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         if (btree_node_is_extents(src))
 395                 nr = bch2_sort_repack_merge(c, btree_bset_first(dst),
 396                                 src, &src_iter,
 397                                 &dst->format,
 398                                 true);
 399         else
 400                 nr = bch2_sort_repack(btree_bset_first(dst),
 401                                 src, &src_iter,
 402                                 &dst->format,
 403                                 true);
 404
 405         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 406                                start_time);
 407
 408         set_btree_bset_end(dst, dst->set);
 409
 410         dst->nr.live_u64s       += nr.live_u64s;
 411         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 412         dst->nr.packed_keys     += nr.packed_keys;
 413         dst->nr.unpacked_keys   += nr.unpacked_keys;
 414
 415         bch2_verify_btree_nr_keys(dst);
 416 }
 417
 418 #define SORT_CRIT       (4096 / sizeof(u64))
 419
 420 /*
 421  * We're about to add another bset to the btree node, so if there's currently
 422  * too many bsets - sort some of them together:
 423  */
 424 static bool btree_node_compact(struct bch_fs *c, struct btree *b)
 425 {
 426         unsigned unwritten_idx;
 427         bool ret = false;
 428
 429         for (unwritten_idx = 0;
 430              unwritten_idx < b->nsets;
 431              unwritten_idx++)
 432                 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
 433                         break;
 434
 435         if (b->nsets - unwritten_idx > 1) {
 436                 btree_node_sort(c, b, unwritten_idx,
 437                                 b->nsets, false);
 438                 ret = true;
 439         }
 440
 441         if (unwritten_idx > 1) {
 442                 btree_node_sort(c, b, 0, unwritten_idx, false);
 443                 ret = true;
 444         }
 445
 446         return ret;
 447 }
 448
 449 void bch2_btree_build_aux_trees(struct btree *b)
 450 {
 451         struct bset_tree *t;
 452
 453         for_each_bset(b, t)
 454                 bch2_bset_build_aux_tree(b, t,
 455                                 !bset_written(b, bset(b, t)) &&
 456                                 t == bset_tree_last(b));
 457 }
 458
 459 /*
 460  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 461  * inserted into
 462  *
 463  * Safe to call if there already is an unwritten bset - will only add a new bset
 464  * if @b doesn't already have one.
 465  *
 466  * Returns true if we sorted (i.e. invalidated iterators
 467  */
 468 void bch2_btree_init_next(struct btree_trans *trans,
 469                           struct btree_iter *iter,
 470                           struct btree *b)
 471 {
 472         struct bch_fs *c = trans->c;
 473         struct btree_node_entry *bne;
 474         bool reinit_iter = false;
 475
 476         EBUG_ON(!(b->c.lock.state.seq & 1));
 477         EBUG_ON(iter && iter->l[b->c.level].b != b);
 478         BUG_ON(bset_written(b, bset(b, &b->set[1])));
 479
 480         if (b->nsets == MAX_BSETS &&
 481             !btree_node_write_in_flight(b)) {
 482                 unsigned log_u64s[] = {
 483                         ilog2(bset_u64s(&b->set[0])),
 484                         ilog2(bset_u64s(&b->set[1])),
 485                         ilog2(bset_u64s(&b->set[2])),
 486                 };
 487
 488                 if (log_u64s[1] >= (log_u64s[0] + log_u64s[2]) / 2) {
 489                         bch2_btree_node_write(c, b, SIX_LOCK_write);
 490                         reinit_iter = true;
 491                 }
 492         }
 493
 494         if (b->nsets == MAX_BSETS &&
 495             btree_node_compact(c, b))
 496                 reinit_iter = true;
 497
 498         BUG_ON(b->nsets >= MAX_BSETS);
 499
 500         bne = want_new_bset(c, b);
 501         if (bne)
 502                 bch2_bset_init_next(c, b, bne);
 503
 504         bch2_btree_build_aux_trees(b);
 505
 506         if (iter && reinit_iter)
 507                 bch2_btree_iter_reinit_node(iter, b);
 508 }
 509
 510 static void btree_pos_to_text(struct printbuf *out, struct bch_fs *c,
 511                           struct btree *b)
 512 {
 513         pr_buf(out, "%s level %u/%u\n  ",
 514                bch2_btree_ids[b->c.btree_id],
 515                b->c.level,
 516                c->btree_roots[b->c.btree_id].level);
 517         bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 518 }
 519
 520 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 521                           struct bch_dev *ca,
 522                           struct btree *b, struct bset *i,
 523                           unsigned offset, int write)
 524 {
 525         pr_buf(out, "error validating btree node ");
 526         if (write)
 527                 pr_buf(out, "before write ");
 528         if (ca)
 529                 pr_buf(out, "on %s ", ca->name);
 530         pr_buf(out, "at btree ");
 531         btree_pos_to_text(out, c, b);
 532
 533         pr_buf(out, "\n  node offset %u", b->written);
 534         if (i)
 535                 pr_buf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 536 }
 537
 538 enum btree_err_type {
 539         BTREE_ERR_FIXABLE,
 540         BTREE_ERR_WANT_RETRY,
 541         BTREE_ERR_MUST_RETRY,
 542         BTREE_ERR_FATAL,
 543 };
 544
 545 enum btree_validate_ret {
 546         BTREE_RETRY_READ = 64,
 547 };
 548
 549 #define btree_err(type, c, ca, b, i, msg, ...)                          \
 550 ({                                                                      \
 551         __label__ out;                                                  \
 552         char _buf[300];                                                 \
 553         char *_buf2 = _buf;                                             \
 554         struct printbuf out = PBUF(_buf);                               \
 555                                                                         \
 556         _buf2 = kmalloc(4096, GFP_ATOMIC);                              \
 557         if (_buf2)                                                      \
 558                 out = _PBUF(_buf2, 4986);                               \
 559                                                                         \
 560         btree_err_msg(&out, c, ca, b, i, b->written, write);            \
 561         pr_buf(&out, ": " msg, ##__VA_ARGS__);                          \
 562                                                                         \
 563         if (type == BTREE_ERR_FIXABLE &&                                \
 564             write == READ &&                                            \
 565             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {             \
 566                 mustfix_fsck_err(c, "%s", _buf2);                       \
 567                 goto out;                                               \
 568         }                                                               \
 569                                                                         \
 570         switch (write) {                                                \
 571         case READ:                                                      \
 572                 bch_err(c, "%s", _buf2);                                \
 573                                                                         \
 574                 switch (type) {                                         \
 575                 case BTREE_ERR_FIXABLE:                                 \
 576                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 577                         goto fsck_err;                                  \
 578                 case BTREE_ERR_WANT_RETRY:                              \
 579                         if (have_retry) {                               \
 580                                 ret = BTREE_RETRY_READ;                 \
 581                                 goto fsck_err;                          \
 582                         }                                               \
 583                         break;                                          \
 584                 case BTREE_ERR_MUST_RETRY:                              \
 585                         ret = BTREE_RETRY_READ;                         \
 586                         goto fsck_err;                                  \
 587                 case BTREE_ERR_FATAL:                                   \
 588                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 589                         goto fsck_err;                                  \
 590                 }                                                       \
 591                 break;                                                  \
 592         case WRITE:                                                     \
 593                 bch_err(c, "corrupt metadata before write: %s", _buf2); \
 594                                                                         \
 595                 if (bch2_fs_inconsistent(c)) {                          \
