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_ge(l.p, bkey_start_pos(&r)));
  87         }
  88 #endif
  89 }
  90
  91 static void set_needs_whiteout(struct bset *i, int v)
  92 {
  93         struct bkey_packed *k;
  94
  95         for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
  96                 k->needs_whiteout = v;
  97 }
  98
  99 static void btree_bounce_free(struct bch_fs *c, size_t size,
 100                               bool used_mempool, void *p)
 101 {
 102         if (used_mempool)
 103                 mempool_free(p, &c->btree_bounce_pool);
 104         else
 105                 vpfree(p, size);
 106 }
 107
 108 static void *btree_bounce_alloc(struct bch_fs *c, size_t size,
 109                                 bool *used_mempool)
 110 {
 111         unsigned flags = memalloc_nofs_save();
 112         void *p;
 113
 114         BUG_ON(size > btree_bytes(c));
 115
 116         *used_mempool = false;
 117         p = vpmalloc(size, __GFP_NOWARN|GFP_NOWAIT);
 118         if (!p) {
 119                 *used_mempool = true;
 120                 p = mempool_alloc(&c->btree_bounce_pool, GFP_NOIO);
 121         }
 122         memalloc_nofs_restore(flags);
 123         return p;
 124 }
 125
 126 static void sort_bkey_ptrs(const struct btree *bt,
 127                            struct bkey_packed **ptrs, unsigned nr)
 128 {
 129         unsigned n = nr, a = nr / 2, b, c, d;
 130
 131         if (!a)
 132                 return;
 133
 134         /* Heap sort: see lib/sort.c: */
 135         while (1) {
 136                 if (a)
 137                         a--;
 138                 else if (--n)
 139                         swap(ptrs[0], ptrs[n]);
 140                 else
 141                         break;
 142
 143                 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
 144                         b = bch2_bkey_cmp_packed(bt,
 145                                             ptrs[c],
 146                                             ptrs[d]) >= 0 ? c : d;
 147                 if (d == n)
 148                         b = c;
 149
 150                 while (b != a &&
 151                        bch2_bkey_cmp_packed(bt,
 152                                        ptrs[a],
 153                                        ptrs[b]) >= 0)
 154                         b = (b - 1) / 2;
 155                 c = b;
 156                 while (b != a) {
 157                         b = (b - 1) / 2;
 158                         swap(ptrs[b], ptrs[c]);
 159                 }
 160         }
 161 }
 162
 163 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
 164 {
 165         struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
 166         bool used_mempool = false;
 167         size_t bytes = b->whiteout_u64s * sizeof(u64);
 168
 169         if (!b->whiteout_u64s)
 170                 return;
 171
 172         new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
 173
 174         ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
 175
 176         for (k = unwritten_whiteouts_start(c, b);
 177              k != unwritten_whiteouts_end(c, b);
 178              k = bkey_next(k))
 179                 *--ptrs = k;
 180
 181         sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
 182
 183         k = new_whiteouts;
 184
 185         while (ptrs != ptrs_end) {
 186                 bkey_copy(k, *ptrs);
 187                 k = bkey_next(k);
 188                 ptrs++;
 189         }
 190
 191         verify_no_dups(b, new_whiteouts,
 192                        (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
 193
 194         memcpy_u64s(unwritten_whiteouts_start(c, b),
 195                     new_whiteouts, b->whiteout_u64s);
 196
 197         btree_bounce_free(c, bytes, used_mempool, new_whiteouts);
 198 }
 199
 200 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
 201                                 bool compacting, enum compact_mode mode)
 202 {
 203         if (!bset_dead_u64s(b, t))
 204                 return false;
 205
 206         switch (mode) {
 207         case COMPACT_LAZY:
 208                 return should_compact_bset_lazy(b, t) ||
 209                         (compacting && !bset_written(b, bset(b, t)));
 210         case COMPACT_ALL:
 211                 return true;
 212         default:
 213                 BUG();
 214         }
 215 }
 216
 217 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
 218 {
 219         struct bset_tree *t;
 220         bool ret = false;
 221
 222         for_each_bset(b, t) {
 223                 struct bset *i = bset(b, t);
 224                 struct bkey_packed *k, *n, *out, *start, *end;
 225                 struct btree_node_entry *src = NULL, *dst = NULL;
 226
 227                 if (t != b->set && !bset_written(b, i)) {
 228                         src = container_of(i, struct btree_node_entry, keys);
 229                         dst = max(write_block(b),
 230                                   (void *) btree_bkey_last(b, t - 1));
 231                 }
 232
 233                 if (src != dst)
 234                         ret = true;
 235
 236                 if (!should_compact_bset(b, t, ret, mode)) {
 237                         if (src != dst) {
 238                                 memmove(dst, src, sizeof(*src) +
 239                                         le16_to_cpu(src->keys.u64s) *
 240                                         sizeof(u64));
 241                                 i = &dst->keys;
 242                                 set_btree_bset(b, t, i);
 243                         }
 244                         continue;
 245                 }
 246
 247                 start   = btree_bkey_first(b, t);
 248                 end     = btree_bkey_last(b, t);
 249
 250                 if (src != dst) {
 251                         memmove(dst, src, sizeof(*src));
 252                         i = &dst->keys;
 253                         set_btree_bset(b, t, i);
 254                 }
 255
 256                 out = i->start;
 257
 258                 for (k = start; k != end; k = n) {
 259                         n = bkey_next(k);
 260
 261                         if (!bkey_deleted(k)) {
 262                                 bkey_copy(out, k);
 263                                 out = bkey_next(out);
 264                         } else {
 265                                 BUG_ON(k->needs_whiteout);
 266                         }
 267                 }
 268
 269                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 270                 set_btree_bset_end(b, t);
 271                 bch2_bset_set_no_aux_tree(b, t);
 272                 ret = true;
 273         }
 274
 275         bch2_verify_btree_nr_keys(b);
 276
 277         bch2_btree_build_aux_trees(b);
 278
 279         return ret;
 280 }
 281
 282 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 283                             enum compact_mode mode)
 284 {
 285         return bch2_drop_whiteouts(b, mode);
 286 }
 287
 288 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 289                             unsigned start_idx,
 290                             unsigned end_idx,
 291                             bool filter_whiteouts)
 292 {
 293         struct btree_node *out;
 294         struct sort_iter sort_iter;
 295         struct bset_tree *t;
 296         struct bset *start_bset = bset(b, &b->set[start_idx]);
 297         bool used_mempool = false;
 298         u64 start_time, seq = 0;
 299         unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
 300         bool sorting_entire_node = start_idx == 0 &&
 301                 end_idx == b->nsets;
 302
 303         sort_iter_init(&sort_iter, b);
 304
 305         for (t = b->set + start_idx;
 306              t < b->set + end_idx;
 307              t++) {
 308                 u64s += le16_to_cpu(bset(b, t)->u64s);
 309                 sort_iter_add(&sort_iter,
 310                               btree_bkey_first(b, t),
 311                               btree_bkey_last(b, t));
 312         }
 313
 314         bytes = sorting_entire_node
 315                 ? btree_bytes(c)
 316                 : __vstruct_bytes(struct btree_node, u64s);
 317
 318         out = btree_bounce_alloc(c, bytes, &used_mempool);
 319
 320         start_time = local_clock();
 321
 322         u64s = bch2_sort_keys(out->keys.start, &sort_iter, filter_whiteouts);
 323
 324         out->keys.u64s = cpu_to_le16(u64s);
 325
 326         BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
 327
 328         if (sorting_entire_node)
 329                 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 330                                        start_time);
 331
 332         /* Make sure we preserve bset journal_seq: */
 333         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 334                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 335         start_bset->journal_seq = cpu_to_le64(seq);
 336
 337         if (sorting_entire_node) {
 338                 unsigned u64s = le16_to_cpu(out->keys.u64s);
 339
 340                 BUG_ON(bytes != btree_bytes(c));
 341
 342                 /*
 343                  * Our temporary buffer is the same size as the btree node's
 344                  * buffer, we can just swap buffers instead of doing a big
 345                  * memcpy()
 346                  */
 347                 *out = *b->data;
 348                 out->keys.u64s = cpu_to_le16(u64s);
 349                 swap(out, b->data);
 350                 set_btree_bset(b, b->set, &b->data->keys);
 351         } else {
 352                 start_bset->u64s = out->keys.u64s;
 353                 memcpy_u64s(start_bset->start,
 354                             out->keys.start,
 355                             le16_to_cpu(out->keys.u64s));
 356         }
 357
 358         for (i = start_idx + 1; i < end_idx; i++)
 359                 b->nr.bset_u64s[start_idx] +=
 360                         b->nr.bset_u64s[i];
 361
 362         b->nsets -= shift;
 363
 364         for (i = start_idx + 1; i < b->nsets; i++) {
 365                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 366                 b->set[i]               = b->set[i + shift];
 367         }
 368
 369         for (i = b->nsets; i < MAX_BSETS; i++)
 370                 b->nr.bset_u64s[i] = 0;
 371
 372         set_btree_bset_end(b, &b->set[start_idx]);
 373         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 374
 375         btree_bounce_free(c, bytes, used_mempool, out);
 376
 377         bch2_verify_btree_nr_keys(b);
 378 }
 379
 380 void bch2_btree_sort_into(struct bch_fs *c,
 381                          struct btree *dst,
 382                          struct btree *src)
 383 {
 384         struct btree_nr_keys nr;
 385         struct btree_node_iter src_iter;
 386         u64 start_time = local_clock();
 387
 388         BUG_ON(dst->nsets != 1);
 389
 390         bch2_bset_set_no_aux_tree(dst, dst->set);
 391
 392         bch2_btree_node_iter_init_from_start(&src_iter, src);
 393
 394         nr = bch2_sort_repack(btree_bset_first(dst),
 395                         src, &src_iter,
 396                         &dst->format,
 397                         true);
 398
 399         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 400                                start_time);
