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 static void verify_no_dups(struct btree *b,
  26                            struct bkey_packed *start,
  27                            struct bkey_packed *end)
  28 {
  29 #ifdef CONFIG_BCACHEFS_DEBUG
  30         struct bkey_packed *k, *p;
  31
  32         if (start == end)
  33                 return;
  34
  35         for (p = start, k = bkey_next_skip_noops(start, end);
  36              k != end;
  37              p = k, k = bkey_next_skip_noops(k, end)) {
  38                 struct bkey l = bkey_unpack_key(b, p);
  39                 struct bkey r = bkey_unpack_key(b, k);
  40
  41                 BUG_ON(bkey_cmp(l.p, bkey_start_pos(&r)) >= 0);
  42         }
  43 #endif
  44 }
  45
  46 static void set_needs_whiteout(struct bset *i, int v)
  47 {
  48         struct bkey_packed *k;
  49
  50         for (k = i->start;
  51              k != vstruct_last(i);
  52              k = bkey_next_skip_noops(k, vstruct_last(i)))
  53                 k->needs_whiteout = v;
  54 }
  55
  56 static void btree_bounce_free(struct bch_fs *c, size_t size,
  57                               bool used_mempool, void *p)
  58 {
  59         if (used_mempool)
  60                 mempool_free(p, &c->btree_bounce_pool);
  61         else
  62                 vpfree(p, size);
  63 }
  64
  65 static void *btree_bounce_alloc(struct bch_fs *c, size_t size,
  66                                 bool *used_mempool)
  67 {
  68         unsigned flags = memalloc_nofs_save();
  69         void *p;
  70
  71         BUG_ON(size > btree_bytes(c));
  72
  73         *used_mempool = false;
  74         p = vpmalloc(size, __GFP_NOWARN|GFP_NOWAIT);
  75         if (!p) {
  76                 *used_mempool = true;
  77                 p = mempool_alloc(&c->btree_bounce_pool, GFP_NOIO);
  78         }
  79         memalloc_nofs_restore(flags);
  80         return p;
  81 }
  82
  83 static void sort_bkey_ptrs(const struct btree *bt,
  84                            struct bkey_packed **ptrs, unsigned nr)
  85 {
  86         unsigned n = nr, a = nr / 2, b, c, d;
  87
  88         if (!a)
  89                 return;
  90
  91         /* Heap sort: see lib/sort.c: */
  92         while (1) {
  93                 if (a)
  94                         a--;
  95                 else if (--n)
  96                         swap(ptrs[0], ptrs[n]);
  97                 else
  98                         break;
  99
 100                 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
 101                         b = bch2_bkey_cmp_packed(bt,
 102                                             ptrs[c],
 103                                             ptrs[d]) >= 0 ? c : d;
 104                 if (d == n)
 105                         b = c;
 106
 107                 while (b != a &&
 108                        bch2_bkey_cmp_packed(bt,
 109                                        ptrs[a],
 110                                        ptrs[b]) >= 0)
 111                         b = (b - 1) / 2;
 112                 c = b;
 113                 while (b != a) {
 114                         b = (b - 1) / 2;
 115                         swap(ptrs[b], ptrs[c]);
 116                 }
 117         }
 118 }
 119
 120 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
 121 {
 122         struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
 123         bool used_mempool = false;
 124         size_t bytes = b->whiteout_u64s * sizeof(u64);
 125
 126         if (!b->whiteout_u64s)
 127                 return;
 128
 129         new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
 130
 131         ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
 132
 133         for (k = unwritten_whiteouts_start(c, b);
 134              k != unwritten_whiteouts_end(c, b);
 135              k = bkey_next(k))
 136                 *--ptrs = k;
 137
 138         sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
 139
 140         k = new_whiteouts;
 141
 142         while (ptrs != ptrs_end) {
 143                 bkey_copy(k, *ptrs);
 144                 k = bkey_next(k);
 145                 ptrs++;
 146         }
 147
 148         verify_no_dups(b, new_whiteouts,
 149                        (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
 150
 151         memcpy_u64s(unwritten_whiteouts_start(c, b),
 152                     new_whiteouts, b->whiteout_u64s);
 153
 154         btree_bounce_free(c, bytes, used_mempool, new_whiteouts);
 155 }
 156
 157 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
 158                                 bool compacting, enum compact_mode mode)
 159 {
 160         if (!bset_dead_u64s(b, t))
 161                 return false;
 162
 163         switch (mode) {
 164         case COMPACT_LAZY:
 165                 return should_compact_bset_lazy(b, t) ||
 166                         (compacting && !bset_written(b, bset(b, t)));
 167         case COMPACT_ALL:
 168                 return true;
 169         default:
 170                 BUG();
 171         }
 172 }
 173
 174 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
 175 {
 176         struct bset_tree *t;
 177         bool ret = false;
 178
 179         for_each_bset(b, t) {
 180                 struct bset *i = bset(b, t);
 181                 struct bkey_packed *k, *n, *out, *start, *end;
 182                 struct btree_node_entry *src = NULL, *dst = NULL;
 183
 184                 if (t != b->set && !bset_written(b, i)) {
 185                         src = container_of(i, struct btree_node_entry, keys);
 186                         dst = max(write_block(b),
 187                                   (void *) btree_bkey_last(b, t - 1));
 188                 }
 189
 190                 if (src != dst)
 191                         ret = true;
 192
 193                 if (!should_compact_bset(b, t, ret, mode)) {
 194                         if (src != dst) {
 195                                 memmove(dst, src, sizeof(*src) +
 196                                         le16_to_cpu(src->keys.u64s) *
 197                                         sizeof(u64));
 198                                 i = &dst->keys;
 199                                 set_btree_bset(b, t, i);
 200                         }
 201                         continue;
 202                 }
 203
 204                 start   = btree_bkey_first(b, t);
 205                 end     = btree_bkey_last(b, t);
 206
 207                 if (src != dst) {
 208                         memmove(dst, src, sizeof(*src));
 209                         i = &dst->keys;
 210                         set_btree_bset(b, t, i);
 211                 }
 212
 213                 out = i->start;
 214
 215                 for (k = start; k != end; k = n) {
 216                         n = bkey_next_skip_noops(k, end);
 217
 218                         if (!bkey_deleted(k)) {
 219                                 bkey_copy(out, k);
 220                                 out = bkey_next(out);
 221                         } else {
 222                                 BUG_ON(k->needs_whiteout);
 223                         }
 224                 }
 225