 596                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 597                         goto fsck_err;                                  \
 598                 }                                                       \
 599                 break;                                                  \
 600         }                                                               \
 601 out:                                                                    \
 602         if (_buf2 != _buf)                                              \
 603                 kfree(_buf2);                                           \
 604         true;                                                           \
 605 })
 606
 607 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 608
 609 /*
 610  * When btree topology repair changes the start or end of a node, that might
 611  * mean we have to drop keys that are no longer inside the node:
 612  */
 613 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
 614 {
 615         struct bset_tree *t;
 616         struct bkey_s_c k;
 617         struct bkey unpacked;
 618         struct btree_node_iter iter;
 619
 620         for_each_bset(b, t) {
 621                 struct bset *i = bset(b, t);
 622                 struct bkey_packed *k;
 623
 624                 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
 625                         if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
 626                                 break;
 627
 628                 if (k != i->start) {
 629                         unsigned shift = (u64 *) k - (u64 *) i->start;
 630
 631                         memmove_u64s_down(i->start, k,
 632                                           (u64 *) vstruct_end(i) - (u64 *) k);
 633                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
 634                         set_btree_bset_end(b, t);
 635                         bch2_bset_set_no_aux_tree(b, t);
 636                 }
 637
 638                 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
 639                         if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
 640                                 break;
 641
 642                 if (k != vstruct_last(i)) {
 643                         i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
 644                         set_btree_bset_end(b, t);
 645                         bch2_bset_set_no_aux_tree(b, t);
 646                 }
 647         }
 648
 649         bch2_btree_build_aux_trees(b);
 650
 651         for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
 652                 BUG_ON(bpos_cmp(k.k->p, b->data->min_key) < 0);
 653                 BUG_ON(bpos_cmp(k.k->p, b->data->max_key) > 0);
 654         }
 655 }
 656
 657 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 658                          struct btree *b, struct bset *i,
 659                          unsigned offset, unsigned sectors,
 660                          int write, bool have_retry)
 661 {
 662         unsigned version = le16_to_cpu(i->version);
 663         const char *err;
 664         char buf1[100];
 665         char buf2[100];
 666         int ret = 0;
 667
 668         btree_err_on((version != BCH_BSET_VERSION_OLD &&
 669                       version < bcachefs_metadata_version_min) ||
 670                      version >= bcachefs_metadata_version_max,
 671                      BTREE_ERR_FATAL, c, ca, b, i,
 672                      "unsupported bset version");
 673
 674         if (btree_err_on(version < c->sb.version_min,
 675                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 676                          "bset version %u older than superblock version_min %u",
 677                          version, c->sb.version_min)) {
 678                 mutex_lock(&c->sb_lock);
 679                 c->disk_sb.sb->version_min = cpu_to_le16(version);
 680                 bch2_write_super(c);
 681                 mutex_unlock(&c->sb_lock);
 682         }
 683
 684         if (btree_err_on(version > c->sb.version,
 685                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 686                          "bset version %u newer than superblock version %u",
 687                          version, c->sb.version)) {
 688                 mutex_lock(&c->sb_lock);
 689                 c->disk_sb.sb->version = cpu_to_le16(version);
 690                 bch2_write_super(c);
 691                 mutex_unlock(&c->sb_lock);
 692         }
 693
 694         btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
 695                      BTREE_ERR_FATAL, c, ca, b, i,
 696                      "BSET_SEPARATE_WHITEOUTS no longer supported");
 697
 698         if (btree_err_on(offset + sectors > c->opts.btree_node_size,
 699                          BTREE_ERR_FIXABLE, c, ca, b, i,
 700                          "bset past end of btree node")) {
 701                 i->u64s = 0;
 702                 return 0;
 703         }
 704
 705         btree_err_on(offset && !i->u64s,
 706                      BTREE_ERR_FIXABLE, c, ca, b, i,
 707                      "empty bset");
 708
 709         btree_err_on(BSET_OFFSET(i) &&
 710                      BSET_OFFSET(i) != offset,
 711                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 712                      "bset at wrong sector offset");
 713
 714         if (!offset) {
 715                 struct btree_node *bn =
 716                         container_of(i, struct btree_node, keys);
 717                 /* These indicate that we read the wrong btree node: */
 718
 719                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 720                         struct bch_btree_ptr_v2 *bp =
 721                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 722
 723                         /* XXX endianness */
 724                         btree_err_on(bp->seq != bn->keys.seq,
 725                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 726                                      "incorrect sequence number (wrong btree node)");
 727                 }
 728
 729                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 730                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 731                              "incorrect btree id");
 732
 733                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 734                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 735                              "incorrect level");
 736
 737                 if (!write)
 738                         compat_btree_node(b->c.level, b->c.btree_id, version,
 739                                           BSET_BIG_ENDIAN(i), write, bn);
 740
 741                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 742                         struct bch_btree_ptr_v2 *bp =
 743                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 744
 745                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 746                                 b->data->min_key = bp->min_key;
 747                                 b->data->max_key = b->key.k.p;
 748                         }
 749
 750                         btree_err_on(bpos_cmp(b->data->min_key, bp->min_key),
 751                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 752                                      "incorrect min_key: got %s should be %s",
 753                                      (bch2_bpos_to_text(&PBUF(buf1), bn->min_key), buf1),
 754                                      (bch2_bpos_to_text(&PBUF(buf2), bp->min_key), buf2));
 755                 }
 756
 757                 btree_err_on(bpos_cmp(bn->max_key, b->key.k.p),
 758                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 759                              "incorrect max key %s",
 760                              (bch2_bpos_to_text(&PBUF(buf1), bn->max_key), buf1));
 761
 762                 if (write)
 763                         compat_btree_node(b->c.level, b->c.btree_id, version,
 764                                           BSET_BIG_ENDIAN(i), write, bn);
 765
 766                 err = bch2_bkey_format_validate(&bn->format);
 767                 btree_err_on(err,
 768                              BTREE_ERR_FATAL, c, ca, b, i,
 769                              "invalid bkey format: %s", err);
 770
 771                 compat_bformat(b->c.level, b->c.btree_id, version,
 772                                BSET_BIG_ENDIAN(i), write,
 773                                &bn->format);
 774         }
 775 fsck_err:
 776         return ret;
 777 }
 778
 779 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 780                          struct bset *i, unsigned *whiteout_u64s,
 781                          int write, bool have_retry)
 782 {
 783         unsigned version = le16_to_cpu(i->version);
 784         struct bkey_packed *k, *prev = NULL;
 785         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 786                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 787         int ret = 0;
 788
 789         for (k = i->start;
 790              k != vstruct_last(i);) {
 791                 struct bkey_s u;
 792                 struct bkey tmp;
 793                 const char *invalid;
 794
