_blacklisted(c,
 993                                         le64_to_cpu(i->journal_seq),
 994                                         true);
 995
 996                 btree_err_on(blacklisted && first,
 997                              BTREE_ERR_FIXABLE, c, ca, b, i,
 998                              "first btree node bset has blacklisted journal seq (%llu)",
 999                              le64_to_cpu(i->journal_seq));
1000
1001                 btree_err_on(blacklisted && ptr_written,
1002                              BTREE_ERR_FIXABLE, c, ca, b, i,
1003                              "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
1004                              le64_to_cpu(i->journal_seq),
1005                              b->written, b->written + sectors, ptr_written);
1006
1007                 b->written += sectors;
1008
1009                 if (blacklisted && !first)
1010                         continue;
1011
1012                 sort_iter_add(iter, i->start,
1013                               vstruct_idx(i, whiteout_u64s));
1014
1015                 sort_iter_add(iter,
1016                               vstruct_idx(i, whiteout_u64s),
1017                               vstruct_last(i));
1018
1019                 nonblacklisted_written = b->written;
1020         }
1021
1022         if (ptr_written) {
1023                 btree_err_on(b->written < ptr_written,
1024                              BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1025                              "btree node data missing: expected %u sectors, found %u",
1026                              ptr_written, b->written);
1027         } else {
1028                 for (bne = write_block(b);
1029                      bset_byte_offset(b, bne) < btree_bytes(c);
1030                      bne = (void *) bne + block_bytes(c))
1031                         btree_err_on(bne->keys.seq == b->data->keys.seq &&
1032                                      !bch2_journal_seq_is_blacklisted(c,
1033                                                                       le64_to_cpu(bne->keys.journal_seq),
1034                                                                       true),
1035                                      BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1036                                      "found bset signature after last bset");
1037
1038                 /*
1039                  * Blacklisted bsets are those that were written after the most recent
1040                  * (flush) journal write. Since there wasn't a flush, they may not have
1041                  * made it to all devices - which means we shouldn't write new bsets
1042                  * after them, as that could leave a gap and then reads from that device
1043                  * wouldn't find all the bsets in that btree node - which means it's
1044                  * important that we start writing new bsets after the most recent _non_
1045                  * blacklisted bset:
1046                  */
1047                 blacklisted_written = b->written;
1048                 b->written = nonblacklisted_written;
1049         }
1050
1051         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1052         sorted->keys.u64s = 0;
1053
1054         set_btree_bset(b, b->set, &b->data->keys);
1055
1056         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1057
1058         u64s = le16_to_cpu(sorted->keys.u64s);
1059         *sorted = *b->data;
1060         sorted->keys.u64s = cpu_to_le16(u64s);
1061         swap(sorted, b->data);
1062         set_btree_bset(b, b->set, &b->data->keys);
1063         b->nsets = 1;
1064
1065         BUG_ON(b->nr.live_u64s != u64s);
1066
1067         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1068
1069         if (updated_range)
1070                 bch2_btree_node_drop_keys_outside_node(b);
1071
1072         i = &b->data->keys;
1073         for (k = i->start; k != vstruct_last(i);) {
1074                 struct bkey tmp;
1075                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1076
1077                 printbuf_reset(&buf);
1078
1079                 if (bch2_bkey_val_invalid(c, u.s_c, READ, &buf) ||
1080                     (bch2_inject_invalid_keys &&
1081                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1082                         printbuf_reset(&buf);
1083
1084                         pr_buf(&buf, "invalid bkey\n  ");
1085                         bch2_bkey_val_to_text(&buf, c, u.s_c);
1086                         pr_buf(&buf, "\n  ");
1087                         bch2_bkey_val_invalid(c, u.s_c, READ, &buf);
1088
1089                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf);
1090
1091                         btree_keys_account_key_drop(&b->nr, 0, k);
1092
1093                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1094                         memmove_u64s_down(k, bkey_next(k),
1095                                           (u64 *) vstruct_end(i) - (u64 *) k);
1096                         set_btree_bset_end(b, b->set);
1097                         continue;
1098                 }
1099
1100                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1101                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1102
1103                         bp.v->mem_ptr = 0;
1104                 }
1105
1106                 k = bkey_next(k);
1107         }
1108
1109         bch2_bset_build_aux_tree(b, b->set, false);
1110
1111         set_needs_whiteout(btree_bset_first(b), true);
1112