 226                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 227                 set_btree_bset_end(b, t);
 228                 bch2_bset_set_no_aux_tree(b, t);
 229                 ret = true;
 230         }
 231
 232         bch2_verify_btree_nr_keys(b);
 233
 234         bch2_btree_build_aux_trees(b);
 235
 236         return ret;
 237 }
 238
 239 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 240                             enum compact_mode mode)
 241 {
 242         return bch2_drop_whiteouts(b, mode);
 243 }
 244
 245 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 246                             struct btree_iter *iter,
 247                             unsigned start_idx,
 248                             unsigned end_idx,
 249                             bool filter_whiteouts)
 250 {
 251         struct btree_node *out;
 252         struct sort_iter sort_iter;
 253         struct bset_tree *t;
 254         struct bset *start_bset = bset(b, &b->set[start_idx]);
 255         bool used_mempool = false;
 256         u64 start_time, seq = 0;
 257         unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
 258         bool sorting_entire_node = start_idx == 0 &&
 259                 end_idx == b->nsets;
 260
 261         sort_iter_init(&sort_iter, b);
 262
 263         for (t = b->set + start_idx;
 264              t < b->set + end_idx;
 265              t++) {
 266                 u64s += le16_to_cpu(bset(b, t)->u64s);
 267                 sort_iter_add(&sort_iter,
 268                               btree_bkey_first(b, t),
 269                               btree_bkey_last(b, t));
 270         }
 271
 272         bytes = sorting_entire_node
 273                 ? btree_bytes(c)
 274                 : __vstruct_bytes(struct btree_node, u64s);
 275
 276         out = btree_bounce_alloc(c, bytes, &used_mempool);
 277
 278         start_time = local_clock();
 279
 280         u64s = bch2_sort_keys(out->keys.start, &sort_iter, filter_whiteouts);
 281
 282         out->keys.u64s = cpu_to_le16(u64s);
 283
 284         BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
 285
 286         if (sorting_entire_node)
 287                 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 288                                        start_time);
 289
 290         /* Make sure we preserve bset journal_seq: */
 291         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 292                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 293         start_bset->journal_seq = cpu_to_le64(seq);
 294
 295         if (sorting_entire_node) {
 296                 unsigned u64s = le16_to_cpu(out->keys.u64s);
 297
 298                 BUG_ON(bytes != btree_bytes(c));
 299
 300                 /*
 301                  * Our temporary buffer is the same size as the btree node's
 302                  * buffer, we can just swap buffers instead of doing a big
 303                  * memcpy()
 304                  */
 305                 *out = *b->data;
 306                 out->keys.u64s = cpu_to_le16(u64s);
 307                 swap(out, b->data);
 308                 set_btree_bset(b, b->set, &b->data->keys);
 309         } else {
 310                 start_bset->u64s = out->keys.u64s;
 311                 memcpy_u64s(start_bset->start,
 312                             out->keys.start,
 313                             le16_to_cpu(out->keys.u64s));
 314         }
 315
 316         for (i = start_idx + 1; i < end_idx; i++)
 317                 b->nr.bset_u64s[start_idx] +=
 318                         b->nr.bset_u64s[i];
 319
 320         b->nsets -= shift;
 321
 322         for (i = start_idx + 1; i < b->nsets; i++) {
 323                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 324                 b->set[i]               = b->set[i + shift];
 325         }
 326
 327         for (i = b->nsets; i < MAX_BSETS; i++)
 328                 b->nr.bset_u64s[i] = 0;
 329
 330         set_btree_bset_end(b, &b->set[start_idx]);
 331         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 332
 333         btree_bounce_free(c, bytes, used_mempool, out);
 334
 335         bch2_verify_btree_nr_keys(b);
 336 }
 337
 338 void bch2_btree_sort_into(struct bch_fs *c,
 339                          struct btree *dst,
 340                          struct btree *src)
 341 {
 342         struct btree_nr_keys nr;
 343         struct btree_node_iter src_iter;
 344         u64 start_time = local_clock();
 345
 346         BUG_ON(dst->nsets != 1);
 347
 348         bch2_bset_set_no_aux_tree(dst, dst->set);
 349
 350         bch2_btree_node_iter_init_from_start(&src_iter, src);
 351
 352         if (btree_node_is_extents(src))
 353                 nr = bch2_sort_repack_merge(c, btree_bset_first(dst),
 354                                 src, &src_iter,
 355                                 &dst->format,
 356                                 true);
 357         else
 358                 nr = bch2_sort_repack(btree_bset_first(dst),
 359                                 src, &src_iter,
 360                                 &dst->format,
 361                                 true);
 362
 363         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 364                                start_time);
 365
 366         set_btree_bset_end(dst, dst->set);
 367
 368         dst->nr.live_u64s       += nr.live_u64s;
 369         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 370         dst->nr.packed_keys     += nr.packed_keys;
 371         dst->nr.unpacked_keys   += nr.unpacked_keys;
 372
 373         bch2_verify_btree_nr_keys(dst);
 374 }
 375
 376 #define SORT_CRIT       (4096 / sizeof(u64))
 377
 378 /*
 379  * We're about to add another bset to the btree node, so if there's currently
 380  * too many bsets - sort some of them together:
 381  */
 382 static bool btree_node_compact(struct bch_fs *c, struct btree *b,
 383                                struct btree_iter *iter)
 384 {
 385         unsigned unwritten_idx;
 386         bool ret = false;
 387
 388         for (unwritten_idx = 0;
 389              unwritten_idx < b->nsets;
 390              unwritten_idx++)
 391                 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
 392                         break;
 393
 394         if (b->nsets - unwritten_idx > 1) {
 395                 btree_node_sort(c, b, iter, unwritten_idx,
 396                                 b->nsets, false);
 397                 ret = true;
 398         }
 399
 400         if (unwritten_idx > 1) {
 401                 btree_node_sort(c, b, iter, 0, unwritten_idx, false);
 402                 ret = true;
 403         }
 404
 405         return ret;
 406 }
 407
 408 void bch2_btree_build_aux_trees(struct btree *b)
 409 {
 410         struct bset_tree *t;
 411
 412         for_each_bset(b, t)
 413                 bch2_bset_build_aux_tree(b, t,
 414                                 !bset_written(b, bset(b, t)) &&
 415                                 t == bset_tree_last(b));
 416 }
 417
 418 /*
 419  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 420  * inserted into
 421  *
 422  * Safe to call if there already is an unwritten bset - will only add a new bset
 423  * if @b doesn't already have one.