 795                 if (btree_err_on(bkey_next(k) > vstruct_last(i),
 796                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 797                                  "key extends past end of bset")) {
 798                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 799                         break;
 800                 }
 801
 802                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 803                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 804                                  "invalid bkey format %u", k->format)) {
 805                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 806                         memmove_u64s_down(k, bkey_next(k),
 807                                           (u64 *) vstruct_end(i) - (u64 *) k);
 808                         continue;
 809                 }
 810
 811                 /* XXX: validate k->u64s */
 812                 if (!write)
 813                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 814                                     BSET_BIG_ENDIAN(i), write,
 815                                     &b->format, k);
 816
 817                 u = __bkey_disassemble(b, k, &tmp);
 818
 819                 invalid = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b)) ?:
 820                         (!updated_range ?  bch2_bkey_in_btree_node(b, u.s_c) : NULL) ?:
 821                         (write ? bch2_bkey_val_invalid(c, u.s_c) : NULL);
 822                 if (invalid) {
 823                         char buf[160];
 824
 825                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 826                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 827                                   "invalid bkey: %s\n%s", invalid, buf);
 828
 829                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 830                         memmove_u64s_down(k, bkey_next(k),
 831                                           (u64 *) vstruct_end(i) - (u64 *) k);
 832                         continue;
 833                 }
 834
 835                 if (write)
 836                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 837                                     BSET_BIG_ENDIAN(i), write,
 838                                     &b->format, k);
 839
 840                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 841                         char buf1[80];
 842                         char buf2[80];
 843                         struct bkey up = bkey_unpack_key(b, prev);
 844
 845                         bch2_bkey_to_text(&PBUF(buf1), &up);
 846                         bch2_bkey_to_text(&PBUF(buf2), u.k);
 847
 848                         bch2_dump_bset(c, b, i, 0);
 849
 850                         if (btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 851                                       "keys out of order: %s > %s",
 852                                       buf1, buf2)) {
 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         return ret;
 865 }
 866
 867 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 868                               struct btree *b, bool have_retry)
 869 {
 870         struct btree_node_entry *bne;
 871         struct sort_iter *iter;
 872         struct btree_node *sorted;
 873         struct bkey_packed *k;
 874         struct bch_extent_ptr *ptr;
 875         struct bset *i;
 876         bool used_mempool, blacklisted;
 877         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 878                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 879         unsigned u64s;
 880         unsigned blacklisted_written, nonblacklisted_written = 0;
 881         unsigned ptr_written = btree_ptr_sectors_written(&b->key);
 882         int ret, retry_read = 0, write = READ;
 883
 884         b->version_ondisk = U16_MAX;
 885
 886         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
 887         sort_iter_init(iter, b);
 888         iter->size = (btree_blocks(c) + 1) * 2;
 889
 890         if (bch2_meta_read_fault("btree"))
 891                 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 892                           "dynamic fault");
 893
 894         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 895                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 896                      "bad magic");
 897
 898         btree_err_on(!b->data->keys.seq,
 899                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 900                      "bad btree header");
 901
 902         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 903                 struct bch_btree_ptr_v2 *bp =
 904                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 905
 906                 btree_err_on(b->data->keys.seq != bp->seq,
 907                              BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 908                              "got wrong btree node (seq %llx want %llx)",
 909                              b->data->keys.seq, bp->seq);
 910         }
 911
 912         while (b->written < (ptr_written ?: c->opts.btree_node_size)) {
 913                 unsigned sectors, whiteout_u64s = 0;
 914                 struct nonce nonce;
 915                 struct bch_csum csum;
 916                 bool first = !b->written;
 917
 918                 if (!b->written) {
 919                         i = &b->data->keys;
 920
 921                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 922                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 923                                      "unknown checksum type %llu",
 924                                      BSET_CSUM_TYPE(i));
 925
 926                         nonce = btree_nonce(i, b->written << 9);
 927                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 928
 929                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 930                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 931                                      "invalid checksum");
 932
 933                         bset_encrypt(c, i, b->written << 9);
 934
 935                         btree_err_on(btree_node_is_extents(b) &&
 936                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
 937                                      BTREE_ERR_FATAL, c, NULL, b, NULL,
 938                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
 939
 940                         sectors = vstruct_sectors(b->data, c->block_bits);
 941                 } else {
 942                         bne = write_block(b);
 943                         i = &bne->keys;
 944
 945                         if (i->seq != b->data->keys.seq)
 946                                 break;
 947
 948                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 949                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 950                                      "unknown checksum type %llu",
 951                                      BSET_CSUM_TYPE(i));
 952
 953                         nonce = btree_nonce(i, b->written << 9);
 954                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
 955
 956                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
 957                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 958                                      "invalid checksum");
 959
 960                         bset_encrypt(c, i, b->written << 9);
 961
 962                         sectors = vstruct_sectors(bne, c->block_bits);
 963                 }
 964
 965                 b->version_ondisk = min(b->version_ondisk,
 966                                         le16_to_cpu(i->version));
 967
 968                 ret = validate_bset(c, ca, b, i, b->written, sectors,
 969                                     READ, have_retry);
 970                 if (ret)
 971                         goto fsck_err;
 972
 973                 if (!b->written)
 974                         btree_node_set_format(b, b->data->format);
 975
 976                 ret = validate_bset_keys(c, b, i, &whiteout_u64s,
 977                                     READ, have_retry);
 978                 if (ret)
 979                         goto fsck_err;
 980
 981                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 982
 983                 b->written += sectors;
 984
 985                 blacklisted = bch2_journal_seq_is_blacklisted(c,
 986                                         le64_to_cpu(i->journal_seq),
 987                                         true);
 988
 989                 btree_err_on(blacklisted && first,
 990                              BTREE_ERR_FIXABLE, c, ca, b, i,
 991                              "first btree node bset has blacklisted journal seq");
 992
 993                 btree_err_on(blacklisted && ptr_written,
 994                              BTREE_ERR_FIXABLE, c, ca, b, i,
 995                              "found blacklisted bset in btree node with sectors_written");
 996                 if (blacklisted && !first)
 997                         continue;
 998
 999                 sort_iter_add(iter, i->start,
1000                               vstruct_idx(i, whiteout_u64s));
1001
1002                 sort_iter_add(iter,
