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         prt_printf(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         prt_printf(out, "error validating btree node ");
 517         if (write)
 518                 prt_printf(out, "before write ");
 519         if (ca)
 520                 prt_printf(out, "on %s ", ca->name);
 521         prt_printf(out, "at btree ");
 522         btree_pos_to_text(out, c, b);
 523
 524         prt_printf(out, "\n  node offset %u", b->written);
 525         if (i)
 526                 prt_printf(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         prt_printf(&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_ERR_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_ERR_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_ERR_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                 }
 620
 621                 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
 622                         if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
 623                                 break;
 624
 625                 if (k != vstruct_last(i)) {
 626                         i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
 627                         set_btree_bset_end(b, t);
 628                 }
 629         }
 630
 631         /*
 632          * Always rebuild search trees: eytzinger search tree nodes directly
 633          * depend on the values of min/max key:
 634          */
 635         bch2_bset_set_no_aux_tree(b, b->set);
 636         bch2_btree_build_aux_trees(b);
 637
 638         for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
 639                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 640                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 641         }
 642 }
 643
 644 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 645                          struct btree *b, struct bset *i,
 646                          unsigned offset, unsigned sectors,
 647                          int write, bool have_retry)
 648 {
 649         unsigned version = le16_to_cpu(i->version);
 650         const char *err;
 651         struct printbuf buf1 = PRINTBUF;
 652         struct printbuf buf2 = PRINTBUF;
 653         int ret = 0;
 654
 655         btree_err_on((version != BCH_BSET_VERSION_OLD &&
 656                       version < bcachefs_metadata_version_min) ||
 657                      version >= bcachefs_metadata_version_max,
 658                      BTREE_ERR_FATAL, c, ca, b, i,
 659                      "unsupported bset version");
 660
 661         if (btree_err_on(version < c->sb.version_min,
 662                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 663                          "bset version %u older than superblock version_min %u",
 664                          version, c->sb.version_min)) {
 665                 mutex_lock(&c->sb_lock);
 666                 c->disk_sb.sb->version_min = cpu_to_le16(version);
 667                 bch2_write_super(c);
 668                 mutex_unlock(&c->sb_lock);
 669         }
 670
 671         if (btree_err_on(version > c->sb.version,
 672                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 673                          "bset version %u newer than superblock version %u",
 674                          version, c->sb.version)) {
 675                 mutex_lock(&c->sb_lock);
 676                 c->disk_sb.sb->version = cpu_to_le16(version);
 677                 bch2_write_super(c);
 678                 mutex_unlock(&c->sb_lock);
 679         }
 680
 681         btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
 682                      BTREE_ERR_FATAL, c, ca, b, i,
 683                      "BSET_SEPARATE_WHITEOUTS no longer supported");
 684
 685         if (btree_err_on(offset + sectors > btree_sectors(c),
 686                          BTREE_ERR_FIXABLE, c, ca, b, i,
 687                          "bset past end of btree node")) {
 688                 i->u64s = 0;
 689                 ret = 0;
 690                 goto out;
 691         }
 692
 693         btree_err_on(offset && !i->u64s,
 694                      BTREE_ERR_FIXABLE, c, ca, b, i,
 695                      "empty bset");
 696
 697         btree_err_on(BSET_OFFSET(i) &&
 698                      BSET_OFFSET(i) != offset,
 699                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 700                      "bset at wrong sector offset");
 701
 702         if (!offset) {
 703                 struct btree_node *bn =
 704                         container_of(i, struct btree_node, keys);
 705                 /* These indicate that we read the wrong btree node: */
 706
 707                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 708                         struct bch_btree_ptr_v2 *bp =
 709                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 710
 711                         /* XXX endianness */
 712                         btree_err_on(bp->seq != bn->keys.seq,
 713                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 714                                      "incorrect sequence number (wrong btree node)");
 715                 }
 716
 717                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 718                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 719                              "incorrect btree id");
 720
 721                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 722                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 723                              "incorrect level");
 724
 725                 if (!write)
 726                         compat_btree_node(b->c.level, b->c.btree_id, version,
 727                                           BSET_BIG_ENDIAN(i), write, bn);
 728
 729                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 730                         struct bch_btree_ptr_v2 *bp =
 731                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 732
 733                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 734                                 b->data->min_key = bp->min_key;
 735                                 b->data->max_key = b->key.k.p;
 736                         }
 737
 738                         btree_err_on(bpos_cmp(b->data->min_key, bp->min_key),
 739                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 740                                      "incorrect min_key: got %s should be %s",
 741                                      (printbuf_reset(&buf1),
 742                                       bch2_bpos_to_text(&buf1, bn->min_key), buf1.buf),
 743                                      (printbuf_reset(&buf2),
 744                                       bch2_bpos_to_text(&buf2, bp->min_key), buf2.buf));
 745                 }
 746
 747                 btree_err_on(bpos_cmp(bn->max_key, b->key.k.p),
 748                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 749                              "incorrect max key %s",
 750                              (printbuf_reset(&buf1),
 751                               bch2_bpos_to_text(&buf1, bn->max_key), buf1.buf));
 752