1113         btree_node_reset_sib_u64s(b);
1114
1115         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1116                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1117
1118                 if (ca->mi.state != BCH_MEMBER_STATE_rw)
1119                         set_btree_node_need_rewrite(b);
1120         }
1121
1122         if (!ptr_written)
1123                 set_btree_node_need_rewrite(b);
1124 out:
1125         mempool_free(iter, &c->fill_iter);
1126         printbuf_exit(&buf);
1127         return retry_read;
1128 fsck_err:
1129         if (ret == BTREE_RETRY_READ) {
1130                 retry_read = 1;
1131         } else {
1132                 bch2_inconsistent_error(c);
1133                 set_btree_node_read_error(b);
1134         }
1135         goto out;
1136 }
1137
1138 static void btree_node_read_work(struct work_struct *work)
1139 {
1140         struct btree_read_bio *rb =
1141                 container_of(work, struct btree_read_bio, work);
1142         struct bch_fs *c        = rb->c;
1143         struct btree *b         = rb->b;
1144         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1145         struct bio *bio         = &rb->bio;
1146         struct bch_io_failures failed = { .nr = 0 };
1147         struct printbuf buf = PRINTBUF;
1148         bool saw_error = false;
1149         bool can_retry;
1150
1151         goto start;
1152         while (1) {
1153                 bch_info(c, "retrying read");
1154                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1155                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1156                 bio_reset(bio);
1157                 bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1158                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1159                 bio->bi_iter.bi_size    = btree_bytes(c);
1160
1161                 if (rb->have_ioref) {
1162                         bio_set_dev(bio, ca->disk_sb.bdev);
1163                         submit_bio_wait(bio);
1164                 } else {
1165                         bio->bi_status = BLK_STS_REMOVED;
1166                 }
1167 start:
1168                 printbuf_reset(&buf);
1169                 btree_pos_to_text(&buf, c, b);
1170                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1171                                    bch2_blk_status_to_str(bio->bi_status), buf.buf);
1172                 if (rb->have_ioref)
1173                         percpu_ref_put(&ca->io_ref);
1174                 rb->have_ioref = false;
1175
1176                 bch2_mark_io_failure(&failed, &rb->pick);
1177
1178                 can_retry = bch2_bkey_pick_read_device(c,
1179                                 bkey_i_to_s_c(&b->key),
1180                                 &failed, &rb->pick) > 0;
1181
1182                 if (!bio->bi_status &&
1183                     !bch2_btree_node_read_done(c, ca, b, can_retry))
1184                         break;
1185
1186                 saw_error = true;
1187
1188                 if (!can_retry) {
1189                         set_btree_node_read_error(b);
1190                         break;
1191                 }
1192         }
1193
1194         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1195                                rb->start_time);
1196         bio_put(&rb->bio);
1197         printbuf_exit(&buf);
1198
1199         if (saw_error && !btree_node_read_error(b))
1200                 bch2_btree_node_rewrite_async(c, b);
1201
1202         clear_btree_node_read_in_flight(b);
1203         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1204 }
1205
1206 static void btree_node_read_endio(struct bio *bio)
1207 {
1208         struct btree_read_bio *rb =
1209                 container_of(bio, struct btree_read_bio, bio);
1210         struct bch_fs *c        = rb->c;
1211
1212         if (rb->have_ioref) {
1213                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1214                 bch2_latency_acct(ca, rb->start_time, READ);
1215         }
1216
1217         queue_work(c->io_complete_wq, &rb->work);
1218 }
1219
1220 struct btree_node_read_all {
1221         struct closure          cl;
1222         struct bch_fs           *c;
1223         struct btree            *b;
1224         unsigned                nr;
1225         void                    *buf[BCH_REPLICAS_MAX];
1226         struct bio              *bio[BCH_REPLICAS_MAX];
1227         int                     err[BCH_REPLICAS_MAX];
1228 };
1229
1230 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1231 {
1232         struct btree_node *bn = data;
1233         struct btree_node_entry *bne;
1234         unsigned offset = 0;
1235
1236         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1237                 return 0;
1238
1239         while (offset < btree_sectors(c)) {
1240                 if (!offset) {
1241                         offset += vstruct_sectors(bn, c->block_bits);
1242                 } else {
1243                         bne = data + (offset << 9);
1244                         if (bne->keys.seq != bn->keys.seq)
1245                                 break;
1246                         offset += vstruct_sectors(bne, c->block_bits);
1247                 }
1248         }
1249
1250         return offset;
1251 }
1252
1253 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1254 {
1255         struct btree_node *bn = data;
1256         struct btree_node_entry *bne;
1257
1258         if (!offset)
1259                 return false;
1260