 401
 402         set_btree_bset_end(dst, dst->set);
 403
 404         dst->nr.live_u64s       += nr.live_u64s;
 405         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 406         dst->nr.packed_keys     += nr.packed_keys;
 407         dst->nr.unpacked_keys   += nr.unpacked_keys;
 408
 409         bch2_verify_btree_nr_keys(dst);
 410 }
 411
 412 #define SORT_CRIT       (4096 / sizeof(u64))
 413
 414 /*
 415  * We're about to add another bset to the btree node, so if there's currently
 416  * too many bsets - sort some of them together:
 417  */
 418 static bool btree_node_compact(struct bch_fs *c, struct btree *b)
 419 {
 420         unsigned unwritten_idx;
 421         bool ret = false;
 422
 423         for (unwritten_idx = 0;
 424              unwritten_idx < b->nsets;
 425              unwritten_idx++)
 426                 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
 427                         break;
 428
 429         if (b->nsets - unwritten_idx > 1) {
 430                 btree_node_sort(c, b, unwritten_idx,
 431                                 b->nsets, false);
 432                 ret = true;
 433         }
 434
 435         if (unwritten_idx > 1) {
 436                 btree_node_sort(c, b, 0, unwritten_idx, false);
 437                 ret = true;
 438         }
 439
 440         return ret;
 441 }
 442
 443 void bch2_btree_build_aux_trees(struct btree *b)
 444 {
 445         struct bset_tree *t;
 446
 447         for_each_bset(b, t)
 448                 bch2_bset_build_aux_tree(b, t,
 449                                 !bset_written(b, bset(b, t)) &&
 450                                 t == bset_tree_last(b));
 451 }
 452
 453 /*
 454  * If we have MAX_BSETS (3) bsets, should we sort them all down to just one?
 455  *
 456  * The first bset is going to be of similar order to the size of the node, the
 457  * last bset is bounded by btree_write_set_buffer(), which is set to keep the
 458  * memmove on insert from being too expensive: the middle bset should, ideally,
 459  * be the geometric mean of the first and the last.
 460  *
 461  * Returns true if the middle bset is greater than that geometric mean:
 462  */
 463 static inline bool should_compact_all(struct bch_fs *c, struct btree *b)
 464 {
 465         unsigned mid_u64s_bits =
 466                 (ilog2(btree_max_u64s(c)) + BTREE_WRITE_SET_U64s_BITS) / 2;
 467
 468         return bset_u64s(&b->set[1]) > 1U << mid_u64s_bits;
 469 }
 470
 471 /*
 472  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 473  * inserted into
 474  *
 475  * Safe to call if there already is an unwritten bset - will only add a new bset
 476  * if @b doesn't already have one.
 477  *
 478  * Returns true if we sorted (i.e. invalidated iterators
 479  */
 480 void bch2_btree_init_next(struct btree_trans *trans, struct btree *b)
 481 {
 482         struct bch_fs *c = trans->c;
 483         struct btree_node_entry *bne;
 484         bool reinit_iter = false;
 485
 486         EBUG_ON(!(b->c.lock.state.seq & 1));
 487         BUG_ON(bset_written(b, bset(b, &b->set[1])));
 488         BUG_ON(btree_node_just_written(b));
 489
 490         if (b->nsets == MAX_BSETS &&
 491             !btree_node_write_in_flight(b) &&
 492             should_compact_all(c, b)) {
 493                 bch2_btree_node_write(c, b, SIX_LOCK_write,
 494                                       BTREE_WRITE_init_next_bset);
 495                 reinit_iter = true;
 496         }
 497
 498         if (b->nsets == MAX_BSETS &&
 499             btree_node_compact(c, b))
 500                 reinit_iter = true;
 501
 502         BUG_ON(b->nsets >= MAX_BSETS);
 503
 504         bne = want_new_bset(c, b);
 505         if (bne)
 506                 bch2_bset_init_next(c, b, bne);
 507
 508         bch2_btree_build_aux_trees(b);
 509
 510         if (reinit_iter)
 511                 bch2_trans_node_reinit_iter(trans, b);
 512 }
 513
 514 static void btree_pos_to_text(struct printbuf *out, struct bch_fs *c,
 515                           struct btree *b)
 516 {
 517         prt_printf(out, "%s level %u/%u\n  ",
 518                bch2_btree_ids[b->c.btree_id],
 519                b->c.level,
 520                c->btree_roots[b->c.btree_id].level);
 521         bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 522 }
 523
 524 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 525                           struct bch_dev *ca,
 526                           struct btree *b, struct bset *i,
 527                           unsigned offset, int write)
 528 {
 529         prt_printf(out, bch2_log_msg(c, ""));
 530         if (!write)
 531                 prt_str(out, "error validating btree node ");
 532         else
 533                 prt_str(out, "corrupt btree node before write ");
 534         if (ca)
 535                 prt_printf(out, "on %s ", ca->name);
 536         prt_printf(out, "at btree ");
 537         btree_pos_to_text(out, c, b);
 538
 539         prt_printf(out, "\n  node offset %u", b->written);
 540         if (i)
 541                 prt_printf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 542         prt_str(out, ": ");
 543 }
 544
 545 enum btree_err_type {
 546         BTREE_ERR_FIXABLE,
 547         BTREE_ERR_WANT_RETRY,
 548         BTREE_ERR_MUST_RETRY,
 549         BTREE_ERR_FATAL,
 550 };
 551
 552 enum btree_validate_ret {
 553         BTREE_RETRY_READ = 64,
 554 };
 555
 556 #define btree_err(type, c, ca, b, i, msg, ...)                          \
 557 ({                                                                      \
 558         __label__ out;                                                  \
 559         struct printbuf out = PRINTBUF;                                 \
 560                                                                         \
 561         btree_err_msg(&out, c, ca, b, i, b->written, write);            \
 562         prt_printf(&out, msg, ##__VA_ARGS__);                           \
 563                                                                         \
 564         if (type == BTREE_ERR_FIXABLE &&                                \
 565             write == READ &&                                            \
 566             !test_bit(BCH_FS_INITIAL_GC_DONE, &c->flags)) {             \
 567                 mustfix_fsck_err(c, "%s", out.buf);                     \
 568                 goto out;                                               \
 569         }                                                               \
 570                                                                         \
 571         bch2_print_string_as_lines(KERN_ERR, out.buf);                  \
 572                                                                         \
 573         switch (write) {                                                \
 574         case READ:                                                      \
 575                 switch (type) {                                         \
 576                 case BTREE_ERR_FIXABLE:                                 \
 577                         ret = -BCH_ERR_fsck_errors_not_fixed;           \
 578                         goto fsck_err;                                  \
 579                 case BTREE_ERR_WANT_RETRY:                              \
 580                         if (have_retry) {                               \
 581                                 ret = BTREE_RETRY_READ;                 \
 582                                 goto fsck_err;                          \
 583                         }                                               \
 584                         break;                                          \
 585                 case BTREE_ERR_MUST_RETRY:                              \
 586                         ret = BTREE_RETRY_READ;                         \
 587                         goto fsck_err;                                  \
 588                 case BTREE_ERR_FATAL:                                   \
 589                         ret = -BCH_ERR_fsck_errors_not_fixed;           \
 590                         goto fsck_err;                                  \
 591                 }                                                       \
 592                 break;                                                  \
 593         case WRITE:                                                     \
 594                 if (bch2_fs_inconsistent(c)) {                          \
 595                         ret = -BCH_ERR_fsck_errors_not_fixed;           \
 596                         goto fsck_err;                                  \
 597                 }                                                       \
 598                 break;                                                  \
 599         }                                                               \
 600 out:                                                                    \
 601         printbuf_exit(&out);                                            \
 602         true;                                                           \
 603 })
 604
 605 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 606
 607 /*
 608  * When btree topology repair changes the start or end of a node, that might
 609  * mean we have to drop keys that are no longer inside the node:
 610  */
 611 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
 612 {
 613         struct bset_tree *t;
 614         struct bkey_s_c k;
 615         struct bkey unpacked;
 616         struct btree_node_iter iter;
 617
 618         for_each_bset(b, t) {
 619                 struct bset *i = bset(b, t);
 620                 struct bkey_packed *k;
 621
 622                 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
 623                         if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
 624                                 break;
 625
 626                 if (k != i->start) {
 627                         unsigned shift = (u64 *) k - (u64 *) i->start;
 628
 629                         memmove_u64s_down(i->start, k,
 630                                           (u64 *) vstruct_end(i) - (u64 *) k);
 631                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
 632                         set_btree_bset_end(b, t);
 633                 }
 634
 635                 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
 636                         if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
 637                                 break;
 638
 639                 if (k != vstruct_last(i)) {
 640                         i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
 641                         set_btree_bset_end(b, t);
 642                 }
 643         }
 644
 645         /*
 646          * Always rebuild search trees: eytzinger search tree nodes directly
 647          * depend on the values of min/max key:
 648          */
 649         bch2_bset_set_no_aux_tree(b, b->set);
 650         bch2_btree_build_aux_trees(b);
 651
 652         for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
 653                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 654                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 655         }
 656 }
 657
