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 #include "trace.h"
  22
  23 #include <linux/sched/mm.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         bch2_assert_btree_nodes_not_locked();
  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         bch2_assert_btree_nodes_not_locked();
  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         bch2_assert_btree_nodes_not_locked();
  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_p_next(start);
  81              k != end;
  82              p = k, k = bkey_p_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_p_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_noprof(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_noprof(size, __GFP_NOWARN|GFP_NOWAIT);
 118         if (!p) {
 119                 *used_mempool = true;
 120                 p = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
 121         }
 122         memalloc_nofs_restore(flags);
 123         return p;
 124 }
 125 #define btree_bounce_alloc(_c, _size, _used_mempool)            \
 126         alloc_hooks(btree_bounce_alloc_noprof(_c, _size, _used_mempool))
 127
 128 static void sort_bkey_ptrs(const struct btree *bt,
 129                            struct bkey_packed **ptrs, unsigned nr)
 130 {
 131         unsigned n = nr, a = nr / 2, b, c, d;
 132
 133         if (!a)
 134                 return;
 135
 136         /* Heap sort: see lib/sort.c: */
 137         while (1) {
 138                 if (a)
 139                         a--;
 140                 else if (--n)
 141                         swap(ptrs[0], ptrs[n]);
 142                 else
 143                         break;
 144
 145                 for (b = a; c = 2 * b + 1, (d = c + 1) < n;)
 146                         b = bch2_bkey_cmp_packed(bt,
 147                                             ptrs[c],
 148                                             ptrs[d]) >= 0 ? c : d;
 149                 if (d == n)
 150                         b = c;
 151
 152                 while (b != a &&
 153                        bch2_bkey_cmp_packed(bt,
 154                                        ptrs[a],
 155                                        ptrs[b]) >= 0)
 156                         b = (b - 1) / 2;
 157                 c = b;
 158                 while (b != a) {
 159                         b = (b - 1) / 2;
 160                         swap(ptrs[b], ptrs[c]);
 161                 }
 162         }
 163 }
 164
 165 static void bch2_sort_whiteouts(struct bch_fs *c, struct btree *b)
 166 {
 167         struct bkey_packed *new_whiteouts, **ptrs, **ptrs_end, *k;
 168         bool used_mempool = false;
 169         size_t bytes = b->whiteout_u64s * sizeof(u64);
 170
 171         if (!b->whiteout_u64s)
 172                 return;
 173
 174         new_whiteouts = btree_bounce_alloc(c, bytes, &used_mempool);
 175
 176         ptrs = ptrs_end = ((void *) new_whiteouts + bytes);
 177
 178         for (k = unwritten_whiteouts_start(c, b);
 179              k != unwritten_whiteouts_end(c, b);
 180              k = bkey_p_next(k))
 181                 *--ptrs = k;
 182
 183         sort_bkey_ptrs(b, ptrs, ptrs_end - ptrs);
 184
 185         k = new_whiteouts;
 186
 187         while (ptrs != ptrs_end) {
 188                 bkey_copy(k, *ptrs);
 189                 k = bkey_p_next(k);
 190                 ptrs++;
 191         }
 192
 193         verify_no_dups(b, new_whiteouts,
 194                        (void *) ((u64 *) new_whiteouts + b->whiteout_u64s));
 195
 196         memcpy_u64s(unwritten_whiteouts_start(c, b),
 197                     new_whiteouts, b->whiteout_u64s);
 198
 199         btree_bounce_free(c, bytes, used_mempool, new_whiteouts);
 200 }
 201
 202 static bool should_compact_bset(struct btree *b, struct bset_tree *t,
 203                                 bool compacting, enum compact_mode mode)
 204 {
 205         if (!bset_dead_u64s(b, t))
 206                 return false;
 207
 208         switch (mode) {
 209         case COMPACT_LAZY:
 210                 return should_compact_bset_lazy(b, t) ||
 211                         (compacting && !bset_written(b, bset(b, t)));
 212         case COMPACT_ALL:
 213                 return true;
 214         default:
 215                 BUG();
 216         }
 217 }
 218
 219 static bool bch2_drop_whiteouts(struct btree *b, enum compact_mode mode)