1003                               vstruct_idx(i, whiteout_u64s),
1004                               vstruct_last(i));
1005
1006                 nonblacklisted_written = b->written;
1007         }
1008
1009         if (ptr_written) {
1010                 btree_err_on(b->written < ptr_written,
1011                              BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1012                              "btree node data missing: expected %u sectors, found %u",
1013                              ptr_written, b->written);
1014         } else {
1015                 for (bne = write_block(b);
1016                      bset_byte_offset(b, bne) < btree_bytes(c);
1017                      bne = (void *) bne + block_bytes(c))
1018                         btree_err_on(bne->keys.seq == b->data->keys.seq &&
1019                                      !bch2_journal_seq_is_blacklisted(c,
1020                                                                       le64_to_cpu(bne->keys.journal_seq),
1021                                                                       true),
1022                                      BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1023                                      "found bset signature after last bset");
1024
1025                 /*
1026                  * Blacklisted bsets are those that were written after the most recent
1027                  * (flush) journal write. Since there wasn't a flush, they may not have
1028                  * made it to all devices - which means we shouldn't write new bsets
1029                  * after them, as that could leave a gap and then reads from that device
1030                  * wouldn't find all the bsets in that btree node - which means it's
1031                  * important that we start writing new bsets after the most recent _non_
1032                  * blacklisted bset:
1033                  */
1034                 blacklisted_written = b->written;
1035                 b->written = nonblacklisted_written;
1036         }
1037
1038         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1039         sorted->keys.u64s = 0;
1040
1041         set_btree_bset(b, b->set, &b->data->keys);
1042
1043         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1044
1045         u64s = le16_to_cpu(sorted->keys.u64s);
1046         *sorted = *b->data;
1047         sorted->keys.u64s = cpu_to_le16(u64s);
1048         swap(sorted, b->data);
1049         set_btree_bset(b, b->set, &b->data->keys);
1050         b->nsets = 1;
1051
1052         BUG_ON(b->nr.live_u64s != u64s);
1053
1054         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1055
1056         if (updated_range)
1057                 bch2_btree_node_drop_keys_outside_node(b);
1058
1059         i = &b->data->keys;
1060         for (k = i->start; k != vstruct_last(i);) {
1061                 struct bkey tmp;
1062                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1063                 const char *invalid = bch2_bkey_val_invalid(c, u.s_c);
1064
1065                 if (invalid ||
1066                     (bch2_inject_invalid_keys &&
1067                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1068                         char buf[160];
1069
1070                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
1071                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
1072                                   "invalid bkey %s: %s", buf, invalid);
1073
1074                         btree_keys_account_key_drop(&b->nr, 0, k);
1075
1076                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1077                         memmove_u64s_down(k, bkey_next(k),
1078                                           (u64 *) vstruct_end(i) - (u64 *) k);
1079                         set_btree_bset_end(b, b->set);
1080                         continue;
1081                 }
1082
1083                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1084                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1085
1086                         bp.v->mem_ptr = 0;
1087                 }
1088
1089                 k = bkey_next(k);
1090         }
1091
1092         bch2_bset_build_aux_tree(b, b->set, false);
1093
1094         set_needs_whiteout(btree_bset_first(b), true);
1095
1096         btree_node_reset_sib_u64s(b);
1097
1098         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1099                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1100
1101                 if (ca->mi.state != BCH_MEMBER_STATE_rw)
1102                         set_btree_node_need_rewrite(b);
1103         }
1104
1105         if (!ptr_written)
1106                 set_btree_node_need_rewrite(b);
1107 out:
1108         mempool_free(iter, &c->fill_iter);
1109         return retry_read;
1110 fsck_err:
1111         if (ret == BTREE_RETRY_READ) {
1112                 retry_read = 1;
1113         } else {
1114                 bch2_inconsistent_error(c);
1115                 set_btree_node_read_error(b);
1116         }
1117         goto out;
1118 }
1119
1120 static void btree_node_read_work(struct work_struct *work)
1121 {
1122         struct btree_read_bio *rb =
1123                 container_of(work, struct btree_read_bio, work);
1124         struct bch_fs *c        = rb->c;
1125         struct btree *b         = rb->b;
1126         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1127         struct bio *bio         = &rb->bio;
1128         struct bch_io_failures failed = { .nr = 0 };
1129         char buf[200];
1130         struct printbuf out;
1131         bool saw_error = false;
1132         bool can_retry;
1133
1134         goto start;
1135         while (1) {
1136                 bch_info(c, "retrying read");
1137                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1138                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1139                 bio_reset(bio);
1140                 bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1141                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1142                 bio->bi_iter.bi_size    = btree_bytes(c);
1143
1144                 if (rb->have_ioref) {
1145                         bio_set_dev(bio, ca->disk_sb.bdev);
1146                         submit_bio_wait(bio);
1147                 } else {
1148                         bio->bi_status = BLK_STS_REMOVED;
1149                 }
1150 start:
1151                 out = PBUF(buf);
1152                 btree_pos_to_text(&out, c, b);
1153                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1154                                    bch2_blk_status_to_str(bio->bi_status), buf);
1155                 if (rb->have_ioref)
1156                         percpu_ref_put(&ca->io_ref);
1157                 rb->have_ioref = false;
1158
1159                 bch2_mark_io_failure(&failed, &rb->pick);
1160
1161                 can_retry = bch2_bkey_pick_read_device(c,
1162                                 bkey_i_to_s_c(&b->key),
1163                                 &failed, &rb->pick) > 0;
1164
1165                 if (!bio->bi_status &&
1166                     !bch2_btree_node_read_done(c, ca, b, can_retry))
1167                         break;
1168
1169                 saw_error = true;
1170
1171                 if (!can_retry) {
1172                         set_btree_node_read_error(b);
1173                         break;
1174                 }
1175         }
1176
1177         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1178                                rb->start_time);
1179         bio_put(&rb->bio);
1180
1181         if (saw_error && !btree_node_read_error(b))
1182                 bch2_btree_node_rewrite_async(c, b);
1183
1184         clear_btree_node_read_in_flight(b);
1185         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1186 }
1187
1188 static void btree_node_read_endio(struct bio *bio)
1189 {
1190         struct btree_read_bio *rb =
1191                 container_of(bio, struct btree_read_bio, bio);
1192         struct bch_fs *c        = rb->c;
1193
1194         if (rb->have_ioref) {
1195                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1196                 bch2_latency_acct(ca, rb->start_time, READ);
1197         }
1198
1199         queue_work(c->io_complete_wq, &rb->work);
1200 }
1201
1202 struct btree_node_read_all {
1203         struct closure          cl;
1204         struct bch_fs           *c;
1205         struct btree            *b;
1206         unsigned                nr;
1207         void                    *buf[BCH_REPLICAS_MAX];
1208         struct bio              *bio[BCH_REPLICAS_MAX];
1209         int                     err[BCH_REPLICAS_MAX];
1210 };
1211
1212 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1213 {
1214         struct btree_node *bn = data;
1215         struct btree_node_entry *bne;
1216         unsigned offset = 0;
1217
1218         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1219                 return 0;
1220
1221         while (offset < c->opts.btree_node_size) {
1222                 if (!offset) {
1223                         offset += vstruct_sectors(bn, c->block_bits);