 658 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 659                          struct btree *b, struct bset *i,
 660                          unsigned offset, unsigned sectors,
 661                          int write, bool have_retry)
 662 {
 663         unsigned version = le16_to_cpu(i->version);
 664         const char *err;
 665         struct printbuf buf1 = PRINTBUF;
 666         struct printbuf buf2 = PRINTBUF;
 667         int ret = 0;
 668
 669         btree_err_on((version != BCH_BSET_VERSION_OLD &&
 670                       version < bcachefs_metadata_version_min) ||
 671                      version >= bcachefs_metadata_version_max,
 672                      BTREE_ERR_FATAL, c, ca, b, i,
 673                      "unsupported bset version");
 674
 675         if (btree_err_on(version < c->sb.version_min,
 676                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 677                          "bset version %u older than superblock version_min %u",
 678                          version, c->sb.version_min)) {
 679                 mutex_lock(&c->sb_lock);
 680                 c->disk_sb.sb->version_min = cpu_to_le16(version);
 681                 bch2_write_super(c);
 682                 mutex_unlock(&c->sb_lock);
 683         }
 684
 685         if (btree_err_on(version > c->sb.version,
 686                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 687                          "bset version %u newer than superblock version %u",
 688                          version, c->sb.version)) {
 689                 mutex_lock(&c->sb_lock);
 690                 c->disk_sb.sb->version = cpu_to_le16(version);
 691                 bch2_write_super(c);
 692                 mutex_unlock(&c->sb_lock);
 693         }
 694
 695         btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
 696                      BTREE_ERR_FATAL, c, ca, b, i,
 697                      "BSET_SEPARATE_WHITEOUTS no longer supported");
 698
 699         if (btree_err_on(offset + sectors > btree_sectors(c),
 700                          BTREE_ERR_FIXABLE, c, ca, b, i,
 701                          "bset past end of btree node")) {
 702                 i->u64s = 0;
 703                 ret = 0;
 704                 goto out;
 705         }
 706
 707         btree_err_on(offset && !i->u64s,
 708                      BTREE_ERR_FIXABLE, c, ca, b, i,
 709                      "empty bset");
 710
 711         btree_err_on(BSET_OFFSET(i) &&
 712                      BSET_OFFSET(i) != offset,
 713                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 714                      "bset at wrong sector offset");
 715
 716         if (!offset) {
 717                 struct btree_node *bn =
 718                         container_of(i, struct btree_node, keys);
 719                 /* These indicate that we read the wrong btree node: */
 720
 721                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 722                         struct bch_btree_ptr_v2 *bp =
 723                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 724
 725                         /* XXX endianness */
 726                         btree_err_on(bp->seq != bn->keys.seq,
 727                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 728                                      "incorrect sequence number (wrong btree node)");
 729                 }
 730
 731                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 732                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 733                              "incorrect btree id");
 734
 735                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 736                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 737                              "incorrect level");
 738
 739                 if (!write)
 740                         compat_btree_node(b->c.level, b->c.btree_id, version,
 741                                           BSET_BIG_ENDIAN(i), write, bn);
 742
 743                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 744                         struct bch_btree_ptr_v2 *bp =
 745                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 746
 747                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 748                                 b->data->min_key = bp->min_key;
 749                                 b->data->max_key = b->key.k.p;
 750                         }
 751
 752                         btree_err_on(!bpos_eq(b->data->min_key, bp->min_key),
 753                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 754                                      "incorrect min_key: got %s should be %s",
 755                                      (printbuf_reset(&buf1),
 756                                       bch2_bpos_to_text(&buf1, bn->min_key), buf1.buf),
 757                                      (printbuf_reset(&buf2),
 758                                       bch2_bpos_to_text(&buf2, bp->min_key), buf2.buf));
 759                 }
 760
 761                 btree_err_on(!bpos_eq(bn->max_key, b->key.k.p),
 762                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 763                              "incorrect max key %s",
 764                              (printbuf_reset(&buf1),
 765                               bch2_bpos_to_text(&buf1, bn->max_key), buf1.buf));
 766
 767                 if (write)
 768                         compat_btree_node(b->c.level, b->c.btree_id, version,
 769                                           BSET_BIG_ENDIAN(i), write, bn);
 770
 771                 err = bch2_bkey_format_validate(&bn->format);
 772                 btree_err_on(err,
 773                              BTREE_ERR_FATAL, c, ca, b, i,
 774                              "invalid bkey format: %s", err);
 775
 776                 compat_bformat(b->c.level, b->c.btree_id, version,
 777                                BSET_BIG_ENDIAN(i), write,
 778                                &bn->format);
 779         }
 780 out:
 781 fsck_err:
 782         printbuf_exit(&buf2);
 783         printbuf_exit(&buf1);
 784         return ret;
 785 }
 786
 787 static int bset_key_invalid(struct bch_fs *c, struct btree *b,
 788                             struct bkey_s_c k,
 789                             bool updated_range, int rw,
 790                             struct printbuf *err)
 791 {
 792         return __bch2_bkey_invalid(c, k, btree_node_type(b), READ, err) ?:
 793                 (!updated_range ? bch2_bkey_in_btree_node(b, k, err) : 0) ?:
 794                 (rw == WRITE ? bch2_bkey_val_invalid(c, k, READ, err) : 0);
 795 }
 796
 797 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 798                          struct bset *i, int write, bool have_retry)
 799 {
 800         unsigned version = le16_to_cpu(i->version);
 801         struct bkey_packed *k, *prev = NULL;
 802         struct printbuf buf = PRINTBUF;
 803         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 804                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 805         int ret = 0;
 806
 807         for (k = i->start;
 808              k != vstruct_last(i);) {
 809                 struct bkey_s u;
 810                 struct bkey tmp;
 811
 812                 if (btree_err_on(bkey_next(k) > vstruct_last(i),
 813                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 814                                  "key extends past end of bset")) {
 815                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 816                         break;
 817                 }
 818
 819                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 820                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 821                                  "invalid bkey format %u", k->format)) {
 822                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 823                         memmove_u64s_down(k, bkey_next(k),
 824                                           (u64 *) vstruct_end(i) - (u64 *) k);
 825                         continue;
 826                 }
 827
 828                 /* XXX: validate k->u64s */
 829                 if (!write)
 830                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 831                                     BSET_BIG_ENDIAN(i), write,
 832                                     &b->format, k);
 833
 834                 u = __bkey_disassemble(b, k, &tmp);
 835
 836                 printbuf_reset(&buf);
 837                 if (bset_key_invalid(c, b, u.s_c, updated_range, write, &buf)) {
 838                         printbuf_reset(&buf);
 839                         prt_printf(&buf, "invalid bkey:  ");
 840                         bset_key_invalid(c, b, u.s_c, updated_range, write, &buf);
 841                         prt_printf(&buf, "\n  ");
 842                         bch2_bkey_val_to_text(&buf, c, u.s_c);
 843
 844                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf);
 845
 846                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 847                         memmove_u64s_down(k, bkey_next(k),
 848                                           (u64 *) vstruct_end(i) - (u64 *) k);
 849                         continue;
 850                 }
 851
 852                 if (write)
 853                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 854                                     BSET_BIG_ENDIAN(i), write,
 855                                     &b->format, k);
 856
 857                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 858                         struct bkey up = bkey_unpack_key(b, prev);
 859
 860                         printbuf_reset(&buf);
 861                         prt_printf(&buf, "keys out of order: ");
 862                         bch2_bkey_to_text(&buf, &up);
 863                         prt_printf(&buf, " > ");
 864                         bch2_bkey_to_text(&buf, u.k);
 865
 866                         bch2_dump_bset(c, b, i, 0);
 867
 868                         if (btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf)) {
 869                                 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 870                                 memmove_u64s_down(k, bkey_next(k),
 871                                                   (u64 *) vstruct_end(i) - (u64 *) k);
 872                                 continue;
 873                         }
 874                 }
 875
 876                 prev = k;
 877                 k = bkey_next(k);
 878         }
 879 fsck_err:
 880         printbuf_exit(&buf);
 881         return ret;
 882 }
 883
 884 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 885                               struct btree *b, bool have_retry)
 886 {
 887         struct btree_node_entry *bne;
 888         struct sort_iter *iter;
 889         struct btree_node *sorted;
 890         struct bkey_packed *k;
 891         struct bch_extent_ptr *ptr;
 892         struct bset *i;
 893         bool used_mempool, blacklisted;
 894         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 895                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 896         unsigned u64s;
 897         unsigned blacklisted_written, nonblacklisted_written = 0;