 753                 if (write)
 754                         compat_btree_node(b->c.level, b->c.btree_id, version,
 755                                           BSET_BIG_ENDIAN(i), write, bn);
 756
 757                 err = bch2_bkey_format_validate(&bn->format);
 758                 btree_err_on(err,
 759                              BTREE_ERR_FATAL, c, ca, b, i,
 760                              "invalid bkey format: %s", err);
 761
 762                 compat_bformat(b->c.level, b->c.btree_id, version,
 763                                BSET_BIG_ENDIAN(i), write,
 764                                &bn->format);
 765         }
 766 out:
 767 fsck_err:
 768         printbuf_exit(&buf2);
 769         printbuf_exit(&buf1);
 770         return ret;
 771 }
 772
 773 static int bset_key_invalid(struct bch_fs *c, struct btree *b,
 774                             struct bkey_s_c k,
 775                             bool updated_range, int rw,
 776                             struct printbuf *err)
 777 {
 778         return __bch2_bkey_invalid(c, k, btree_node_type(b), READ, err) ?:
 779                 (!updated_range ? bch2_bkey_in_btree_node(b, k, err) : 0) ?:
 780                 (rw == WRITE ? bch2_bkey_val_invalid(c, k, READ, err) : 0);
 781 }
 782
 783 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 784                          struct bset *i, int write, bool have_retry)
 785 {
 786         unsigned version = le16_to_cpu(i->version);
 787         struct bkey_packed *k, *prev = NULL;
 788         struct printbuf buf = PRINTBUF;
 789         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 790                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 791         int ret = 0;
 792
 793         for (k = i->start;
 794              k != vstruct_last(i);) {
 795                 struct bkey_s u;
 796                 struct bkey tmp;
 797
 798                 if (btree_err_on(bkey_next(k) > vstruct_last(i),
 799                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 800                                  "key extends past end of bset")) {
 801                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 802                         break;
 803                 }
 804
 805                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 806                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 807                                  "invalid bkey format %u", k->format)) {
 808                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 809                         memmove_u64s_down(k, bkey_next(k),
 810                                           (u64 *) vstruct_end(i) - (u64 *) k);
 811                         continue;
 812                 }
 813
 814                 /* XXX: validate k->u64s */
 815                 if (!write)
 816                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 817                                     BSET_BIG_ENDIAN(i), write,
 818                                     &b->format, k);
 819
 820                 u = __bkey_disassemble(b, k, &tmp);
 821
 822                 printbuf_reset(&buf);
 823                 if (bset_key_invalid(c, b, u.s_c, updated_range, write, &buf)) {
 824                         printbuf_reset(&buf);
 825                         prt_printf(&buf, "invalid bkey:  ");
 826                         bset_key_invalid(c, b, u.s_c, updated_range, write, &buf);
 827                         prt_printf(&buf, "\n  ");
 828                         bch2_bkey_val_to_text(&buf, c, u.s_c);
 829
 830                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf);
 831
 832                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 833                         memmove_u64s_down(k, bkey_next(k),
 834                                           (u64 *) vstruct_end(i) - (u64 *) k);
 835                         continue;
 836                 }
 837
 838                 if (write)
 839                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 840                                     BSET_BIG_ENDIAN(i), write,
 841                                     &b->format, k);
 842
 843                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 844                         struct bkey up = bkey_unpack_key(b, prev);
 845
 846                         printbuf_reset(&buf);
 847                         prt_printf(&buf, "keys out of order: ");
 848                         bch2_bkey_to_text(&buf, &up);
 849                         prt_printf(&buf, " > ");
 850                         bch2_bkey_to_text(&buf, u.k);
 851
 852                         bch2_dump_bset(c, b, i, 0);
 853
 854                         if (btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf)) {
 855                                 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 856                                 memmove_u64s_down(k, bkey_next(k),
 857                                                   (u64 *) vstruct_end(i) - (u64 *) k);
 858                                 continue;
 859                         }
 860                 }
 861
 862                 prev = k;
 863                 k = bkey_next(k);
 864         }
 865 fsck_err:
 866         printbuf_exit(&buf);
 867         return ret;
 868 }
 869
 870 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 871                               struct btree *b, bool have_retry)
 872 {
 873         struct btree_node_entry *bne;
 874         struct sort_iter *iter;
 875         struct btree_node *sorted;
 876         struct bkey_packed *k;
 877         struct bch_extent_ptr *ptr;
 878         struct bset *i;
 879         bool used_mempool, blacklisted;
 880         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 881                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 882         unsigned u64s;
 883         unsigned blacklisted_written, nonblacklisted_written = 0;
 884         unsigned ptr_written = btree_ptr_sectors_written(&b->key);
 885         struct printbuf buf = PRINTBUF;
 886         int ret, retry_read = 0, write = READ;
 887
 888         b->version_ondisk = U16_MAX;
 889         /* We might get called multiple times on read retry: */
 890         b->written = 0;
 891
 892         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
 893         sort_iter_init(iter, b);
 894         iter->size = (btree_blocks(c) + 1) * 2;
 895
 896         if (bch2_meta_read_fault("btree"))
 897                 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 898                           "dynamic fault");
 899
 900         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 901                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 902                      "bad magic: want %llx, got %llx",
 903                      bset_magic(c), le64_to_cpu(b->data->magic));
 904
 905         btree_err_on(!b->data->keys.seq,
 906                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 907                      "bad btree header: seq 0");
 908
 909         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 910                 struct bch_btree_ptr_v2 *bp =
 911                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 912