1224                 } else {
1225                         bne = data + (offset << 9);
1226                         if (bne->keys.seq != bn->keys.seq)
1227                                 break;
1228                         offset += vstruct_sectors(bne, c->block_bits);
1229                 }
1230         }
1231
1232         return offset;
1233 }
1234
1235 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1236 {
1237         struct btree_node *bn = data;
1238         struct btree_node_entry *bne;
1239
1240         if (!offset)
1241                 return false;
1242
1243         while (offset < c->opts.btree_node_size) {
1244                 bne = data + (offset << 9);
1245                 if (bne->keys.seq == bn->keys.seq)
1246                         return true;
1247                 offset++;
1248         }
1249
1250         return false;
1251         return offset;
1252 }
1253
1254 static void btree_node_read_all_replicas_done(struct closure *cl)
1255 {
1256         struct btree_node_read_all *ra =
1257                 container_of(cl, struct btree_node_read_all, cl);
1258         struct bch_fs *c = ra->c;
1259         struct btree *b = ra->b;
1260         bool dump_bset_maps = false;
1261         bool have_retry = false;
1262         int ret = 0, best = -1, write = READ;
1263         unsigned i, written, written2;
1264         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1265                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1266
1267         for (i = 0; i < ra->nr; i++) {
1268                 struct btree_node *bn = ra->buf[i];
1269
1270                 if (ra->err[i])
1271                         continue;
1272
1273                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1274                     (seq && seq != bn->keys.seq))
1275                         continue;
1276
1277                 if (best < 0) {
1278                         best = i;
1279                         written = btree_node_sectors_written(c, bn);
1280                         continue;
1281                 }
1282
1283                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1284                 if (btree_err_on(written2 != written, BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1285                                  "btree node sectors written mismatch: %u != %u",
1286                                  written, written2) ||
1287                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1288                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1289                                  "found bset signature after last bset") ||
1290                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1291                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1292                                  "btree node replicas content mismatch"))
1293                         dump_bset_maps = true;
1294
1295                 if (written2 > written) {
1296                         written = written2;
1297                         best = i;
1298                 }
1299         }
1300 fsck_err:
1301         if (dump_bset_maps) {
1302                 for (i = 0; i < ra->nr; i++) {
1303                         char buf[200];
1304                         struct printbuf out = PBUF(buf);
1305                         struct btree_node *bn = ra->buf[i];
1306                         struct btree_node_entry *bne = NULL;
1307                         unsigned offset = 0, sectors;
1308                         bool gap = false;
1309
1310                         if (ra->err[i])
1311                                 continue;
1312
1313                         while (offset < c->opts.btree_node_size) {
1314                                 if (!offset) {
1315                                         sectors = vstruct_sectors(bn, c->block_bits);
1316                                 } else {
1317                                         bne = ra->buf[i] + (offset << 9);
1318                                         if (bne->keys.seq != bn->keys.seq)
1319                                                 break;
1320                                         sectors = vstruct_sectors(bne, c->block_bits);
1321                                 }
1322
1323                                 pr_buf(&out, " %u-%u", offset, offset + sectors);
1324                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1325                                                         le64_to_cpu(bne->keys.journal_seq), false))
1326                                         pr_buf(&out, "*");
1327                                 offset += sectors;
1328                         }
1329
1330                         while (offset < c->opts.btree_node_size) {
1331                                 bne = ra->buf[i] + (offset << 9);
1332                                 if (bne->keys.seq == bn->keys.seq) {
1333                                         if (!gap)
1334                                                 pr_buf(&out, " GAP");
1335                                         gap = true;
1336
1337                                         sectors = vstruct_sectors(bne, c->block_bits);
1338                                         pr_buf(&out, " %u-%u", offset, offset + sectors);
1339                                         if (bch2_journal_seq_is_blacklisted(c,
1340                                                         le64_to_cpu(bne->keys.journal_seq), false))
1341                                                 pr_buf(&out, "*");
1342                                 }
1343                                 offset++;
1344                         }
1345
1346                         bch_err(c, "replica %u:%s", i, buf);
1347                 }
1348         }
1349
1350         if (best >= 0) {
1351                 memcpy(b->data, ra->buf[best], btree_bytes(c));
1352                 ret = bch2_btree_node_read_done(c, NULL, b, false);
1353         } else {
1354                 ret = -1;
1355         }
1356
1357         if (ret)
1358                 set_btree_node_read_error(b);
1359
1360         for (i = 0; i < ra->nr; i++) {
1361                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1362                 bio_put(ra->bio[i]);
1363         }
1364
1365         closure_debug_destroy(&ra->cl);
1366         kfree(ra);
1367
1368         clear_btree_node_read_in_flight(b);
1369         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1370 }
1371
1372 static void btree_node_read_all_replicas_endio(struct bio *bio)
1373 {
1374         struct btree_read_bio *rb =
1375                 container_of(bio, struct btree_read_bio, bio);
1376         struct bch_fs *c        = rb->c;
1377         struct btree_node_read_all *ra = rb->ra;
1378
1379         if (rb->have_ioref) {
1380                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1381                 bch2_latency_acct(ca, rb->start_time, READ);
1382         }
1383
1384         ra->err[rb->idx] = bio->bi_status;
1385         closure_put(&ra->cl);
1386 }
1387
1388 /*
1389  * XXX This allocates multiple times from the same mempools, and can deadlock
1390  * under sufficient memory pressure (but is only a debug path)
1391  */
1392 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1393 {
1394         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1395         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1396         const union bch_extent_entry *entry;
1397         struct extent_ptr_decoded pick;
1398         struct btree_node_read_all *ra;
1399         unsigned i;
1400
1401         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1402         if (!ra)
1403                 return -ENOMEM;
1404
1405         closure_init(&ra->cl, NULL);
1406         ra->c   = c;
1407         ra->b   = b;
1408         ra->nr  = bch2_bkey_nr_ptrs(k);
1409
1410         for (i = 0; i < ra->nr; i++) {
1411                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1412                 ra->bio[i] = bio_alloc_bioset(GFP_NOFS, buf_pages(ra->buf[i],
1413                                                                   btree_bytes(c)),
1414                                               &c->btree_bio);
1415         }
1416
1417         i = 0;
1418         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1419                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1420                 struct btree_read_bio *rb =
1421                         container_of(ra->bio[i], struct btree_read_bio, bio);
1422                 rb->c                   = c;
1423                 rb->b                   = b;
1424                 rb->ra                  = ra;
1425                 rb->start_time          = local_clock();
1426                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1427                 rb->idx                 = i;
1428                 rb->pick                = pick;
1429                 rb->bio.bi_opf          = REQ_OP_READ|REQ_SYNC|REQ_META;
1430                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1431                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1432                 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1433
1434                 if (rb->have_ioref) {