 424  *
 425  * Returns true if we sorted (i.e. invalidated iterators
 426  */
 427 void bch2_btree_init_next(struct bch_fs *c, struct btree *b,
 428                           struct btree_iter *iter)
 429 {
 430         struct btree_node_entry *bne;
 431         bool did_sort;
 432
 433         EBUG_ON(!(b->c.lock.state.seq & 1));
 434         EBUG_ON(iter && iter->l[b->c.level].b != b);
 435
 436         did_sort = btree_node_compact(c, b, iter);
 437
 438         bne = want_new_bset(c, b);
 439         if (bne)
 440                 bch2_bset_init_next(c, b, bne);
 441
 442         bch2_btree_build_aux_trees(b);
 443
 444         if (iter && did_sort)
 445                 bch2_btree_iter_reinit_node(iter, b);
 446 }
 447
 448 static void btree_pos_to_text(struct printbuf *out, struct bch_fs *c,
 449                           struct btree *b)
 450 {
 451         pr_buf(out, "%s level %u/%u\n  ",
 452                bch2_btree_ids[b->c.btree_id],
 453                b->c.level,
 454                c->btree_roots[b->c.btree_id].level);
 455         bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 456 }
 457
 458 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 459                           struct bch_dev *ca,
 460                           struct btree *b, struct bset *i,
 461                           unsigned offset, int write)
 462 {
 463         pr_buf(out, "error validating btree node ");
 464         if (write)
 465                 pr_buf(out, "before write ");
 466         if (ca)
 467                 pr_buf(out, "on %s ", ca->name);
 468         pr_buf(out, "at btree ");
 469         btree_pos_to_text(out, c, b);
 470
 471         pr_buf(out, "\n  node offset %u", b->written);
 472         if (i)
 473                 pr_buf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 474 }
 475
 476 enum btree_err_type {
 477         BTREE_ERR_FIXABLE,
 478         BTREE_ERR_WANT_RETRY,
 479         BTREE_ERR_MUST_RETRY,
 480         BTREE_ERR_FATAL,
 481 };
 482
 483 enum btree_validate_ret {
 484         BTREE_RETRY_READ = 64,
 485 };
 486
 487 #define btree_err(type, c, ca, b, i, msg, ...)                          \
 488 ({                                                                      \
 489         __label__ out;                                                  \
 490         char _buf[300];                                                 \
 491         char *buf2 = _buf;                                              \
 492         struct printbuf out = PBUF(_buf);                               \
 493                                                                         \
 494         buf2 = kmalloc(4096, GFP_ATOMIC);                               \
 495         if (buf2)                                                       \
 496                 out = _PBUF(buf2, 4986);                                \
 497                                                                         \
 498         btree_err_msg(&out, c, ca, b, i, b->written, write);            \
 499         pr_buf(&out, ": " msg, ##__VA_ARGS__);                          \
 500                                                                         \
 501         if (type == BTREE_ERR_FIXABLE &&                                \
 502             write == READ &&                                            \
 503             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {             \
 504                 mustfix_fsck_err(c, "%s", buf2);                        \
 505                 goto out;                                               \
 506         }                                                               \
 507                                                                         \
 508         switch (write) {                                                \
 509         case READ:                                                      \
 510                 bch_err(c, "%s", buf2);                                 \
 511                                                                         \
 512                 switch (type) {                                         \
 513                 case BTREE_ERR_FIXABLE:                                 \
 514                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 515                         goto fsck_err;                                  \
 516                 case BTREE_ERR_WANT_RETRY:                              \
 517                         if (have_retry) {                               \
 518                                 ret = BTREE_RETRY_READ;                 \
 519                                 goto fsck_err;                          \
 520                         }                                               \
 521                         break;                                          \
 522                 case BTREE_ERR_MUST_RETRY:                              \
 523                         ret = BTREE_RETRY_READ;                         \
 524                         goto fsck_err;                                  \
 525                 case BTREE_ERR_FATAL:                                   \
 526                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 527                         goto fsck_err;                                  \
 528                 }                                                       \
 529                 break;                                                  \
 530         case WRITE:                                                     \
 531                 bch_err(c, "corrupt metadata before write: %s", buf2);  \
 532                                                                         \
 533                 if (bch2_fs_inconsistent(c)) {                          \
 534                         ret = BCH_FSCK_ERRORS_NOT_FIXED;                \
 535                         goto fsck_err;                                  \
 536                 }                                                       \
 537                 break;                                                  \
 538         }                                                               \
 539 out:                                                                    \
 540         if (buf2 != _buf)                                               \
 541                 kfree(buf2);                                            \
 542         true;                                                           \
 543 })
 544
 545 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 546
 547 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 548                          struct btree *b, struct bset *i,
 549                          unsigned sectors, int write, bool have_retry)
 550 {
 551         unsigned version = le16_to_cpu(i->version);
 552         const char *err;
 553         int ret = 0;
 554
 555         btree_err_on((version != BCH_BSET_VERSION_OLD &&
 556                       version < bcachefs_metadata_version_min) ||
 557                      version >= bcachefs_metadata_version_max,
 558                      BTREE_ERR_FATAL, c, ca, b, i,
 559                      "unsupported bset version");
 560
 561         if (btree_err_on(b->written + sectors > c->opts.btree_node_size,
 562                          BTREE_ERR_FIXABLE, c, ca, b, i,
 563                          "bset past end of btree node")) {
 564                 i->u64s = 0;
 565                 return 0;
 566         }
 567
 568         btree_err_on(b->written && !i->u64s,
 569                      BTREE_ERR_FIXABLE, c, ca, b, i,
 570                      "empty bset");
 571
 572         if (!b->written) {
 573                 struct btree_node *bn =
 574                         container_of(i, struct btree_node, keys);
 575                 /* These indicate that we read the wrong btree node: */
 576
 577                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 578                         struct bch_btree_ptr_v2 *bp =
 579                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 580
 581                         /* XXX endianness */
 582                         btree_err_on(bp->seq != bn->keys.seq,
 583                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 584                                      "incorrect sequence number (wrong btree node)");
 585                 }
 586
 587                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 588                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 589                              "incorrect btree id");
 590
 591                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 592                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 593                              "incorrect level");
 594
 595                 if (BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN) {
 596                         u64 *p = (u64 *) &bn->ptr;
 597
 598                         *p = swab64(*p);
 599                 }
 600
 601                 if (!write)