 898         unsigned ptr_written = btree_ptr_sectors_written(&b->key);
 899         struct printbuf buf = PRINTBUF;
 900         int ret, retry_read = 0, write = READ;
 901
 902         b->version_ondisk = U16_MAX;
 903         /* We might get called multiple times on read retry: */
 904         b->written = 0;
 905
 906         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
 907         sort_iter_init(iter, b);
 908         iter->size = (btree_blocks(c) + 1) * 2;
 909
 910         if (bch2_meta_read_fault("btree"))
 911                 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 912                           "dynamic fault");
 913
 914         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 915                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 916                      "bad magic: want %llx, got %llx",
 917                      bset_magic(c), le64_to_cpu(b->data->magic));
 918
 919         btree_err_on(!b->data->keys.seq,
 920                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 921                      "bad btree header: seq 0");
 922
 923         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 924                 struct bch_btree_ptr_v2 *bp =
 925                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 926
 927                 btree_err_on(b->data->keys.seq != bp->seq,
 928                              BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 929                              "got wrong btree node (seq %llx want %llx)",
 930                              b->data->keys.seq, bp->seq);
 931         }
 932
 933         while (b->written < (ptr_written ?: btree_sectors(c))) {
 934                 unsigned sectors;
 935                 struct nonce nonce;
 936                 struct bch_csum csum;
 937                 bool first = !b->written;
 938
 939                 if (!b->written) {
 940                         i = &b->data->keys;
 941
 942                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 943                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 944                                      "unknown checksum type %llu",
 945                                      BSET_CSUM_TYPE(i));
 946
 947                         nonce = btree_nonce(i, b->written << 9);
 948                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 949
 950                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 951                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 952                                      "invalid checksum");
 953
 954                         ret = bset_encrypt(c, i, b->written << 9);
 955                         if (bch2_fs_fatal_err_on(ret, c,
 956                                         "error decrypting btree node: %i", ret))
 957                                 goto fsck_err;
 958
 959                         btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
 960                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
 961                                      BTREE_ERR_FATAL, c, NULL, b, NULL,
 962                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
 963
 964                         sectors = vstruct_sectors(b->data, c->block_bits);
 965                 } else {
 966                         bne = write_block(b);
 967                         i = &bne->keys;
 968
 969                         if (i->seq != b->data->keys.seq)
 970                                 break;
 971
 972                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 973                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 974                                      "unknown checksum type %llu",
 975                                      BSET_CSUM_TYPE(i));
 976
 977                         nonce = btree_nonce(i, b->written << 9);
 978                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
 979
 980                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
 981                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 982                                      "invalid checksum");
 983
 984                         ret = bset_encrypt(c, i, b->written << 9);
 985                         if (bch2_fs_fatal_err_on(ret, c,
 986                                         "error decrypting btree node: %i\n", ret))
 987                                 goto fsck_err;
 988
 989                         sectors = vstruct_sectors(bne, c->block_bits);
 990                 }
 991
 992                 b->version_ondisk = min(b->version_ondisk,
 993                                         le16_to_cpu(i->version));
 994
 995                 ret = validate_bset(c, ca, b, i, b->written, sectors,
 996                                     READ, have_retry);
 997                 if (ret)
 998                         goto fsck_err;
 999
1000                 if (!b->written)
1001                         btree_node_set_format(b, b->data->format);
1002
1003                 ret = validate_bset_keys(c, b, i, READ, have_retry);
1004                 if (ret)
1005                         goto fsck_err;
1006
1007                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1008
1009                 blacklisted = bch2_journal_seq_is_blacklisted(c,
1010                                         le64_to_cpu(i->journal_seq),
1011                                         true);
1012
1013                 btree_err_on(blacklisted && first,
1014                              BTREE_ERR_FIXABLE, c, ca, b, i,
1015                              "first btree node bset has blacklisted journal seq (%llu)",
1016                              le64_to_cpu(i->journal_seq));
1017
1018                 btree_err_on(blacklisted && ptr_written,
1019                              BTREE_ERR_FIXABLE, c, ca, b, i,
1020                              "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
1021                              le64_to_cpu(i->journal_seq),
1022                              b->written, b->written + sectors, ptr_written);
1023
1024                 b->written += sectors;
1025
1026                 if (blacklisted && !first)
1027                         continue;
1028
1029                 sort_iter_add(iter,
1030                               vstruct_idx(i, 0),
1031                               vstruct_last(i));
1032
1033                 nonblacklisted_written = b->written;
1034         }
1035
1036         if (ptr_written) {
1037                 btree_err_on(b->written < ptr_written,
1038                              BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1039                              "btree node data missing: expected %u sectors, found %u",
1040                              ptr_written, b->written);
1041         } else {
1042                 for (bne = write_block(b);
1043                      bset_byte_offset(b, bne) < btree_bytes(c);
1044                      bne = (void *) bne + block_bytes(c))
1045                         btree_err_on(bne->keys.seq == b->data->keys.seq &&
1046                                      !bch2_journal_seq_is_blacklisted(c,
1047                                                                       le64_to_cpu(bne->keys.journal_seq),
1048                                                                       true),
1049                                      BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1050                                      "found bset signature after last bset");
1051
1052                 /*
1053                  * Blacklisted bsets are those that were written after the most recent
1054                  * (flush) journal write. Since there wasn't a flush, they may not have
1055                  * made it to all devices - which means we shouldn't write new bsets
1056                  * after them, as that could leave a gap and then reads from that device
1057                  * wouldn't find all the bsets in that btree node - which means it's
1058                  * important that we start writing new bsets after the most recent _non_
1059                  * blacklisted bset:
1060                  */
1061                 blacklisted_written = b->written;
1062                 b->written = nonblacklisted_written;
1063         }
1064
1065         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1066         sorted->keys.u64s = 0;
1067
1068         set_btree_bset(b, b->set, &b->data->keys);
1069
1070         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1071
1072         u64s = le16_to_cpu(sorted->keys.u64s);
1073         *sorted = *b->data;
1074         sorted->keys.u64s = cpu_to_le16(u64s);
1075         swap(sorted, b->data);
1076         set_btree_bset(b, b->set, &b->data->keys);
1077         b->nsets = 1;
1078
1079         BUG_ON(b->nr.live_u64s != u64s);
1080
1081         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1082
1083         if (updated_range)
1084                 bch2_btree_node_drop_keys_outside_node(b);
1085
1086         i = &b->data->keys;
1087         for (k = i->start; k != vstruct_last(i);) {
1088                 struct bkey tmp;
1089                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1090
1091                 printbuf_reset(&buf);
1092
1093                 if (bch2_bkey_val_invalid(c, u.s_c, READ, &buf) ||
1094                     (bch2_inject_invalid_keys &&
1095                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1096                         printbuf_reset(&buf);
1097
1098                         prt_printf(&buf, "invalid bkey: ");
1099                         bch2_bkey_val_invalid(c, u.s_c, READ, &buf);
1100                         prt_printf(&buf, "\n  ");
1101                         bch2_bkey_val_to_text(&buf, c, u.s_c);
1102
1103                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf);
1104
1105                         btree_keys_account_key_drop(&b->nr, 0, k);
1106
1107                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1108                         memmove_u64s_down(k, bkey_next(k),
1109                                           (u64 *) vstruct_end(i) - (u64 *) k);
1110                         set_btree_bset_end(b, b->set);
1111                         continue;
1112                 }
1113
1114                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1115                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1116
1117                         bp.v->mem_ptr = 0;
1118                 }
1119
1120                 k = bkey_next(k);
1121         }
1122
1123         bch2_bset_build_aux_tree(b, b->set, false);
1124
1125         set_needs_whiteout(btree_bset_first(b), true);
1126
1127         btree_node_reset_sib_u64s(b);
1128
1129         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1130                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1131
1132                 if (ca->mi.state != BCH_MEMBER_STATE_rw)
1133                         set_btree_node_need_rewrite(b);
1134         }
1135
1136         if (!ptr_written)
1137                 set_btree_node_need_rewrite(b);
1138 out:
1139         mempool_free(iter, &c->fill_iter);
1140         printbuf_exit(&buf);
1141         return retry_read;
1142 fsck_err:
1143         if (ret == BTREE_RETRY_READ) {
1144                 retry_read = 1;
1145         } else {