1435                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1436                                      bio_sectors(&rb->bio));
1437                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1438
1439                         closure_get(&ra->cl);
1440                         submit_bio(&rb->bio);
1441                 } else {
1442                         ra->err[i] = BLK_STS_REMOVED;
1443                 }
1444
1445                 i++;
1446         }
1447
1448         if (sync) {
1449                 closure_sync(&ra->cl);
1450                 btree_node_read_all_replicas_done(&ra->cl);
1451         } else {
1452                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1453                             c->io_complete_wq);
1454         }
1455
1456         return 0;
1457 }
1458
1459 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1460                           bool sync)
1461 {
1462         struct extent_ptr_decoded pick;
1463         struct btree_read_bio *rb;
1464         struct bch_dev *ca;
1465         struct bio *bio;
1466         char buf[200];
1467         int ret;
1468
1469         btree_pos_to_text(&PBUF(buf), c, b);
1470         trace_btree_read(c, b);
1471
1472         if (bch2_verify_all_btree_replicas &&
1473             !btree_node_read_all_replicas(c, b, sync))
1474                 return;
1475
1476         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1477                                          NULL, &pick);
1478         if (bch2_fs_fatal_err_on(ret <= 0, c,
1479                         "btree node read error: no device to read from\n"
1480                         " at %s", buf)) {
1481                 set_btree_node_read_error(b);
1482                 return;
1483         }
1484
1485         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1486
1487         bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1488                                                    btree_bytes(c)),
1489                                &c->btree_bio);
1490         rb = container_of(bio, struct btree_read_bio, bio);
1491         rb->c                   = c;
1492         rb->b                   = b;
1493         rb->ra                  = NULL;
1494         rb->start_time          = local_clock();
1495         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1496         rb->pick                = pick;
1497         INIT_WORK(&rb->work, btree_node_read_work);
1498         bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1499         bio->bi_iter.bi_sector  = pick.ptr.offset;
1500         bio->bi_end_io          = btree_node_read_endio;
1501         bch2_bio_map(bio, b->data, btree_bytes(c));
1502
1503         if (rb->have_ioref) {
1504                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1505                              bio_sectors(bio));
1506                 bio_set_dev(bio, ca->disk_sb.bdev);
1507
1508                 if (sync) {
1509                         submit_bio_wait(bio);
1510
1511                         btree_node_read_work(&rb->work);
1512                 } else {
1513                         submit_bio(bio);
1514                 }
1515         } else {
1516                 bio->bi_status = BLK_STS_REMOVED;
1517
1518                 if (sync)
1519                         btree_node_read_work(&rb->work);
1520                 else
1521                         queue_work(c->io_complete_wq, &rb->work);
1522         }
1523 }
1524
1525 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1526                         const struct bkey_i *k, unsigned level)
1527 {
1528         struct closure cl;
1529         struct btree *b;
1530         int ret;
1531
1532         closure_init_stack(&cl);
1533
1534         do {
1535                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1536                 closure_sync(&cl);
1537         } while (ret);
1538
1539         b = bch2_btree_node_mem_alloc(c);
1540         bch2_btree_cache_cannibalize_unlock(c);
1541
1542         BUG_ON(IS_ERR(b));
1543
1544         bkey_copy(&b->key, k);
1545         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1546
1547         set_btree_node_read_in_flight(b);
1548
1549         bch2_btree_node_read(c, b, true);
1550
1551         if (btree_node_read_error(b)) {
1552                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1553
1554                 mutex_lock(&c->btree_cache.lock);
1555                 list_move(&b->list, &c->btree_cache.freeable);
1556                 mutex_unlock(&c->btree_cache.lock);
1557
1558                 ret = -EIO;
1559                 goto err;
1560         }
1561
1562         bch2_btree_set_root_for_read(c, b);
1563 err:
1564         six_unlock_write(&b->c.lock);
1565         six_unlock_intent(&b->c.lock);
1566
1567         return ret;
1568 }
1569
1570 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1571                               struct btree_write *w)
1572 {
1573         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1574
1575         do {
1576                 old = new = v;
1577                 if (!(old & 1))
1578                         break;
1579
1580                 new &= ~1UL;
1581         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1582
1583         if (old & 1)
1584                 closure_put(&((struct btree_update *) new)->cl);
1585
1586         bch2_journal_pin_drop(&c->journal, &w->journal);
1587 }
1588
1589 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1590 {
1591         struct btree_write *w = btree_prev_write(b);
1592         unsigned long old, new, v;
1593
1594         bch2_btree_complete_write(c, b, w);
1595
1596         v = READ_ONCE(b->flags);
1597         do {
1598                 old = new = v;
1599
1600                 if (old & (1U << BTREE_NODE_need_write))
1601                         goto do_write;
1602
1603                 new &= ~(1U << BTREE_NODE_write_in_flight);
1604                 new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1605         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1606
1607         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1608         return;
1609
1610 do_write:
1611         six_lock_read(&b->c.lock, NULL, NULL);
1612         v = READ_ONCE(b->flags);
1613         do {
1614                 old = new = v;
1615
1616                 if ((old & (1U << BTREE_NODE_dirty)) &&
1617                     (old & (1U << BTREE_NODE_need_write)) &&
1618                     !(old & (1U << BTREE_NODE_never_write)) &&
1619                     btree_node_may_write(b)) {
1620                         new &= ~(1U << BTREE_NODE_dirty);
1621                         new &= ~(1U << BTREE_NODE_need_write);
1622                         new |=  (1U << BTREE_NODE_write_in_flight);
1623                         new |=  (1U << BTREE_NODE_write_in_flight_inner);
1624                         new |=  (1U << BTREE_NODE_just_written);
1625                         new ^=  (1U << BTREE_NODE_write_idx);
1626                 } else {
1627                         new &= ~(1U << BTREE_NODE_write_in_flight);
1628                         new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1629                 }
1630         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1631
1632         if (new & (1U << BTREE_NODE_write_in_flight))
1633                 __bch2_btree_node_write(c, b, true);
1634
1635         six_unlock_read(&b->c.lock);
1636 }
1637
1638 static void btree_node_write_work(struct work_struct *work)
1639 {
1640         struct btree_write_bio *wbio =
1641                 container_of(work, struct btree_write_bio, work);
1642         struct bch_fs *c        = wbio->wbio.c;
1643         struct btree *b         = wbio->wbio.bio.bi_private;
1644         struct bch_extent_ptr *ptr;
1645         int ret;
1646
1647         btree_bounce_free(c,
1648                 wbio->data_bytes,
1649                 wbio->wbio.used_mempool,
1650                 wbio->data);
1651
1652         bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1653                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1654
1655         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key)))
1656                 goto err;
1657
1658         if (wbio->wbio.first_btree_write) {
1659                 if (wbio->wbio.failed.nr) {
1660
1661                 }
1662         } else {
1663                 ret = bch2_trans_do(c, NULL, NULL, 0,
1664                         bch2_btree_node_update_key_get_iter(&trans, b, &wbio->key,
1665                                                             !wbio->wbio.failed.nr));
1666                 if (ret)
1667                         goto err;
1668         }
1669 out:
1670         bio_put(&wbio->wbio.bio);
1671         btree_node_write_done(c, b);
1672         return;
1673 err:
1674         set_btree_node_noevict(b);
1675         bch2_fs_fatal_error(c, "fatal error writing btree node");
1676         goto out;
1677 }
1678
1679 static void btree_node_write_endio(struct bio *bio)
1680 {
1681         struct bch_write_bio *wbio      = to_wbio(bio);
1682         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1683         struct bch_write_bio *orig      = parent ?: wbio;
1684         struct btree_write_bio *wb      = container_of(orig, struct btree_write_bio, wbio);