 602                         compat_btree_node(b->c.level, b->c.btree_id, version,
 603                                           BSET_BIG_ENDIAN(i), write, bn);
 604
 605                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 606                         struct bch_btree_ptr_v2 *bp =
 607                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 608
 609                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 610                                 b->data->min_key = bp->min_key;
 611                                 b->data->max_key = b->key.k.p;
 612                         }
 613
 614                         btree_err_on(bkey_cmp(b->data->min_key, bp->min_key),
 615                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 616                                      "incorrect min_key: got %llu:%llu should be %llu:%llu",
 617                                      b->data->min_key.inode,
 618                                      b->data->min_key.offset,
 619                                      bp->min_key.inode,
 620                                      bp->min_key.offset);
 621                 }
 622
 623                 btree_err_on(bkey_cmp(bn->max_key, b->key.k.p),
 624                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 625                              "incorrect max key %llu:%llu",
 626                              bn->max_key.inode,
 627                              bn->max_key.offset);
 628
 629                 if (write)
 630                         compat_btree_node(b->c.level, b->c.btree_id, version,
 631                                           BSET_BIG_ENDIAN(i), write, bn);
 632
 633                 /* XXX: ideally we would be validating min_key too */
 634 #if 0
 635                 /*
 636                  * not correct anymore, due to btree node write error
 637                  * handling
 638                  *
 639                  * need to add bn->seq to btree keys and verify
 640                  * against that
 641                  */
 642                 btree_err_on(!extent_contains_ptr(bkey_i_to_s_c_extent(&b->key),
 643                                                   bn->ptr),
 644                              BTREE_ERR_FATAL, c, b, i,
 645                              "incorrect backpointer");
 646 #endif
 647                 err = bch2_bkey_format_validate(&bn->format);
 648                 btree_err_on(err,
 649                              BTREE_ERR_FATAL, c, ca, b, i,
 650                              "invalid bkey format: %s", err);
 651
 652                 compat_bformat(b->c.level, b->c.btree_id, version,
 653                                BSET_BIG_ENDIAN(i), write,
 654                                &bn->format);
 655         }
 656 fsck_err:
 657         return ret;
 658 }
 659
 660 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 661                          struct bset *i, unsigned *whiteout_u64s,
 662                          int write, bool have_retry)
 663 {
 664         unsigned version = le16_to_cpu(i->version);
 665         struct bkey_packed *k, *prev = NULL;
 666         bool seen_non_whiteout = false;
 667         int ret = 0;
 668
 669         if (!BSET_SEPARATE_WHITEOUTS(i)) {
 670                 seen_non_whiteout = true;
 671                 *whiteout_u64s = 0;
 672         }
 673
 674         for (k = i->start;
 675              k != vstruct_last(i);) {
 676                 struct bkey_s u;
 677                 struct bkey tmp;
 678                 const char *invalid;
 679
 680                 if (btree_err_on(bkey_next(k) > vstruct_last(i),
 681                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 682                                  "key extends past end of bset")) {
 683                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 684                         break;
 685                 }
 686
 687                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 688                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 689                                  "invalid bkey format %u", k->format)) {
 690                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 691                         memmove_u64s_down(k, bkey_next(k),
 692                                           (u64 *) vstruct_end(i) - (u64 *) k);
 693                         continue;
 694                 }
 695
 696                 /* XXX: validate k->u64s */
 697                 if (!write)
 698                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 699                                     BSET_BIG_ENDIAN(i), write,
 700                                     &b->format, k);
 701
 702                 u = __bkey_disassemble(b, k, &tmp);
 703
 704                 invalid = __bch2_bkey_invalid(c, u.s_c, btree_node_type(b)) ?:
 705                         bch2_bkey_in_btree_node(b, u.s_c) ?:
 706                         (write ? bch2_bkey_val_invalid(c, u.s_c) : NULL);
 707                 if (invalid) {
 708                         char buf[160];
 709
 710                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 711                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 712                                   "invalid bkey: %s\n%s", invalid, buf);
 713
 714                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 715                         memmove_u64s_down(k, bkey_next(k),
 716                                           (u64 *) vstruct_end(i) - (u64 *) k);
 717                         continue;
 718                 }
 719
 720                 if (write)
 721                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 722                                     BSET_BIG_ENDIAN(i), write,
 723                                     &b->format, k);
 724
 725                 /*
 726                  * with the separate whiteouts thing (used for extents), the
 727                  * second set of keys actually can have whiteouts too, so we
 728                  * can't solely go off bkey_deleted()...
 729                  */
 730
 731                 if (!seen_non_whiteout &&
 732                     (!bkey_deleted(k) ||
 733                      (prev && bkey_iter_cmp(b, prev, k) > 0))) {
 734                         *whiteout_u64s = k->_data - i->_data;
 735                         seen_non_whiteout = true;
 736                 } else if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 737                         char buf1[80];
 738                         char buf2[80];
 739                         struct bkey up = bkey_unpack_key(b, prev);
 740
 741                         bch2_bkey_to_text(&PBUF(buf1), &up);
 742                         bch2_bkey_to_text(&PBUF(buf2), u.k);
 743
 744                         bch2_dump_bset(c, b, i, 0);
 745                         btree_err(BTREE_ERR_FATAL, c, NULL, b, i,
 746                                   "keys out of order: %s > %s",
 747                                   buf1, buf2);
 748                         /* XXX: repair this */
 749                 }
 750
 751                 prev = k;
 752                 k = bkey_next_skip_noops(k, vstruct_last(i));
 753         }
 754 fsck_err:
 755         return ret;
 756 }
 757
 758 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 759                               struct btree *b, bool have_retry)
 760 {
 761         struct btree_node_entry *bne;
 762         struct sort_iter *iter;
 763         struct btree_node *sorted;
 764         struct bkey_packed *k;
 765         struct bch_extent_ptr *ptr;
 766         struct bset *i;
 767         bool used_mempool, blacklisted;
 768         unsigned u64s;
 769         int ret, retry_read = 0, write = READ;
 770
 771         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
 772         sort_iter_init(iter, b);
 773         iter->size = (btree_blocks(c) + 1) * 2;
 774
 775         if (bch2_meta_read_fault("btree"))
 776                 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 777                           "dynamic fault");
 778
 779         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 780                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 781                      "bad magic");
 782
 783         btree_err_on(!b->data->keys.seq,
 784                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 785                      "bad btree header");
 786
 787         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 788                 struct bch_btree_ptr_v2 *bp =
 789                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 790
 791                 btree_err_on(b->data->keys.seq != bp->seq,
 792                              BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 793                              "got wrong btree node (seq %llx want %llx)",
 794                              b->data->keys.seq, bp->seq);
 795         }
 796
 797         while (b->written < c->opts.btree_node_size) {
 798                 unsigned sectors, whiteout_u64s = 0;
