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, "%s"),
 530                    write == READ
 531                    ? "error validating btree node "
 532                    : "corrupt btree node before write ");
 533         if (ca)
 534                 prt_printf(out, "on %s ", ca->name);
 535         prt_printf(out, "at btree ");
 536         btree_pos_to_text(out, c, b);
 537
 538         prt_printf(out, "\n  node offset %u", b->written);
 539         if (i)
 540                 prt_printf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 541         prt_str(out, ": ");
 542 }
 543
 544 enum btree_err_type {
 545         /*
 546          * We can repair this locally, and we're after the checksum check so
 547          * there's no need to try another replica:
 548          */
 549         BTREE_ERR_FIXABLE,
 550         /*
 551          * We can repair this if we have to, but we should try reading another
 552          * replica if we can:
 553          */
 554         BTREE_ERR_WANT_RETRY,
 555         /*
 556          * Read another replica if we have one, otherwise consider the whole
 557          * node bad:
 558          */
 559         BTREE_ERR_MUST_RETRY,
 560         BTREE_ERR_BAD_NODE,
 561         BTREE_ERR_INCOMPATIBLE,
 562 };
 563
 564 enum btree_validate_ret {
 565         BTREE_RETRY_READ = 64,
 566 };
 567
 568 static int __btree_err(enum btree_err_type type,
 569                        struct bch_fs *c,
 570                        struct bch_dev *ca,
 571                        struct btree *b,
 572                        struct bset *i,
 573                        int write,
 574                        bool have_retry,
 575                        const char *fmt, ...)
 576 {
 577         struct printbuf out = PRINTBUF;
 578         va_list args;
 579         int ret = -BCH_ERR_fsck_fix;
 580
 581         btree_err_msg(&out, c, ca, b, i, b->written, write);
 582
 583         va_start(args, fmt);
 584         prt_vprintf(&out, fmt, args);
 585         va_end(args);
 586
 587         if (write == WRITE) {
 588                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 589                 ret = c->opts.errors == BCH_ON_ERROR_continue
 590                         ? 0
 591                         : -BCH_ERR_fsck_errors_not_fixed;
 592                 goto out;
 593         }
 594
 595         if (!have_retry && type == BTREE_ERR_WANT_RETRY)
 596                 type = BTREE_ERR_FIXABLE;
 597         if (!have_retry && type == BTREE_ERR_MUST_RETRY)
 598                 type = BTREE_ERR_BAD_NODE;
 599
 600         switch (type) {
 601         case BTREE_ERR_FIXABLE:
 602                 mustfix_fsck_err(c, "%s", out.buf);
 603                 ret = -BCH_ERR_fsck_fix;
 604                 break;
 605         case BTREE_ERR_WANT_RETRY:
 606         case BTREE_ERR_MUST_RETRY:
 607                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 608                 ret = BTREE_RETRY_READ;
 609                 break;
 610         case BTREE_ERR_BAD_NODE:
 611                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 612                 bch2_topology_error(c);
 613                 ret = -BCH_ERR_need_topology_repair;
 614                 break;
 615         case BTREE_ERR_INCOMPATIBLE:
 616                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 617                 ret = -BCH_ERR_fsck_errors_not_fixed;
 618                 break;
 619         default:
 620                 BUG();
 621         }
 622 out:
 623 fsck_err:
 624         printbuf_exit(&out);
 625         return ret;
 626 }
 627
 628 #define btree_err(type, c, ca, b, i, msg, ...)                          \
 629 ({                                                                      \
 630         int _ret = __btree_err(type, c, ca, b, i, write, have_retry, msg, ##__VA_ARGS__);\
 631                                                                         \
 632         if (_ret != -BCH_ERR_fsck_fix)                                  \
 633                 goto fsck_err;                                          \
 634         *saw_error = true;                                              \
 635 })
 636
 637 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 638
 639 /*
 640  * When btree topology repair changes the start or end of a node, that might
 641  * mean we have to drop keys that are no longer inside the node:
 642  */
 643 __cold
 644 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
 645 {
 646         struct bset_tree *t;
 647         struct bkey_s_c k;
 648         struct bkey unpacked;
 649         struct btree_node_iter iter;
 650
 651         for_each_bset(b, t) {
 652                 struct bset *i = bset(b, t);
 653                 struct bkey_packed *k;
 654
 655                 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
 656                         if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
 657                                 break;
 658
 659                 if (k != i->start) {
 660                         unsigned shift = (u64 *) k - (u64 *) i->start;
 661
 662                         memmove_u64s_down(i->start, k,
 663                                           (u64 *) vstruct_end(i) - (u64 *) k);
 664                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
 665                         set_btree_bset_end(b, t);
 666                 }
 667
 668                 for (k = i->start; k != vstruct_last(i); k = bkey_next(k))
 669                         if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
 670                                 break;
 671
 672                 if (k != vstruct_last(i)) {
 673                         i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
 674                         set_btree_bset_end(b, t);
 675                 }
 676         }
 677
 678         /*
 679          * Always rebuild search trees: eytzinger search tree nodes directly
 680          * depend on the values of min/max key:
 681          */
 682         bch2_bset_set_no_aux_tree(b, b->set);
 683         bch2_btree_build_aux_trees(b);
 684
 685         for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
 686                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 687                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 688         }
 689 }
 690
 691 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 692                          struct btree *b, struct bset *i,
 693                          unsigned offset, unsigned sectors,
 694                          int write, bool have_retry, bool *saw_error)
 695 {
 696         unsigned version = le16_to_cpu(i->version);
 697         const char *err;
 698         struct printbuf buf1 = PRINTBUF;
 699         struct printbuf buf2 = PRINTBUF;
 700         int ret = 0;
 701
 702         btree_err_on((version != BCH_BSET_VERSION_OLD &&
 703                       version < bcachefs_metadata_version_min) ||
 704                      version >= bcachefs_metadata_version_max,
 705                      BTREE_ERR_INCOMPATIBLE, c, ca, b, i,
 706                      "unsupported bset version");
 707
 708         if (btree_err_on(version < c->sb.version_min,
 709                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 710                          "bset version %u older than superblock version_min %u",
 711                          version, c->sb.version_min)) {
 712                 mutex_lock(&c->sb_lock);
 713                 c->disk_sb.sb->version_min = cpu_to_le16(version);
 714                 bch2_write_super(c);
 715                 mutex_unlock(&c->sb_lock);
 716         }
 717
 718         if (btree_err_on(version > c->sb.version,
 719                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 720                          "bset version %u newer than superblock version %u",
 721                          version, c->sb.version)) {
 722                 mutex_lock(&c->sb_lock);
 723                 c->disk_sb.sb->version = cpu_to_le16(version);
 724                 bch2_write_super(c);
 725                 mutex_unlock(&c->sb_lock);
 726         }
 727
 728         btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
 729                      BTREE_ERR_INCOMPATIBLE, c, ca, b, i,
 730                      "BSET_SEPARATE_WHITEOUTS no longer supported");
 731
 732         if (btree_err_on(offset + sectors > btree_sectors(c),
 733                          BTREE_ERR_FIXABLE, c, ca, b, i,
 734                          "bset past end of btree node")) {
 735                 i->u64s = 0;
 736                 ret = 0;
 737                 goto out;
 738         }
 739
 740         btree_err_on(offset && !i->u64s,
 741                      BTREE_ERR_FIXABLE, c, ca, b, i,
 742                      "empty bset");
 743
 744         btree_err_on(BSET_OFFSET(i) &&
 745                      BSET_OFFSET(i) != offset,
 746                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 747                      "bset at wrong sector offset");
 748
 749         if (!offset) {
 750                 struct btree_node *bn =