1146                 bch2_inconsistent_error(c);
1147                 set_btree_node_read_error(b);
1148         }
1149         goto out;
1150 }
1151
1152 static void btree_node_read_work(struct work_struct *work)
1153 {
1154         struct btree_read_bio *rb =
1155                 container_of(work, struct btree_read_bio, work);
1156         struct bch_fs *c        = rb->c;
1157         struct btree *b         = rb->b;
1158         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1159         struct bio *bio         = &rb->bio;
1160         struct bch_io_failures failed = { .nr = 0 };
1161         struct printbuf buf = PRINTBUF;
1162         bool saw_error = false;
1163         bool retry = false;
1164         bool can_retry;
1165
1166         goto start;
1167         while (1) {
1168                 retry = true;
1169                 bch_info(c, "retrying read");
1170                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1171                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1172                 bio_reset(bio, NULL, REQ_OP_READ|REQ_SYNC|REQ_META);
1173                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1174                 bio->bi_iter.bi_size    = btree_bytes(c);
1175
1176                 if (rb->have_ioref) {
1177                         bio_set_dev(bio, ca->disk_sb.bdev);
1178                         submit_bio_wait(bio);
1179                 } else {
1180                         bio->bi_status = BLK_STS_REMOVED;
1181                 }
1182 start:
1183                 printbuf_reset(&buf);
1184                 btree_pos_to_text(&buf, c, b);
1185                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1186                                    bch2_blk_status_to_str(bio->bi_status), buf.buf);
1187                 if (rb->have_ioref)
1188                         percpu_ref_put(&ca->io_ref);
1189                 rb->have_ioref = false;
1190
1191                 bch2_mark_io_failure(&failed, &rb->pick);
1192
1193                 can_retry = bch2_bkey_pick_read_device(c,
1194                                 bkey_i_to_s_c(&b->key),
1195                                 &failed, &rb->pick) > 0;
1196
1197                 if (!bio->bi_status &&
1198                     !bch2_btree_node_read_done(c, ca, b, can_retry)) {
1199                         if (retry)
1200                                 bch_info(c, "retry success");
1201                         break;
1202                 }
1203
1204                 saw_error = true;
1205
1206                 if (!can_retry) {
1207                         set_btree_node_read_error(b);
1208                         break;
1209                 }
1210         }
1211
1212         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1213                                rb->start_time);
1214         bio_put(&rb->bio);
1215         printbuf_exit(&buf);
1216
1217         if (saw_error && !btree_node_read_error(b))
1218                 bch2_btree_node_rewrite_async(c, b);
1219
1220         clear_btree_node_read_in_flight(b);
1221         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1222 }
1223
1224 static void btree_node_read_endio(struct bio *bio)
1225 {
1226         struct btree_read_bio *rb =
1227                 container_of(bio, struct btree_read_bio, bio);
1228         struct bch_fs *c        = rb->c;
1229
1230         if (rb->have_ioref) {
1231                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1232
1233                 bch2_latency_acct(ca, rb->start_time, READ);
1234         }
1235
1236         queue_work(c->io_complete_wq, &rb->work);
1237 }
1238
1239 struct btree_node_read_all {
1240         struct closure          cl;
1241         struct bch_fs           *c;
1242         struct btree            *b;
1243         unsigned                nr;
1244         void                    *buf[BCH_REPLICAS_MAX];
1245         struct bio              *bio[BCH_REPLICAS_MAX];
1246         int                     err[BCH_REPLICAS_MAX];
1247 };
1248
1249 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1250 {
1251         struct btree_node *bn = data;
1252         struct btree_node_entry *bne;
1253         unsigned offset = 0;
1254
1255         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1256                 return 0;
1257
1258         while (offset < btree_sectors(c)) {
1259                 if (!offset) {
1260                         offset += vstruct_sectors(bn, c->block_bits);
1261                 } else {
1262                         bne = data + (offset << 9);
1263                         if (bne->keys.seq != bn->keys.seq)
1264                                 break;
1265                         offset += vstruct_sectors(bne, c->block_bits);
1266                 }
1267         }
1268
1269         return offset;
1270 }
1271
1272 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1273 {
1274         struct btree_node *bn = data;
1275         struct btree_node_entry *bne;
1276
1277         if (!offset)
1278                 return false;
1279
1280         while (offset < btree_sectors(c)) {
1281                 bne = data + (offset << 9);
1282                 if (bne->keys.seq == bn->keys.seq)
1283                         return true;
1284                 offset++;
1285         }
1286
1287         return false;
1288         return offset;
1289 }
1290
1291 static void btree_node_read_all_replicas_done(struct closure *cl)
1292 {
1293         struct btree_node_read_all *ra =
1294                 container_of(cl, struct btree_node_read_all, cl);
1295         struct bch_fs *c = ra->c;
1296         struct btree *b = ra->b;
1297         struct printbuf buf = PRINTBUF;
1298         bool dump_bset_maps = false;
1299         bool have_retry = false;
1300         int ret = 0, best = -1, write = READ;
1301         unsigned i, written = 0, written2 = 0;
1302         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1303                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1304
1305         for (i = 0; i < ra->nr; i++) {
1306                 struct btree_node *bn = ra->buf[i];
1307
1308                 if (ra->err[i])
1309                         continue;
1310
1311                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1312                     (seq && seq != bn->keys.seq))
1313                         continue;
1314
1315                 if (best < 0) {
1316                         best = i;
1317                         written = btree_node_sectors_written(c, bn);
1318                         continue;
1319                 }
1320
1321                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1322                 if (btree_err_on(written2 != written, BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1323                                  "btree node sectors written mismatch: %u != %u",
1324                                  written, written2) ||
1325                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1326                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1327                                  "found bset signature after last bset") ||
1328                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1329                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1330                                  "btree node replicas content mismatch"))
1331                         dump_bset_maps = true;
1332
1333                 if (written2 > written) {
1334                         written = written2;
1335                         best = i;
1336                 }
1337         }
1338 fsck_err:
1339         if (dump_bset_maps) {
1340                 for (i = 0; i < ra->nr; i++) {
1341                         struct btree_node *bn = ra->buf[i];
1342                         struct btree_node_entry *bne = NULL;
1343                         unsigned offset = 0, sectors;
1344                         bool gap = false;
1345
1346                         if (ra->err[i])
1347                                 continue;
1348
1349                         printbuf_reset(&buf);
1350
1351                         while (offset < btree_sectors(c)) {
1352                                 if (!offset) {
1353                                         sectors = vstruct_sectors(bn, c->block_bits);
1354                                 } else {
1355                                         bne = ra->buf[i] + (offset << 9);
1356                                         if (bne->keys.seq != bn->keys.seq)
1357                                                 break;
1358                                         sectors = vstruct_sectors(bne, c->block_bits);
1359                                 }
1360
1361                                 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1362                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1363                                                         le64_to_cpu(bne->keys.journal_seq), false))
1364                                         prt_printf(&buf, "*");
1365                                 offset += sectors;
1366                         }
1367
1368                         while (offset < btree_sectors(c)) {
1369                                 bne = ra->buf[i] + (offset << 9);
1370                                 if (bne->keys.seq == bn->keys.seq) {
1371                                         if (!gap)
1372                                                 prt_printf(&buf, " GAP");
1373                                         gap = true;
1374
1375                                         sectors = vstruct_sectors(bne, c->block_bits);
1376                                         prt_printf(&buf, " %u-%u", offset, offset + sectors);
1377                                         if (bch2_journal_seq_is_blacklisted(c,
1378                                                         le64_to_cpu(bne->keys.journal_seq), false))
1379                                                 prt_printf(&buf, "*");
1380                                 }
1381                                 offset++;
1382                         }
1383
1384                         bch_err(c, "replica %u:%s", i, buf.buf);
1385                 }
1386         }
1387
1388         if (best >= 0) {
1389                 memcpy(b->data, ra->buf[best], btree_bytes(c));
1390                 ret = bch2_btree_node_read_done(c, NULL, b, false);
1391         } else {
1392                 ret = -1;
1393         }
1394
1395         if (ret)
1396                 set_btree_node_read_error(b);
1397
1398         for (i = 0; i < ra->nr; i++) {
1399                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1400                 bio_put(ra->bio[i]);
1401         }
1402
1403         closure_debug_destroy(&ra->cl);
1404         kfree(ra);
1405         printbuf_exit(&buf);
1406
1407         clear_btree_node_read_in_flight(b);
1408         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1409 }
1410
1411 static void btree_node_read_all_replicas_endio(struct bio *bio)
1412 {
1413         struct btree_read_bio *rb =
1414                 container_of(bio, struct btree_read_bio, bio);
1415         struct bch_fs *c        = rb->c;
1416         struct btree_node_read_all *ra = rb->ra;
1417
1418         if (rb->have_ioref) {
1419                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1420