 799                 struct nonce nonce;
 800                 struct bch_csum csum;
 801                 bool first = !b->written;
 802
 803                 if (!b->written) {
 804                         i = &b->data->keys;
 805
 806                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 807                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 808                                      "unknown checksum type %llu",
 809                                      BSET_CSUM_TYPE(i));
 810
 811                         nonce = btree_nonce(i, b->written << 9);
 812                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 813
 814                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 815                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 816                                      "invalid checksum");
 817
 818                         bset_encrypt(c, i, b->written << 9);
 819
 820                         btree_err_on(btree_node_is_extents(b) &&
 821                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
 822                                      BTREE_ERR_FATAL, c, NULL, b, NULL,
 823                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
 824
 825                         sectors = vstruct_sectors(b->data, c->block_bits);
 826                 } else {
 827                         bne = write_block(b);
 828                         i = &bne->keys;
 829
 830                         if (i->seq != b->data->keys.seq)
 831                                 break;
 832
 833                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 834                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 835                                      "unknown checksum type %llu",
 836                                      BSET_CSUM_TYPE(i));
 837
 838                         nonce = btree_nonce(i, b->written << 9);
 839                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
 840
 841                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
 842                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 843                                      "invalid checksum");
 844
 845                         bset_encrypt(c, i, b->written << 9);
 846
 847                         sectors = vstruct_sectors(bne, c->block_bits);
 848                 }
 849
 850                 ret = validate_bset(c, ca, b, i, sectors,
 851                                     READ, have_retry);
 852                 if (ret)
 853                         goto fsck_err;
 854
 855                 if (!b->written)
 856                         btree_node_set_format(b, b->data->format);
 857
 858                 ret = validate_bset_keys(c, b, i, &whiteout_u64s,
 859                                     READ, have_retry);
 860                 if (ret)
 861                         goto fsck_err;
 862
 863                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
 864
 865                 b->written += sectors;
 866
 867                 blacklisted = bch2_journal_seq_is_blacklisted(c,
 868                                         le64_to_cpu(i->journal_seq),
 869                                         true);
 870
 871                 btree_err_on(blacklisted && first,
 872                              BTREE_ERR_FIXABLE, c, ca, b, i,
 873                              "first btree node bset has blacklisted journal seq");
 874                 if (blacklisted && !first)
 875                         continue;
 876
 877                 sort_iter_add(iter, i->start,
 878                               vstruct_idx(i, whiteout_u64s));
 879
 880                 sort_iter_add(iter,
 881                               vstruct_idx(i, whiteout_u64s),
 882                               vstruct_last(i));
 883         }
 884
 885         for (bne = write_block(b);
 886              bset_byte_offset(b, bne) < btree_bytes(c);
 887              bne = (void *) bne + block_bytes(c))
 888                 btree_err_on(bne->keys.seq == b->data->keys.seq,
 889                              BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
 890                              "found bset signature after last bset");
 891
 892         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
 893         sorted->keys.u64s = 0;
 894
 895         set_btree_bset(b, b->set, &b->data->keys);
 896
 897         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
 898
 899         u64s = le16_to_cpu(sorted->keys.u64s);
 900         *sorted = *b->data;
 901         sorted->keys.u64s = cpu_to_le16(u64s);
 902         swap(sorted, b->data);
 903         set_btree_bset(b, b->set, &b->data->keys);
 904         b->nsets = 1;
 905
 906         BUG_ON(b->nr.live_u64s != u64s);
 907
 908         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
 909
 910         i = &b->data->keys;
 911         for (k = i->start; k != vstruct_last(i);) {
 912                 struct bkey tmp;
 913                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
 914                 const char *invalid = bch2_bkey_val_invalid(c, u.s_c);
 915
 916                 if (invalid ||
 917                     (bch2_inject_invalid_keys &&
 918                      !bversion_cmp(u.k->version, MAX_VERSION))) {
 919                         char buf[160];
 920
 921                         bch2_bkey_val_to_text(&PBUF(buf), c, u.s_c);
 922                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i,
 923                                   "invalid bkey %s: %s", buf, invalid);
 924
 925                         btree_keys_account_key_drop(&b->nr, 0, k);
 926
 927                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 928                         memmove_u64s_down(k, bkey_next(k),
 929                                           (u64 *) vstruct_end(i) - (u64 *) k);
 930                         set_btree_bset_end(b, b->set);
 931                         continue;
 932                 }
 933
 934                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
 935                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
 936
 937                         bp.v->mem_ptr = 0;
 938                 }
 939
 940                 k = bkey_next_skip_noops(k, vstruct_last(i));
 941         }
 942
 943         bch2_bset_build_aux_tree(b, b->set, false);
 944
 945         set_needs_whiteout(btree_bset_first(b), true);
 946
 947         btree_node_reset_sib_u64s(b);
 948
 949         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
 950                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 951
 952                 if (ca->mi.state != BCH_MEMBER_STATE_RW)
 953                         set_btree_node_need_rewrite(b);
 954         }
 955 out:
 956         mempool_free(iter, &c->fill_iter);
 957         return retry_read;
 958 fsck_err:
 959         if (ret == BTREE_RETRY_READ) {
 960                 retry_read = 1;
 961         } else {
 962                 bch2_inconsistent_error(c);
 963                 set_btree_node_read_error(b);
 964         }
 965         goto out;
 966 }
 967
 968 static void btree_node_read_work(struct work_struct *work)
 969 {
 970         struct btree_read_bio *rb =
 971                 container_of(work, struct btree_read_bio, work);
 972         struct bch_fs *c        = rb->c;
 973         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
 974         struct btree *b         = rb->bio.bi_private;
 975         struct bio *bio         = &rb->bio;
 976         struct bch_io_failures failed = { .nr = 0 };
 977         char buf[200];
 978         struct printbuf out;
 979         bool can_retry;
 980
 981         goto start;
 982         while (1) {
 983                 bch_info(c, "retrying read");
 984                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
 985                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
 986                 bio_reset(bio);
 987                 bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
 988                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
 989                 bio->bi_iter.bi_size    = btree_bytes(c);
 990
 991                 if (rb->have_ioref) {
 992                         bio_set_dev(bio, ca->disk_sb.bdev);
 993                         submit_bio_wait(bio);
 994                 } else {
 995                         bio->bi_status = BLK_STS_REMOVED;
 996                 }
 997 start:
 998                 out = PBUF(buf);
 999                 btree_pos_to_text(&out, c, b);
1000                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1001                                    bch2_blk_status_to_str(bio->bi_status), buf);
1002                 if (rb->have_ioref)
1003                         percpu_ref_put(&ca->io_ref);
1004                 rb->have_ioref = false;
1005
1006                 bch2_mark_io_failure(&failed, &rb->pick);
1007
1008                 can_retry = bch2_bkey_pick_read_device(c,
1009                                 bkey_i_to_s_c(&b->key),