1685         struct bch_fs *c                = wbio->c;
1686         struct btree *b                 = wbio->bio.bi_private;
1687         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1688         unsigned long flags;
1689
1690         if (wbio->have_ioref)
1691                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1692
1693         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1694                                bch2_blk_status_to_str(bio->bi_status)) ||
1695             bch2_meta_write_fault("btree")) {
1696                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1697                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1698                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1699         }
1700
1701         if (wbio->have_ioref)
1702                 percpu_ref_put(&ca->io_ref);
1703
1704         if (parent) {
1705                 bio_put(bio);
1706                 bio_endio(&parent->bio);
1707                 return;
1708         }
1709
1710         clear_btree_node_write_in_flight_inner(b);
1711         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1712         INIT_WORK(&wb->work, btree_node_write_work);
1713         queue_work(c->btree_io_complete_wq, &wb->work);
1714 }
1715
1716 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1717                                    struct bset *i, unsigned sectors)
1718 {
1719         unsigned whiteout_u64s = 0;
1720         int ret;
1721
1722         if (bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), BKEY_TYPE_btree))
1723                 return -1;
1724
1725         ret = validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false) ?:
1726                 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false);
1727         if (ret) {
1728                 bch2_inconsistent_error(c);
1729                 dump_stack();
1730         }
1731
1732         return ret;
1733 }
1734
1735 static void btree_write_submit(struct work_struct *work)
1736 {
1737         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1738         struct bch_extent_ptr *ptr;
1739         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1740
1741         bkey_copy(&tmp.k, &wbio->key);
1742
1743         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1744                 ptr->offset += wbio->sector_offset;
1745
1746         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree, &tmp.k);
1747 }
1748
1749 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, bool already_started)
1750 {
1751         struct btree_write_bio *wbio;
1752         struct bset_tree *t;
1753         struct bset *i;
1754         struct btree_node *bn = NULL;
1755         struct btree_node_entry *bne = NULL;
1756         struct sort_iter sort_iter;
1757         struct nonce nonce;
1758         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1759         u64 seq = 0;
1760         bool used_mempool;
1761         unsigned long old, new;
1762         bool validate_before_checksum = false;
1763         void *data;
1764
1765         if (already_started)
1766                 goto do_write;
1767
1768         if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
1769                 return;
1770
1771         /*
1772          * We may only have a read lock on the btree node - the dirty bit is our
1773          * "lock" against racing with other threads that may be trying to start
1774          * a write, we do a write iff we clear the dirty bit. Since setting the
1775          * dirty bit requires a write lock, we can't race with other threads
1776          * redirtying it:
1777          */
1778         do {
1779                 old = new = READ_ONCE(b->flags);
1780
1781                 if (!(old & (1 << BTREE_NODE_dirty)))
1782                         return;
1783
1784                 if (!btree_node_may_write(b))
1785                         return;
1786
1787                 if (old & (1 << BTREE_NODE_never_write))
1788                         return;
1789
1790                 BUG_ON(old & (1 << BTREE_NODE_write_in_flight));
1791
1792                 new &= ~(1 << BTREE_NODE_dirty);
1793                 new &= ~(1 << BTREE_NODE_need_write);
1794                 new |=  (1 << BTREE_NODE_write_in_flight);
1795                 new |=  (1 << BTREE_NODE_write_in_flight_inner);
1796                 new |=  (1 << BTREE_NODE_just_written);
1797                 new ^=  (1 << BTREE_NODE_write_idx);
1798         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1799
1800         if (new & (1U << BTREE_NODE_need_write))
1801                 return;
1802 do_write:
1803         atomic_dec(&c->btree_cache.dirty);
1804
1805         BUG_ON(btree_node_fake(b));
1806         BUG_ON((b->will_make_reachable != 0) != !b->written);
1807
1808         BUG_ON(b->written >= c->opts.btree_node_size);
1809         BUG_ON(b->written & (c->opts.block_size - 1));
1810         BUG_ON(bset_written(b, btree_bset_last(b)));
1811         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1812         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1813
1814         bch2_sort_whiteouts(c, b);
1815
1816         sort_iter_init(&sort_iter, b);
1817
1818         bytes = !b->written
1819                 ? sizeof(struct btree_node)
1820                 : sizeof(struct btree_node_entry);
1821
1822         bytes += b->whiteout_u64s * sizeof(u64);
1823
1824         for_each_bset(b, t) {
1825                 i = bset(b, t);
1826
1827                 if (bset_written(b, i))
1828                         continue;
1829
1830                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1831                 sort_iter_add(&sort_iter,
1832                               btree_bkey_first(b, t),
1833                               btree_bkey_last(b, t));
1834                 seq = max(seq, le64_to_cpu(i->journal_seq));
1835         }
1836
1837         BUG_ON(b->written && !seq);
1838
1839         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1840         bytes += 8;
1841
1842         /* buffer must be a multiple of the block size */
1843         bytes = round_up(bytes, block_bytes(c));
1844
1845         data = btree_bounce_alloc(c, bytes, &used_mempool);
1846
1847         if (!b->written) {
1848                 bn = data;
1849                 *bn = *b->data;
1850                 i = &bn->keys;
1851         } else {
1852                 bne = data;
1853                 bne->keys = b->data->keys;
1854                 i = &bne->keys;
1855         }
1856
1857         i->journal_seq  = cpu_to_le64(seq);
1858         i->u64s         = 0;
1859
1860         sort_iter_add(&sort_iter,
1861                       unwritten_whiteouts_start(c, b),
1862                       unwritten_whiteouts_end(c, b));
1863         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1864
1865         b->whiteout_u64s = 0;
1866
1867         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1868         le16_add_cpu(&i->u64s, u64s);
1869
1870         set_needs_whiteout(i, false);
1871
1872         /* do we have data to write? */
1873         if (b->written && !i->u64s)
1874                 goto nowrite;
1875
1876         bytes_to_write = vstruct_end(i) - data;
1877         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1878
1879         memset(data + bytes_to_write, 0,
1880                (sectors_to_write << 9) - bytes_to_write);
1881
1882         BUG_ON(b->written + sectors_to_write > c->opts.btree_node_size);
1883         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1884         BUG_ON(i->seq != b->data->keys.seq);
1885
1886         i->version = c->sb.version < bcachefs_metadata_version_new_versioning
1887                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1888                 : cpu_to_le16(c->sb.version);
1889         SET_BSET_OFFSET(i, b->written);
1890         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1891
1892         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1893                 validate_before_checksum = true;
1894
1895         /* validate_bset will be modifying: */
1896         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
1897                 validate_before_checksum = true;
1898
1899         /* if we're going to be encrypting, check metadata validity first: */
1900         if (validate_before_checksum &&
1901             validate_bset_for_write(c, b, i, sectors_to_write))
1902                 goto err;
1903
1904         bset_encrypt(c, i, b->written << 9);
1905
1906         nonce = btree_nonce(i, b->written << 9);
1907
1908         if (bn)
1909                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1910         else
1911                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1912
1913         /* if we're not encrypting, check metadata after checksumming: */
1914         if (!validate_before_checksum &&
1915             validate_bset_for_write(c, b, i, sectors_to_write))
1916                 goto err;
1917
1918         /*
1919          * We handle btree write errors by immediately halting the journal -
1920          * after we've done that, we can't issue any subsequent btree writes
1921          * because they might have pointers to new nodes that failed to write.