 751                         container_of(i, struct btree_node, keys);
 752                 /* These indicate that we read the wrong btree node: */
 753
 754                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 755                         struct bch_btree_ptr_v2 *bp =
 756                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 757
 758                         /* XXX endianness */
 759                         btree_err_on(bp->seq != bn->keys.seq,
 760                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 761                                      "incorrect sequence number (wrong btree node)");
 762                 }
 763
 764                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 765                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 766                              "incorrect btree id");
 767
 768                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 769                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 770                              "incorrect level");
 771
 772                 if (!write)
 773                         compat_btree_node(b->c.level, b->c.btree_id, version,
 774                                           BSET_BIG_ENDIAN(i), write, bn);
 775
 776                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 777                         struct bch_btree_ptr_v2 *bp =
 778                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 779
 780                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 781                                 b->data->min_key = bp->min_key;
 782                                 b->data->max_key = b->key.k.p;
 783                         }
 784
 785                         btree_err_on(!bpos_eq(b->data->min_key, bp->min_key),
 786                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 787                                      "incorrect min_key: got %s should be %s",
 788                                      (printbuf_reset(&buf1),
 789                                       bch2_bpos_to_text(&buf1, bn->min_key), buf1.buf),
 790                                      (printbuf_reset(&buf2),
 791                                       bch2_bpos_to_text(&buf2, bp->min_key), buf2.buf));
 792                 }
 793
 794                 btree_err_on(!bpos_eq(bn->max_key, b->key.k.p),
 795                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 796                              "incorrect max key %s",
 797                              (printbuf_reset(&buf1),
 798                               bch2_bpos_to_text(&buf1, bn->max_key), buf1.buf));
 799
 800                 if (write)
 801                         compat_btree_node(b->c.level, b->c.btree_id, version,
 802                                           BSET_BIG_ENDIAN(i), write, bn);
 803
 804                 err = bch2_bkey_format_validate(&bn->format);
 805                 btree_err_on(err,
 806                              BTREE_ERR_BAD_NODE, c, ca, b, i,
 807                              "invalid bkey format: %s", err);
 808
 809                 compat_bformat(b->c.level, b->c.btree_id, version,
 810                                BSET_BIG_ENDIAN(i), write,
 811                                &bn->format);
 812         }
 813 out:
 814 fsck_err:
 815         printbuf_exit(&buf2);
 816         printbuf_exit(&buf1);
 817         return ret;
 818 }
 819
 820 static int bset_key_invalid(struct bch_fs *c, struct btree *b,
 821                             struct bkey_s_c k,
 822                             bool updated_range, int rw,
 823                             struct printbuf *err)
 824 {
 825         return __bch2_bkey_invalid(c, k, btree_node_type(b), READ, err) ?:
 826                 (!updated_range ? bch2_bkey_in_btree_node(b, k, err) : 0) ?:
 827                 (rw == WRITE ? bch2_bkey_val_invalid(c, k, READ, err) : 0);
 828 }
 829
 830 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 831                          struct bset *i, int write,
 832                          bool have_retry, bool *saw_error)
 833 {
 834         unsigned version = le16_to_cpu(i->version);
 835         struct bkey_packed *k, *prev = NULL;
 836         struct printbuf buf = PRINTBUF;
 837         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 838                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 839         int ret = 0;
 840
 841         for (k = i->start;
 842              k != vstruct_last(i);) {
 843                 struct bkey_s u;
 844                 struct bkey tmp;
 845
 846                 if (btree_err_on(bkey_next(k) > vstruct_last(i),
 847                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 848                                  "key extends past end of bset")) {
 849                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 850                         break;
 851                 }
 852
 853                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 854                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 855                                  "invalid bkey format %u", k->format)) {
 856                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 857                         memmove_u64s_down(k, bkey_next(k),
 858                                           (u64 *) vstruct_end(i) - (u64 *) k);
 859                         continue;
 860                 }
 861
 862                 /* XXX: validate k->u64s */
 863                 if (!write)
 864                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 865                                     BSET_BIG_ENDIAN(i), write,
 866                                     &b->format, k);
 867
 868                 u = __bkey_disassemble(b, k, &tmp);
 869
 870                 printbuf_reset(&buf);
 871                 if (bset_key_invalid(c, b, u.s_c, updated_range, write, &buf)) {
 872                         printbuf_reset(&buf);
 873                         prt_printf(&buf, "invalid bkey:  ");
 874                         bset_key_invalid(c, b, u.s_c, updated_range, write, &buf);
 875                         prt_printf(&buf, "\n  ");
 876                         bch2_bkey_val_to_text(&buf, c, u.s_c);
 877
 878                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf);
 879
 880                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 881                         memmove_u64s_down(k, bkey_next(k),
 882                                           (u64 *) vstruct_end(i) - (u64 *) k);
 883                         continue;
 884                 }
 885
 886                 if (write)
 887                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 888                                     BSET_BIG_ENDIAN(i), write,
 889                                     &b->format, k);
 890
 891                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 892                         struct bkey up = bkey_unpack_key(b, prev);
 893
 894                         printbuf_reset(&buf);
 895                         prt_printf(&buf, "keys out of order: ");
 896                         bch2_bkey_to_text(&buf, &up);
 897                         prt_printf(&buf, " > ");
 898                         bch2_bkey_to_text(&buf, u.k);
 899
 900                         bch2_dump_bset(c, b, i, 0);
 901
 902                         if (btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf)) {
 903                                 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 904                                 memmove_u64s_down(k, bkey_next(k),
 905                                                   (u64 *) vstruct_end(i) - (u64 *) k);
 906                                 continue;
 907                         }
 908                 }
 909
 910                 prev = k;
 911                 k = bkey_next(k);
 912         }
 913 fsck_err:
 914         printbuf_exit(&buf);
 915         return ret;
 916 }
 917
 918 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 919                               struct btree *b, bool have_retry, bool *saw_error)
 920 {
 921         struct btree_node_entry *bne;
 922         struct sort_iter *iter;
 923         struct btree_node *sorted;
 924         struct bkey_packed *k;
 925         struct bch_extent_ptr *ptr;
 926         struct bset *i;
 927         bool used_mempool, blacklisted;
 928         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 929                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 930         unsigned u64s;
 931         unsigned blacklisted_written, nonblacklisted_written = 0;
 932         unsigned ptr_written = btree_ptr_sectors_written(&b->key);
 933         struct printbuf buf = PRINTBUF;
 934         int ret = 0, retry_read = 0, write = READ;
 935
 936         b->version_ondisk = U16_MAX;
 937         /* We might get called multiple times on read retry: */
 938         b->written = 0;
 939
 940         iter = mempool_alloc(&c->fill_iter, GFP_NOIO);
 941         sort_iter_init(iter, b);
 942         iter->size = (btree_blocks(c) + 1) * 2;