1010                                 &failed, &rb->pick) > 0;
1011
1012                 if (!bio->bi_status &&
1013                     !bch2_btree_node_read_done(c, ca, b, can_retry))
1014                         break;
1015
1016                 if (!can_retry) {
1017                         set_btree_node_read_error(b);
1018                         break;
1019                 }
1020         }
1021
1022         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1023                                rb->start_time);
1024         bio_put(&rb->bio);
1025         clear_btree_node_read_in_flight(b);
1026         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1027 }
1028
1029 static void btree_node_read_endio(struct bio *bio)
1030 {
1031         struct btree_read_bio *rb =
1032                 container_of(bio, struct btree_read_bio, bio);
1033         struct bch_fs *c        = rb->c;
1034
1035         if (rb->have_ioref) {
1036                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1037                 bch2_latency_acct(ca, rb->start_time, READ);
1038         }
1039
1040         queue_work(system_unbound_wq, &rb->work);
1041 }
1042
1043 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1044                           bool sync)
1045 {
1046         struct extent_ptr_decoded pick;
1047         struct btree_read_bio *rb;
1048         struct bch_dev *ca;
1049         struct bio *bio;
1050         int ret;
1051
1052         trace_btree_read(c, b);
1053
1054         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1055                                          NULL, &pick);
1056         if (bch2_fs_fatal_err_on(ret <= 0, c,
1057                         "btree node read error: no device to read from")) {
1058                 set_btree_node_read_error(b);
1059                 return;
1060         }
1061
1062         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1063
1064         bio = bio_alloc_bioset(GFP_NOIO, buf_pages(b->data,
1065                                                    btree_bytes(c)),
1066                                &c->btree_bio);
1067         rb = container_of(bio, struct btree_read_bio, bio);
1068         rb->c                   = c;
1069         rb->start_time          = local_clock();
1070         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1071         rb->pick                = pick;
1072         INIT_WORK(&rb->work, btree_node_read_work);
1073         bio->bi_opf             = REQ_OP_READ|REQ_SYNC|REQ_META;
1074         bio->bi_iter.bi_sector  = pick.ptr.offset;
1075         bio->bi_end_io          = btree_node_read_endio;
1076         bio->bi_private         = b;
1077         bch2_bio_map(bio, b->data, btree_bytes(c));
1078
1079         set_btree_node_read_in_flight(b);
1080
1081         if (rb->have_ioref) {
1082                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1083                              bio_sectors(bio));
1084                 bio_set_dev(bio, ca->disk_sb.bdev);
1085
1086                 if (sync) {
1087                         submit_bio_wait(bio);
1088
1089                         bio->bi_private = b;
1090                         btree_node_read_work(&rb->work);
1091                 } else {
1092                         submit_bio(bio);
1093                 }
1094         } else {
1095                 bio->bi_status = BLK_STS_REMOVED;
1096
1097                 if (sync)
1098                         btree_node_read_work(&rb->work);
1099                 else
1100                         queue_work(system_unbound_wq, &rb->work);
1101
1102         }
1103 }
1104
1105 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1106                         const struct bkey_i *k, unsigned level)
1107 {
1108         struct closure cl;
1109         struct btree *b;
1110         int ret;
1111
1112         closure_init_stack(&cl);
1113
1114         do {
1115                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1116                 closure_sync(&cl);
1117         } while (ret);
1118
1119         b = bch2_btree_node_mem_alloc(c);
1120         bch2_btree_cache_cannibalize_unlock(c);
1121
1122         BUG_ON(IS_ERR(b));
1123
1124         bkey_copy(&b->key, k);
1125         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1126
1127         bch2_btree_node_read(c, b, true);
1128
1129         if (btree_node_read_error(b)) {
1130                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1131
1132                 mutex_lock(&c->btree_cache.lock);
1133                 list_move(&b->list, &c->btree_cache.freeable);
1134                 mutex_unlock(&c->btree_cache.lock);
1135
1136                 ret = -EIO;
1137                 goto err;
1138         }
1139
1140         bch2_btree_set_root_for_read(c, b);
1141 err:
1142         six_unlock_write(&b->c.lock);
1143         six_unlock_intent(&b->c.lock);
1144
1145         return ret;
1146 }
1147
1148 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1149                               struct btree_write *w)
1150 {
1151         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1152
1153         do {
1154                 old = new = v;
1155                 if (!(old & 1))
1156                         break;
1157
1158                 new &= ~1UL;
1159         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1160
1161         if (old & 1)
1162                 closure_put(&((struct btree_update *) new)->cl);
1163
1164         bch2_journal_pin_drop(&c->journal, &w->journal);
1165 }
1166
1167 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1168 {
1169         struct btree_write *w = btree_prev_write(b);
1170
1171         bch2_btree_complete_write(c, b, w);
1172         btree_node_io_unlock(b);
1173 }
1174
1175 static void bch2_btree_node_write_error(struct bch_fs *c,
1176                                         struct btree_write_bio *wbio)
1177 {
1178         struct btree *b         = wbio->wbio.bio.bi_private;
1179         struct bkey_buf k;
1180         struct bch_extent_ptr *ptr;
1181         struct btree_trans trans;
1182         struct btree_iter *iter;
1183         int ret;
1184
1185         bch2_bkey_buf_init(&k);
1186         bch2_trans_init(&trans, c, 0, 0);
1187
1188         iter = bch2_trans_get_node_iter(&trans, b->c.btree_id, b->key.k.p,
1189                                         BTREE_MAX_DEPTH, b->c.level, 0);
1190 retry:
1191         ret = bch2_btree_iter_traverse(iter);
1192         if (ret)
1193                 goto err;
1194
1195         /* has node been freed? */
1196         if (iter->l[b->c.level].b != b) {
1197                 /* node has been freed: */
1198                 BUG_ON(!btree_node_dying(b));
1199                 goto out;
1200         }
1201
1202         BUG_ON(!btree_node_hashed(b));
1203
1204         bch2_bkey_buf_copy(&k, c, &b->key);
1205
1206         bch2_bkey_drop_ptrs(bkey_i_to_s(k.k), ptr,
1207                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1208
1209         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(k.k)))
1210                 goto err;
1211
1212         ret = bch2_btree_node_update_key(c, iter, b, k.k);
1213         if (ret == -EINTR)
1214                 goto retry;
1215         if (ret)
1216                 goto err;
1217 out:
1218         bch2_trans_exit(&trans);
1219         bch2_bkey_buf_exit(&k, c);
1220         bio_put(&wbio->wbio.bio);
1221         btree_node_write_done(c, b);
1222         return;
1223 err:
1224         set_btree_node_noevict(b);
1225         bch2_fs_fatal_error(c, "fatal error writing btree node");
1226         goto out;
1227 }
1228
1229 void bch2_btree_write_error_work(struct work_struct *work)
1230 {
1231         struct bch_fs *c = container_of(work, struct bch_fs,
1232                                         btree_write_error_work);
1233         struct bio *bio;
1234
1235         while (1) {
1236                 spin_lock_irq(&c->btree_write_error_lock);
1237                 bio = bio_list_pop(&c->btree_write_error_list);
1238                 spin_unlock_irq(&c->btree_write_error_lock);
1239
1240                 if (!bio)
1241                         break;
1242
1243                 bch2_btree_node_write_error(c,
1244                         container_of(bio, struct btree_write_bio, wbio.bio));
1245         }
1246 }
1247
1248 static void btree_node_write_work(struct work_struct *work)
1249 {
1250         struct btree_write_bio *wbio =
1251                 container_of(work, struct btree_write_bio, work);
1252         struct bch_fs *c        = wbio->wbio.c;
1253         struct btree *b         = wbio->wbio.bio.bi_private;
1254
1255         btree_bounce_free(c,
1256                 wbio->bytes,
1257                 wbio->wbio.used_mempool,
1258                 wbio->data);
1259
1260         if (wbio->wbio.failed.nr) {
1261                 unsigned long flags;
1262
1263                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1264                 bio_list_add(&c->btree_write_error_list, &wbio->wbio.bio);
1265                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1266