1261         while (offset < btree_sectors(c)) {
1262                 bne = data + (offset << 9);
1263                 if (bne->keys.seq == bn->keys.seq)
1264                         return true;
1265                 offset++;
1266         }
1267
1268         return false;
1269         return offset;
1270 }
1271
1272 static void btree_node_read_all_replicas_done(struct closure *cl)
1273 {
1274         struct btree_node_read_all *ra =
1275                 container_of(cl, struct btree_node_read_all, cl);
1276         struct bch_fs *c = ra->c;
1277         struct btree *b = ra->b;
1278         struct printbuf buf = PRINTBUF;
1279         bool dump_bset_maps = false;
1280         bool have_retry = false;
1281         int ret = 0, best = -1, write = READ;
1282         unsigned i, written = 0, written2 = 0;
1283         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1284                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1285
1286         for (i = 0; i < ra->nr; i++) {
1287                 struct btree_node *bn = ra->buf[i];
1288
1289                 if (ra->err[i])
1290                         continue;
1291
1292                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1293                     (seq && seq != bn->keys.seq))
1294                         continue;
1295
1296                 if (best < 0) {
1297                         best = i;
1298                         written = btree_node_sectors_written(c, bn);
1299                         continue;
1300                 }
1301
1302                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1303                 if (btree_err_on(written2 != written, BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1304                                  "btree node sectors written mismatch: %u != %u",
1305                                  written, written2) ||
1306                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1307                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1308                                  "found bset signature after last bset") ||
1309                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1310                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1311                                  "btree node replicas content mismatch"))
1312                         dump_bset_maps = true;
1313
1314                 if (written2 > written) {
1315                         written = written2;
1316                         best = i;
1317                 }
1318         }
1319 fsck_err:
1320         if (dump_bset_maps) {
1321                 for (i = 0; i < ra->nr; i++) {
1322                         struct btree_node *bn = ra->buf[i];
1323                         struct btree_node_entry *bne = NULL;
1324                         unsigned offset = 0, sectors;
1325                         bool gap = false;
1326
1327                         if (ra->err[i])
1328                                 continue;
1329
1330                         printbuf_reset(&buf);
1331
1332                         while (offset < btree_sectors(c)) {
1333                                 if (!offset) {
1334                                         sectors = vstruct_sectors(bn, c->block_bits);
1335                                 } else {
1336                                         bne = ra->buf[i] + (offset << 9);
1337                                         if (bne->keys.seq != bn->keys.seq)
1338                                                 break;
1339                                         sectors = vstruct_sectors(bne, c->block_bits);
1340                                 }
1341
1342                                 pr_buf(&buf, " %u-%u", offset, offset + sectors);
1343                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1344                                                         le64_to_cpu(bne->keys.journal_seq), false))
1345                                         pr_buf(&buf, "*");
1346                                 offset += sectors;
1347                         }
1348
1349                         while (offset < btree_sectors(c)) {
1350                                 bne = ra->buf[i] + (offset << 9);
1351                                 if (bne->keys.seq == bn->keys.seq) {
1352                                         if (!gap)
1353                                                 pr_buf(&buf, " GAP");
1354                                         gap = true;
1355
1356                                         sectors = vstruct_sectors(bne, c->block_bits);
1357                                         pr_buf(&buf, " %u-%u", offset, offset + sectors);
1358                                         if (bch2_journal_seq_is_blacklisted(c,
1359                                                         le64_to_cpu(bne->keys.journal_seq), false))
1360                                                 pr_buf(&buf, "*");
1361                                 }
1362                                 offset++;
1363                         }
1364
1365                         bch_err(c, "replica %u:%s", i, buf.buf);
1366                 }
1367         }
1368
1369         if (best >= 0) {
1370                 memcpy(b->data, ra->buf[best], btree_bytes(c));
1371                 ret = bch2_btree_node_read_done(c, NULL, b, false);
1372         } else {
1373                 ret = -1;
1374         }
1375
1376         if (ret)
1377                 set_btree_node_read_error(b);
1378
1379         for (i = 0; i < ra->nr; i++) {