1421                 bch2_latency_acct(ca, rb->start_time, READ);
1422         }
1423
1424         ra->err[rb->idx] = bio->bi_status;
1425         closure_put(&ra->cl);
1426 }
1427
1428 /*
1429  * XXX This allocates multiple times from the same mempools, and can deadlock
1430  * under sufficient memory pressure (but is only a debug path)
1431  */
1432 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1433 {
1434         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1435         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1436         const union bch_extent_entry *entry;
1437         struct extent_ptr_decoded pick;
1438         struct btree_node_read_all *ra;
1439         unsigned i;
1440
1441         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1442         if (!ra)
1443                 return -ENOMEM;
1444
1445         closure_init(&ra->cl, NULL);
1446         ra->c   = c;
1447         ra->b   = b;
1448         ra->nr  = bch2_bkey_nr_ptrs(k);
1449
1450         for (i = 0; i < ra->nr; i++) {
1451                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1452                 ra->bio[i] = bio_alloc_bioset(NULL,
1453                                               buf_pages(ra->buf[i], btree_bytes(c)),
1454                                               REQ_OP_READ|REQ_SYNC|REQ_META,
1455                                               GFP_NOFS,
1456                                               &c->btree_bio);
1457         }
1458
1459         i = 0;
1460         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1461                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1462                 struct btree_read_bio *rb =
1463                         container_of(ra->bio[i], struct btree_read_bio, bio);
1464                 rb->c                   = c;
1465                 rb->b                   = b;
1466                 rb->ra                  = ra;
1467                 rb->start_time          = local_clock();
1468                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1469                 rb->idx                 = i;
1470                 rb->pick                = pick;
1471                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1472                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1473                 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1474
1475                 if (rb->have_ioref) {
1476                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1477                                      bio_sectors(&rb->bio));
1478                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1479
1480                         closure_get(&ra->cl);
1481                         submit_bio(&rb->bio);
1482                 } else {
1483                         ra->err[i] = BLK_STS_REMOVED;
1484                 }
1485
1486                 i++;
1487         }
1488
1489         if (sync) {
1490                 closure_sync(&ra->cl);
1491                 btree_node_read_all_replicas_done(&ra->cl);
1492         } else {
1493                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1494                             c->io_complete_wq);
1495         }
1496
1497         return 0;
1498 }
1499
1500 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1501                           bool sync)
1502 {
1503         struct extent_ptr_decoded pick;
1504         struct btree_read_bio *rb;
1505         struct bch_dev *ca;
1506         struct bio *bio;
1507         int ret;
1508
1509         trace_and_count(c, btree_node_read, c, b);
1510
1511         if (bch2_verify_all_btree_replicas &&
1512             !btree_node_read_all_replicas(c, b, sync))
1513                 return;
1514
1515         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1516                                          NULL, &pick);
1517
1518         if (ret <= 0) {
1519                 struct printbuf buf = PRINTBUF;
1520
1521                 prt_str(&buf, "btree node read error: no device to read from\n at ");
1522                 btree_pos_to_text(&buf, c, b);
1523                 bch_err(c, "%s", buf.buf);
1524
1525                 if (test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags))
1526                         bch2_fatal_error(c);
1527
1528                 set_btree_node_read_error(b);
1529                 clear_btree_node_read_in_flight(b);
1530                 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1531                 printbuf_exit(&buf);
1532                 return;
1533         }
1534
1535         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1536
1537         bio = bio_alloc_bioset(NULL,
1538                                buf_pages(b->data, btree_bytes(c)),
1539                                REQ_OP_READ|REQ_SYNC|REQ_META,
1540                                GFP_NOIO,
1541                                &c->btree_bio);
1542         rb = container_of(bio, struct btree_read_bio, bio);
1543         rb->c                   = c;
1544         rb->b                   = b;
1545         rb->ra                  = NULL;
1546         rb->start_time          = local_clock();
1547         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1548         rb->pick                = pick;
1549         INIT_WORK(&rb->work, btree_node_read_work);
1550         bio->bi_iter.bi_sector  = pick.ptr.offset;
1551         bio->bi_end_io          = btree_node_read_endio;
1552         bch2_bio_map(bio, b->data, btree_bytes(c));
1553
1554         if (rb->have_ioref) {
1555                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1556                              bio_sectors(bio));
1557                 bio_set_dev(bio, ca->disk_sb.bdev);
1558
1559                 if (sync) {
1560                         submit_bio_wait(bio);
1561
1562                         btree_node_read_work(&rb->work);
1563                 } else {
1564                         submit_bio(bio);
1565                 }
1566         } else {
1567                 bio->bi_status = BLK_STS_REMOVED;
1568
1569                 if (sync)
1570                         btree_node_read_work(&rb->work);
1571                 else
1572                         queue_work(c->io_complete_wq, &rb->work);
1573         }
1574 }
1575
1576 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1577                         const struct bkey_i *k, unsigned level)
1578 {
1579         struct closure cl;
1580         struct btree *b;
1581         int ret;
1582
1583         closure_init_stack(&cl);
1584
1585         do {
1586                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1587                 closure_sync(&cl);
1588         } while (ret);
1589
1590         b = bch2_btree_node_mem_alloc(c, level != 0);
1591         bch2_btree_cache_cannibalize_unlock(c);
1592
1593         BUG_ON(IS_ERR(b));
1594
1595         bkey_copy(&b->key, k);
1596         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1597
1598         set_btree_node_read_in_flight(b);
1599
1600         bch2_btree_node_read(c, b, true);
1601
1602         if (btree_node_read_error(b)) {
1603                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1604
1605                 mutex_lock(&c->btree_cache.lock);
1606                 list_move(&b->list, &c->btree_cache.freeable);
1607                 mutex_unlock(&c->btree_cache.lock);
1608
1609                 ret = -EIO;
1610                 goto err;
1611         }
1612
1613         bch2_btree_set_root_for_read(c, b);
1614 err:
1615         six_unlock_write(&b->c.lock);
1616         six_unlock_intent(&b->c.lock);
1617
1618         return ret;
1619 }
1620
1621 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1622                               struct btree_write *w)
1623 {
1624         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1625
1626         do {
1627                 old = new = v;
1628                 if (!(old & 1))
1629                         break;
1630
1631                 new &= ~1UL;
1632         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1633
1634         if (old & 1)
1635                 closure_put(&((struct btree_update *) new)->cl);
1636
1637         bch2_journal_pin_drop(&c->journal, &w->journal);
1638 }
1639
1640 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1641 {
1642         struct btree_write *w = btree_prev_write(b);
1643         unsigned long old, new, v;
1644         unsigned type = 0;
1645
1646         bch2_btree_complete_write(c, b, w);
1647
1648         v = READ_ONCE(b->flags);
1649         do {
1650                 old = new = v;
1651
1652                 if ((old & (1U << BTREE_NODE_dirty)) &&
1653                     (old & (1U << BTREE_NODE_need_write)) &&
1654                     !(old & (1U << BTREE_NODE_never_write)) &&
1655                     !(old & (1U << BTREE_NODE_write_blocked)) &&
1656                     !(old & (1U << BTREE_NODE_will_make_reachable))) {
1657                         new &= ~(1U << BTREE_NODE_dirty);
1658                         new &= ~(1U << BTREE_NODE_need_write);
1659                         new |=  (1U << BTREE_NODE_write_in_flight);
1660                         new |=  (1U << BTREE_NODE_write_in_flight_inner);
1661                         new |=  (1U << BTREE_NODE_just_written);
1662                         new ^=  (1U << BTREE_NODE_write_idx);
1663
1664                         type = new & BTREE_WRITE_TYPE_MASK;
1665                         new &= ~BTREE_WRITE_TYPE_MASK;
1666                 } else {
1667                         new &= ~(1U << BTREE_NODE_write_in_flight);
1668                         new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1669                 }
1670         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1671
1672         if (new & (1U << BTREE_NODE_write_in_flight))
1673                 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED|type);
1674         else
1675                 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1676 }
1677
1678 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1679 {
1680         struct btree_trans trans;
1681
1682         bch2_trans_init(&trans, c, 0, 0);
1683
1684         btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
1685         __btree_node_write_done(c, b);
1686         six_unlock_read(&b->c.lock);
1687
1688         bch2_trans_exit(&trans);
1689 }
1690
1691 static void btree_node_write_work(struct work_struct *work)
1692 {
1693         struct btree_write_bio *wbio =
1694                 container_of(work, struct btree_write_bio, work);
1695         struct bch_fs *c        = wbio->wbio.c;
1696         struct btree *b         = wbio->wbio.bio.bi_private;
1697         struct bch_extent_ptr *ptr;
1698         int ret;
1699
1700         btree_bounce_free(c,
1701                 wbio->data_bytes,
1702                 wbio->wbio.used_mempool,
1703                 wbio->data);
1704
1705         bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1706                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1707
1708         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key)))
1709                 goto err;
1710
1711         if (wbio->wbio.first_btree_write) {
1712                 if (wbio->wbio.failed.nr) {
1713
1714                 }
1715         } else {
1716                 ret = bch2_trans_do(c, NULL, NULL, 0,