 943
 944         if (bch2_meta_read_fault("btree"))
 945                 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 946                           "dynamic fault");
 947
 948         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 949                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 950                      "bad magic: want %llx, got %llx",
 951                      bset_magic(c), le64_to_cpu(b->data->magic));
 952
 953         btree_err_on(!b->data->keys.seq,
 954                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 955                      "bad btree header: seq 0");
 956
 957         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 958                 struct bch_btree_ptr_v2 *bp =
 959                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 960
 961                 btree_err_on(b->data->keys.seq != bp->seq,
 962                              BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 963                              "got wrong btree node (seq %llx want %llx)",
 964                              b->data->keys.seq, bp->seq);
 965         }
 966
 967         while (b->written < (ptr_written ?: btree_sectors(c))) {
 968                 unsigned sectors;
 969                 struct nonce nonce;
 970                 struct bch_csum csum;
 971                 bool first = !b->written;
 972
 973                 if (!b->written) {
 974                         i = &b->data->keys;
 975
 976                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 977                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 978                                      "unknown checksum type %llu",
 979                                      BSET_CSUM_TYPE(i));
 980
 981                         nonce = btree_nonce(i, b->written << 9);
 982                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 983
 984                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 985                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 986                                      "invalid checksum");
 987
 988                         ret = bset_encrypt(c, i, b->written << 9);
 989                         if (bch2_fs_fatal_err_on(ret, c,
 990                                         "error decrypting btree node: %i", ret))
 991                                 goto fsck_err;
 992
 993                         btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
 994                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
 995                                      BTREE_ERR_INCOMPATIBLE, c, NULL, b, NULL,
 996                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
 997
 998                         sectors = vstruct_sectors(b->data, c->block_bits);
 999                 } else {
1000                         bne = write_block(b);
1001                         i = &bne->keys;
1002
1003                         if (i->seq != b->data->keys.seq)
1004                                 break;
1005
1006                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1007                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
1008                                      "unknown checksum type %llu",
1009                                      BSET_CSUM_TYPE(i));
1010
1011                         nonce = btree_nonce(i, b->written << 9);
1012                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1013
1014                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
1015                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
1016                                      "invalid checksum");
1017
1018                         ret = bset_encrypt(c, i, b->written << 9);
1019                         if (bch2_fs_fatal_err_on(ret, c,
1020                                         "error decrypting btree node: %i\n", ret))
1021                                 goto fsck_err;
1022
1023                         sectors = vstruct_sectors(bne, c->block_bits);
1024                 }
1025
1026                 b->version_ondisk = min(b->version_ondisk,
1027                                         le16_to_cpu(i->version));
1028
1029                 ret = validate_bset(c, ca, b, i, b->written, sectors,
1030                                     READ, have_retry, saw_error);
1031                 if (ret)
1032                         goto fsck_err;
1033
1034                 if (!b->written)
1035                         btree_node_set_format(b, b->data->format);
1036
1037                 ret = validate_bset_keys(c, b, i, READ, have_retry, saw_error);
1038                 if (ret)
1039                         goto fsck_err;
1040
1041                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1042
1043                 blacklisted = bch2_journal_seq_is_blacklisted(c,
1044                                         le64_to_cpu(i->journal_seq),
1045                                         true);
1046
1047                 btree_err_on(blacklisted && first,
1048                              BTREE_ERR_FIXABLE, c, ca, b, i,
1049                              "first btree node bset has blacklisted journal seq (%llu)",
1050                              le64_to_cpu(i->journal_seq));
1051
1052                 btree_err_on(blacklisted && ptr_written,
1053                              BTREE_ERR_FIXABLE, c, ca, b, i,
1054                              "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
1055                              le64_to_cpu(i->journal_seq),
1056                              b->written, b->written + sectors, ptr_written);
1057
1058                 b->written += sectors;
1059
1060                 if (blacklisted && !first)
1061                         continue;
1062
1063                 sort_iter_add(iter,
1064                               vstruct_idx(i, 0),
1065                               vstruct_last(i));
1066
1067                 nonblacklisted_written = b->written;
1068         }
1069
1070         if (ptr_written) {
1071                 btree_err_on(b->written < ptr_written,
1072                              BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1073                              "btree node data missing: expected %u sectors, found %u",
1074                              ptr_written, b->written);
1075         } else {
1076                 for (bne = write_block(b);
1077                      bset_byte_offset(b, bne) < btree_bytes(c);
1078                      bne = (void *) bne + block_bytes(c))
1079                         btree_err_on(bne->keys.seq == b->data->keys.seq &&
1080                                      !bch2_journal_seq_is_blacklisted(c,
1081                                                                       le64_to_cpu(bne->keys.journal_seq),
1082                                                                       true),
1083                                      BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1084                                      "found bset signature after last bset");
1085
1086                 /*
1087                  * Blacklisted bsets are those that were written after the most recent
1088                  * (flush) journal write. Since there wasn't a flush, they may not have
1089                  * made it to all devices - which means we shouldn't write new bsets
1090                  * after them, as that could leave a gap and then reads from that device
1091                  * wouldn't find all the bsets in that btree node - which means it's
1092                  * important that we start writing new bsets after the most recent _non_
1093                  * blacklisted bset:
1094                  */
1095                 blacklisted_written = b->written;
1096                 b->written = nonblacklisted_written;
1097         }
1098
1099         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1100         sorted->keys.u64s = 0;
1101
1102         set_btree_bset(b, b->set, &b->data->keys);
1103
1104         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1105
1106         u64s = le16_to_cpu(sorted->keys.u64s);
1107         *sorted = *b->data;
1108         sorted->keys.u64s = cpu_to_le16(u64s);
1109         swap(sorted, b->data);
1110         set_btree_bset(b, b->set, &b->data->keys);
1111         b->nsets = 1;
1112
1113         BUG_ON(b->nr.live_u64s != u64s);
1114
1115         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1116
1117         if (updated_range)
1118                 bch2_btree_node_drop_keys_outside_node(b);
1119
1120         i = &b->data->keys;
1121         for (k = i->start; k != vstruct_last(i);) {
1122                 struct bkey tmp;
1123                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1124
1125                 printbuf_reset(&buf);
1126
1127                 if (bch2_bkey_val_invalid(c, u.s_c, READ, &buf) ||
1128                     (bch2_inject_invalid_keys &&
1129                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1130                         printbuf_reset(&buf);
1131