1380                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1381                 bio_put(ra->bio[i]);
1382         }
1383
1384         closure_debug_destroy(&ra->cl);
1385         kfree(ra);
1386         printbuf_exit(&buf);
1387
1388         clear_btree_node_read_in_flight(b);
1389         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1390 }
1391
1392 static void btree_node_read_all_replicas_endio(struct bio *bio)
1393 {
1394         struct btree_read_bio *rb =
1395                 container_of(bio, struct btree_read_bio, bio);
1396         struct bch_fs *c        = rb->c;
1397         struct btree_node_read_all *ra = rb->ra;
1398
1399         if (rb->have_ioref) {
1400                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1401                 bch2_latency_acct(ca, rb->start_time, READ);
1402         }
1403
1404         ra->err[rb->idx] = bio->bi_status;
1405         closure_put(&ra->cl);
1406 }
1407
1408 /*
1409  * XXX This allocates multiple times from the same mempools, and can deadlock
1410  * under sufficient memory pressure (but is only a debug path)
1411  */
1412 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1413 {
1414         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1415         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1416         const union bch_extent_entry *entry;
1417         struct extent_ptr_decoded pick;
1418         struct btree_node_read_all *ra;
1419         unsigned i;
1420
1421         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1422         if (!ra)
1423                 return -ENOMEM;
1424
1425         closure_init(&ra->cl, NULL);
1426         ra->c   = c;
1427         ra->b   = b;
1428         ra->nr  = bch2_bkey_nr_ptrs(k);
1429
1430         for (i = 0; i < ra->nr; i++) {
1431                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1432                 ra->bio[i] = bio_alloc_bioset(GFP_NOFS, buf_pages(ra->buf[i],
1433                                                                   btree_bytes(c)),
1434                                               &c->btree_bio);
1435         }
1436
1437         i = 0;
1438         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1439                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1440                 struct btree_read_bio *rb =
1441                         container_of(ra->bio[i], struct btree_read_bio, bio);
1442                 rb->c                   = c;
1443                 rb->b                   = b;
1444                 rb->ra                  = ra;
1445                 rb->start_time          = local_clock();
1446                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1447                 rb->idx                 = i;
1448                 rb->pick                = pick;
1449                 rb->bio.bi_opf          = REQ_OP_READ|REQ_SYNC|REQ_META;
1450                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1451                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1452                 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1453
1454                 if (rb->have_ioref) {
1455                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1456                                      bio_sectors(&rb->bio));
1457                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1458
1459                         closure_get(&ra->cl);
1460                         submit_bio(&rb->bio);
1461                 } else {
1462                         ra->err[i] = BLK_STS_REMOVED;
1463                 }
1464
1465                 i++;
1466         }
1467
1468         if (sync) {
1469                 closure_sync(&ra->cl);
1470                 btree_node_read_all_replicas_done(&ra->cl);
1471         } else {
1472                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1473                             c->io_complete_wq);
1474         }
1475
1476         return 0;
1477 }
1478
1479 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1480                           bool sync)
1481 {
1482         struct extent_ptr_decoded pick;
1483         struct btree_read_bio *rb;
1484         struct bch_dev *ca;
1485         struct bio *bio;
1486         struct printbuf buf = PRINTBUF;
1487         int ret;
1488
1489         btree_pos_to_text(&buf, c, b);
1490         trace_btree_read(c, b);
1491
1492         if (bch2_verify_all_btree_replicas &&
1493             !btree_node_read_all_replicas(c, b, sync))
1494                 goto out;
1495
1496         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1497                                          NULL, &pick);
1498         if (bch2_fs_fatal_err_on(ret <= 0, c,
1499                         "btree node read error: no device to read from\n"
1500                         " at %s", buf.buf)) {
1501                 set_btree_node_read_error(b);
1502                 goto out;
1503         }
1504
1505         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1506
1507         bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1508                                                    btree_bytes(c)),
1509                                &c->btree_bio);
1510         rb = container_of(bio, struct btree_read_bio, bio);
1511         rb->c                   = c;
1512         rb->b                   = b;
1513         rb->ra                  = NULL;
1514         rb->start_time          = local_clock();
1515         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1516         rb->pick                = pick;
1517         INIT_WORK(&rb->work, btree_node_read_work);