1717                         bch2_btree_node_update_key_get_iter(&trans, b, &wbio->key,
1718                                                             !wbio->wbio.failed.nr));
1719                 if (ret)
1720                         goto err;
1721         }
1722 out:
1723         bio_put(&wbio->wbio.bio);
1724         btree_node_write_done(c, b);
1725         return;
1726 err:
1727         set_btree_node_noevict(b);
1728         bch2_fs_fatal_error(c, "fatal error writing btree node");
1729         goto out;
1730 }
1731
1732 static void btree_node_write_endio(struct bio *bio)
1733 {
1734         struct bch_write_bio *wbio      = to_wbio(bio);
1735         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1736         struct bch_write_bio *orig      = parent ?: wbio;
1737         struct btree_write_bio *wb      = container_of(orig, struct btree_write_bio, wbio);
1738         struct bch_fs *c                = wbio->c;
1739         struct btree *b                 = wbio->bio.bi_private;
1740         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1741         unsigned long flags;
1742
1743         if (wbio->have_ioref)
1744                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1745
1746         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1747                                bch2_blk_status_to_str(bio->bi_status)) ||
1748             bch2_meta_write_fault("btree")) {
1749                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1750                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1751                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1752         }
1753
1754         if (wbio->have_ioref)
1755                 percpu_ref_put(&ca->io_ref);
1756
1757         if (parent) {
1758                 bio_put(bio);
1759                 bio_endio(&parent->bio);
1760                 return;
1761         }
1762
1763         clear_btree_node_write_in_flight_inner(b);
1764         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1765         INIT_WORK(&wb->work, btree_node_write_work);
1766         queue_work(c->btree_io_complete_wq, &wb->work);
1767 }
1768
1769 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1770                                    struct bset *i, unsigned sectors)
1771 {
1772         struct printbuf buf = PRINTBUF;
1773         int ret;
1774
1775         ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key),
1776                                 BKEY_TYPE_btree, WRITE, &buf);
1777
1778         if (ret)
1779                 bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf);
1780         printbuf_exit(&buf);
1781         if (ret)
1782                 return ret;
1783
1784         ret = validate_bset_keys(c, b, i, WRITE, false) ?:
1785                 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false);
1786         if (ret) {
1787                 bch2_inconsistent_error(c);
1788                 dump_stack();
1789         }
1790
1791         return ret;
1792 }
1793
1794 static void btree_write_submit(struct work_struct *work)
1795 {
1796         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1797         struct bch_extent_ptr *ptr;
1798         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1799
1800         bkey_copy(&tmp.k, &wbio->key);
1801
1802         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1803                 ptr->offset += wbio->sector_offset;
1804
1805         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree,
1806                                   &tmp.k, false);
1807 }
1808
1809 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1810 {
1811         struct btree_write_bio *wbio;
1812         struct bset_tree *t;
1813         struct bset *i;
1814         struct btree_node *bn = NULL;
1815         struct btree_node_entry *bne = NULL;
1816         struct sort_iter sort_iter;
1817         struct nonce nonce;
1818         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1819         u64 seq = 0;
1820         bool used_mempool;
1821         unsigned long old, new;
1822         bool validate_before_checksum = false;
1823         enum btree_write_type type = flags & BTREE_WRITE_TYPE_MASK;
1824         void *data;
1825         int ret;
1826
1827         if (flags & BTREE_WRITE_ALREADY_STARTED)
1828                 goto do_write;
1829
1830         /*
1831          * We may only have a read lock on the btree node - the dirty bit is our
1832          * "lock" against racing with other threads that may be trying to start
1833          * a write, we do a write iff we clear the dirty bit. Since setting the
1834          * dirty bit requires a write lock, we can't race with other threads
1835          * redirtying it:
1836          */
1837         do {
1838                 old = new = READ_ONCE(b->flags);
1839
1840                 if (!(old & (1 << BTREE_NODE_dirty)))
1841                         return;
1842
1843                 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
1844                     !(old & (1 << BTREE_NODE_need_write)))
1845                         return;
1846
1847                 if (old &
1848                     ((1 << BTREE_NODE_never_write)|
1849                      (1 << BTREE_NODE_write_blocked)))
1850                         return;
1851
1852                 if (b->written &&
1853                     (old & (1 << BTREE_NODE_will_make_reachable)))
1854                         return;
1855
1856                 if (old & (1 << BTREE_NODE_write_in_flight))
1857                         return;
1858
1859                 if (flags & BTREE_WRITE_ONLY_IF_NEED)
1860                         type = new & BTREE_WRITE_TYPE_MASK;
1861                 new &= ~BTREE_WRITE_TYPE_MASK;
1862
1863                 new &= ~(1 << BTREE_NODE_dirty);
1864                 new &= ~(1 << BTREE_NODE_need_write);
1865                 new |=  (1 << BTREE_NODE_write_in_flight);
1866                 new |=  (1 << BTREE_NODE_write_in_flight_inner);
1867                 new |=  (1 << BTREE_NODE_just_written);
1868                 new ^=  (1 << BTREE_NODE_write_idx);
1869         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1870
1871         if (new & (1U << BTREE_NODE_need_write))
1872                 return;
1873 do_write:
1874         BUG_ON((type == BTREE_WRITE_initial) != (b->written == 0));
1875
1876         atomic_dec(&c->btree_cache.dirty);
1877
1878         BUG_ON(btree_node_fake(b));
1879         BUG_ON((b->will_make_reachable != 0) != !b->written);
1880
1881         BUG_ON(b->written >= btree_sectors(c));
1882         BUG_ON(b->written & (block_sectors(c) - 1));
1883         BUG_ON(bset_written(b, btree_bset_last(b)));
1884         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1885         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1886
1887         bch2_sort_whiteouts(c, b);
1888
1889         sort_iter_init(&sort_iter, b);
1890
1891         bytes = !b->written
1892                 ? sizeof(struct btree_node)
1893                 : sizeof(struct btree_node_entry);
1894
1895         bytes += b->whiteout_u64s * sizeof(u64);
1896
1897         for_each_bset(b, t) {
1898                 i = bset(b, t);
1899
1900                 if (bset_written(b, i))
1901                         continue;
1902
1903                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1904                 sort_iter_add(&sort_iter,
1905                               btree_bkey_first(b, t),
1906                               btree_bkey_last(b, t));
1907                 seq = max(seq, le64_to_cpu(i->journal_seq));
1908         }
1909
1910         BUG_ON(b->written && !seq);
1911
1912         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1913         bytes += 8;
1914
1915         /* buffer must be a multiple of the block size */
1916         bytes = round_up(bytes, block_bytes(c));
1917
1918         data = btree_bounce_alloc(c, bytes, &used_mempool);
1919
1920         if (!b->written) {
1921                 bn = data;
1922                 *bn = *b->data;
1923                 i = &bn->keys;
1924         } else {
1925                 bne = data;
1926                 bne->keys = b->data->keys;
1927                 i = &bne->keys;
1928         }
1929
1930         i->journal_seq  = cpu_to_le64(seq);
1931         i->u64s         = 0;
1932
1933         sort_iter_add(&sort_iter,
1934                       unwritten_whiteouts_start(c, b),
1935                       unwritten_whiteouts_end(c, b));
1936         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1937
1938         b->whiteout_u64s = 0;
1939
1940         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1941         le16_add_cpu(&i->u64s, u64s);
1942
1943         BUG_ON(!b->written && i->u64s != b->data->keys.u64s);
1944
1945         set_needs_whiteout(i, false);
1946
1947         /* do we have data to write? */
1948         if (b->written && !i->u64s)
1949                 goto nowrite;
1950
1951         bytes_to_write = vstruct_end(i) - data;
1952         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1953
1954         if (!b->written &&
1955             b->key.k.type == KEY_TYPE_btree_ptr_v2)
1956                 BUG_ON(btree_ptr_sectors_written(&b->key) != sectors_to_write);
1957
1958         memset(data + bytes_to_write, 0,
1959                (sectors_to_write << 9) - bytes_to_write);
1960
1961         BUG_ON(b->written + sectors_to_write > btree_sectors(c));
1962         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
1963         BUG_ON(i->seq != b->data->keys.seq);
1964
1965         i->version = c->sb.version < bcachefs_metadata_version_bkey_renumber
1966                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
1967                 : cpu_to_le16(c->sb.version);
1968         SET_BSET_OFFSET(i, b->written);
1969         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
1970
1971         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
1972                 validate_before_checksum = true;
1973
1974         /* validate_bset will be modifying: */
1975         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
1976                 validate_before_checksum = true;
1977
1978         /* if we're going to be encrypting, check metadata validity first: */
1979         if (validate_before_checksum &&
1980             validate_bset_for_write(c, b, i, sectors_to_write))
1981                 goto err;
1982
1983         ret = bset_encrypt(c, i, b->written << 9);
1984         if (bch2_fs_fatal_err_on(ret, c,
1985                         "error encrypting btree node: %i\n", ret))
1986                 goto err;
1987
1988         nonce = btree_nonce(i, b->written << 9);
1989
1990         if (bn)
1991                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
1992         else
1993                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1994
1995         /* if we're not encrypting, check metadata after checksumming: */
1996         if (!validate_before_checksum &&
1997             validate_bset_for_write(c, b, i, sectors_to_write))
1998                 goto err;
1999
2000         /*
2001          * We handle btree write errors by immediately halting the journal -
2002          * after we've done that, we can't issue any subsequent btree writes
2003          * because they might have pointers to new nodes that failed to write.