1132                         prt_printf(&buf, "invalid bkey: ");
1133                         bch2_bkey_val_invalid(c, u.s_c, READ, &buf);
1134                         prt_printf(&buf, "\n  ");
1135                         bch2_bkey_val_to_text(&buf, c, u.s_c);
1136
1137                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf);
1138
1139                         btree_keys_account_key_drop(&b->nr, 0, k);
1140
1141                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1142                         memmove_u64s_down(k, bkey_next(k),
1143                                           (u64 *) vstruct_end(i) - (u64 *) k);
1144                         set_btree_bset_end(b, b->set);
1145                         continue;
1146                 }
1147
1148                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1149                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1150
1151                         bp.v->mem_ptr = 0;
1152                 }
1153
1154                 k = bkey_next(k);
1155         }
1156
1157         bch2_bset_build_aux_tree(b, b->set, false);
1158
1159         set_needs_whiteout(btree_bset_first(b), true);
1160
1161         btree_node_reset_sib_u64s(b);
1162
1163         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1164                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1165
1166                 if (ca->mi.state != BCH_MEMBER_STATE_rw)
1167                         set_btree_node_need_rewrite(b);
1168         }
1169
1170         if (!ptr_written)
1171                 set_btree_node_need_rewrite(b);
1172 out:
1173         mempool_free(iter, &c->fill_iter);
1174         printbuf_exit(&buf);
1175         return retry_read;
1176 fsck_err:
1177         if (ret == BTREE_RETRY_READ)
1178                 retry_read = 1;
1179         else
1180                 set_btree_node_read_error(b);
1181         goto out;
1182 }
1183
1184 static void btree_node_read_work(struct work_struct *work)
1185 {
1186         struct btree_read_bio *rb =
1187                 container_of(work, struct btree_read_bio, work);
1188         struct bch_fs *c        = rb->c;
1189         struct btree *b         = rb->b;
1190         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1191         struct bio *bio         = &rb->bio;
1192         struct bch_io_failures failed = { .nr = 0 };
1193         struct printbuf buf = PRINTBUF;
1194         bool saw_error = false;
1195         bool retry = false;
1196         bool can_retry;
1197
1198         goto start;
1199         while (1) {
1200                 retry = true;
1201                 bch_info(c, "retrying read");
1202                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1203                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1204                 bio_reset(bio, NULL, REQ_OP_READ|REQ_SYNC|REQ_META);
1205                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1206                 bio->bi_iter.bi_size    = btree_bytes(c);
1207
1208                 if (rb->have_ioref) {
1209                         bio_set_dev(bio, ca->disk_sb.bdev);
1210                         submit_bio_wait(bio);
1211                 } else {
1212                         bio->bi_status = BLK_STS_REMOVED;
1213                 }
1214 start:
1215                 printbuf_reset(&buf);
1216                 btree_pos_to_text(&buf, c, b);
1217                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1218                                    bch2_blk_status_to_str(bio->bi_status), buf.buf);
1219                 if (rb->have_ioref)
1220                         percpu_ref_put(&ca->io_ref);
1221                 rb->have_ioref = false;
1222
1223                 bch2_mark_io_failure(&failed, &rb->pick);
1224
1225                 can_retry = bch2_bkey_pick_read_device(c,
1226                                 bkey_i_to_s_c(&b->key),
1227                                 &failed, &rb->pick) > 0;
1228
1229                 if (!bio->bi_status &&
1230                     !bch2_btree_node_read_done(c, ca, b, can_retry, &saw_error)) {
1231                         if (retry)
1232                                 bch_info(c, "retry success");
1233                         break;
1234                 }
1235
1236                 saw_error = true;
1237
1238                 if (!can_retry) {
1239                         set_btree_node_read_error(b);
1240                         break;
1241                 }
1242         }
1243
1244         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1245                                rb->start_time);
1246         bio_put(&rb->bio);
1247         printbuf_exit(&buf);
1248
1249         if (!btree_node_read_error(b) &&
1250             (saw_error ||
1251              btree_node_need_rewrite(b))) {
1252                 struct printbuf buf = PRINTBUF;
1253
1254                 bch2_bpos_to_text(&buf, b->key.k.p);
1255                 bch_info(c, "%s: rewriting btree node at btree=%s level=%u %s due to error",
1256                          __func__, bch2_btree_ids[b->c.btree_id], b->c.level, buf.buf);
1257                 printbuf_exit(&buf);
1258
1259                 bch2_btree_node_rewrite_async(c, b);
1260         }
1261
1262         clear_btree_node_read_in_flight(b);
1263         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1264 }
1265
1266 static void btree_node_read_endio(struct bio *bio)
1267 {
1268         struct btree_read_bio *rb =
1269                 container_of(bio, struct btree_read_bio, bio);
1270         struct bch_fs *c        = rb->c;
1271
1272         if (rb->have_ioref) {
1273                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1274
1275                 bch2_latency_acct(ca, rb->start_time, READ);
1276         }
1277
1278         queue_work(c->io_complete_wq, &rb->work);
1279 }
1280
1281 struct btree_node_read_all {
1282         struct closure          cl;
1283         struct bch_fs           *c;
1284         struct btree            *b;
1285         unsigned                nr;
1286         void                    *buf[BCH_REPLICAS_MAX];
1287         struct bio              *bio[BCH_REPLICAS_MAX];
1288         int                     err[BCH_REPLICAS_MAX];
1289 };
1290
1291 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1292 {
1293         struct btree_node *bn = data;
1294         struct btree_node_entry *bne;
1295         unsigned offset = 0;
1296
1297         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1298                 return 0;
1299
1300         while (offset < btree_sectors(c)) {
1301                 if (!offset) {
1302                         offset += vstruct_sectors(bn, c->block_bits);
1303                 } else {
1304                         bne = data + (offset << 9);
1305                         if (bne->keys.seq != bn->keys.seq)
1306                                 break;
1307                         offset += vstruct_sectors(bne, c->block_bits);
1308                 }
1309         }
1310
1311         return offset;
1312 }
1313
1314 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1315 {
1316         struct btree_node *bn = data;
1317         struct btree_node_entry *bne;
1318
1319         if (!offset)
1320                 return false;
1321
1322         while (offset < btree_sectors(c)) {
1323                 bne = data + (offset << 9);
1324                 if (bne->keys.seq == bn->keys.seq)
1325                         return true;
1326                 offset++;
1327         }
1328
1329         return false;
1330         return offset;
1331 }
1332
1333 static void btree_node_read_all_replicas_done(struct closure *cl)
1334 {
1335         struct btree_node_read_all *ra =
1336                 container_of(cl, struct btree_node_read_all, cl);
1337         struct bch_fs *c = ra->c;
1338         struct btree *b = ra->b;
1339         struct printbuf buf = PRINTBUF;
1340         bool dump_bset_maps = false;
1341         bool have_retry = false;
1342         int ret = 0, best = -1, write = READ;
1343         unsigned i, written = 0, written2 = 0;
1344         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1345                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1346         bool _saw_error = false, *saw_error = &_saw_error;
1347
1348         for (i = 0; i < ra->nr; i++) {
1349                 struct btree_node *bn = ra->buf[i];
1350
1351                 if (ra->err[i])
1352                         continue;
1353
1354                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1355                     (seq && seq != bn->keys.seq))
1356                         continue;
1357
1358                 if (best < 0) {
1359                         best = i;
1360                         written = btree_node_sectors_written(c, bn);
1361                         continue;
1362                 }
1363
1364                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1365                 if (btree_err_on(written2 != written, BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1366                                  "btree node sectors written mismatch: %u != %u",