1518         bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1519         bio->bi_iter.bi_sector  = pick.ptr.offset;
1520         bio->bi_end_io          = btree_node_read_endio;
1521         bch2_bio_map(bio, b->data, btree_bytes(c));
1522
1523         if (rb->have_ioref) {
1524                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1525                              bio_sectors(bio));
1526                 bio_set_dev(bio, ca->disk_sb.bdev);
1527
1528                 if (sync) {
1529                         submit_bio_wait(bio);
1530
1531                         btree_node_read_work(&rb->work);
1532                 } else {
1533                         submit_bio(bio);
1534                 }
1535         } else {
1536                 bio->bi_status = BLK_STS_REMOVED;
1537
1538                 if (sync)
1539                         btree_node_read_work(&rb->work);
1540                 else
1541                         queue_work(c->io_complete_wq, &rb->work);
1542         }
1543 out:
1544         printbuf_exit(&buf);
1545 }
1546
1547 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1548                         const struct bkey_i *k, unsigned level)
1549 {
1550         struct closure cl;
1551         struct btree *b;
1552         int ret;
1553
1554         closure_init_stack(&cl);
1555
1556         do {
1557                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1558                 closure_sync(&cl);
1559         } while (ret);
1560
1561         b = bch2_btree_node_mem_alloc(c, level != 0);
1562         bch2_btree_cache_cannibalize_unlock(c);
1563
1564         BUG_ON(IS_ERR(b));
1565
1566         bkey_copy(&b->key, k);
1567         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1568
1569         set_btree_node_read_in_flight(b);
1570
1571         bch2_btree_node_read(c, b, true);
1572
1573         if (btree_node_read_error(b)) {
1574                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1575
1576                 mutex_lock(&c->btree_cache.lock);
1577                 list_move(&b->list, &c->btree_cache.freeable);
1578                 mutex_unlock(&c->btree_cache.lock);
1579
1580                 ret = -EIO;
1581                 goto err;
1582         }
1583
1584         bch2_btree_set_root_for_read(c, b);
1585 err:
1586         six_unlock_write(&b->c.lock);
1587         six_unlock_intent(&b->c.lock);
1588
1589         return ret;
1590 }
1591
1592 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1593                               struct btree_write *w)
1594 {
1595         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1596
1597         do {
1598                 old = new = v;
1599                 if (!(old & 1))
1600                         break;
1601
1602                 new &= ~1UL;
1603         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1604
1605         if (old & 1)
1606                 closure_put(&((struct btree_update *) new)->cl);
1607
1608         bch2_journal_pin_drop(&c->journal, &w->journal);
1609 }
1610
1611 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1612 {
1613         struct btree_write *w = btree_prev_write(b);
1614         unsigned long old, new, v;
1615
1616         bch2_btree_complete_write(c, b, w);
1617
1618         v = READ_ONCE(b->flags);
1619         do {
1620                 old = new = v;
1621
1622                 if ((old & (1U << BTREE_NODE_dirty)) &&
1623                     (old & (1U << BTREE_NODE_need_write)) &&
1624                     !(old & (1U << BTREE_NODE_never_write)) &&
1625                     !(old & (1U << BTREE_NODE_write_blocked)) &&
1626                     !(old & (1U << BTREE_NODE_will_make_reachable))) {
1627                         new &= ~(1U << BTREE_NODE_dirty);
1628                         new &= ~(1U << BTREE_NODE_need_write);
1629                         new |=  (1U << BTREE_NODE_write_in_flight);
1630                         new |=  (1U << BTREE_NODE_write_in_flight_inner);
1631                         new |=  (1U << BTREE_NODE_just_written);
1632                         new ^=  (1U << BTREE_NODE_write_idx);
1633                 } else {
1634                         new &= ~(1U << BTREE_NODE_write_in_flight);
1635                         new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1636                 }
1637         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1638
1639         if (new & (1U << BTREE_NODE_write_in_flight))
1640                 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED);
1641         else
1642                 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1643 }
1644
1645 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1646 {
1647         six_lock_read(&b->c.lock, NULL, NULL);
1648         __btree_node_write_done(c, b);
1649         six_unlock_read(&b->c.lock);
1650 }
1651
1652 static void btree_node_write_work(struct work_struct *work)
1653 {
1654         struct btree_write_bio *wbio =
1655                 container_of(work, struct btree_write_bio, work);
1656         struct bch_fs *c        = wbio->wbio.c;
1657         struct btree *b         = wbio->wbio.bio.bi_private;
1658         struct bch_extent_ptr *ptr;
1659         int ret;
1660
1661         btree_bounce_free(c,
1662                 wbio->data_bytes,
1663                 wbio->wbio.used_mempool,
1664                 wbio->data);
1665
1666         bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1667                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1668