 913                 btree_err_on(b->data->keys.seq != bp->seq,
 914                              BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 915                              "got wrong btree node (seq %llx want %llx)",
 916                              b->data->keys.seq, bp->seq);
 917         }
 918
 919         while (b->written < (ptr_written ?: btree_sectors(c))) {
 920                 unsigned sectors;
 921                 struct nonce nonce;
 922                 struct bch_csum csum;
 923                 bool first = !b->written;
 924
 925                 if (!b->written) {
 926                         i = &b->data->keys;
 927
 928                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 929                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 930                                      "unknown checksum type %llu",
 931                                      BSET_CSUM_TYPE(i));
 932
 933                         nonce = btree_nonce(i, b->written << 9);
 934                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 935
 936                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 937                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 938                                      "invalid checksum");
 939
 940                         ret = bset_encrypt(c, i, b->written << 9);
 941                         if (bch2_fs_fatal_err_on(ret, c,
 942                                         "error decrypting btree node: %i", ret))
 943                                 goto fsck_err;
 944
 945                         btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
 946                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
 947                                      BTREE_ERR_FATAL, c, NULL, b, NULL,
 948                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
 949
 950                         sectors = vstruct_sectors(b->data, c->block_bits);
 951                 } else {
 952                         bne = write_block(b);
 953                         i = &bne->keys;
 954
 955                         if (i->seq != b->data->keys.seq)
 956                                 break;
 957
 958                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 959                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 960                                      "unknown checksum type %llu",
 961                                      BSET_CSUM_TYPE(i));
 962
 963                         nonce = btree_nonce(i, b->written << 9);
 964                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
 965
 966                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
 967                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 968                                      "invalid checksum");
 969
 970                         ret = bset_encrypt(c, i, b->written << 9);
 971                         if (bch2_fs_fatal_err_on(ret, c,
 972                                         "error decrypting btree node: %i\n", ret))
 973                                 goto fsck_err;
 974
 975                         sectors = vstruct_sectors(bne, c->block_bits);
 976                 }
 977
 978                 b->version_ondisk = min(b->version_ondisk,
 979                                         le16_to_cpu(i->version));
 980
 981                 ret = validate_bset(c, ca, b, i, b->written, sectors,
 982                                     READ, have_retry);
 983                 if (ret)
 984                         goto fsck_err;
 985
 986                 if (!b->written)
 987                         btree_node_set_format(b, b->data->format);
 988
 989                 ret = validate_bset_keys(c, b, i, READ, have_retry);
 990                 if (ret)
 991                         goto fsck_err;
 992
 993                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 994
 995                 blacklisted = bch2_journal_seq_is_blacklisted(c,
 996                                         le64_to_cpu(i->journal_seq),
 997                                         true);
 998
 999                 btree_err_on(blacklisted && first,
1000                              BTREE_ERR_FIXABLE, c, ca, b, i,
1001                              "first btree node bset has blacklisted journal seq (%llu)",
1002                              le64_to_cpu(i->journal_seq));
1003
1004                 btree_err_on(blacklisted && ptr_written,
1005                              BTREE_ERR_FIXABLE, c, ca, b, i,
1006                              "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
1007                              le64_to_cpu(i->journal_seq),
1008                              b->written, b->written + sectors, ptr_written);
1009
1010                 b->written += sectors;
1011
1012                 if (blacklisted && !first)
1013                         continue;
1014
1015                 sort_iter_add(iter,
1016                               vstruct_idx(i, 0),
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                         prt_printf(&buf, "invalid bkey: ");
1085                         bch2_bkey_val_invalid(c, u.s_c, READ, &buf);
1086                         prt_printf(&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, NULL, REQ_OP_READ|REQ_SYNC|REQ_META);
1159                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1160                 bio->bi_iter.bi_size    = btree_bytes(c);
1161
1162                 if (rb->have_ioref) {
1163                         bio_set_dev(bio, ca->disk_sb.bdev);
1164                         submit_bio_wait(bio);
1165                 } else {
1166                         bio->bi_status = BLK_STS_REMOVED;
1167                 }
1168 start:
1169                 printbuf_reset(&buf);
1170                 btree_pos_to_text(&buf, c, b);
1171                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1172                                    bch2_blk_status_to_str(bio->bi_status), buf.buf);
1173                 if (rb->have_ioref)
1174                         percpu_ref_put(&ca->io_ref);
1175                 rb->have_ioref = false;
1176
1177                 bch2_mark_io_failure(&failed, &rb->pick);
1178
1179                 can_retry = bch2_bkey_pick_read_device(c,
1180                                 bkey_i_to_s_c(&b->key),
1181                                 &failed, &rb->pick) > 0;
1182
1183                 if (!bio->bi_status &&
1184                     !bch2_btree_node_read_done(c, ca, b, can_retry)) {
1185                         if (retry)
1186                                 bch_info(c, "retry success");
1187                         break;
1188                 }
1189
1190                 saw_error = true;
1191
1192                 if (!can_retry) {
1193                         set_btree_node_read_error(b);
1194                         break;
1195                 }
1196         }
1197
1198         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1199                                rb->start_time);
1200         bio_put(&rb->bio);
1201         printbuf_exit(&buf);
1202
1203         if (saw_error && !btree_node_read_error(b))
1204                 bch2_btree_node_rewrite_async(c, b);
1205