1922          *
1923          * Furthermore, there's no point in doing any more btree writes because
1924          * with the journal stopped, we're never going to update the journal to
1925          * reflect that those writes were done and the data flushed from the
1926          * journal:
1927          *
1928          * Also on journal error, the pending write may have updates that were
1929          * never journalled (interior nodes, see btree_update_nodes_written()) -
1930          * it's critical that we don't do the write in that case otherwise we
1931          * will have updates visible that weren't in the journal:
1932          *
1933          * Make sure to update b->written so bch2_btree_init_next() doesn't
1934          * break:
1935          */
1936         if (bch2_journal_error(&c->journal) ||
1937             c->opts.nochanges)
1938                 goto err;
1939
1940         trace_btree_write(b, bytes_to_write, sectors_to_write);
1941
1942         wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1943                                 buf_pages(data, sectors_to_write << 9),
1944                                 &c->btree_bio),
1945                             struct btree_write_bio, wbio.bio);
1946         wbio_init(&wbio->wbio.bio);
1947         wbio->data                      = data;
1948         wbio->data_bytes                = bytes;
1949         wbio->sector_offset             = b->written;
1950         wbio->wbio.c                    = c;
1951         wbio->wbio.used_mempool         = used_mempool;
1952         wbio->wbio.first_btree_write    = !b->written;
1953         wbio->wbio.bio.bi_opf           = REQ_OP_WRITE|REQ_META;
1954         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
1955         wbio->wbio.bio.bi_private       = b;
1956
1957         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
1958
1959         bkey_copy(&wbio->key, &b->key);
1960
1961         b->written += sectors_to_write;
1962
1963         if (wbio->wbio.first_btree_write &&
1964             b->key.k.type == KEY_TYPE_btree_ptr_v2)
1965                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
1966                         cpu_to_le16(b->written);
1967
1968         if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
1969                 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
1970                         cpu_to_le16(b->written);
1971
1972         atomic64_inc(&c->btree_writes_nr);
1973         atomic64_add(sectors_to_write, &c->btree_writes_sectors);
1974
1975         INIT_WORK(&wbio->work, btree_write_submit);
1976         queue_work(c->io_complete_wq, &wbio->work);
1977         return;
1978 err:
1979         set_btree_node_noevict(b);
1980         if (!b->written &&
1981             b->key.k.type == KEY_TYPE_btree_ptr_v2)
1982                 bkey_i_to_btree_ptr_v2(&b->key)->v.sectors_written =
1983                         cpu_to_le16(sectors_to_write);
1984         b->written += sectors_to_write;
1985 nowrite:
1986         btree_bounce_free(c, bytes, used_mempool, data);
1987         btree_node_write_done(c, b);
1988 }
1989
1990 /*
1991  * Work that must be done with write lock held:
1992  */
1993 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
1994 {
1995         bool invalidated_iter = false;
1996         struct btree_node_entry *bne;
1997         struct bset_tree *t;
1998
1999         if (!btree_node_just_written(b))
2000                 return false;
2001
2002         BUG_ON(b->whiteout_u64s);
2003
2004         clear_btree_node_just_written(b);
2005
2006         /*
2007          * Note: immediately after write, bset_written() doesn't work - the
2008          * amount of data we had to write after compaction might have been
2009          * smaller than the offset of the last bset.
2010          *
2011          * However, we know that all bsets have been written here, as long as
2012          * we're still holding the write lock:
2013          */
2014
2015         /*
2016          * XXX: decide if we really want to unconditionally sort down to a
2017          * single bset:
2018          */
2019         if (b->nsets > 1) {
2020                 btree_node_sort(c, b, 0, b->nsets, true);
2021                 invalidated_iter = true;
2022         } else {
2023                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2024         }
2025
2026         for_each_bset(b, t)
2027                 set_needs_whiteout(bset(b, t), true);
2028
2029         bch2_btree_verify(c, b);
2030
2031         /*
2032          * If later we don't unconditionally sort down to a single bset, we have
2033          * to ensure this is still true:
2034          */
2035         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2036
2037         bne = want_new_bset(c, b);
2038         if (bne)
2039                 bch2_bset_init_next(c, b, bne);
2040
2041         bch2_btree_build_aux_trees(b);
2042
2043         return invalidated_iter;
2044 }
2045
2046 /*
2047  * Use this one if the node is intent locked:
2048  */
2049 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2050                            enum six_lock_type lock_type_held)
2051 {
2052         if (lock_type_held == SIX_LOCK_intent ||
2053             (lock_type_held == SIX_LOCK_read &&
2054              six_lock_tryupgrade(&b->c.lock))) {
2055                 __bch2_btree_node_write(c, b, false);
2056
2057                 /* don't cycle lock unnecessarily: */
2058                 if (btree_node_just_written(b) &&
2059                     six_trylock_write(&b->c.lock)) {
2060                         bch2_btree_post_write_cleanup(c, b);
2061                         six_unlock_write(&b->c.lock);
2062                 }
2063
2064                 if (lock_type_held == SIX_LOCK_read)
2065                         six_lock_downgrade(&b->c.lock);
2066         } else {
2067                 __bch2_btree_node_write(c, b, false);
2068                 if (lock_type_held == SIX_LOCK_write &&
2069                     btree_node_just_written(b))
2070                         bch2_btree_post_write_cleanup(c, b);
2071         }
2072 }
2073
2074 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2075 {
2076         struct bucket_table *tbl;
2077         struct rhash_head *pos;
2078         struct btree *b;
2079         unsigned i;
2080 restart:
2081         rcu_read_lock();
2082         for_each_cached_btree(b, c, tbl, i, pos)
2083                 if (test_bit(flag, &b->flags)) {
2084                         rcu_read_unlock();
2085                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2086                         goto restart;
2087
2088                 }
2089         rcu_read_unlock();
2090 }
2091
2092 void bch2_btree_flush_all_reads(struct bch_fs *c)
2093 {
2094         __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2095 }
2096
2097 void bch2_btree_flush_all_writes(struct bch_fs *c)
2098 {
2099         __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2100 }
2101
2102 void bch2_dirty_btree_nodes_to_text(struct printbuf *out, struct bch_fs *c)
2103 {
2104         struct bucket_table *tbl;
2105         struct rhash_head *pos;
2106         struct btree *b;
2107         unsigned i;
2108
2109         rcu_read_lock();
2110         for_each_cached_btree(b, c, tbl, i, pos) {
2111                 unsigned long flags = READ_ONCE(b->flags);
2112
2113                 if (!(flags & (1 << BTREE_NODE_dirty)))
2114                         continue;
2115
2116                 pr_buf(out, "%p d %u n %u l %u w %u b %u r %u:%lu\n",
2117                        b,
2118                        (flags & (1 << BTREE_NODE_dirty)) != 0,
2119                        (flags & (1 << BTREE_NODE_need_write)) != 0,
2120                        b->c.level,
2121                        b->written,
2122                        !list_empty_careful(&b->write_blocked),
2123                        b->will_make_reachable != 0,
2124                        b->will_make_reachable & 1);
2125         }
2126         rcu_read_unlock();
2127 }