 220 {
 221         struct bset_tree *t;
 222         bool ret = false;
 223
 224         for_each_bset(b, t) {
 225                 struct bset *i = bset(b, t);
 226                 struct bkey_packed *k, *n, *out, *start, *end;
 227                 struct btree_node_entry *src = NULL, *dst = NULL;
 228
 229                 if (t != b->set && !bset_written(b, i)) {
 230                         src = container_of(i, struct btree_node_entry, keys);
 231                         dst = max(write_block(b),
 232                                   (void *) btree_bkey_last(b, t - 1));
 233                 }
 234
 235                 if (src != dst)
 236                         ret = true;
 237
 238                 if (!should_compact_bset(b, t, ret, mode)) {
 239                         if (src != dst) {
 240                                 memmove(dst, src, sizeof(*src) +
 241                                         le16_to_cpu(src->keys.u64s) *
 242                                         sizeof(u64));
 243                                 i = &dst->keys;
 244                                 set_btree_bset(b, t, i);
 245                         }
 246                         continue;
 247                 }
 248
 249                 start   = btree_bkey_first(b, t);
 250                 end     = btree_bkey_last(b, t);
 251
 252                 if (src != dst) {
 253                         memmove(dst, src, sizeof(*src));
 254                         i = &dst->keys;
 255                         set_btree_bset(b, t, i);
 256                 }
 257
 258                 out = i->start;
 259
 260                 for (k = start; k != end; k = n) {
 261                         n = bkey_p_next(k);
 262
 263                         if (!bkey_deleted(k)) {
 264                                 bkey_copy(out, k);
 265                                 out = bkey_p_next(out);
 266                         } else {
 267                                 BUG_ON(k->needs_whiteout);
 268                         }
 269                 }
 270
 271                 i->u64s = cpu_to_le16((u64 *) out - i->_data);
 272                 set_btree_bset_end(b, t);
 273                 bch2_bset_set_no_aux_tree(b, t);
 274                 ret = true;
 275         }
 276
 277         bch2_verify_btree_nr_keys(b);
 278
 279         bch2_btree_build_aux_trees(b);
 280
 281         return ret;
 282 }
 283
 284 bool bch2_compact_whiteouts(struct bch_fs *c, struct btree *b,
 285                             enum compact_mode mode)
 286 {
 287         return bch2_drop_whiteouts(b, mode);
 288 }
 289
 290 static void btree_node_sort(struct bch_fs *c, struct btree *b,
 291                             unsigned start_idx,
 292                             unsigned end_idx,
 293                             bool filter_whiteouts)
 294 {
 295         struct btree_node *out;
 296         struct sort_iter sort_iter;
 297         struct bset_tree *t;
 298         struct bset *start_bset = bset(b, &b->set[start_idx]);
 299         bool used_mempool = false;
 300         u64 start_time, seq = 0;
 301         unsigned i, u64s = 0, bytes, shift = end_idx - start_idx - 1;
 302         bool sorting_entire_node = start_idx == 0 &&
 303                 end_idx == b->nsets;
 304
 305         sort_iter_init(&sort_iter, b);
 306
 307         for (t = b->set + start_idx;
 308              t < b->set + end_idx;
 309              t++) {
 310                 u64s += le16_to_cpu(bset(b, t)->u64s);
 311                 sort_iter_add(&sort_iter,
 312                               btree_bkey_first(b, t),
 313                               btree_bkey_last(b, t));
 314         }
 315
 316         bytes = sorting_entire_node
 317                 ? btree_bytes(c)
 318                 : __vstruct_bytes(struct btree_node, u64s);
 319
 320         out = btree_bounce_alloc(c, bytes, &used_mempool);
 321
 322         start_time = local_clock();
 323
 324         u64s = bch2_sort_keys(out->keys.start, &sort_iter, filter_whiteouts);
 325
 326         out->keys.u64s = cpu_to_le16(u64s);
 327
 328         BUG_ON(vstruct_end(&out->keys) > (void *) out + bytes);
 329
 330         if (sorting_entire_node)
 331                 bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 332                                        start_time);
 333
 334         /* Make sure we preserve bset journal_seq: */