1206         clear_btree_node_read_in_flight(b);
1207         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1208 }
1209
1210 static void btree_node_read_endio(struct bio *bio)
1211 {
1212         struct btree_read_bio *rb =
1213                 container_of(bio, struct btree_read_bio, bio);
1214         struct bch_fs *c        = rb->c;
1215
1216         if (rb->have_ioref) {
1217                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1218                 bch2_latency_acct(ca, rb->start_time, READ);
1219         }
1220
1221         queue_work(c->io_complete_wq, &rb->work);
1222 }
1223
1224 struct btree_node_read_all {
1225         struct closure          cl;
1226         struct bch_fs           *c;
1227         struct btree            *b;
1228         unsigned                nr;
1229         void                    *buf[BCH_REPLICAS_MAX];
1230         struct bio              *bio[BCH_REPLICAS_MAX];
1231         int                     err[BCH_REPLICAS_MAX];
1232 };
1233
1234 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1235 {
1236         struct btree_node *bn = data;
1237         struct btree_node_entry *bne;
1238         unsigned offset = 0;
1239
1240         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1241                 return 0;
1242
1243         while (offset < btree_sectors(c)) {
1244                 if (!offset) {
1245                         offset += vstruct_sectors(bn, c->block_bits);
1246                 } else {
1247                         bne = data + (offset << 9);
1248                         if (bne->keys.seq != bn->keys.seq)
1249                                 break;
1250                         offset += vstruct_sectors(bne, c->block_bits);
1251                 }
1252         }
1253
1254         return offset;
1255 }
1256
1257 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1258 {
1259         struct btree_node *bn = data;
1260         struct btree_node_entry *bne;
1261
1262         if (!offset)
1263                 return false;
1264
1265         while (offset < btree_sectors(c)) {
1266                 bne = data + (offset << 9);
1267                 if (bne->keys.seq == bn->keys.seq)
1268                         return true;
1269                 offset++;
1270         }
1271
1272         return false;
1273         return offset;
1274 }
1275
1276 static void btree_node_read_all_replicas_done(struct closure *cl)
1277 {
1278         struct btree_node_read_all *ra =
1279                 container_of(cl, struct btree_node_read_all, cl);
1280         struct bch_fs *c = ra->c;
1281         struct btree *b = ra->b;
1282         struct printbuf buf = PRINTBUF;
1283         bool dump_bset_maps = false;
1284         bool have_retry = false;
1285         int ret = 0, best = -1, write = READ;
1286         unsigned i, written = 0, written2 = 0;
1287         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1288                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1289
1290         for (i = 0; i < ra->nr; i++) {
1291                 struct btree_node *bn = ra->buf[i];
1292
1293                 if (ra->err[i])
1294                         continue;
1295
1296                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1297                     (seq && seq != bn->keys.seq))
1298                         continue;
1299
1300                 if (best < 0) {
1301                         best = i;
1302                         written = btree_node_sectors_written(c, bn);
1303                         continue;
1304                 }
1305
1306                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1307                 if (btree_err_on(written2 != written, BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1308                                  "btree node sectors written mismatch: %u != %u",
1309                                  written, written2) ||
1310                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1311                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1312                                  "found bset signature after last bset") ||
1313                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1314                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1315                                  "btree node replicas content mismatch"))
1316                         dump_bset_maps = true;
1317
1318                 if (written2 > written) {
1319                         written = written2;
1320                         best = i;
1321                 }
1322         }
1323 fsck_err:
1324         if (dump_bset_maps) {
1325                 for (i = 0; i < ra->nr; i++) {
1326                         struct btree_node *bn = ra->buf[i];
1327                         struct btree_node_entry *bne = NULL;
1328                         unsigned offset = 0, sectors;
1329                         bool gap = false;
1330
1331                         if (ra->err[i])
1332                                 continue;
1333
1334                         printbuf_reset(&buf);
1335
1336                         while (offset < btree_sectors(c)) {
1337                                 if (!offset) {
1338                                         sectors = vstruct_sectors(bn, c->block_bits);
1339                                 } else {
1340                                         bne = ra->buf[i] + (offset << 9);
1341                                         if (bne->keys.seq != bn->keys.seq)
1342                                                 break;
1343                                         sectors = vstruct_sectors(bne, c->block_bits);
1344                                 }
1345
1346                                 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1347                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1348                                                         le64_to_cpu(bne->keys.journal_seq), false))
1349                                         prt_printf(&buf, "*");
1350                                 offset += sectors;
1351                         }
1352
1353                         while (offset < btree_sectors(c)) {
1354                                 bne = ra->buf[i] + (offset << 9);
1355                                 if (bne->keys.seq == bn->keys.seq) {
1356                                         if (!gap)
1357                                                 prt_printf(&buf, " GAP");
1358                                         gap = true;
1359
1360                                         sectors = vstruct_sectors(bne, c->block_bits);
1361                                         prt_printf(&buf, " %u-%u", offset, offset + sectors);
1362                                         if (bch2_journal_seq_is_blacklisted(c,
1363                                                         le64_to_cpu(bne->keys.journal_seq), false))
1364                                                 prt_printf(&buf, "*");
1365                                 }
1366                                 offset++;
1367                         }
1368
1369                         bch_err(c, "replica %u:%s", i, buf.buf);
1370                 }
1371         }
1372
1373         if (best >= 0) {