1669         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key)))
1670                 goto err;
1671
1672         if (wbio->wbio.first_btree_write) {
1673                 if (wbio->wbio.failed.nr) {
1674
1675                 }
1676         } else {
1677                 ret = bch2_trans_do(c, NULL, NULL, 0,
1678                         bch2_btree_node_update_key_get_iter(&trans, b, &wbio->key,
1679                                                             !wbio->wbio.failed.nr));
1680                 if (ret)
1681                         goto err;
1682         }
1683 out:
1684         bio_put(&wbio->wbio.bio);
1685         btree_node_write_done(c, b);
1686         return;
1687 err:
1688         set_btree_node_noevict(b);
1689         bch2_fs_fatal_error(c, "fatal error writing btree node");
1690         goto out;
1691 }
1692
1693 static void btree_node_write_endio(struct bio *bio)
1694 {
1695         struct bch_write_bio *wbio      = to_wbio(bio);
1696         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1697         struct bch_write_bio *orig      = parent ?: wbio;
1698         struct btree_write_bio *wb      = container_of(orig, struct btree_write_bio, wbio);
1699         struct bch_fs *c                = wbio->c;
1700         struct btree *b                 = wbio->bio.bi_private;
1701         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1702         unsigned long flags;
1703
1704         if (wbio->have_ioref)
1705                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1706
1707         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1708                                bch2_blk_status_to_str(bio->bi_status)) ||
1709             bch2_meta_write_fault("btree")) {
1710                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1711                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1712                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1713         }
1714
1715         if (wbio->have_ioref)
1716                 percpu_ref_put(&ca->io_ref);
1717
1718         if (parent) {
1719                 bio_put(bio);
1720                 bio_endio(&parent->bio);
1721                 return;
1722         }
1723
1724         clear_btree_node_write_in_flight_inner(b);
1725         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1726         INIT_WORK(&wb->work, btree_node_write_work);
1727         queue_work(c->btree_io_complete_wq, &wb->work);
1728 }
1729
1730 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1731                                    struct bset *i, unsigned sectors)
1732 {
1733         unsigned whiteout_u64s = 0;
1734         struct printbuf buf = PRINTBUF;
1735         int ret;
1736
1737         ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key),
1738                                 BKEY_TYPE_btree, WRITE, &buf);
1739
1740         if (ret)
1741                 bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf);
1742         printbuf_exit(&buf);
1743         if (ret)
1744                 return ret;
1745
1746         ret = validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false) ?:
1747                 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false);
1748         if (ret) {
1749                 bch2_inconsistent_error(c);
1750                 dump_stack();
1751         }
1752
1753         return ret;
1754 }
1755
1756 static void btree_write_submit(struct work_struct *work)
1757 {
1758         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1759         struct bch_extent_ptr *ptr;
1760         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1761
1762         bkey_copy(&tmp.k, &wbio->key);
1763
1764         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1765                 ptr->offset += wbio->sector_offset;
1766
1767         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree, &tmp.k);
1768 }
1769
1770 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1771 {
1772         struct btree_write_bio *wbio;
1773         struct bset_tree *t;
1774         struct bset *i;
1775         struct btree_node *bn = NULL;
1776         struct btree_node_entry *bne = NULL;
1777         struct sort_iter sort_iter;
1778         struct nonce nonce;
1779         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1780         u64 seq = 0;
1781         bool used_mempool;
1782         unsigned long old, new;
1783         bool validate_before_checksum = false;
1784         void *data;
1785         int ret;
1786
1787         if (flags & BTREE_WRITE_ALREADY_STARTED)
1788                 goto do_write;
1789
1790         /*
1791          * We may only have a read lock on the btree node - the dirty bit is our
1792          * "lock" against racing with other threads that may be trying to start
1793          * a write, we do a write iff we clear the dirty bit. Since setting the
1794          * dirty bit requires a write lock, we can't race with other threads
1795          * redirtying it:
1796          */
1797         do {
1798                 old = new = READ_ONCE(b->flags);
1799
1800                 if (!(old & (1 << BTREE_NODE_dirty)))
1801                         return;
1802
1803                 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
1804                     !(old & (1 << BTREE_NODE_need_write)))
1805                         return;
1806
1807                 if (old &
1808                     ((1 << BTREE_NODE_never_write)|
1809                      (1 << BTREE_NODE_write_blocked)))
1810                         return;
1811