 335         for (t = b->set + start_idx; t < b->set + end_idx; t++)
 336                 seq = max(seq, le64_to_cpu(bset(b, t)->journal_seq));
 337         start_bset->journal_seq = cpu_to_le64(seq);
 338
 339         if (sorting_entire_node) {
 340                 unsigned u64s = le16_to_cpu(out->keys.u64s);
 341
 342                 BUG_ON(bytes != btree_bytes(c));
 343
 344                 /*
 345                  * Our temporary buffer is the same size as the btree node's
 346                  * buffer, we can just swap buffers instead of doing a big
 347                  * memcpy()
 348                  */
 349                 *out = *b->data;
 350                 out->keys.u64s = cpu_to_le16(u64s);
 351                 swap(out, b->data);
 352                 set_btree_bset(b, b->set, &b->data->keys);
 353         } else {
 354                 start_bset->u64s = out->keys.u64s;
 355                 memcpy_u64s(start_bset->start,
 356                             out->keys.start,
 357                             le16_to_cpu(out->keys.u64s));
 358         }
 359
 360         for (i = start_idx + 1; i < end_idx; i++)
 361                 b->nr.bset_u64s[start_idx] +=
 362                         b->nr.bset_u64s[i];
 363
 364         b->nsets -= shift;
 365
 366         for (i = start_idx + 1; i < b->nsets; i++) {
 367                 b->nr.bset_u64s[i]      = b->nr.bset_u64s[i + shift];
 368                 b->set[i]               = b->set[i + shift];
 369         }
 370
 371         for (i = b->nsets; i < MAX_BSETS; i++)
 372                 b->nr.bset_u64s[i] = 0;
 373
 374         set_btree_bset_end(b, &b->set[start_idx]);
 375         bch2_bset_set_no_aux_tree(b, &b->set[start_idx]);
 376
 377         btree_bounce_free(c, bytes, used_mempool, out);
 378
 379         bch2_verify_btree_nr_keys(b);
 380 }
 381
 382 void bch2_btree_sort_into(struct bch_fs *c,
 383                          struct btree *dst,
 384                          struct btree *src)
 385 {
 386         struct btree_nr_keys nr;
 387         struct btree_node_iter src_iter;
 388         u64 start_time = local_clock();
 389
 390         BUG_ON(dst->nsets != 1);
 391
 392         bch2_bset_set_no_aux_tree(dst, dst->set);
 393
 394         bch2_btree_node_iter_init_from_start(&src_iter, src);
 395
 396         nr = bch2_sort_repack(btree_bset_first(dst),
 397                         src, &src_iter,
 398                         &dst->format,
 399                         true);
 400
 401         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_sort],
 402                                start_time);
 403
 404         set_btree_bset_end(dst, dst->set);
 405
 406         dst->nr.live_u64s       += nr.live_u64s;
 407         dst->nr.bset_u64s[0]    += nr.bset_u64s[0];
 408         dst->nr.packed_keys     += nr.packed_keys;
 409         dst->nr.unpacked_keys   += nr.unpacked_keys;
 410
 411         bch2_verify_btree_nr_keys(dst);
 412 }
 413
 414 #define SORT_CRIT       (4096 / sizeof(u64))
 415
 416 /*
 417  * We're about to add another bset to the btree node, so if there's currently
 418  * too many bsets - sort some of them together:
 419  */
 420 static bool btree_node_compact(struct bch_fs *c, struct btree *b)
 421 {
 422         unsigned unwritten_idx;
 423         bool ret = false;
 424
 425         for (unwritten_idx = 0;
 426              unwritten_idx < b->nsets;
 427              unwritten_idx++)
 428                 if (!bset_written(b, bset(b, &b->set[unwritten_idx])))
 429                         break;
 430
 431         if (b->nsets - unwritten_idx > 1) {
 432                 btree_node_sort(c, b, unwritten_idx,
 433                                 b->nsets, false);
 434                 ret = true;
 435         }
 436
 437         if (unwritten_idx > 1) {
 438                 btree_node_sort(c, b, 0, unwritten_idx, false);
 439                 ret = true;
 440         }
 441
 442         return ret;
 443 }
 444
 445 void bch2_btree_build_aux_trees(struct btree *b)
 446 {
 447         struct bset_tree *t;
 448
 449         for_each_bset(b, t)
 450                 bch2_bset_build_aux_tree(b, t,
 451                                 !bset_written(b, bset(b, t)) &&
 452                                 t == bset_tree_last(b));
 453 }
 454
 455 /*
 456  * If we have MAX_BSETS (3) bsets, should we sort them all down to just one?