1374                 memcpy(b->data, ra->buf[best], btree_bytes(c));
1375                 ret = bch2_btree_node_read_done(c, NULL, b, false);
1376         } else {
1377                 ret = -1;
1378         }
1379
1380         if (ret)
1381                 set_btree_node_read_error(b);
1382
1383         for (i = 0; i < ra->nr; i++) {
1384                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1385                 bio_put(ra->bio[i]);
1386         }
1387
1388         closure_debug_destroy(&ra->cl);
1389         kfree(ra);
1390         printbuf_exit(&buf);
1391
1392         clear_btree_node_read_in_flight(b);
1393         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1394 }
1395
1396 static void btree_node_read_all_replicas_endio(struct bio *bio)
1397 {
1398         struct btree_read_bio *rb =
1399                 container_of(bio, struct btree_read_bio, bio);
1400         struct bch_fs *c        = rb->c;
1401         struct btree_node_read_all *ra = rb->ra;
1402
1403         if (rb->have_ioref) {
1404                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1405                 bch2_latency_acct(ca, rb->start_time, READ);
1406         }
1407
1408         ra->err[rb->idx] = bio->bi_status;
1409         closure_put(&ra->cl);
1410 }
1411
1412 /*
1413  * XXX This allocates multiple times from the same mempools, and can deadlock
1414  * under sufficient memory pressure (but is only a debug path)
1415  */
1416 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1417 {
1418         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1419         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1420         const union bch_extent_entry *entry;
1421         struct extent_ptr_decoded pick;
1422         struct btree_node_read_all *ra;
1423         unsigned i;
1424
1425         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1426         if (!ra)
1427                 return -ENOMEM;
1428
1429         closure_init(&ra->cl, NULL);
1430         ra->c   = c;
1431         ra->b   = b;
1432         ra->nr  = bch2_bkey_nr_ptrs(k);
1433
1434         for (i = 0; i < ra->nr; i++) {
1435                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1436                 ra->bio[i] = bio_alloc_bioset(NULL,
1437                                               buf_pages(ra->buf[i], btree_bytes(c)),
1438                                               REQ_OP_READ|REQ_SYNC|REQ_META,
1439                                               GFP_NOFS,
1440                                               &c->btree_bio);
1441         }
1442
1443         i = 0;
1444         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1445                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1446                 struct btree_read_bio *rb =
1447                         container_of(ra->bio[i], struct btree_read_bio, bio);
1448                 rb->c                   = c;
1449                 rb->b                   = b;
1450                 rb->ra                  = ra;
1451                 rb->start_time          = local_clock();
1452                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1453                 rb->idx                 = i;
1454                 rb->pick                = pick;
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         int ret;
1492
1493         trace_and_count(c, btree_node_read, c, b);
1494
1495         if (bch2_verify_all_btree_replicas &&
1496             !btree_node_read_all_replicas(c, b, sync))
1497                 return;
1498
1499         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1500                                          NULL, &pick);
1501
1502         if (ret <= 0) {
1503                 struct printbuf buf = PRINTBUF;
1504
1505                 prt_str(&buf, "btree node read error: no device to read from\n at ");
1506                 btree_pos_to_text(&buf, c, b);
1507                 bch_err(c, "%s", buf.buf);
1508
1509                 if (test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags))
1510                         bch2_fatal_error(c);
1511
1512                 set_btree_node_read_error(b);
1513                 clear_btree_node_read_in_flight(b);
1514                 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1515                 printbuf_exit(&buf);
1516                 return;
1517         }
1518
1519         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1520
1521         bio = bio_alloc_bioset(NULL,
1522                                buf_pages(b->data, btree_bytes(c)),
1523                                REQ_OP_READ|REQ_SYNC|REQ_META,
1524                                GFP_NOIO,
1525                                &c->btree_bio);
1526         rb = container_of(bio, struct btree_read_bio, bio);
1527         rb->c                   = c;
1528         rb->b                   = b;
1529         rb->ra                  = NULL;
1530         rb->start_time          = local_clock();
1531         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1532         rb->pick                = pick;
1533         INIT_WORK(&rb->work, btree_node_read_work);
1534         bio->bi_iter.bi_sector  = pick.ptr.offset;
1535         bio->bi_end_io          = btree_node_read_endio;
1536         bch2_bio_map(bio, b->data, btree_bytes(c));
1537
1538         if (rb->have_ioref) {
1539                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1540                              bio_sectors(bio));
1541                 bio_set_dev(bio, ca->disk_sb.bdev);
1542
1543                 if (sync) {
1544                         submit_bio_wait(bio);
1545
1546                         btree_node_read_work(&rb->work);
1547                 } else {
1548                         submit_bio(bio);
1549                 }
1550         } else {
1551                 bio->bi_status = BLK_STS_REMOVED;
1552
1553                 if (sync)
1554                         btree_node_read_work(&rb->work);
1555                 else
1556                         queue_work(c->io_complete_wq, &rb->work);
1557         }
1558 }
1559
1560 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1561                         const struct bkey_i *k, unsigned level)
1562 {
1563         struct closure cl;
1564         struct btree *b;
1565         int ret;
1566
1567         closure_init_stack(&cl);
1568
1569         do {
1570                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1571                 closure_sync(&cl);
1572         } while (ret);
1573
1574         b = bch2_btree_node_mem_alloc(c, level != 0);
1575         bch2_btree_cache_cannibalize_unlock(c);
1576
1577         BUG_ON(IS_ERR(b));
1578
1579         bkey_copy(&b->key, k);
1580         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1581
1582         set_btree_node_read_in_flight(b);
1583
1584         bch2_btree_node_read(c, b, true);
1585
1586         if (btree_node_read_error(b)) {
1587                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1588
1589                 mutex_lock(&c->btree_cache.lock);
1590                 list_move(&b->list, &c->btree_cache.freeable);
1591                 mutex_unlock(&c->btree_cache.lock);
1592
1593                 ret = -EIO;
1594                 goto err;
1595         }
1596
