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:  ");
 824                         bset_key_invalid(c, b, u.s_c, updated_range, write, &buf);
 825                         pr_buf(&buf, "\n  ");
 826                         bch2_bkey_val_to_text(&buf, c, u.s_c);
 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: ");
1085                         bch2_bkey_val_invalid(c, u.s_c, READ, &buf);
1086                         pr_buf(&buf, "\n  ");
1087                         bch2_bkey_val_to_text(&buf, c, u.s_c);
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 retry = false;
1150         bool can_retry;
1151
1152         goto start;
1153         while (1) {
1154                 retry = true;
1155                 bch_info(c, "retrying read");
1156                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1157                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1158                 bio_reset(bio);
1159                 bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1160                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1161                 bio->bi_iter.bi_size    = btree_bytes(c);
1162
1163                 if (rb->have_ioref) {
1164                         bio_set_dev(bio, ca->disk_sb.bdev);
1165                         submit_bio_wait(bio);
1166                 } else {
1167                         bio->bi_status = BLK_STS_REMOVED;
1168                 }
1169 start:
1170                 printbuf_reset(&buf);
1171                 btree_pos_to_text(&buf, c, b);
1172                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1173                                    bch2_blk_status_to_str(bio->bi_status), buf.buf);
1174                 if (rb->have_ioref)
1175                         percpu_ref_put(&ca->io_ref);
1176                 rb->have_ioref = false;
1177
1178                 bch2_mark_io_failure(&failed, &rb->pick);
1179
1180                 can_retry = bch2_bkey_pick_read_device(c,
1181                                 bkey_i_to_s_c(&b->key),
1182                                 &failed, &rb->pick) > 0;
1183
1184                 if (!bio->bi_status &&
1185                     !bch2_btree_node_read_done(c, ca, b, can_retry)) {
1186                         if (retry)
1187                                 bch_info(c, "retry success");
1188                         break;
1189                 }
1190
1191                 saw_error = true;
1192
1193                 if (!can_retry) {
1194                         set_btree_node_read_error(b);
1195                         break;
1196                 }
1197         }
1198
1199         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1200                                rb->start_time);
1201         bio_put(&rb->bio);
1202         printbuf_exit(&buf);
1203
1204         if (saw_error && !btree_node_read_error(b))
1205                 bch2_btree_node_rewrite_async(c, b);
1206
1207         clear_btree_node_read_in_flight(b);
1208         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1209 }
1210
1211 static void btree_node_read_endio(struct bio *bio)
1212 {
1213         struct btree_read_bio *rb =
1214                 container_of(bio, struct btree_read_bio, bio);
1215         struct bch_fs *c        = rb->c;
1216
1217         if (rb->have_ioref) {
1218                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1219                 bch2_latency_acct(ca, rb->start_time, READ);
1220         }
1221
1222         queue_work(c->io_complete_wq, &rb->work);
1223 }
1224
1225 struct btree_node_read_all {
1226         struct closure          cl;
1227         struct bch_fs           *c;
1228         struct btree            *b;
1229         unsigned                nr;
1230         void                    *buf[BCH_REPLICAS_MAX];
1231         struct bio              *bio[BCH_REPLICAS_MAX];
1232         int                     err[BCH_REPLICAS_MAX];
1233 };
1234
1235 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1236 {
1237         struct btree_node *bn = data;
1238         struct btree_node_entry *bne;
1239         unsigned offset = 0;
1240
1241         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1242                 return 0;
1243
1244         while (offset < btree_sectors(c)) {
1245                 if (!offset) {
1246                         offset += vstruct_sectors(bn, c->block_bits);
1247                 } else {
1248                         bne = data + (offset << 9);
1249                         if (bne->keys.seq != bn->keys.seq)
1250                                 break;
1251                         offset += vstruct_sectors(bne, c->block_bits);
1252                 }
1253         }
1254
1255         return offset;
1256 }
1257
1258 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1259 {
1260         struct btree_node *bn = data;
1261         struct btree_node_entry *bne;
1262
1263         if (!offset)
1264                 return false;
1265
1266         while (offset < btree_sectors(c)) {
1267                 bne = data + (offset << 9);
1268                 if (bne->keys.seq == bn->keys.seq)
1269                         return true;
1270                 offset++;
1271         }
1272
1273         return false;