1367                                  written, written2) ||
1368                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1369                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1370                                  "found bset signature after last bset") ||
1371                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1372                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1373                                  "btree node replicas content mismatch"))
1374                         dump_bset_maps = true;
1375
1376                 if (written2 > written) {
1377                         written = written2;
1378                         best = i;
1379                 }
1380         }
1381 fsck_err:
1382         if (dump_bset_maps) {
1383                 for (i = 0; i < ra->nr; i++) {
1384                         struct btree_node *bn = ra->buf[i];
1385                         struct btree_node_entry *bne = NULL;
1386                         unsigned offset = 0, sectors;
1387                         bool gap = false;
1388
1389                         if (ra->err[i])
1390                                 continue;
1391
1392                         printbuf_reset(&buf);
1393
1394                         while (offset < btree_sectors(c)) {
1395                                 if (!offset) {
1396                                         sectors = vstruct_sectors(bn, c->block_bits);
1397                                 } else {
1398                                         bne = ra->buf[i] + (offset << 9);
1399                                         if (bne->keys.seq != bn->keys.seq)
1400                                                 break;
1401                                         sectors = vstruct_sectors(bne, c->block_bits);
1402                                 }
1403
1404                                 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1405                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1406                                                         le64_to_cpu(bne->keys.journal_seq), false))
1407                                         prt_printf(&buf, "*");
1408                                 offset += sectors;
1409                         }
1410
1411                         while (offset < btree_sectors(c)) {
1412                                 bne = ra->buf[i] + (offset << 9);
1413                                 if (bne->keys.seq == bn->keys.seq) {
1414                                         if (!gap)
1415                                                 prt_printf(&buf, " GAP");
1416                                         gap = true;
1417
1418                                         sectors = vstruct_sectors(bne, c->block_bits);
1419                                         prt_printf(&buf, " %u-%u", offset, offset + sectors);
1420                                         if (bch2_journal_seq_is_blacklisted(c,
1421                                                         le64_to_cpu(bne->keys.journal_seq), false))
1422                                                 prt_printf(&buf, "*");
1423                                 }
1424                                 offset++;
1425                         }
1426
1427                         bch_err(c, "replica %u:%s", i, buf.buf);
1428                 }
1429         }
1430
1431         if (best >= 0) {
1432                 memcpy(b->data, ra->buf[best], btree_bytes(c));
1433                 ret = bch2_btree_node_read_done(c, NULL, b, false, saw_error);
1434         } else {
1435                 ret = -1;
1436         }
1437
1438         if (ret)
1439                 set_btree_node_read_error(b);
1440         else if (*saw_error)
1441                 bch2_btree_node_rewrite_async(c, b);
1442
1443         for (i = 0; i < ra->nr; i++) {
1444                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1445                 bio_put(ra->bio[i]);
1446         }
1447
1448         closure_debug_destroy(&ra->cl);
1449         kfree(ra);
1450         printbuf_exit(&buf);
1451
1452         clear_btree_node_read_in_flight(b);
1453         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1454 }
1455
1456 static void btree_node_read_all_replicas_endio(struct bio *bio)
1457 {
1458         struct btree_read_bio *rb =
1459                 container_of(bio, struct btree_read_bio, bio);
1460         struct bch_fs *c        = rb->c;
1461         struct btree_node_read_all *ra = rb->ra;
1462
1463         if (rb->have_ioref) {
1464                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1465
1466                 bch2_latency_acct(ca, rb->start_time, READ);
1467         }
1468
1469         ra->err[rb->idx] = bio->bi_status;
1470         closure_put(&ra->cl);
1471 }
1472
1473 /*
1474  * XXX This allocates multiple times from the same mempools, and can deadlock
1475  * under sufficient memory pressure (but is only a debug path)
1476  */
1477 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1478 {
1479         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1480         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1481         const union bch_extent_entry *entry;
1482         struct extent_ptr_decoded pick;
1483         struct btree_node_read_all *ra;
1484         unsigned i;
1485
1486         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1487         if (!ra)
1488                 return -ENOMEM;
1489
1490         closure_init(&ra->cl, NULL);
1491         ra->c   = c;
1492         ra->b   = b;
1493         ra->nr  = bch2_bkey_nr_ptrs(k);
1494
1495         for (i = 0; i < ra->nr; i++) {
1496                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1497                 ra->bio[i] = bio_alloc_bioset(NULL,
1498                                               buf_pages(ra->buf[i], btree_bytes(c)),
1499                                               REQ_OP_READ|REQ_SYNC|REQ_META,
1500                                               GFP_NOFS,
1501                                               &c->btree_bio);
1502         }
1503
1504         i = 0;
1505         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1506                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1507                 struct btree_read_bio *rb =
1508                         container_of(ra->bio[i], struct btree_read_bio, bio);
1509                 rb->c                   = c;
1510                 rb->b                   = b;
1511                 rb->ra                  = ra;
1512                 rb->start_time          = local_clock();
1513                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1514                 rb->idx                 = i;
1515                 rb->pick                = pick;
1516                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1517                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1518                 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1519
1520                 if (rb->have_ioref) {
1521                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1522                                      bio_sectors(&rb->bio));
1523                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1524
1525                         closure_get(&ra->cl);
1526                         submit_bio(&rb->bio);
1527                 } else {
1528                         ra->err[i] = BLK_STS_REMOVED;
1529                 }
1530
1531                 i++;
1532         }
1533
1534         if (sync) {
1535                 closure_sync(&ra->cl);
1536                 btree_node_read_all_replicas_done(&ra->cl);
1537         } else {
1538                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1539                             c->io_complete_wq);
1540         }
1541
1542         return 0;
1543 }
1544
1545 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1546                           bool sync)
1547 {
1548         struct extent_ptr_decoded pick;
1549         struct btree_read_bio *rb;
1550         struct bch_dev *ca;
1551         struct bio *bio;
1552         int ret;
1553
1554         trace_and_count(c, btree_node_read, c, b);
1555
1556         if (bch2_verify_all_btree_replicas &&
1557             !btree_node_read_all_replicas(c, b, sync))
1558                 return;
1559
1560         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1561                                          NULL, &pick);
1562
1563         if (ret <= 0) {
1564                 struct printbuf buf = PRINTBUF;
1565
1566                 prt_str(&buf, "btree node read error: no device to read from\n at ");
1567                 btree_pos_to_text(&buf, c, b);
1568                 bch_err(c, "%s", buf.buf);
1569
1570                 if (test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags))
1571                         bch2_fatal_error(c);
1572
1573                 set_btree_node_read_error(b);
1574                 clear_btree_node_read_in_flight(b);
1575                 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1576                 printbuf_exit(&buf);
1577                 return;