1267                 queue_work(c->wq, &c->btree_write_error_work);
1268                 return;
1269         }
1270
1271         bio_put(&wbio->wbio.bio);
1272         btree_node_write_done(c, b);
1273 }
1274
1275 static void btree_node_write_endio(struct bio *bio)
1276 {
1277         struct bch_write_bio *wbio      = to_wbio(bio);
1278         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1279         struct bch_write_bio *orig      = parent ?: wbio;
1280         struct bch_fs *c                = wbio->c;
1281         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1282         unsigned long flags;
1283
1284         if (wbio->have_ioref)
1285                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1286
1287         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1288                                bch2_blk_status_to_str(bio->bi_status)) ||
1289             bch2_meta_write_fault("btree")) {
1290                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1291                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1292                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1293         }
1294
1295         if (wbio->have_ioref)
1296                 percpu_ref_put(&ca->io_ref);
1297
1298         if (parent) {
1299                 bio_put(bio);
1300                 bio_endio(&parent->bio);
1301         } else {
1302                 struct btree_write_bio *wb =
1303                         container_of(orig, struct btree_write_bio, wbio);
1304
1305                 INIT_WORK(&wb->work, btree_node_write_work);
1306                 queue_work(system_unbound_wq, &wb->work);
1307         }
1308 }
1309
1310 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1311                                    struct bset *i, unsigned sectors)
1312 {
1313         unsigned whiteout_u64s = 0;
1314         int ret;
1315
1316         if (bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key), BKEY_TYPE_BTREE))
1317                 return -1;
1318
1319         ret = validate_bset(c, NULL, b, i, sectors, WRITE, false) ?:
1320                 validate_bset_keys(c, b, i, &whiteout_u64s, WRITE, false);
1321         if (ret) {
1322                 bch2_inconsistent_error(c);
1323                 dump_stack();
1324         }
1325
1326         return ret;
1327 }
1328
1329 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1330                             enum six_lock_type lock_type_held)
1331 {
1332         struct btree_write_bio *wbio;
1333         struct bset_tree *t;
1334         struct bset *i;
1335         struct btree_node *bn = NULL;
1336         struct btree_node_entry *bne = NULL;
1337         struct bkey_buf k;
1338         struct bch_extent_ptr *ptr;
1339         struct sort_iter sort_iter;
1340         struct nonce nonce;
1341         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1342         u64 seq = 0;
1343         bool used_mempool;
1344         unsigned long old, new;
1345         bool validate_before_checksum = false;
1346         void *data;
1347
1348         bch2_bkey_buf_init(&k);
1349
1350         if (test_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags))
1351                 return;
1352
1353         /*
1354          * We may only have a read lock on the btree node - the dirty bit is our
1355          * "lock" against racing with other threads that may be trying to start
1356          * a write, we do a write iff we clear the dirty bit. Since setting the
1357          * dirty bit requires a write lock, we can't race with other threads
1358          * redirtying it:
1359          */
1360         do {
1361                 old = new = READ_ONCE(b->flags);
1362
1363                 if (!(old & (1 << BTREE_NODE_dirty)))
1364                         return;
1365
1366                 if (!btree_node_may_write(b))
1367                         return;
1368
1369                 if (old & (1 << BTREE_NODE_never_write))
1370                         return;
1371
1372                 if (old & (1 << BTREE_NODE_write_in_flight)) {
1373                         btree_node_wait_on_io(b);
1374                         continue;
1375                 }
1376
1377                 new &= ~(1 << BTREE_NODE_dirty);
1378                 new &= ~(1 << BTREE_NODE_need_write);
1379                 new |=  (1 << BTREE_NODE_write_in_flight);
1380                 new |=  (1 << BTREE_NODE_just_written);
1381                 new ^=  (1 << BTREE_NODE_write_idx);
1382         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1383
1384         atomic_dec(&c->btree_cache.dirty);
1385
1386         BUG_ON(btree_node_fake(b));
1387         BUG_ON((b->will_make_reachable != 0) != !b->written);
1388
1389         BUG_ON(b->written >= c->opts.btree_node_size);
1390         BUG_ON(b->written & (c->opts.block_size - 1));
1391         BUG_ON(bset_written(b, btree_bset_last(b)));
1392         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1393         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1394
1395         bch2_sort_whiteouts(c, b);
1396
1397         sort_iter_init(&sort_iter, b);
1398
1399         bytes = !b->written
1400                 ? sizeof(struct btree_node)
1401                 : sizeof(struct btree_node_entry);
1402
1403         bytes += b->whiteout_u64s * sizeof(u64);
1404
1405         for_each_bset(b, t) {
1406                 i = bset(b, t);
1407
1408                 if (bset_written(b, i))
1409                         continue;
1410
1411                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1412                 sort_iter_add(&sort_iter,
1413                               btree_bkey_first(b, t),
1414                               btree_bkey_last(b, t));
1415                 seq = max(seq, le64_to_cpu(i->journal_seq));
1416         }
1417
1418         BUG_ON(b->written && !seq);
1419
1420         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1421         bytes += 8;
1422
1423         data = btree_bounce_alloc(c, bytes, &used_mempool);
1424
1425         if (!b->written) {
1426                 bn = data;
1427                 *bn = *b->data;
1428                 i = &bn->keys;
1429         } else {
1430                 bne = data;
1431                 bne->keys = b->data->keys;
1432                 i = &bne->keys;
1433         }
1434
1435         i->journal_seq  = cpu_to_le64(seq);
1436         i->u64s         = 0;
1437
1438         sort_iter_add(&sort_iter,
1439                       unwritten_whiteouts_start(c, b),
1440                       unwritten_whiteouts_end(c, b));
1441         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1442
1443         b->whiteout_u64s = 0;
1444
1445         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1446         le16_add_cpu(&i->u64s, u64s);
1447
1448         set_needs_whiteout(i, false);
1449
1450         /* do we have data to write? */
1451         if (b->written && !i->u64s)
1452                 goto nowrite;
1453
1454         bytes_to_write = vstruct_end(i) - data;
1455         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1456
1457         memset(data + bytes_to_write, 0,
1458                (sectors_to_write << 9) - bytes_to_write);
1459
1460         BUG_ON(b->written + sectors_to_write > c->opts.btree_node_size);
1461         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1462         BUG_ON(i->seq != b->data->keys.seq);
1463
1464         i->version = c->sb.version < bcachefs_metadata_version_new_versioning
1465                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1466                 : cpu_to_le16(c->sb.version);
1467         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1468
1469         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1470                 validate_before_checksum = true;
1471
1472         /* validate_bset will be modifying: */
1473         if (le16_to_cpu(i->version) <= bcachefs_metadata_version_inode_btree_change)
1474                 validate_before_checksum = true;
1475
1476         /* if we're going to be encrypting, check metadata validity first: */
1477         if (validate_before_checksum &&
1478             validate_bset_for_write(c, b, i, sectors_to_write))
1479                 goto err;
1480
1481         bset_encrypt(c, i, b->written << 9);
1482
1483         nonce = btree_nonce(i, b->written << 9);
1484
1485         if (bn)
1486                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1487         else
1488                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1489
1490         /* if we're not encrypting, check metadata after checksumming: */
1491         if (!validate_before_checksum &&
1492             validate_bset_for_write(c, b, i, sectors_to_write))
1493                 goto err;
1494
1495         /*
1496          * We handle btree write errors by immediately halting the journal -
1497          * after we've done that, we can't issue any subsequent btree writes
1498          * because they might have pointers to new nodes that failed to write.