1274         return offset;
1275 }
1276
1277 static void btree_node_read_all_replicas_done(struct closure *cl)
1278 {
1279         struct btree_node_read_all *ra =
1280                 container_of(cl, struct btree_node_read_all, cl);
1281         struct bch_fs *c = ra->c;
1282         struct btree *b = ra->b;
1283         struct printbuf buf = PRINTBUF;
1284         bool dump_bset_maps = false;
1285         bool have_retry = false;
1286         int ret = 0, best = -1, write = READ;
1287         unsigned i, written = 0, written2 = 0;
1288         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1289                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1290
1291         for (i = 0; i < ra->nr; i++) {
1292                 struct btree_node *bn = ra->buf[i];
1293
1294                 if (ra->err[i])
1295                         continue;
1296
1297                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1298                     (seq && seq != bn->keys.seq))
1299                         continue;
1300
1301                 if (best < 0) {
1302                         best = i;
1303                         written = btree_node_sectors_written(c, bn);
1304                         continue;
1305                 }
1306
1307                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1308                 if (btree_err_on(written2 != written, BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1309                                  "btree node sectors written mismatch: %u != %u",
1310                                  written, written2) ||
1311                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1312                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1313                                  "found bset signature after last bset") ||
1314                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1315                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1316                                  "btree node replicas content mismatch"))
1317                         dump_bset_maps = true;
1318
1319                 if (written2 > written) {
1320                         written = written2;
1321                         best = i;
1322                 }
1323         }
1324 fsck_err:
1325         if (dump_bset_maps) {
1326                 for (i = 0; i < ra->nr; i++) {
1327                         struct btree_node *bn = ra->buf[i];
1328                         struct btree_node_entry *bne = NULL;
1329                         unsigned offset = 0, sectors;
1330                         bool gap = false;
1331
1332                         if (ra->err[i])
1333                                 continue;
1334
1335                         printbuf_reset(&buf);
1336
1337                         while (offset < btree_sectors(c)) {
1338                                 if (!offset) {
1339                                         sectors = vstruct_sectors(bn, c->block_bits);
1340                                 } else {
1341                                         bne = ra->buf[i] + (offset << 9);
1342                                         if (bne->keys.seq != bn->keys.seq)
1343                                                 break;
1344                                         sectors = vstruct_sectors(bne, c->block_bits);
1345                                 }
1346
1347                                 pr_buf(&buf, " %u-%u", offset, offset + sectors);
1348                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1349                                                         le64_to_cpu(bne->keys.journal_seq), false))
1350                                         pr_buf(&buf, "*");
1351                                 offset += sectors;
1352                         }
1353
1354                         while (offset < btree_sectors(c)) {
1355                                 bne = ra->buf[i] + (offset << 9);
1356                                 if (bne->keys.seq == bn->keys.seq) {
1357                                         if (!gap)
1358                                                 pr_buf(&buf, " GAP");
1359                                         gap = true;
1360
1361                                         sectors = vstruct_sectors(bne, c->block_bits);
1362                                         pr_buf(&buf, " %u-%u", offset, offset + sectors);
1363                                         if (bch2_journal_seq_is_blacklisted(c,
1364                                                         le64_to_cpu(bne->keys.journal_seq), false))
1365                                                 pr_buf(&buf, "*");
1366                                 }
1367                                 offset++;
1368                         }
1369
1370                         bch_err(c, "replica %u:%s", i, buf.buf);
1371                 }
1372         }
1373
1374         if (best >= 0) {
1375                 memcpy(b->data, ra->buf[best], btree_bytes(c));
1376                 ret = bch2_btree_node_read_done(c, NULL, b, false);
1377         } else {
1378                 ret = -1;
1379         }
1380
1381         if (ret)
1382                 set_btree_node_read_error(b);
1383