1597         bch2_btree_set_root_for_read(c, b);
1598 err:
1599         six_unlock_write(&b->c.lock);
1600         six_unlock_intent(&b->c.lock);
1601
1602         return ret;
1603 }
1604
1605 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1606                               struct btree_write *w)
1607 {
1608         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1609
1610         do {
1611                 old = new = v;
1612                 if (!(old & 1))
1613                         break;
1614
1615                 new &= ~1UL;
1616         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1617
1618         if (old & 1)
1619                 closure_put(&((struct btree_update *) new)->cl);
1620
1621         bch2_journal_pin_drop(&c->journal, &w->journal);
1622 }
1623
1624 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1625 {
1626         struct btree_write *w = btree_prev_write(b);
1627         unsigned long old, new, v;
1628
1629         bch2_btree_complete_write(c, b, w);
1630
1631         v = READ_ONCE(b->flags);
1632         do {
1633                 old = new = v;
1634
1635                 if ((old & (1U << BTREE_NODE_dirty)) &&
1636                     (old & (1U << BTREE_NODE_need_write)) &&
1637                     !(old & (1U << BTREE_NODE_never_write)) &&
1638                     !(old & (1U << BTREE_NODE_write_blocked)) &&
1639                     !(old & (1U << BTREE_NODE_will_make_reachable))) {
1640                         new &= ~(1U << BTREE_NODE_dirty);
1641                         new &= ~(1U << BTREE_NODE_need_write);
1642                         new |=  (1U << BTREE_NODE_write_in_flight);
1643                         new |=  (1U << BTREE_NODE_write_in_flight_inner);
1644                         new |=  (1U << BTREE_NODE_just_written);
1645                         new ^=  (1U << BTREE_NODE_write_idx);
1646                 } else {
1647                         new &= ~(1U << BTREE_NODE_write_in_flight);
1648                         new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1649                 }
1650         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1651
1652         if (new & (1U << BTREE_NODE_write_in_flight))
1653                 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED);
1654         else
1655                 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1656 }
1657
1658 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1659 {
1660         struct btree_trans trans;
1661
1662         bch2_trans_init(&trans, c, 0, 0);
1663
1664         btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
1665         __btree_node_write_done(c, b);
1666         six_unlock_read(&b->c.lock);
1667
1668         bch2_trans_exit(&trans);
1669 }
1670
1671 static void btree_node_write_work(struct work_struct *work)
1672 {
1673         struct btree_write_bio *wbio =
1674                 container_of(work, struct btree_write_bio, work);
1675         struct bch_fs *c        = wbio->wbio.c;
1676         struct btree *b         = wbio->wbio.bio.bi_private;
1677         struct bch_extent_ptr *ptr;
1678         int ret;
1679
1680         btree_bounce_free(c,
1681                 wbio->data_bytes,
1682                 wbio->wbio.used_mempool,
1683                 wbio->data);
1684
1685         bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1686                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1687
1688         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key)))
1689                 goto err;
1690
1691         if (wbio->wbio.first_btree_write) {
1692                 if (wbio->wbio.failed.nr) {
1693
1694                 }
1695         } else {
1696                 ret = bch2_trans_do(c, NULL, NULL, 0,
1697                         bch2_btree_node_update_key_get_iter(&trans, b, &wbio->key,
1698                                                             !wbio->wbio.failed.nr));
1699                 if (ret)
1700                         goto err;
1701         }
1702 out:
1703         bio_put(&wbio->wbio.bio);
1704         btree_node_write_done(c, b);
1705         return;
1706 err:
1707         set_btree_node_noevict(b);
1708         bch2_fs_fatal_error(c, "fatal error writing btree node");
1709         goto out;
1710 }
1711
1712 static void btree_node_write_endio(struct bio *bio)
1713 {
1714         struct bch_write_bio *wbio      = to_wbio(bio);
1715         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1716         struct bch_write_bio *orig      = parent ?: wbio;
1717         struct btree_write_bio *wb      = container_of(orig, struct btree_write_bio, wbio);
1718         struct bch_fs *c                = wbio->c;
1719         struct btree *b                 = wbio->bio.bi_private;
1720         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1721         unsigned long flags;
1722
1723         if (wbio->have_ioref)
1724                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1725
1726         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1727                                bch2_blk_status_to_str(bio->bi_status)) ||
1728             bch2_meta_write_fault("btree")) {
1729                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1730                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1731                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1732         }
1733
1734         if (wbio->have_ioref)
1735                 percpu_ref_put(&ca->io_ref);
1736
1737         if (parent) {
1738                 bio_put(bio);
1739                 bio_endio(&parent->bio);
1740                 return;
1741         }
1742
1743         clear_btree_node_write_in_flight_inner(b);
1744         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1745         INIT_WORK(&wb->work, btree_node_write_work);
1746         queue_work(c->btree_io_complete_wq, &wb->work);
1747 }
1748
1749 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1750                                    struct bset *i, unsigned sectors)
1751 {
1752         struct printbuf buf = PRINTBUF;
1753         int ret;
1754
1755         ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key),
1756                                 BKEY_TYPE_btree, WRITE, &buf);
1757
1758         if (ret)
1759                 bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf);
1760         printbuf_exit(&buf);
1761         if (ret)
1762                 return ret;
1763
1764         ret = validate_bset_keys(c, b, i, WRITE, false) ?:
1765                 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false);
1766         if (ret) {
1767                 bch2_inconsistent_error(c);
1768                 dump_stack();
1769         }
1770
1771         return ret;
1772 }
1773
1774 static void btree_write_submit(struct work_struct *work)
1775 {
1776         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1777         struct bch_extent_ptr *ptr;
1778         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1779
1780         bkey_copy(&tmp.k, &wbio->key);
1781
1782         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1783                 ptr->offset += wbio->sector_offset;
1784
1785         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree, &tmp.k);