1499          *
1500          * Furthermore, there's no point in doing any more btree writes because
1501          * with the journal stopped, we're never going to update the journal to
1502          * reflect that those writes were done and the data flushed from the
1503          * journal:
1504          *
1505          * Also on journal error, the pending write may have updates that were
1506          * never journalled (interior nodes, see btree_update_nodes_written()) -
1507          * it's critical that we don't do the write in that case otherwise we
1508          * will have updates visible that weren't in the journal:
1509          *
1510          * Make sure to update b->written so bch2_btree_init_next() doesn't
1511          * break:
1512          */
1513         if (bch2_journal_error(&c->journal) ||
1514             c->opts.nochanges)
1515                 goto err;
1516
1517         trace_btree_write(b, bytes_to_write, sectors_to_write);
1518
1519         wbio = container_of(bio_alloc_bioset(GFP_NOIO,
1520                                 buf_pages(data, sectors_to_write << 9),
1521                                 &c->btree_bio),
1522                             struct btree_write_bio, wbio.bio);
1523         wbio_init(&wbio->wbio.bio);
1524         wbio->data                      = data;
1525         wbio->bytes                     = bytes;
1526         wbio->wbio.used_mempool         = used_mempool;
1527         wbio->wbio.bio.bi_opf           = REQ_OP_WRITE|REQ_META;
1528         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
1529         wbio->wbio.bio.bi_private       = b;
1530
1531         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
1532
1533         /*
1534          * If we're appending to a leaf node, we don't technically need FUA -
1535          * this write just needs to be persisted before the next journal write,
1536          * which will be marked FLUSH|FUA.
1537          *
1538          * Similarly if we're writing a new btree root - the pointer is going to
1539          * be in the next journal entry.
1540          *
1541          * But if we're writing a new btree node (that isn't a root) or
1542          * appending to a non leaf btree node, we need either FUA or a flush
1543          * when we write the parent with the new pointer. FUA is cheaper than a
1544          * flush, and writes appending to leaf nodes aren't blocking anything so
1545          * just make all btree node writes FUA to keep things sane.
1546          */
1547
1548         bch2_bkey_buf_copy(&k, c, &b->key);
1549
1550         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(k.k)), ptr)
1551                 ptr->offset += b->written;
1552
1553         b->written += sectors_to_write;
1554
1555         /* XXX: submitting IO with btree locks held: */
1556         bch2_submit_wbio_replicas(&wbio->wbio, c, BCH_DATA_btree, k.k);
1557         bch2_bkey_buf_exit(&k, c);
1558         return;
1559 err:
1560         set_btree_node_noevict(b);
1561         b->written += sectors_to_write;
1562 nowrite:
1563         btree_bounce_free(c, bytes, used_mempool, data);
1564         btree_node_write_done(c, b);
1565 }
1566
1567 /*
1568  * Work that must be done with write lock held:
1569  */
1570 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
1571 {
1572         bool invalidated_iter = false;
1573         struct btree_node_entry *bne;
1574         struct bset_tree *t;
1575
1576         if (!btree_node_just_written(b))
1577                 return false;
1578
1579         BUG_ON(b->whiteout_u64s);
1580
1581         clear_btree_node_just_written(b);
1582
1583         /*
1584          * Note: immediately after write, bset_written() doesn't work - the
1585          * amount of data we had to write after compaction might have been
1586          * smaller than the offset of the last bset.
1587          *
1588          * However, we know that all bsets have been written here, as long as
1589          * we're still holding the write lock:
1590          */
1591
1592         /*
1593          * XXX: decide if we really want to unconditionally sort down to a
1594          * single bset:
1595          */
1596         if (b->nsets > 1) {
1597                 btree_node_sort(c, b, NULL, 0, b->nsets, true);
1598                 invalidated_iter = true;
1599         } else {
1600                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
1601         }
1602
1603         for_each_bset(b, t)
1604                 set_needs_whiteout(bset(b, t), true);
1605
1606         bch2_btree_verify(c, b);
1607
1608         /*
1609          * If later we don't unconditionally sort down to a single bset, we have
1610          * to ensure this is still true:
1611          */
1612         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
1613
1614         bne = want_new_bset(c, b);
1615         if (bne)
1616                 bch2_bset_init_next(c, b, bne);
1617
1618         bch2_btree_build_aux_trees(b);
1619
1620         return invalidated_iter;
1621 }
1622
1623 /*
1624  * Use this one if the node is intent locked:
1625  */
1626 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
1627                           enum six_lock_type lock_type_held)
1628 {
1629         BUG_ON(lock_type_held == SIX_LOCK_write);
1630
1631         if (lock_type_held == SIX_LOCK_intent ||
1632             six_lock_tryupgrade(&b->c.lock)) {
1633                 __bch2_btree_node_write(c, b, SIX_LOCK_intent);
1634
1635                 /* don't cycle lock unnecessarily: */
1636                 if (btree_node_just_written(b) &&
1637                     six_trylock_write(&b->c.lock)) {
1638                         bch2_btree_post_write_cleanup(c, b);
1639                         six_unlock_write(&b->c.lock);
1640                 }
1641
1642                 if (lock_type_held == SIX_LOCK_read)
1643                         six_lock_downgrade(&b->c.lock);
1644         } else {
1645                 __bch2_btree_node_write(c, b, SIX_LOCK_read);
1646         }
1647 }
1648
1649 static void __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
1650 {
1651         struct bucket_table *tbl;
1652         struct rhash_head *pos;
1653         struct btree *b;
1654         unsigned i;
1655 restart:
1656         rcu_read_lock();
1657         for_each_cached_btree(b, c, tbl, i, pos)
1658                 if (test_bit(flag, &b->flags)) {
1659                         rcu_read_unlock();
1660                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
1661                         goto restart;
1662
1663                 }
1664         rcu_read_unlock();
1665 }
1666
1667 void bch2_btree_flush_all_reads(struct bch_fs *c)
1668 {
1669         __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
1670 }
1671
1672 void bch2_btree_flush_all_writes(struct bch_fs *c)
1673 {
1674         __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
1675 }
1676
1677 void bch2_dirty_btree_nodes_to_text(struct printbuf *out, struct bch_fs *c)
1678 {
1679         struct bucket_table *tbl;
1680         struct rhash_head *pos;
1681         struct btree *b;
1682         unsigned i;
1683
1684         rcu_read_lock();
1685         for_each_cached_btree(b, c, tbl, i, pos) {
1686                 unsigned long flags = READ_ONCE(b->flags);
1687
1688                 if (!(flags & (1 << BTREE_NODE_dirty)))
1689                         continue;
1690
1691                 pr_buf(out, "%p d %u n %u l %u w %u b %u r %u:%lu\n",
1692                        b,
1693                        (flags & (1 << BTREE_NODE_dirty)) != 0,
1694                        (flags & (1 << BTREE_NODE_need_write)) != 0,
1695                        b->c.level,
1696                        b->written,
1697                        !list_empty_careful(&b->write_blocked),
1698                        b->will_make_reachable != 0,
1699                        b->will_make_reachable & 1);
1700         }
1701         rcu_read_unlock();
1702 }