1384         for (i = 0; i < ra->nr; i++) {
1385                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1386                 bio_put(ra->bio[i]);
1387         }
1388
1389         closure_debug_destroy(&ra->cl);
1390         kfree(ra);
1391         printbuf_exit(&buf);
1392
1393         clear_btree_node_read_in_flight(b);
1394         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1395 }
1396
1397 static void btree_node_read_all_replicas_endio(struct bio *bio)
1398 {
1399         struct btree_read_bio *rb =
1400                 container_of(bio, struct btree_read_bio, bio);
1401         struct bch_fs *c        = rb->c;
1402         struct btree_node_read_all *ra = rb->ra;
1403
1404         if (rb->have_ioref) {
1405                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1406                 bch2_latency_acct(ca, rb->start_time, READ);
1407         }
1408
1409         ra->err[rb->idx] = bio->bi_status;
1410         closure_put(&ra->cl);
1411 }
1412
1413 /*
1414  * XXX This allocates multiple times from the same mempools, and can deadlock
1415  * under sufficient memory pressure (but is only a debug path)
1416  */
1417 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1418 {
1419         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1420         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1421         const union bch_extent_entry *entry;
1422         struct extent_ptr_decoded pick;
1423         struct btree_node_read_all *ra;
1424         unsigned i;
1425
1426         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1427         if (!ra)
1428                 return -ENOMEM;
1429
1430         closure_init(&ra->cl, NULL);
1431         ra->c   = c;
1432         ra->b   = b;
1433         ra->nr  = bch2_bkey_nr_ptrs(k);
1434
1435         for (i = 0; i < ra->nr; i++) {
1436                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1437                 ra->bio[i] = bio_alloc_bioset(GFP_NOFS, buf_pages(ra->buf[i],
1438                                                                   btree_bytes(c)),
1439                                               &c->btree_bio);
1440         }
1441
1442         i = 0;
1443         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1444                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1445                 struct btree_read_bio *rb =
1446                         container_of(ra->bio[i], struct btree_read_bio, bio);
1447                 rb->c                   = c;
1448                 rb->b                   = b;
1449                 rb->ra                  = ra;
1450                 rb->start_time          = local_clock();
1451                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1452                 rb->idx                 = i;
1453                 rb->pick                = pick;
1454                 rb->bio.bi_opf          = REQ_OP_READ|REQ_SYNC|REQ_META;
1455                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1456                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1457                 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1458
1459                 if (rb->have_ioref) {
1460                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1461                                      bio_sectors(&rb->bio));
1462                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1463
1464                         closure_get(&ra->cl);
1465                         submit_bio(&rb->bio);
1466                 } else {
1467                         ra->err[i] = BLK_STS_REMOVED;
1468                 }
1469
1470                 i++;
1471         }
1472
1473         if (sync) {
1474                 closure_sync(&ra->cl);
1475                 btree_node_read_all_replicas_done(&ra->cl);
1476         } else {
1477                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1478                             c->io_complete_wq);
1479         }
1480
1481         return 0;
1482 }
1483
1484 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1485                           bool sync)
1486 {
1487         struct extent_ptr_decoded pick;
1488         struct btree_read_bio *rb;
1489         struct bch_dev *ca;
1490         struct bio *bio;
1491         struct printbuf buf = PRINTBUF;
1492         int ret;
1493
1494         btree_pos_to_text(&buf, c, b);
1495         trace_btree_read(c, b);
1496
1497         if (bch2_verify_all_btree_replicas &&
1498             !btree_node_read_all_replicas(c, b, sync))
1499                 goto out;
1500
1501         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1502                                          NULL, &pick);
1503         if (bch2_fs_fatal_err_on(ret <= 0, c,
1504                         "btree node read error: no device to read from\n"
1505                         " at %s", buf.buf)) {
1506                 set_btree_node_read_error(b);
1507                 goto out;
1508         }
1509
1510         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1511
1512         bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1513                                                    btree_bytes(c)),
1514                                &c->btree_bio);