 457  *
 458  * The first bset is going to be of similar order to the size of the node, the
 459  * last bset is bounded by btree_write_set_buffer(), which is set to keep the
 460  * memmove on insert from being too expensive: the middle bset should, ideally,
 461  * be the geometric mean of the first and the last.
 462  *
 463  * Returns true if the middle bset is greater than that geometric mean:
 464  */
 465 static inline bool should_compact_all(struct bch_fs *c, struct btree *b)
 466 {
 467         unsigned mid_u64s_bits =
 468                 (ilog2(btree_max_u64s(c)) + BTREE_WRITE_SET_U64s_BITS) / 2;
 469
 470         return bset_u64s(&b->set[1]) > 1U << mid_u64s_bits;
 471 }
 472
 473 /*
 474  * @bch_btree_init_next - initialize a new (unwritten) bset that can then be
 475  * inserted into
 476  *
 477  * Safe to call if there already is an unwritten bset - will only add a new bset
 478  * if @b doesn't already have one.
 479  *
 480  * Returns true if we sorted (i.e. invalidated iterators
 481  */
 482 void bch2_btree_init_next(struct btree_trans *trans, struct btree *b)
 483 {
 484         struct bch_fs *c = trans->c;
 485         struct btree_node_entry *bne;
 486         bool reinit_iter = false;
 487
 488         EBUG_ON(!six_lock_counts(&b->c.lock).n[SIX_LOCK_write]);
 489         BUG_ON(bset_written(b, bset(b, &b->set[1])));
 490         BUG_ON(btree_node_just_written(b));
 491
 492         if (b->nsets == MAX_BSETS &&
 493             !btree_node_write_in_flight(b) &&
 494             should_compact_all(c, b)) {
 495                 bch2_btree_node_write(c, b, SIX_LOCK_write,
 496                                       BTREE_WRITE_init_next_bset);
 497                 reinit_iter = true;
 498         }
 499
 500         if (b->nsets == MAX_BSETS &&
 501             btree_node_compact(c, b))
 502                 reinit_iter = true;
 503
 504         BUG_ON(b->nsets >= MAX_BSETS);
 505
 506         bne = want_new_bset(c, b);
 507         if (bne)
 508                 bch2_bset_init_next(c, b, bne);
 509
 510         bch2_btree_build_aux_trees(b);
 511
 512         if (reinit_iter)
 513                 bch2_trans_node_reinit_iter(trans, b);
 514 }
 515
 516 static void btree_pos_to_text(struct printbuf *out, struct bch_fs *c,
 517                           struct btree *b)
 518 {
 519         prt_printf(out, "%s level %u/%u\n  ",
 520                bch2_btree_ids[b->c.btree_id],
 521                b->c.level,
 522                bch2_btree_id_root(c, b->c.btree_id)->level);
 523         bch2_bkey_val_to_text(out, c, bkey_i_to_s_c(&b->key));
 524 }
 525
 526 static void btree_err_msg(struct printbuf *out, struct bch_fs *c,
 527                           struct bch_dev *ca,
 528                           struct btree *b, struct bset *i,
 529                           unsigned offset, int write)
 530 {
 531         prt_printf(out, bch2_log_msg(c, "%s"),
 532                    write == READ
 533                    ? "error validating btree node "
 534                    : "corrupt btree node before write ");
 535         if (ca)
 536                 prt_printf(out, "on %s ", ca->name);
 537         prt_printf(out, "at btree ");
 538         btree_pos_to_text(out, c, b);
 539
 540         prt_printf(out, "\n  node offset %u", b->written);
 541         if (i)
 542                 prt_printf(out, " bset u64s %u", le16_to_cpu(i->u64s));
 543         prt_str(out, ": ");
 544 }
 545
 546 enum btree_err_type {
 547         /*
 548          * We can repair this locally, and we're after the checksum check so
 549          * there's no need to try another replica:
 550          */
 551         BTREE_ERR_FIXABLE,
 552         /*
 553          * We can repair this if we have to, but we should try reading another
 554          * replica if we can:
 555          */
 556         BTREE_ERR_WANT_RETRY,
 557         /*
 558          * Read another replica if we have one, otherwise consider the whole
 559          * node bad:
 560          */
 561         BTREE_ERR_MUST_RETRY,
 562         BTREE_ERR_BAD_NODE,
 563         BTREE_ERR_INCOMPATIBLE,
 564 };
 565
 566 enum btree_validate_ret {
 567         BTREE_RETRY_READ = 64,
 568 };
 569
 570 static int __btree_err(enum btree_err_type type,
 571                        struct bch_fs *c,
 572                        struct bch_dev *ca,
 573                        struct btree *b,
 574                        struct bset *i,
 575                        int write,
 576                        bool have_retry,
 577                        const char *fmt, ...)