2004          *
2005          * Furthermore, there's no point in doing any more btree writes because
2006          * with the journal stopped, we're never going to update the journal to
2007          * reflect that those writes were done and the data flushed from the
2008          * journal:
2009          *
2010          * Also on journal error, the pending write may have updates that were
2011          * never journalled (interior nodes, see btree_update_nodes_written()) -
2012          * it's critical that we don't do the write in that case otherwise we
2013          * will have updates visible that weren't in the journal:
2014          *
2015          * Make sure to update b->written so bch2_btree_init_next() doesn't
2016          * break:
2017          */
2018         if (bch2_journal_error(&c->journal) ||
2019             c->opts.nochanges)
2020                 goto err;
2021
2022         trace_and_count(c, btree_node_write, b, bytes_to_write, sectors_to_write);
2023
2024         wbio = container_of(bio_alloc_bioset(NULL,
2025                                 buf_pages(data, sectors_to_write << 9),
2026                                 REQ_OP_WRITE|REQ_META,
2027                                 GFP_NOIO,
2028                                 &c->btree_bio),
2029                             struct btree_write_bio, wbio.bio);
2030         wbio_init(&wbio->wbio.bio);
2031         wbio->data                      = data;
2032         wbio->data_bytes                = bytes;
2033         wbio->sector_offset             = b->written;
2034         wbio->wbio.c                    = c;
2035         wbio->wbio.used_mempool         = used_mempool;
2036         wbio->wbio.first_btree_write    = !b->written;
2037         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
2038         wbio->wbio.bio.bi_private       = b;
2039
2040         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
2041
2042         bkey_copy(&wbio->key, &b->key);
2043
2044         b->written += sectors_to_write;
2045
2046         if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
2047                 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2048                         cpu_to_le16(b->written);
2049
2050         atomic64_inc(&c->btree_write_stats[type].nr);
2051         atomic64_add(bytes_to_write, &c->btree_write_stats[type].bytes);
2052
2053         INIT_WORK(&wbio->work, btree_write_submit);
2054         queue_work(c->io_complete_wq, &wbio->work);
2055         return;
2056 err:
2057         set_btree_node_noevict(b);
2058         b->written += sectors_to_write;
2059 nowrite:
2060         btree_bounce_free(c, bytes, used_mempool, data);
2061         __btree_node_write_done(c, b);
2062 }
2063
2064 /*
2065  * Work that must be done with write lock held:
2066  */
2067 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2068 {
2069         bool invalidated_iter = false;
2070         struct btree_node_entry *bne;
2071         struct bset_tree *t;
2072
2073         if (!btree_node_just_written(b))
2074                 return false;
2075
2076         BUG_ON(b->whiteout_u64s);
2077
2078         clear_btree_node_just_written(b);
2079
2080         /*
2081          * Note: immediately after write, bset_written() doesn't work - the
2082          * amount of data we had to write after compaction might have been
2083          * smaller than the offset of the last bset.
2084          *
2085          * However, we know that all bsets have been written here, as long as
2086          * we're still holding the write lock:
2087          */
2088
2089         /*
2090          * XXX: decide if we really want to unconditionally sort down to a
2091          * single bset:
2092          */
2093         if (b->nsets > 1) {
2094                 btree_node_sort(c, b, 0, b->nsets, true);
2095                 invalidated_iter = true;
2096         } else {
2097                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2098         }
2099
2100         for_each_bset(b, t)
2101                 set_needs_whiteout(bset(b, t), true);
2102
2103         bch2_btree_verify(c, b);
2104
2105         /*
2106          * If later we don't unconditionally sort down to a single bset, we have
2107          * to ensure this is still true:
2108          */
2109         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2110
2111         bne = want_new_bset(c, b);
2112         if (bne)
2113                 bch2_bset_init_next(c, b, bne);
2114
2115         bch2_btree_build_aux_trees(b);
2116
2117         return invalidated_iter;
2118 }
2119
2120 /*
2121  * Use this one if the node is intent locked:
2122  */
2123 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2124                            enum six_lock_type lock_type_held,
2125                            unsigned flags)
2126 {
2127         if (lock_type_held == SIX_LOCK_intent ||
2128             (lock_type_held == SIX_LOCK_read &&
2129              six_lock_tryupgrade(&b->c.lock))) {
2130                 __bch2_btree_node_write(c, b, flags);
2131
2132                 /* don't cycle lock unnecessarily: */
2133                 if (btree_node_just_written(b) &&
2134                     six_trylock_write(&b->c.lock)) {
2135                         bch2_btree_post_write_cleanup(c, b);
2136                         six_unlock_write(&b->c.lock);
2137                 }
2138
2139                 if (lock_type_held == SIX_LOCK_read)
2140                         six_lock_downgrade(&b->c.lock);
2141         } else {
2142                 __bch2_btree_node_write(c, b, flags);
2143                 if (lock_type_held == SIX_LOCK_write &&
2144                     btree_node_just_written(b))
2145                         bch2_btree_post_write_cleanup(c, b);
2146         }
2147 }
2148
2149 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2150 {
2151         struct bucket_table *tbl;
2152         struct rhash_head *pos;
2153         struct btree *b;
2154         unsigned i;
2155         bool ret = false;
2156 restart:
2157         rcu_read_lock();
2158         for_each_cached_btree(b, c, tbl, i, pos)
2159                 if (test_bit(flag, &b->flags)) {
2160                         rcu_read_unlock();
2161                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2162                         ret = true;
2163                         goto restart;
2164                 }
2165         rcu_read_unlock();
2166
2167         return ret;
2168 }
2169
2170 bool bch2_btree_flush_all_reads(struct bch_fs *c)
2171 {
2172         return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2173 }
2174
2175 bool bch2_btree_flush_all_writes(struct bch_fs *c)
2176 {
2177         return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2178 }
2179
2180 const char * const bch2_btree_write_types[] = {
2181 #define x(t, n) [n] = #t,
2182         BCH_BTREE_WRITE_TYPES()
2183         NULL
2184 };
2185
2186 void bch2_btree_write_stats_to_text(struct printbuf *out, struct bch_fs *c)
2187 {
2188         printbuf_tabstop_push(out, 20);
2189         printbuf_tabstop_push(out, 10);
2190
2191         prt_tab(out);
2192         prt_str(out, "nr");
2193         prt_tab(out);
2194         prt_str(out, "size");
2195         prt_newline(out);
2196
2197         for (unsigned i = 0; i < BTREE_WRITE_TYPE_NR; i++) {
2198                 u64 nr          = atomic64_read(&c->btree_write_stats[i].nr);
2199                 u64 bytes       = atomic64_read(&c->btree_write_stats[i].bytes);
2200
2201                 prt_printf(out, "%s:", bch2_btree_write_types[i]);
2202                 prt_tab(out);
2203                 prt_u64(out, nr);
2204                 prt_tab(out);
2205                 prt_human_readable_u64(out, nr ? div64_u64(bytes, nr) : 0);
2206                 prt_newline(out);
2207         }
2208 }