1515         rb = container_of(bio, struct btree_read_bio, bio);
1516         rb->c                   = c;
1517         rb->b                   = b;
1518         rb->ra                  = NULL;
1519         rb->start_time          = local_clock();
1520         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1521         rb->pick                = pick;
1522         INIT_WORK(&rb->work, btree_node_read_work);
1523         bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1524         bio->bi_iter.bi_sector  = pick.ptr.offset;
1525         bio->bi_end_io          = btree_node_read_endio;
1526         bch2_bio_map(bio, b->data, btree_bytes(c));
1527
1528         if (rb->have_ioref) {
1529                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1530                              bio_sectors(bio));
1531                 bio_set_dev(bio, ca->disk_sb.bdev);
1532
1533                 if (sync) {
1534                         submit_bio_wait(bio);
1535
1536                         btree_node_read_work(&rb->work);
1537                 } else {
1538                         submit_bio(bio);
1539                 }
1540         } else {
1541                 bio->bi_status = BLK_STS_REMOVED;
1542
1543                 if (sync)
1544                         btree_node_read_work(&rb->work);
1545                 else
1546                         queue_work(c->io_complete_wq, &rb->work);
1547         }
1548 out:
1549         printbuf_exit(&buf);
1550 }
1551
1552 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1553                         const struct bkey_i *k, unsigned level)
1554 {
1555         struct closure cl;
1556         struct btree *b;
1557         int ret;
1558
1559         closure_init_stack(&cl);
1560
1561         do {
1562                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1563                 closure_sync(&cl);
1564         } while (ret);
1565
1566         b = bch2_btree_node_mem_alloc(c, level != 0);
1567         bch2_btree_cache_cannibalize_unlock(c);
1568
1569         BUG_ON(IS_ERR(b));
1570
1571         bkey_copy(&b->key, k);
1572         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1573
1574         set_btree_node_read_in_flight(b);
1575
1576         bch2_btree_node_read(c, b, true);
1577
1578         if (btree_node_read_error(b)) {
1579                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1580
1581                 mutex_lock(&c->btree_cache.lock);
1582                 list_move(&b->list, &c->btree_cache.freeable);
1583                 mutex_unlock(&c->btree_cache.lock);
1584
1585                 ret = -EIO;
1586                 goto err;
1587         }
1588
1589         bch2_btree_set_root_for_read(c, b);
1590 err:
1591         six_unlock_write(&b->c.lock);
1592         six_unlock_intent(&b->c.lock);
1593
1594         return ret;
1595 }
1596
1597 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1598                               struct btree_write *w)
1599 {
1600         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1601
1602         do {
1603                 old = new = v;
1604                 if (!(old & 1))
1605                         break;
1606
1607                 new &= ~1UL;
1608         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1609
1610         if (old & 1)
1611                 closure_put(&((struct btree_update *) new)->cl);
1612
1613         bch2_journal_pin_drop(&c->journal, &w->journal);
1614 }
1615
1616 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1617 {
1618         struct btree_write *w = btree_prev_write(b);
1619         unsigned long old, new, v;
1620
1621         bch2_btree_complete_write(c, b, w);
1622
1623         v = READ_ONCE(b->flags);
1624         do {
1625                 old = new = v;
1626
1627                 if ((old & (1U << BTREE_NODE_dirty)) &&
1628                     (old & (1U << BTREE_NODE_need_write)) &&
1629                     !(old & (1U << BTREE_NODE_never_write)) &&
1630                     !(old & (1U << BTREE_NODE_write_blocked)) &&
1631                     !(old & (1U << BTREE_NODE_will_make_reachable))) {
1632                         new &= ~(1U << BTREE_NODE_dirty);
1633                         new &= ~(1U << BTREE_NODE_need_write);
1634                         new |=  (1U << BTREE_NODE_write_in_flight);
1635                         new |=  (1U << BTREE_NODE_write_in_flight_inner);
1636                         new |=  (1U << BTREE_NODE_just_written);
1637                         new ^=  (1U << BTREE_NODE_write_idx);
1638                 } else {
1639                         new &= ~(1U << BTREE_NODE_write_in_flight);
1640                         new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1641                 }
1642         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1643
1644         if (new & (1U << BTREE_NODE_write_in_flight))
1645                 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED);
1646         else
1647                 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1648 }
1649
1650 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1651 {
1652         six_lock_read(&b->c.lock, NULL, NULL);
1653         __btree_node_write_done(c, b);