 578 {
 579         struct printbuf out = PRINTBUF;
 580         va_list args;
 581         int ret = -BCH_ERR_fsck_fix;
 582
 583         btree_err_msg(&out, c, ca, b, i, b->written, write);
 584
 585         va_start(args, fmt);
 586         prt_vprintf(&out, fmt, args);
 587         va_end(args);
 588
 589         if (write == WRITE) {
 590                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 591                 ret = c->opts.errors == BCH_ON_ERROR_continue
 592                         ? 0
 593                         : -BCH_ERR_fsck_errors_not_fixed;
 594                 goto out;
 595         }
 596
 597         if (!have_retry && type == BTREE_ERR_WANT_RETRY)
 598                 type = BTREE_ERR_FIXABLE;
 599         if (!have_retry && type == BTREE_ERR_MUST_RETRY)
 600                 type = BTREE_ERR_BAD_NODE;
 601
 602         switch (type) {
 603         case BTREE_ERR_FIXABLE:
 604                 mustfix_fsck_err(c, "%s", out.buf);
 605                 ret = -BCH_ERR_fsck_fix;
 606                 break;
 607         case BTREE_ERR_WANT_RETRY:
 608         case BTREE_ERR_MUST_RETRY:
 609                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 610                 ret = BTREE_RETRY_READ;
 611                 break;
 612         case BTREE_ERR_BAD_NODE:
 613                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 614                 bch2_topology_error(c);
 615                 ret = -BCH_ERR_need_topology_repair;
 616                 break;
 617         case BTREE_ERR_INCOMPATIBLE:
 618                 bch2_print_string_as_lines(KERN_ERR, out.buf);
 619                 ret = -BCH_ERR_fsck_errors_not_fixed;
 620                 break;
 621         default:
 622                 BUG();
 623         }
 624 out:
 625 fsck_err:
 626         printbuf_exit(&out);
 627         return ret;
 628 }
 629
 630 #define btree_err(type, c, ca, b, i, msg, ...)                          \
 631 ({                                                                      \
 632         int _ret = __btree_err(type, c, ca, b, i, write, have_retry, msg, ##__VA_ARGS__);\
 633                                                                         \
 634         if (_ret != -BCH_ERR_fsck_fix)                                  \
 635                 goto fsck_err;                                          \
 636         *saw_error = true;                                              \
 637 })
 638
 639 #define btree_err_on(cond, ...) ((cond) ? btree_err(__VA_ARGS__) : false)
 640
 641 /*
 642  * When btree topology repair changes the start or end of a node, that might
 643  * mean we have to drop keys that are no longer inside the node:
 644  */
 645 __cold
 646 void bch2_btree_node_drop_keys_outside_node(struct btree *b)
 647 {
 648         struct bset_tree *t;
 649         struct bkey_s_c k;
 650         struct bkey unpacked;
 651         struct btree_node_iter iter;
 652
 653         for_each_bset(b, t) {
 654                 struct bset *i = bset(b, t);
 655                 struct bkey_packed *k;
 656
 657                 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
 658                         if (bkey_cmp_left_packed(b, k, &b->data->min_key) >= 0)
 659                                 break;
 660
 661                 if (k != i->start) {
 662                         unsigned shift = (u64 *) k - (u64 *) i->start;
 663
 664                         memmove_u64s_down(i->start, k,
 665                                           (u64 *) vstruct_end(i) - (u64 *) k);
 666                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - shift);
 667                         set_btree_bset_end(b, t);
 668                 }
 669
 670                 for (k = i->start; k != vstruct_last(i); k = bkey_p_next(k))
 671                         if (bkey_cmp_left_packed(b, k, &b->data->max_key) > 0)
 672                                 break;
 673
 674                 if (k != vstruct_last(i)) {
 675                         i->u64s = cpu_to_le16((u64 *) k - (u64 *) i->start);
 676                         set_btree_bset_end(b, t);
 677                 }
 678         }
 679
 680         /*
 681          * Always rebuild search trees: eytzinger search tree nodes directly
 682          * depend on the values of min/max key:
 683          */
 684         bch2_bset_set_no_aux_tree(b, b->set);
 685         bch2_btree_build_aux_trees(b);
 686
 687         for_each_btree_node_key_unpack(b, k, &iter, &unpacked) {
 688                 BUG_ON(bpos_lt(k.k->p, b->data->min_key));
 689                 BUG_ON(bpos_gt(k.k->p, b->data->max_key));
 690         }
 691 }
 692
 693 static int validate_bset(struct bch_fs *c, struct bch_dev *ca,
 694                          struct btree *b, struct bset *i,
 695                          unsigned offset, unsigned sectors,
 696                          int write, bool have_retry, bool *saw_error)