1578         }
1579
1580         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1581
1582         bio = bio_alloc_bioset(NULL,
1583                                buf_pages(b->data, btree_bytes(c)),
1584                                REQ_OP_READ|REQ_SYNC|REQ_META,
1585                                GFP_NOIO,
1586                                &c->btree_bio);
1587         rb = container_of(bio, struct btree_read_bio, bio);
1588         rb->c                   = c;
1589         rb->b                   = b;
1590         rb->ra                  = NULL;
1591         rb->start_time          = local_clock();
1592         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1593         rb->pick                = pick;
1594         INIT_WORK(&rb->work, btree_node_read_work);
1595         bio->bi_iter.bi_sector  = pick.ptr.offset;
1596         bio->bi_end_io          = btree_node_read_endio;
1597         bch2_bio_map(bio, b->data, btree_bytes(c));
1598
1599         if (rb->have_ioref) {
1600                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1601                              bio_sectors(bio));
1602                 bio_set_dev(bio, ca->disk_sb.bdev);
1603
1604                 if (sync) {
1605                         submit_bio_wait(bio);
1606
1607                         btree_node_read_work(&rb->work);
1608                 } else {
1609                         submit_bio(bio);
1610                 }
1611         } else {
1612                 bio->bi_status = BLK_STS_REMOVED;
1613
1614                 if (sync)
1615                         btree_node_read_work(&rb->work);
1616                 else
1617                         queue_work(c->io_complete_wq, &rb->work);
1618         }
1619 }
1620
1621 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1622                         const struct bkey_i *k, unsigned level)
1623 {
1624         struct closure cl;
1625         struct btree *b;
1626         int ret;
1627
1628         closure_init_stack(&cl);
1629
1630         do {
1631                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1632                 closure_sync(&cl);
1633         } while (ret);
1634
1635         b = bch2_btree_node_mem_alloc(c, level != 0);
1636         bch2_btree_cache_cannibalize_unlock(c);
1637
1638         BUG_ON(IS_ERR(b));
1639
1640         bkey_copy(&b->key, k);
1641         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1642
1643         set_btree_node_read_in_flight(b);
1644
1645         bch2_btree_node_read(c, b, true);
1646
1647         if (btree_node_read_error(b)) {
1648                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1649
1650                 mutex_lock(&c->btree_cache.lock);
1651                 list_move(&b->list, &c->btree_cache.freeable);
1652                 mutex_unlock(&c->btree_cache.lock);
1653
1654                 ret = -EIO;
1655                 goto err;
1656         }
1657
1658         bch2_btree_set_root_for_read(c, b);
1659 err:
1660         six_unlock_write(&b->c.lock);
1661         six_unlock_intent(&b->c.lock);
1662
1663         return ret;
1664 }
1665
1666 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1667                               struct btree_write *w)
1668 {
1669         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1670
1671         do {
1672                 old = new = v;
1673                 if (!(old & 1))
1674                         break;
1675
1676                 new &= ~1UL;
1677         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1678
1679         if (old & 1)
1680                 closure_put(&((struct btree_update *) new)->cl);
1681
1682         bch2_journal_pin_drop(&c->journal, &w->journal);
1683 }
1684
1685 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1686 {
1687         struct btree_write *w = btree_prev_write(b);
1688         unsigned long old, new, v;
1689         unsigned type = 0;
1690
1691         bch2_btree_complete_write(c, b, w);
1692
1693         v = READ_ONCE(b->flags);
1694         do {
1695                 old = new = v;
1696
1697                 if ((old & (1U << BTREE_NODE_dirty)) &&
1698                     (old & (1U << BTREE_NODE_need_write)) &&
1699                     !(old & (1U << BTREE_NODE_never_write)) &&
1700                     !(old & (1U << BTREE_NODE_write_blocked)) &&
1701                     !(old & (1U << BTREE_NODE_will_make_reachable))) {
1702                         new &= ~(1U << BTREE_NODE_dirty);
1703                         new &= ~(1U << BTREE_NODE_need_write);
1704                         new |=  (1U << BTREE_NODE_write_in_flight);
1705                         new |=  (1U << BTREE_NODE_write_in_flight_inner);
1706                         new |=  (1U << BTREE_NODE_just_written);
1707                         new ^=  (1U << BTREE_NODE_write_idx);
1708
1709                         type = new & BTREE_WRITE_TYPE_MASK;
1710                         new &= ~BTREE_WRITE_TYPE_MASK;
1711                 } else {
1712                         new &= ~(1U << BTREE_NODE_write_in_flight);
1713                         new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1714                 }
1715         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1716
1717         if (new & (1U << BTREE_NODE_write_in_flight))
1718                 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED|type);
1719         else
1720                 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1721 }
1722
1723 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1724 {
1725         struct btree_trans trans;
1726
1727         bch2_trans_init(&trans, c, 0, 0);
1728
1729         btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
1730         __btree_node_write_done(c, b);
1731         six_unlock_read(&b->c.lock);
1732
1733         bch2_trans_exit(&trans);
1734 }
1735
1736 static void btree_node_write_work(struct work_struct *work)
1737 {
1738         struct btree_write_bio *wbio =
1739                 container_of(work, struct btree_write_bio, work);
1740         struct bch_fs *c        = wbio->wbio.c;
1741         struct btree *b         = wbio->wbio.bio.bi_private;
1742         struct bch_extent_ptr *ptr;
1743         int ret;
1744
1745         btree_bounce_free(c,
1746                 wbio->data_bytes,
1747                 wbio->wbio.used_mempool,
1748                 wbio->data);
1749
1750         bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1751                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1752
1753         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key)))
1754                 goto err;
1755
1756         if (wbio->wbio.first_btree_write) {
1757                 if (wbio->wbio.failed.nr) {
1758
1759                 }
1760         } else {
1761                 ret = bch2_trans_do(c, NULL, NULL, 0,
1762                         bch2_btree_node_update_key_get_iter(&trans, b, &wbio->key,
1763                                                             !wbio->wbio.failed.nr));
1764                 if (ret)
1765                         goto err;
1766         }
1767 out:
1768         bio_put(&wbio->wbio.bio);
1769         btree_node_write_done(c, b);
1770         return;
1771 err:
1772         set_btree_node_noevict(b);
1773         bch2_fs_fatal_error(c, "fatal error writing btree node");
1774         goto out;
1775 }
1776
1777 static void btree_node_write_endio(struct bio *bio)
1778 {
1779         struct bch_write_bio *wbio      = to_wbio(bio);
1780         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1781         struct bch_write_bio *orig      = parent ?: wbio;
1782         struct btree_write_bio *wb      = container_of(orig, struct btree_write_bio, wbio);
1783         struct bch_fs *c                = wbio->c;
1784         struct btree *b                 = wbio->bio.bi_private;
1785         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1786         unsigned long flags;
1787
1788         if (wbio->have_ioref)
1789                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1790
1791         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1792                                bch2_blk_status_to_str(bio->bi_status)) ||
1793             bch2_meta_write_fault("btree")) {
1794                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1795                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1796                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1797         }
1798
1799         if (wbio->have_ioref)
1800                 percpu_ref_put(&ca->io_ref);
1801
1802         if (parent) {
1803                 bio_put(bio);
1804                 bio_endio(&parent->bio);
1805                 return;
1806         }
1807
1808         clear_btree_node_write_in_flight_inner(b);
1809         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1810         INIT_WORK(&wb->work, btree_node_write_work);
1811         queue_work(c->btree_io_complete_wq, &wb->work);
1812 }
1813
1814 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1815                                    struct bset *i, unsigned sectors)
1816 {