1654         six_unlock_read(&b->c.lock);
1655 }
1656
1657 static void btree_node_write_work(struct work_struct *work)
1658 {
1659         struct btree_write_bio *wbio =
1660                 container_of(work, struct btree_write_bio, work);
1661         struct bch_fs *c        = wbio->wbio.c;
1662         struct btree *b         = wbio->wbio.bio.bi_private;
1663         struct bch_extent_ptr *ptr;
1664         int ret;
1665
1666         btree_bounce_free(c,
1667                 wbio->data_bytes,
1668                 wbio->wbio.used_mempool,
1669                 wbio->data);
1670
1671         bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1672                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1673
1674         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key)))
1675                 goto err;
1676
1677         if (wbio->wbio.first_btree_write) {
1678                 if (wbio->wbio.failed.nr) {
1679
1680                 }
1681         } else {
1682                 ret = bch2_trans_do(c, NULL, NULL, 0,
1683                         bch2_btree_node_update_key_get_iter(&trans, b, &wbio->key,
1684                                                             !wbio->wbio.failed.nr));
1685                 if (ret)
1686                         goto err;
1687         }
1688 out:
1689         bio_put(&wbio->wbio.bio);
1690         btree_node_write_done(c, b);
1691         return;
1692 err:
1693         set_btree_node_noevict(b);
1694         bch2_fs_fatal_error(c, "fatal error writing btree node");
1695         goto out;
1696 }
1697
1698 static void btree_node_write_endio(struct bio *bio)
1699 {
1700         struct bch_write_bio *wbio      = to_wbio(bio);
1701         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1702         struct bch_write_bio *orig      = parent ?: wbio;
1703         struct btree_write_bio *wb      = container_of(orig, struct btree_write_bio, wbio);
1704         struct bch_fs *c                = wbio->c;
1705         struct btree *b                 = wbio->bio.bi_private;
1706         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1707         unsigned long flags;
1708
1709         if (wbio->have_ioref)
1710                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1711
1712         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1713                                bch2_blk_status_to_str(bio->bi_status)) ||
1714             bch2_meta_write_fault("btree")) {
1715                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1716                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1717                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1718         }
1719
1720         if (wbio->have_ioref)
1721                 percpu_ref_put(&ca->io_ref);
1722
1723         if (parent) {
1724                 bio_put(bio);
1725                 bio_endio(&parent->bio);
1726                 return;
1727         }
1728
1729         clear_btree_node_write_in_flight_inner(b);
1730         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1731         INIT_WORK(&wb->work, btree_node_write_work);
1732         queue_work(c->btree_io_complete_wq, &wb->work);
1733 }
1734
1735 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1736                                    struct bset *i, unsigned sectors)
1737 {
1738         unsigned whiteout_u64s = 0;
1739         struct printbuf buf = PRINTBUF;
1740         int ret;
1741
1742         ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key),
1743                                 BKEY_TYPE_btree, WRITE, &buf);
1744
1745         if (ret)
1746                 bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf);
1747         printbuf_exit(&buf);
1748         if (ret)
1749                 return ret;
1750
1751         ret = validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false) ?:
1752                 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false);
1753         if (ret) {
1754                 bch2_inconsistent_error(c);
1755                 dump_stack();
1756         }
1757
1758         return ret;
1759 }
1760
1761 static void btree_write_submit(struct work_struct *work)
1762 {
1763         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1764         struct bch_extent_ptr *ptr;
1765         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1766
1767         bkey_copy(&tmp.k, &wbio->key);
1768
1769         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1770                 ptr->offset += wbio->sector_offset;
1771
1772         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree, &tmp.k);
1773 }
1774
1775 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1776 {
1777         struct btree_write_bio *wbio;
1778         struct bset_tree *t;
1779         struct bset *i;
1780         struct btree_node *bn = NULL;
1781         struct btree_node_entry *bne = NULL;
1782         struct sort_iter sort_iter;
1783         struct nonce nonce;
1784         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1785         u64 seq = 0;
1786         bool used_mempool;
1787         unsigned long old, new;
1788         bool validate_before_checksum = false;
1789         void *data;