 697 {
 698         unsigned version = le16_to_cpu(i->version);
 699         const char *err;
 700         struct printbuf buf1 = PRINTBUF;
 701         struct printbuf buf2 = PRINTBUF;
 702         int ret = 0;
 703
 704         btree_err_on(!bch2_version_compatible(version),
 705                      BTREE_ERR_INCOMPATIBLE, c, ca, b, i,
 706                      "unsupported bset version %u", 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_p_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_p_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_p_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_p_next(k),
 905                                                   (u64 *) vstruct_end(i) - (u64 *) k);
 906                                 continue;
 907                         }
 908                 }
 909
 910                 prev = k;
 911                 k = bkey_p_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_NOFS);
 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_p_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_p_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 (saw_error && !btree_node_read_error(b)) {
1250                 struct printbuf buf = PRINTBUF;
1251
1252                 bch2_bpos_to_text(&buf, b->key.k.p);
1253                 bch_info(c, "%s: rewriting btree node at btree=%s level=%u %s due to error",
1254                          __func__, bch2_btree_ids[b->c.btree_id], b->c.level, buf.buf);
1255                 printbuf_exit(&buf);
1256
1257                 bch2_btree_node_rewrite_async(c, b);
1258         }
1259
1260         clear_btree_node_read_in_flight(b);
1261         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1262 }
1263
1264 static void btree_node_read_endio(struct bio *bio)
1265 {
1266         struct btree_read_bio *rb =
1267                 container_of(bio, struct btree_read_bio, bio);
1268         struct bch_fs *c        = rb->c;
1269
1270         if (rb->have_ioref) {
1271                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1272
1273                 bch2_latency_acct(ca, rb->start_time, READ);
1274         }
1275
1276         queue_work(c->io_complete_wq, &rb->work);
1277 }
1278
1279 struct btree_node_read_all {
1280         struct closure          cl;
1281         struct bch_fs           *c;
1282         struct btree            *b;
1283         unsigned                nr;
1284         void                    *buf[BCH_REPLICAS_MAX];
1285         struct bio              *bio[BCH_REPLICAS_MAX];
1286         blk_status_t            err[BCH_REPLICAS_MAX];
1287 };
1288
1289 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1290 {
1291         struct btree_node *bn = data;
1292         struct btree_node_entry *bne;
1293         unsigned offset = 0;
1294
1295         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1296                 return 0;
1297
1298         while (offset < btree_sectors(c)) {
1299                 if (!offset) {
1300                         offset += vstruct_sectors(bn, c->block_bits);
1301                 } else {
1302                         bne = data + (offset << 9);
1303                         if (bne->keys.seq != bn->keys.seq)
1304                                 break;
1305                         offset += vstruct_sectors(bne, c->block_bits);
1306                 }
1307         }
1308
1309         return offset;
1310 }
1311
1312 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1313 {
1314         struct btree_node *bn = data;
1315         struct btree_node_entry *bne;
1316
1317         if (!offset)
1318                 return false;
1319
1320         while (offset < btree_sectors(c)) {
1321                 bne = data + (offset << 9);
1322                 if (bne->keys.seq == bn->keys.seq)
1323                         return true;
1324                 offset++;
1325         }
1326
1327         return false;
1328         return offset;
1329 }
1330
1331 static void btree_node_read_all_replicas_done(struct closure *cl)
1332 {
1333         struct btree_node_read_all *ra =
1334                 container_of(cl, struct btree_node_read_all, cl);
1335         struct bch_fs *c = ra->c;
1336         struct btree *b = ra->b;
1337         struct printbuf buf = PRINTBUF;
1338         bool dump_bset_maps = false;
1339         bool have_retry = false;
1340         int ret = 0, best = -1, write = READ;
1341         unsigned i, written = 0, written2 = 0;
1342         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1343                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1344         bool _saw_error = false, *saw_error = &_saw_error;
1345
1346         for (i = 0; i < ra->nr; i++) {
1347                 struct btree_node *bn = ra->buf[i];
1348
1349                 if (ra->err[i])
1350                         continue;
1351