1817         struct printbuf buf = PRINTBUF;
1818         bool saw_error;
1819         int ret;
1820
1821         ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key),
1822                                 BKEY_TYPE_btree, WRITE, &buf);
1823
1824         if (ret)
1825                 bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf);
1826         printbuf_exit(&buf);
1827         if (ret)
1828                 return ret;
1829
1830         ret = validate_bset_keys(c, b, i, WRITE, false, &saw_error) ?:
1831                 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false, &saw_error);
1832         if (ret) {
1833                 bch2_inconsistent_error(c);
1834                 dump_stack();
1835         }
1836
1837         return ret;
1838 }
1839
1840 static void btree_write_submit(struct work_struct *work)
1841 {
1842         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1843         struct bch_extent_ptr *ptr;
1844         __BKEY_PADDED(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1845
1846         bkey_copy(&tmp.k, &wbio->key);
1847
1848         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1849                 ptr->offset += wbio->sector_offset;
1850
1851         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree,
1852                                   &tmp.k, false);
1853 }
1854
1855 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1856 {
1857         struct btree_write_bio *wbio;
1858         struct bset_tree *t;
1859         struct bset *i;
1860         struct btree_node *bn = NULL;
1861         struct btree_node_entry *bne = NULL;
1862         struct sort_iter sort_iter;
1863         struct nonce nonce;
1864         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1865         u64 seq = 0;
1866         bool used_mempool;
1867         unsigned long old, new;
1868         bool validate_before_checksum = false;
1869         enum btree_write_type type = flags & BTREE_WRITE_TYPE_MASK;
1870         void *data;
1871         int ret;
1872
1873         if (flags & BTREE_WRITE_ALREADY_STARTED)
1874                 goto do_write;
1875
1876         /*
1877          * We may only have a read lock on the btree node - the dirty bit is our
1878          * "lock" against racing with other threads that may be trying to start
1879          * a write, we do a write iff we clear the dirty bit. Since setting the
1880          * dirty bit requires a write lock, we can't race with other threads
1881          * redirtying it:
1882          */
1883         do {
1884                 old = new = READ_ONCE(b->flags);
1885
1886                 if (!(old & (1 << BTREE_NODE_dirty)))
1887                         return;
1888
1889                 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
1890                     !(old & (1 << BTREE_NODE_need_write)))
1891                         return;
1892
1893                 if (old &
1894                     ((1 << BTREE_NODE_never_write)|
1895                      (1 << BTREE_NODE_write_blocked)))
1896                         return;
1897
1898                 if (b->written &&
1899                     (old & (1 << BTREE_NODE_will_make_reachable)))
1900                         return;
1901
1902                 if (old & (1 << BTREE_NODE_write_in_flight))
1903                         return;
1904
1905                 if (flags & BTREE_WRITE_ONLY_IF_NEED)
1906                         type = new & BTREE_WRITE_TYPE_MASK;
1907                 new &= ~BTREE_WRITE_TYPE_MASK;
1908
1909                 new &= ~(1 << BTREE_NODE_dirty);
1910                 new &= ~(1 << BTREE_NODE_need_write);
1911                 new |=  (1 << BTREE_NODE_write_in_flight);
1912                 new |=  (1 << BTREE_NODE_write_in_flight_inner);
1913                 new |=  (1 << BTREE_NODE_just_written);
1914                 new ^=  (1 << BTREE_NODE_write_idx);
1915         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1916
1917         if (new & (1U << BTREE_NODE_need_write))
1918                 return;
1919 do_write:
1920         BUG_ON((type == BTREE_WRITE_initial) != (b->written == 0));
1921
1922         atomic_dec(&c->btree_cache.dirty);
1923
1924         BUG_ON(btree_node_fake(b));
1925         BUG_ON((b->will_make_reachable != 0) != !b->written);
1926
1927         BUG_ON(b->written >= btree_sectors(c));
1928         BUG_ON(b->written & (block_sectors(c) - 1));
1929         BUG_ON(bset_written(b, btree_bset_last(b)));
1930         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1931         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1932
1933         bch2_sort_whiteouts(c, b);
1934
1935         sort_iter_init(&sort_iter, b);
1936
1937         bytes = !b->written
1938                 ? sizeof(struct btree_node)
1939                 : sizeof(struct btree_node_entry);
1940
1941         bytes += b->whiteout_u64s * sizeof(u64);
1942
1943         for_each_bset(b, t) {
1944                 i = bset(b, t);
1945
1946                 if (bset_written(b, i))
1947                         continue;
1948
1949                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1950                 sort_iter_add(&sort_iter,
1951                               btree_bkey_first(b, t),
1952                               btree_bkey_last(b, t));
1953                 seq = max(seq, le64_to_cpu(i->journal_seq));
1954         }
1955
1956         BUG_ON(b->written && !seq);
1957
1958         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1959         bytes += 8;
1960
1961         /* buffer must be a multiple of the block size */
1962         bytes = round_up(bytes, block_bytes(c));
1963
1964         data = btree_bounce_alloc(c, bytes, &used_mempool);
1965
1966         if (!b->written) {
1967                 bn = data;
1968                 *bn = *b->data;
1969                 i = &bn->keys;
1970         } else {
1971                 bne = data;
1972                 bne->keys = b->data->keys;
1973                 i = &bne->keys;
1974         }
1975
1976         i->journal_seq  = cpu_to_le64(seq);
1977         i->u64s         = 0;
1978
1979         sort_iter_add(&sort_iter,
1980                       unwritten_whiteouts_start(c, b),
1981                       unwritten_whiteouts_end(c, b));
1982         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1983
1984         b->whiteout_u64s = 0;
1985
1986         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1987         le16_add_cpu(&i->u64s, u64s);
1988
1989         BUG_ON(!b->written && i->u64s != b->data->keys.u64s);
1990
1991         set_needs_whiteout(i, false);
1992
1993         /* do we have data to write? */
1994         if (b->written && !i->u64s)
1995                 goto nowrite;
1996
1997         bytes_to_write = vstruct_end(i) - data;
1998         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
1999
2000         if (!b->written &&
2001             b->key.k.type == KEY_TYPE_btree_ptr_v2)
2002                 BUG_ON(btree_ptr_sectors_written(&b->key) != sectors_to_write);
2003
2004         memset(data + bytes_to_write, 0,
2005                (sectors_to_write << 9) - bytes_to_write);
2006
2007         BUG_ON(b->written + sectors_to_write > btree_sectors(c));
2008         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
2009         BUG_ON(i->seq != b->data->keys.seq);
2010
2011         i->version = c->sb.version < bcachefs_metadata_version_bkey_renumber
2012                 ? cpu_to_le16(BCH_BSET_VERSION_OLD)
2013                 : cpu_to_le16(c->sb.version);
2014         SET_BSET_OFFSET(i, b->written);
2015         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
2016
2017         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
2018                 validate_before_checksum = true;
2019
2020         /* validate_bset will be modifying: */
2021         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
2022                 validate_before_checksum = true;
2023
2024         /* if we're going to be encrypting, check metadata validity first: */
2025         if (validate_before_checksum &&
2026             validate_bset_for_write(c, b, i, sectors_to_write))
2027                 goto err;
2028
2029         ret = bset_encrypt(c, i, b->written << 9);
2030         if (bch2_fs_fatal_err_on(ret, c,
2031                         "error encrypting btree node: %i\n", ret))
2032                 goto err;
2033
2034         nonce = btree_nonce(i, b->written << 9);
2035
2036         if (bn)
2037                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
2038         else
2039                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
2040
2041         /* if we're not encrypting, check metadata after checksumming: */
2042         if (!validate_before_checksum &&
2043             validate_bset_for_write(c, b, i, sectors_to_write))
2044                 goto err;
2045
2046         /*
2047          * We handle btree write errors by immediately halting the journal -
2048          * after we've done that, we can't issue any subsequent btree writes
2049          * because they might have pointers to new nodes that failed to write.