1790         int ret;
1791
1792         if (flags & BTREE_WRITE_ALREADY_STARTED)
1793                 goto do_write;
1794
1795         /*
1796          * We may only have a read lock on the btree node - the dirty bit is our
1797          * "lock" against racing with other threads that may be trying to start
1798          * a write, we do a write iff we clear the dirty bit. Since setting the
1799          * dirty bit requires a write lock, we can't race with other threads
1800          * redirtying it:
1801          */
1802         do {
1803                 old = new = READ_ONCE(b->flags);
1804
1805                 if (!(old & (1 << BTREE_NODE_dirty)))
1806                         return;
1807
1808                 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
1809                     !(old & (1 << BTREE_NODE_need_write)))
1810                         return;
1811
1812                 if (old &
1813                     ((1 << BTREE_NODE_never_write)|
1814                      (1 << BTREE_NODE_write_blocked)))
1815                         return;
1816
1817                 if (b->written &&
1818                     (old & (1 << BTREE_NODE_will_make_reachable)))
1819                         return;
1820
1821                 if (old & (1 << BTREE_NODE_write_in_flight))
1822                         return;
1823
1824                 new &= ~(1 << BTREE_NODE_dirty);
1825                 new &= ~(1 << BTREE_NODE_need_write);
1826                 new |=  (1 << BTREE_NODE_write_in_flight);
1827                 new |=  (1 << BTREE_NODE_write_in_flight_inner);
1828                 new |=  (1 << BTREE_NODE_just_written);
1829                 new ^=  (1 << BTREE_NODE_write_idx);
1830         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1831
1832         if (new & (1U << BTREE_NODE_need_write))
1833                 return;
1834 do_write:
1835         atomic_dec(&c->btree_cache.dirty);
1836
1837         BUG_ON(btree_node_fake(b));
1838         BUG_ON((b->will_make_reachable != 0) != !b->written);
1839
1840         BUG_ON(b->written >= btree_sectors(c));
1841         BUG_ON(b->written & (block_sectors(c) - 1));
1842         BUG_ON(bset_written(b, btree_bset_last(b)));
1843         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1844         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1845
1846         bch2_sort_whiteouts(c, b);
1847
1848         sort_iter_init(&sort_iter, b);
1849
1850         bytes = !b->written
1851                 ? sizeof(struct btree_node)
1852                 : sizeof(struct btree_node_entry);
1853
1854         bytes += b->whiteout_u64s * sizeof(u64);
1855
1856         for_each_bset(b, t) {
1857                 i = bset(b, t);
1858
1859                 if (bset_written(b, i))
1860                         continue;
1861
1862                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1863                 sort_iter_add(&sort_iter,
1864                               btree_bkey_first(b, t),
1865                               btree_bkey_last(b, t));
1866                 seq = max(seq, le64_to_cpu(i->journal_seq));
1867         }
1868
1869         BUG_ON(b->written && !seq);
1870
1871         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1872         bytes += 8;
1873
1874         /* buffer must be a multiple of the block size */
1875         bytes = round_up(bytes, block_bytes(c));
1876
1877         data = btree_bounce_alloc(c, bytes, &used_mempool);
1878
1879         if (!b->written) {
1880                 bn = data;
1881                 *bn = *b->data;
1882                 i = &bn->keys;
1883         } else {
1884                 bne = data;
1885                 bne->keys = b->data->keys;
1886                 i = &bne->keys;
1887         }
1888
1889         i->journal_seq  = cpu_to_le64(seq);
1890         i->u64s         = 0;
1891
1892         sort_iter_add(&sort_iter,
1893                       unwritten_whiteouts_start(c, b),
1894                       unwritten_whiteouts_end(c, b));
1895         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1896
1897         b->whiteout_u64s = 0;
1898
1899         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1900         le16_add_cpu(&i->u64s, u64s);
1901
1902         set_needs_whiteout(i, false);
1903
1904         /* do we have data to write? */
1905         if (b->written && !i->u64s)
1906                 goto nowrite;
1907
1908         bytes_to_write = vstruct_end(i) - data;
1909         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1910
1911         memset(data + bytes_to_write, 0,
1912                (sectors_to_write << 9) - bytes_to_write);
1913
1914         BUG_ON(b->written + sectors_to_write > btree_sectors(c));
1915         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1916         BUG_ON(i->seq != b->data->keys.seq);
1917
1918         i->version = c->sb.version < bcachefs_metadata_version_bkey_renumber
1919                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1920                 : cpu_to_le16(c->sb.version);
1921         SET_BSET_OFFSET(i, b->written);
1922         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1923
1924         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1925                 validate_before_checksum = true;