1812                 if (b->written &&
1813                     (old & (1 << BTREE_NODE_will_make_reachable)))
1814                         return;
1815
1816                 if (old & (1 << BTREE_NODE_write_in_flight))
1817                         return;
1818
1819                 new &= ~(1 << BTREE_NODE_dirty);
1820                 new &= ~(1 << BTREE_NODE_need_write);
1821                 new |=  (1 << BTREE_NODE_write_in_flight);
1822                 new |=  (1 << BTREE_NODE_write_in_flight_inner);
1823                 new |=  (1 << BTREE_NODE_just_written);
1824                 new ^=  (1 << BTREE_NODE_write_idx);
1825         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1826
1827         if (new & (1U << BTREE_NODE_need_write))
1828                 return;
1829 do_write:
1830         atomic_dec(&c->btree_cache.dirty);
1831
1832         BUG_ON(btree_node_fake(b));
1833         BUG_ON((b->will_make_reachable != 0) != !b->written);
1834
1835         BUG_ON(b->written >= btree_sectors(c));
1836         BUG_ON(b->written & (block_sectors(c) - 1));
1837         BUG_ON(bset_written(b, btree_bset_last(b)));
1838         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1839         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1840
1841         bch2_sort_whiteouts(c, b);
1842
1843         sort_iter_init(&sort_iter, b);
1844
1845         bytes = !b->written
1846                 ? sizeof(struct btree_node)
1847                 : sizeof(struct btree_node_entry);
1848
1849         bytes += b->whiteout_u64s * sizeof(u64);
1850
1851         for_each_bset(b, t) {
1852                 i = bset(b, t);
1853
1854                 if (bset_written(b, i))
1855                         continue;
1856
1857                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1858                 sort_iter_add(&sort_iter,
1859                               btree_bkey_first(b, t),
1860                               btree_bkey_last(b, t));
1861                 seq = max(seq, le64_to_cpu(i->journal_seq));
1862         }
1863
1864         BUG_ON(b->written && !seq);
1865
1866         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1867         bytes += 8;
1868
1869         /* buffer must be a multiple of the block size */
1870         bytes = round_up(bytes, block_bytes(c));
1871
1872         data = btree_bounce_alloc(c, bytes, &used_mempool);
1873
1874         if (!b->written) {
1875                 bn = data;
1876                 *bn = *b->data;
1877                 i = &bn->keys;
1878         } else {
1879                 bne = data;
1880                 bne->keys = b->data->keys;
1881                 i = &bne->keys;
1882         }
1883
1884         i->journal_seq  = cpu_to_le64(seq);
1885         i->u64s         = 0;
1886
1887         sort_iter_add(&sort_iter,
1888                       unwritten_whiteouts_start(c, b),
1889                       unwritten_whiteouts_end(c, b));
1890         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1891
1892         b->whiteout_u64s = 0;
1893
1894         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1895         le16_add_cpu(&i->u64s, u64s);
1896
1897         set_needs_whiteout(i, false);
1898
1899         /* do we have data to write? */
1900         if (b->written && !i->u64s)
1901                 goto nowrite;
1902
1903         bytes_to_write = vstruct_end(i) - data;
1904         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1905
1906         memset(data + bytes_to_write, 0,
1907                (sectors_to_write << 9) - bytes_to_write);
1908
1909         BUG_ON(b->written + sectors_to_write > btree_sectors(c));
1910         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1911         BUG_ON(i->seq != b->data->keys.seq);
1912
1913         i->version = c->sb.version < bcachefs_metadata_version_bkey_renumber
1914                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1915                 : cpu_to_le16(c->sb.version);
1916         SET_BSET_OFFSET(i, b->written);
1917         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1918
1919         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1920                 validate_before_checksum = true;
1921
1922         /* validate_bset will be modifying: */
1923         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
1924                 validate_before_checksum = true;
1925
1926         /* if we're going to be encrypting, check metadata validity first: */
1927         if (validate_before_checksum &&
1928             validate_bset_for_write(c, b, i, sectors_to_write))
1929                 goto err;
1930
1931         ret = bset_encrypt(c, i, b->written << 9);
1932         if (bch2_fs_fatal_err_on(ret, c,
1933                         "error encrypting btree node: %i\n", ret))
1934                 goto err;
1935
1936         nonce = btree_nonce(i, b->written << 9);
1937
1938         if (bn)
1939                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1940         else
1941                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1942
1943         /* if we're not encrypting, check metadata after checksumming: */
1944         if (!validate_before_checksum &&
1945             validate_bset_for_write(c, b, i, sectors_to_write))
1946                 goto err;
1947
1948         /*
1949          * We handle btree write errors by immediately halting the journal -
1950          * after we've done that, we can't issue any subsequent btree writes
1951          * because they might have pointers to new nodes that failed to write.