1786 }
1787
1788 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1789 {
1790         struct btree_write_bio *wbio;
1791         struct bset_tree *t;
1792         struct bset *i;
1793         struct btree_node *bn = NULL;
1794         struct btree_node_entry *bne = NULL;
1795         struct sort_iter sort_iter;
1796         struct nonce nonce;
1797         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1798         u64 seq = 0;
1799         bool used_mempool;
1800         unsigned long old, new;
1801         bool validate_before_checksum = false;
1802         void *data;
1803         int ret;
1804
1805         if (flags & BTREE_WRITE_ALREADY_STARTED)
1806                 goto do_write;
1807
1808         /*
1809          * We may only have a read lock on the btree node - the dirty bit is our
1810          * "lock" against racing with other threads that may be trying to start
1811          * a write, we do a write iff we clear the dirty bit. Since setting the
1812          * dirty bit requires a write lock, we can't race with other threads
1813          * redirtying it:
1814          */
1815         do {
1816                 old = new = READ_ONCE(b->flags);
1817
1818                 if (!(old & (1 << BTREE_NODE_dirty)))
1819                         return;
1820
1821                 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
1822                     !(old & (1 << BTREE_NODE_need_write)))
1823                         return;
1824
1825                 if (old &
1826                     ((1 << BTREE_NODE_never_write)|
1827                      (1 << BTREE_NODE_write_blocked)))
1828                         return;
1829
1830                 if (b->written &&
1831                     (old & (1 << BTREE_NODE_will_make_reachable)))
1832                         return;
1833
1834                 if (old & (1 << BTREE_NODE_write_in_flight))
1835                         return;
1836
1837                 new &= ~(1 << BTREE_NODE_dirty);
1838                 new &= ~(1 << BTREE_NODE_need_write);
1839                 new |=  (1 << BTREE_NODE_write_in_flight);
1840                 new |=  (1 << BTREE_NODE_write_in_flight_inner);
1841                 new |=  (1 << BTREE_NODE_just_written);
1842                 new ^=  (1 << BTREE_NODE_write_idx);
1843         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1844
1845         if (new & (1U << BTREE_NODE_need_write))
1846                 return;
1847 do_write:
1848         atomic_dec(&c->btree_cache.dirty);
1849
1850         BUG_ON(btree_node_fake(b));
1851         BUG_ON((b->will_make_reachable != 0) != !b->written);
1852
1853         BUG_ON(b->written >= btree_sectors(c));
1854         BUG_ON(b->written & (block_sectors(c) - 1));
1855         BUG_ON(bset_written(b, btree_bset_last(b)));
1856         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1857         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1858
1859         bch2_sort_whiteouts(c, b);
1860
1861         sort_iter_init(&sort_iter, b);
1862
1863         bytes = !b->written
1864                 ? sizeof(struct btree_node)
1865                 : sizeof(struct btree_node_entry);
1866
1867         bytes += b->whiteout_u64s * sizeof(u64);
1868
1869         for_each_bset(b, t) {
1870                 i = bset(b, t);
1871
1872                 if (bset_written(b, i))
1873                         continue;
1874
1875                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1876                 sort_iter_add(&sort_iter,
1877                               btree_bkey_first(b, t),
1878                               btree_bkey_last(b, t));
1879                 seq = max(seq, le64_to_cpu(i->journal_seq));
1880         }
1881
1882         BUG_ON(b->written && !seq);
1883
1884         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1885         bytes += 8;
1886
1887         /* buffer must be a multiple of the block size */
1888         bytes = round_up(bytes, block_bytes(c));
1889
1890         data = btree_bounce_alloc(c, bytes, &used_mempool);
1891
1892         if (!b->written) {
1893                 bn = data;
1894                 *bn = *b->data;
1895                 i = &bn->keys;
1896         } else {
1897                 bne = data;
1898                 bne->keys = b->data->keys;
1899                 i = &bne->keys;
1900         }
1901
1902         i->journal_seq  = cpu_to_le64(seq);
1903         i->u64s         = 0;
1904
1905         sort_iter_add(&sort_iter,
1906                       unwritten_whiteouts_start(c, b),
1907                       unwritten_whiteouts_end(c, b));
1908         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1909
1910         b->whiteout_u64s = 0;
1911
1912         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1913         le16_add_cpu(&i->u64s, u64s);
1914
1915         set_needs_whiteout(i, false);
1916
1917         /* do we have data to write? */
1918         if (b->written && !i->u64s)
1919                 goto nowrite;
1920
1921         bytes_to_write = vstruct_end(i) - data;
1922         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1923
1924         memset(data + bytes_to_write, 0,
1925                (sectors_to_write << 9) - bytes_to_write);
1926
1927         BUG_ON(b->written + sectors_to_write > btree_sectors(c));
1928         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1929         BUG_ON(i->seq != b->data->keys.seq);
1930
1931         i->version = c->sb.version < bcachefs_metadata_version_bkey_renumber
1932                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1933                 : cpu_to_le16(c->sb.version);
1934         SET_BSET_OFFSET(i, b->written);
1935         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1936
1937         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1938                 validate_before_checksum = true;
1939
1940         /* validate_bset will be modifying: */
1941         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
1942                 validate_before_checksum = true;
1943
1944         /* if we're going to be encrypting, check metadata validity first: */
1945         if (validate_before_checksum &&
1946             validate_bset_for_write(c, b, i, sectors_to_write))
1947                 goto err;
1948
1949         ret = bset_encrypt(c, i, b->written << 9);
1950         if (bch2_fs_fatal_err_on(ret, c,
1951                         "error encrypting btree node: %i\n", ret))
1952                 goto err;
1953
1954         nonce = btree_nonce(i, b->written << 9);
1955
1956         if (bn)
1957                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1958         else
1959                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1960
1961         /* if we're not encrypting, check metadata after checksumming: */
1962         if (!validate_before_checksum &&
1963             validate_bset_for_write(c, b, i, sectors_to_write))
1964                 goto err;
1965
1966         /*
1967          * We handle btree write errors by immediately halting the journal -
1968          * after we've done that, we can't issue any subsequent btree writes
1969          * because they might have pointers to new nodes that failed to write.