1926
1927         /* validate_bset will be modifying: */
1928         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
1929                 validate_before_checksum = true;
1930
1931         /* if we're going to be encrypting, check metadata validity first: */
1932         if (validate_before_checksum &&
1933             validate_bset_for_write(c, b, i, sectors_to_write))
1934                 goto err;
1935
1936         ret = bset_encrypt(c, i, b->written << 9);
1937         if (bch2_fs_fatal_err_on(ret, c,
1938                         "error encrypting btree node: %i\n", ret))
1939                 goto err;
1940
1941         nonce = btree_nonce(i, b->written << 9);
1942
1943         if (bn)
1944                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1945         else
1946                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1947
1948         /* if we're not encrypting, check metadata after checksumming: */
1949         if (!validate_before_checksum &&
1950             validate_bset_for_write(c, b, i, sectors_to_write))
1951                 goto err;
1952
1953         /*
1954          * We handle btree write errors by immediately halting the journal -
1955          * after we've done that, we can't issue any subsequent btree writes
1956          * because they might have pointers to new nodes that failed to write.
1957          *
1958          * Furthermore, there's no point in doing any more btree writes because
1959          * with the journal stopped, we're never going to update the journal to
1960          * reflect that those writes were done and the data flushed from the
1961          * journal:
1962          *
1963          * Also on journal error, the pending write may have updates that were
1964          * never journalled (interior nodes, see btree_update_nodes_written()) -
1965          * it's critical that we don't do the write in that case otherwise we
1966          * will have updates visible that weren't in the journal:
1967          *
1968          * Make sure to update b->written so bch2_btree_init_next() doesn't
1969          * break:
1970          */
1971         if (bch2_journal_error(&c->journal) ||
1972             c->opts.nochanges)
1973                 goto err;
1974
1975         trace_btree_write(b, bytes_to_write, sectors_to_write);
1976
1977         wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1978                                 buf_pages(data, sectors_to_write << 9),
1979                                 &c->btree_bio),
1980                             struct btree_write_bio, wbio.bio);
1981         wbio_init(&wbio->wbio.bio);
1982         wbio->data                      = data;
1983         wbio->data_bytes                = bytes;
1984         wbio->sector_offset             = b->written;
1985         wbio->wbio.c                    = c;
1986         wbio->wbio.used_mempool         = used_mempool;
1987         wbio->wbio.first_btree_write    = !b->written;
1988         wbio->wbio.bio.bi_opf           = REQ_OP_WRITE|REQ_META;
1989         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
1990         wbio->wbio.bio.bi_private       = b;
1991
1992         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
1993
1994         bkey_copy(&wbio->key, &b->key);
1995
1996         b->written += sectors_to_write;
1997
1998         if (wbio->wbio.first_btree_write &&
1999             b->key.k.type == KEY_TYPE_btree_ptr_v2)
2000                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
2001                         cpu_to_le16(b->written);
2002
2003         if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
2004                 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2005                         cpu_to_le16(b->written);
2006
2007         atomic64_inc(&c->btree_writes_nr);
2008         atomic64_add(sectors_to_write, &c->btree_writes_sectors);
2009
2010         INIT_WORK(&wbio->work, btree_write_submit);
2011         queue_work(c->io_complete_wq, &wbio->work);
2012         return;
2013 err:
2014         set_btree_node_noevict(b);
2015         if (!b->written &&
2016             b->key.k.type == KEY_TYPE_btree_ptr_v2)
2017                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
2018                         cpu_to_le16(sectors_to_write);
2019         b->written += sectors_to_write;
2020 nowrite:
2021         btree_bounce_free(c, bytes, used_mempool, data);
2022         __btree_node_write_done(c, b);
2023 }
2024
2025 /*
2026  * Work that must be done with write lock held:
2027  */
2028 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2029 {
2030         bool invalidated_iter = false;
2031         struct btree_node_entry *bne;
2032         struct bset_tree *t;
2033
2034         if (!btree_node_just_written(b))
2035                 return false;
2036
2037         BUG_ON(b->whiteout_u64s);
2038
2039         clear_btree_node_just_written(b);
2040
2041         /*
2042          * Note: immediately after write, bset_written() doesn't work - the
2043          * amount of data we had to write after compaction might have been
2044          * smaller than the offset of the last bset.