1952          *
1953          * Furthermore, there's no point in doing any more btree writes because
1954          * with the journal stopped, we're never going to update the journal to
1955          * reflect that those writes were done and the data flushed from the
1956          * journal:
1957          *
1958          * Also on journal error, the pending write may have updates that were
1959          * never journalled (interior nodes, see btree_update_nodes_written()) -
1960          * it's critical that we don't do the write in that case otherwise we
1961          * will have updates visible that weren't in the journal:
1962          *
1963          * Make sure to update b->written so bch2_btree_init_next() doesn't
1964          * break:
1965          */
1966         if (bch2_journal_error(&c->journal) ||
1967             c->opts.nochanges)
1968                 goto err;
1969
1970         trace_btree_write(b, bytes_to_write, sectors_to_write);
1971
1972         wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1973                                 buf_pages(data, sectors_to_write << 9),
1974                                 &c->btree_bio),
1975                             struct btree_write_bio, wbio.bio);
1976         wbio_init(&wbio->wbio.bio);
1977         wbio->data                      = data;
1978         wbio->data_bytes                = bytes;
1979         wbio->sector_offset             = b->written;
1980         wbio->wbio.c                    = c;
1981         wbio->wbio.used_mempool         = used_mempool;
1982         wbio->wbio.first_btree_write    = !b->written;
1983         wbio->wbio.bio.bi_opf           = REQ_OP_WRITE|REQ_META;
1984         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
1985         wbio->wbio.bio.bi_private       = b;
1986
1987         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
1988
1989         bkey_copy(&wbio->key, &b->key);
1990
1991         b->written += sectors_to_write;
1992
1993         if (wbio->wbio.first_btree_write &&
1994             b->key.k.type == KEY_TYPE_btree_ptr_v2)
1995                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
1996                         cpu_to_le16(b->written);
1997
1998         if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
1999                 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2000                         cpu_to_le16(b->written);
2001
2002         atomic64_inc(&c->btree_writes_nr);
2003         atomic64_add(sectors_to_write, &c->btree_writes_sectors);
2004
2005         INIT_WORK(&wbio->work, btree_write_submit);
2006         queue_work(c->io_complete_wq, &wbio->work);
2007         return;
2008 err:
2009         set_btree_node_noevict(b);
2010         if (!b->written &&
2011             b->key.k.type == KEY_TYPE_btree_ptr_v2)
2012                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
2013                         cpu_to_le16(sectors_to_write);
2014         b->written += sectors_to_write;
2015 nowrite:
2016         btree_bounce_free(c, bytes, used_mempool, data);
2017         __btree_node_write_done(c, b);
2018 }
2019
2020 /*
2021  * Work that must be done with write lock held:
2022  */
2023 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2024 {
2025         bool invalidated_iter = false;
2026         struct btree_node_entry *bne;
2027         struct bset_tree *t;
2028
2029         if (!btree_node_just_written(b))
2030                 return false;
2031
2032         BUG_ON(b->whiteout_u64s);
2033
2034         clear_btree_node_just_written(b);
2035
2036         /*
2037          * Note: immediately after write, bset_written() doesn't work - the
2038          * amount of data we had to write after compaction might have been
2039          * smaller than the offset of the last bset.
2040          *
2041          * However, we know that all bsets have been written here, as long as
2042          * we're still holding the write lock:
2043          */
2044
2045         /*
2046          * XXX: decide if we really want to unconditionally sort down to a
2047          * single bset:
2048          */
2049         if (b->nsets > 1) {
2050                 btree_node_sort(c, b, 0, b->nsets, true);
2051                 invalidated_iter = true;
2052         } else {
2053                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2054         }
2055
2056         for_each_bset(b, t)
2057                 set_needs_whiteout(bset(b, t), true);
2058
2059         bch2_btree_verify(c, b);
2060
2061         /*
2062          * If later we don't unconditionally sort down to a single bset, we have
2063          * to ensure this is still true:
2064          */
2065         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2066
2067         bne = want_new_bset(c, b);
2068         if (bne)
2069                 bch2_bset_init_next(c, b, bne);
2070
2071         bch2_btree_build_aux_trees(b);
2072
2073         return invalidated_iter;
2074 }
2075
2076 /*
2077  * Use this one if the node is intent locked:
2078  */
2079 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2080                            enum six_lock_type lock_type_held,
2081                            unsigned flags)
2082 {
2083         if (lock_type_held == SIX_LOCK_intent ||
2084             (lock_type_held == SIX_LOCK_read &&
2085              six_lock_tryupgrade(&b->c.lock))) {
2086                 __bch2_btree_node_write(c, b, flags);
2087
2088                 /* don't cycle lock unnecessarily: */
2089                 if (btree_node_just_written(b) &&
2090                     six_trylock_write(&b->c.lock)) {
2091                         bch2_btree_post_write_cleanup(c, b);
2092                         six_unlock_write(&b->c.lock);
2093                 }
2094
2095                 if (lock_type_held == SIX_LOCK_read)
2096                         six_lock_downgrade(&b->c.lock);
2097         } else {
2098                 __bch2_btree_node_write(c, b, flags);
2099                 if (lock_type_held == SIX_LOCK_write &&
2100                     btree_node_just_written(b))
2101                         bch2_btree_post_write_cleanup(c, b);
2102         }
2103 }
2104
2105 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2106 {
2107         struct bucket_table *tbl;
2108         struct rhash_head *pos;
2109         struct btree *b;
2110         unsigned i;
2111 restart:
2112         rcu_read_lock();
2113         for_each_cached_btree(b, c, tbl, i, pos)
2114                 if (test_bit(flag, &b->flags)) {
2115                         rcu_read_unlock();
2116                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2117                         goto restart;
2118                 }
2119         rcu_read_unlock();
2120 }
2121
2122 void bch2_btree_flush_all_reads(struct bch_fs *c)
2123 {
2124         __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2125 }
2126
2127 void bch2_btree_flush_all_writes(struct bch_fs *c)
2128 {
2129         __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2130 }