2050          *
2051          * Furthermore, there's no point in doing any more btree writes because
2052          * with the journal stopped, we're never going to update the journal to
2053          * reflect that those writes were done and the data flushed from the
2054          * journal:
2055          *
2056          * Also on journal error, the pending write may have updates that were
2057          * never journalled (interior nodes, see btree_update_nodes_written()) -
2058          * it's critical that we don't do the write in that case otherwise we
2059          * will have updates visible that weren't in the journal:
2060          *
2061          * Make sure to update b->written so bch2_btree_init_next() doesn't
2062          * break:
2063          */
2064         if (bch2_journal_error(&c->journal) ||
2065             c->opts.nochanges)
2066                 goto err;
2067
2068         trace_and_count(c, btree_node_write, b, bytes_to_write, sectors_to_write);
2069
2070         wbio = container_of(bio_alloc_bioset(NULL,
2071                                 buf_pages(data, sectors_to_write << 9),
2072                                 REQ_OP_WRITE|REQ_META,
2073                                 GFP_NOIO,
2074                                 &c->btree_bio),
2075                             struct btree_write_bio, wbio.bio);
2076         wbio_init(&wbio->wbio.bio);
2077         wbio->data                      = data;
2078         wbio->data_bytes                = bytes;
2079         wbio->sector_offset             = b->written;
2080         wbio->wbio.c                    = c;
2081         wbio->wbio.used_mempool         = used_mempool;
2082         wbio->wbio.first_btree_write    = !b->written;
2083         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
2084         wbio->wbio.bio.bi_private       = b;
2085
2086         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
2087
2088         bkey_copy(&wbio->key, &b->key);
2089
2090         b->written += sectors_to_write;
2091
2092         if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
2093                 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2094                         cpu_to_le16(b->written);
2095
2096         atomic64_inc(&c->btree_write_stats[type].nr);
2097         atomic64_add(bytes_to_write, &c->btree_write_stats[type].bytes);
2098
2099         INIT_WORK(&wbio->work, btree_write_submit);
2100         queue_work(c->io_complete_wq, &wbio->work);
2101         return;
2102 err:
2103         set_btree_node_noevict(b);
2104         b->written += sectors_to_write;
2105 nowrite:
2106         btree_bounce_free(c, bytes, used_mempool, data);
2107         __btree_node_write_done(c, b);
2108 }
2109
2110 /*
2111  * Work that must be done with write lock held:
2112  */
2113 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2114 {
2115         bool invalidated_iter = false;
2116         struct btree_node_entry *bne;
2117         struct bset_tree *t;
2118
2119         if (!btree_node_just_written(b))
2120                 return false;
2121
2122         BUG_ON(b->whiteout_u64s);
2123
2124         clear_btree_node_just_written(b);
2125
2126         /*
2127          * Note: immediately after write, bset_written() doesn't work - the
2128          * amount of data we had to write after compaction might have been
2129          * smaller than the offset of the last bset.
2130          *
2131          * However, we know that all bsets have been written here, as long as
2132          * we're still holding the write lock:
2133          */
2134
2135         /*
2136          * XXX: decide if we really want to unconditionally sort down to a
2137          * single bset:
2138          */
2139         if (b->nsets > 1) {
2140                 btree_node_sort(c, b, 0, b->nsets, true);
2141                 invalidated_iter = true;
2142         } else {
2143                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2144         }
2145
2146         for_each_bset(b, t)
2147                 set_needs_whiteout(bset(b, t), true);
2148
2149         bch2_btree_verify(c, b);
2150
2151         /*
2152          * If later we don't unconditionally sort down to a single bset, we have
2153          * to ensure this is still true:
2154          */
2155         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2156
2157         bne = want_new_bset(c, b);
2158         if (bne)
2159                 bch2_bset_init_next(c, b, bne);
2160
2161         bch2_btree_build_aux_trees(b);
2162
2163         return invalidated_iter;
2164 }
2165
2166 /*
2167  * Use this one if the node is intent locked:
2168  */
2169 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2170                            enum six_lock_type lock_type_held,
2171                            unsigned flags)
2172 {
2173         if (lock_type_held == SIX_LOCK_intent ||
2174             (lock_type_held == SIX_LOCK_read &&
2175              six_lock_tryupgrade(&b->c.lock))) {
2176                 __bch2_btree_node_write(c, b, flags);
2177
2178                 /* don't cycle lock unnecessarily: */
2179                 if (btree_node_just_written(b) &&
2180                     six_trylock_write(&b->c.lock)) {
2181                         bch2_btree_post_write_cleanup(c, b);
2182                         six_unlock_write(&b->c.lock);
2183                 }
2184
2185                 if (lock_type_held == SIX_LOCK_read)
2186                         six_lock_downgrade(&b->c.lock);
2187         } else {
2188                 __bch2_btree_node_write(c, b, flags);
2189                 if (lock_type_held == SIX_LOCK_write &&
2190                     btree_node_just_written(b))
2191                         bch2_btree_post_write_cleanup(c, b);
2192         }
2193 }
2194
2195 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2196 {
2197         struct bucket_table *tbl;
2198         struct rhash_head *pos;
2199         struct btree *b;
2200         unsigned i;
2201         bool ret = false;
2202 restart:
2203         rcu_read_lock();
2204         for_each_cached_btree(b, c, tbl, i, pos)
2205                 if (test_bit(flag, &b->flags)) {
2206                         rcu_read_unlock();
2207                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2208                         ret = true;
2209                         goto restart;
2210                 }
2211         rcu_read_unlock();
2212
2213         return ret;
2214 }
2215
2216 bool bch2_btree_flush_all_reads(struct bch_fs *c)
2217 {
2218         return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2219 }
2220
2221 bool bch2_btree_flush_all_writes(struct bch_fs *c)
2222 {
2223         return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2224 }
2225
2226 const char * const bch2_btree_write_types[] = {
2227 #define x(t, n) [n] = #t,
2228         BCH_BTREE_WRITE_TYPES()
2229         NULL
2230 };
2231
2232 void bch2_btree_write_stats_to_text(struct printbuf *out, struct bch_fs *c)
2233 {
2234         printbuf_tabstop_push(out, 20);
2235         printbuf_tabstop_push(out, 10);
2236
2237         prt_tab(out);
2238         prt_str(out, "nr");
2239         prt_tab(out);
2240         prt_str(out, "size");
2241         prt_newline(out);
2242
2243         for (unsigned i = 0; i < BTREE_WRITE_TYPE_NR; i++) {
2244                 u64 nr          = atomic64_read(&c->btree_write_stats[i].nr);
2245                 u64 bytes       = atomic64_read(&c->btree_write_stats[i].bytes);
2246
2247                 prt_printf(out, "%s:", bch2_btree_write_types[i]);
2248                 prt_tab(out);
2249                 prt_u64(out, nr);
2250                 prt_tab(out);
2251                 prt_human_readable_u64(out, nr ? div64_u64(bytes, nr) : 0);
2252                 prt_newline(out);
2253         }
2254 }