2045          *
2046          * However, we know that all bsets have been written here, as long as
2047          * we're still holding the write lock:
2048          */
2049
2050         /*
2051          * XXX: decide if we really want to unconditionally sort down to a
2052          * single bset:
2053          */
2054         if (b->nsets > 1) {
2055                 btree_node_sort(c, b, 0, b->nsets, true);
2056                 invalidated_iter = true;
2057         } else {
2058                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2059         }
2060
2061         for_each_bset(b, t)
2062                 set_needs_whiteout(bset(b, t), true);
2063
2064         bch2_btree_verify(c, b);
2065
2066         /*
2067          * If later we don't unconditionally sort down to a single bset, we have
2068          * to ensure this is still true:
2069          */
2070         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2071
2072         bne = want_new_bset(c, b);
2073         if (bne)
2074                 bch2_bset_init_next(c, b, bne);
2075
2076         bch2_btree_build_aux_trees(b);
2077
2078         return invalidated_iter;
2079 }
2080
2081 /*
2082  * Use this one if the node is intent locked:
2083  */
2084 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2085                            enum six_lock_type lock_type_held,
2086                            unsigned flags)
2087 {
2088         if (lock_type_held == SIX_LOCK_intent ||
2089             (lock_type_held == SIX_LOCK_read &&
2090              six_lock_tryupgrade(&b->c.lock))) {
2091                 __bch2_btree_node_write(c, b, flags);
2092
2093                 /* don't cycle lock unnecessarily: */
2094                 if (btree_node_just_written(b) &&
2095                     six_trylock_write(&b->c.lock)) {
2096                         bch2_btree_post_write_cleanup(c, b);
2097                         six_unlock_write(&b->c.lock);
2098                 }
2099
2100                 if (lock_type_held == SIX_LOCK_read)
2101                         six_lock_downgrade(&b->c.lock);
2102         } else {
2103                 __bch2_btree_node_write(c, b, flags);
2104                 if (lock_type_held == SIX_LOCK_write &&
2105                     btree_node_just_written(b))
2106                         bch2_btree_post_write_cleanup(c, b);
2107         }
2108 }
2109
2110 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2111 {
2112         struct bucket_table *tbl;
2113         struct rhash_head *pos;
2114         struct btree *b;
2115         unsigned i;
2116         bool ret = false;
2117 restart:
2118         rcu_read_lock();
2119         for_each_cached_btree(b, c, tbl, i, pos)
2120                 if (test_bit(flag, &b->flags)) {
2121                         rcu_read_unlock();
2122                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2123                         ret = true;
2124                         goto restart;
2125                 }
2126         rcu_read_unlock();
2127
2128         return ret;
2129 }
2130
2131 bool bch2_btree_flush_all_reads(struct bch_fs *c)
2132 {
2133         return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2134 }
2135
2136 bool bch2_btree_flush_all_writes(struct bch_fs *c)
2137 {
2138         return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2139 }