1970          *
1971          * Furthermore, there's no point in doing any more btree writes because
1972          * with the journal stopped, we're never going to update the journal to
1973          * reflect that those writes were done and the data flushed from the
1974          * journal:
1975          *
1976          * Also on journal error, the pending write may have updates that were
1977          * never journalled (interior nodes, see btree_update_nodes_written()) -
1978          * it's critical that we don't do the write in that case otherwise we
1979          * will have updates visible that weren't in the journal:
1980          *
1981          * Make sure to update b->written so bch2_btree_init_next() doesn't
1982          * break:
1983          */
1984         if (bch2_journal_error(&c->journal) ||
1985             c->opts.nochanges)
1986                 goto err;
1987
1988         trace_and_count(c, btree_node_write, b, bytes_to_write, sectors_to_write);
1989
1990         wbio = container_of(bio_alloc_bioset(NULL,
1991                                 buf_pages(data, sectors_to_write << 9),
1992                                 REQ_OP_WRITE|REQ_META,
1993                                 GFP_NOIO,
1994                                 &c->btree_bio),
1995                             struct btree_write_bio, wbio.bio);
1996         wbio_init(&wbio->wbio.bio);
1997         wbio->data                      = data;
1998         wbio->data_bytes                = bytes;
1999         wbio->sector_offset             = b->written;
2000         wbio->wbio.c                    = c;
2001         wbio->wbio.used_mempool         = used_mempool;
2002         wbio->wbio.first_btree_write    = !b->written;
2003         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
2004         wbio->wbio.bio.bi_private       = b;
2005
2006         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
2007
2008         bkey_copy(&wbio->key, &b->key);
2009
2010         b->written += sectors_to_write;
2011
2012         if (wbio->wbio.first_btree_write &&
2013             b->key.k.type == KEY_TYPE_btree_ptr_v2)
2014                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
2015                         cpu_to_le16(b->written);
2016
2017         if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
2018                 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2019                         cpu_to_le16(b->written);
2020
2021         atomic64_inc(&c->btree_writes_nr);
2022         atomic64_add(sectors_to_write, &c->btree_writes_sectors);
2023
2024         INIT_WORK(&wbio->work, btree_write_submit);
2025         queue_work(c->io_complete_wq, &wbio->work);
2026         return;
2027 err:
2028         set_btree_node_noevict(b);
2029         if (!b->written &&
2030             b->key.k.type == KEY_TYPE_btree_ptr_v2)
2031                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
2032                         cpu_to_le16(sectors_to_write);
2033         b->written += sectors_to_write;
2034 nowrite:
2035         btree_bounce_free(c, bytes, used_mempool, data);
2036         __btree_node_write_done(c, b);
2037 }
2038
2039 /*
2040  * Work that must be done with write lock held:
2041  */
2042 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2043 {
2044         bool invalidated_iter = false;
2045         struct btree_node_entry *bne;
2046         struct bset_tree *t;
2047
2048         if (!btree_node_just_written(b))
2049                 return false;
2050
2051         BUG_ON(b->whiteout_u64s);
2052
2053         clear_btree_node_just_written(b);
2054
2055         /*
2056          * Note: immediately after write, bset_written() doesn't work - the
2057          * amount of data we had to write after compaction might have been
2058          * smaller than the offset of the last bset.
2059          *
2060          * However, we know that all bsets have been written here, as long as
2061          * we're still holding the write lock:
2062          */
2063
2064         /*
2065          * XXX: decide if we really want to unconditionally sort down to a
2066          * single bset:
2067          */
2068         if (b->nsets > 1) {
2069                 btree_node_sort(c, b, 0, b->nsets, true);
2070                 invalidated_iter = true;
2071         } else {
2072                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2073         }
2074
2075         for_each_bset(b, t)
2076                 set_needs_whiteout(bset(b, t), true);
2077
2078         bch2_btree_verify(c, b);
2079
2080         /*
2081          * If later we don't unconditionally sort down to a single bset, we have
2082          * to ensure this is still true:
2083          */
2084         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2085
2086         bne = want_new_bset(c, b);
2087         if (bne)
2088                 bch2_bset_init_next(c, b, bne);
2089
2090         bch2_btree_build_aux_trees(b);
2091
2092         return invalidated_iter;
2093 }
2094
2095 /*
2096  * Use this one if the node is intent locked:
2097  */
2098 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2099                            enum six_lock_type lock_type_held,
2100                            unsigned flags)
2101 {
2102         if (lock_type_held == SIX_LOCK_intent ||
2103             (lock_type_held == SIX_LOCK_read &&
2104              six_lock_tryupgrade(&b->c.lock))) {
2105                 __bch2_btree_node_write(c, b, flags);
2106
2107                 /* don't cycle lock unnecessarily: */
2108                 if (btree_node_just_written(b) &&
2109                     six_trylock_write(&b->c.lock)) {
2110                         bch2_btree_post_write_cleanup(c, b);
2111                         six_unlock_write(&b->c.lock);
2112                 }
2113
2114                 if (lock_type_held == SIX_LOCK_read)
2115                         six_lock_downgrade(&b->c.lock);
2116         } else {
2117                 __bch2_btree_node_write(c, b, flags);
2118                 if (lock_type_held == SIX_LOCK_write &&
2119                     btree_node_just_written(b))
2120                         bch2_btree_post_write_cleanup(c, b);
2121         }
2122 }
2123
2124 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2125 {
2126         struct bucket_table *tbl;
2127         struct rhash_head *pos;
2128         struct btree *b;
2129         unsigned i;
2130         bool ret = false;
2131 restart:
2132         rcu_read_lock();
2133         for_each_cached_btree(b, c, tbl, i, pos)
2134                 if (test_bit(flag, &b->flags)) {
2135                         rcu_read_unlock();
2136                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2137                         ret = true;
2138                         goto restart;
2139                 }
2140         rcu_read_unlock();
2141
2142         return ret;
2143 }
2144
2145 bool bch2_btree_flush_all_reads(struct bch_fs *c)
2146 {
2147         return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2148 }
2149
2150 bool bch2_btree_flush_all_writes(struct bch_fs *c)
2151 {
2152         return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2153 }