1352                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1353                     (seq && seq != bn->keys.seq))
1354                         continue;
1355
1356                 if (best < 0) {
1357                         best = i;
1358                         written = btree_node_sectors_written(c, bn);
1359                         continue;
1360                 }
1361
1362                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1363                 if (btree_err_on(written2 != written, BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1364                                  "btree node sectors written mismatch: %u != %u",
1365                                  written, written2) ||
1366                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1367                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1368                                  "found bset signature after last bset") ||
1369                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1370                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1371                                  "btree node replicas content mismatch"))
1372                         dump_bset_maps = true;
1373
1374                 if (written2 > written) {
1375                         written = written2;
1376                         best = i;
1377                 }
1378         }
1379 fsck_err:
1380         if (dump_bset_maps) {
1381                 for (i = 0; i < ra->nr; i++) {
1382                         struct btree_node *bn = ra->buf[i];
1383                         struct btree_node_entry *bne = NULL;
1384                         unsigned offset = 0, sectors;
1385                         bool gap = false;
1386
1387                         if (ra->err[i])
1388                                 continue;
1389
1390                         printbuf_reset(&buf);
1391
1392                         while (offset < btree_sectors(c)) {
1393                                 if (!offset) {
1394                                         sectors = vstruct_sectors(bn, c->block_bits);
1395                                 } else {
1396                                         bne = ra->buf[i] + (offset << 9);
1397                                         if (bne->keys.seq != bn->keys.seq)
1398                                                 break;
1399                                         sectors = vstruct_sectors(bne, c->block_bits);
1400                                 }
1401
1402                                 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1403                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1404                                                         le64_to_cpu(bne->keys.journal_seq), false))
1405                                         prt_printf(&buf, "*");
1406                                 offset += sectors;
1407                         }
1408
1409                         while (offset < btree_sectors(c)) {
1410                                 bne = ra->buf[i] + (offset << 9);
1411                                 if (bne->keys.seq == bn->keys.seq) {
1412                                         if (!gap)
1413                                                 prt_printf(&buf, " GAP");
1414                                         gap = true;
1415
1416                                         sectors = vstruct_sectors(bne, c->block_bits);
1417                                         prt_printf(&buf, " %u-%u", offset, offset + sectors);
1418                                         if (bch2_journal_seq_is_blacklisted(c,
1419                                                         le64_to_cpu(bne->keys.journal_seq), false))
1420                                                 prt_printf(&buf, "*");
1421                                 }
1422                                 offset++;
1423                         }
1424
1425                         bch_err(c, "replica %u:%s", i, buf.buf);
1426                 }
1427         }
1428
1429         if (best >= 0) {
1430                 memcpy(b->data, ra->buf[best], btree_bytes(c));
1431                 ret = bch2_btree_node_read_done(c, NULL, b, false, saw_error);
1432         } else {
1433                 ret = -1;
1434         }
1435
1436         if (ret)
1437                 set_btree_node_read_error(b);
1438         else if (*saw_error)
1439                 bch2_btree_node_rewrite_async(c, b);
1440
1441         for (i = 0; i < ra->nr; i++) {
1442                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1443                 bio_put(ra->bio[i]);
1444         }
1445
1446         closure_debug_destroy(&ra->cl);
1447         kfree(ra);
1448         printbuf_exit(&buf);
1449
1450         clear_btree_node_read_in_flight(b);
1451         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1452 }
1453
1454 static void btree_node_read_all_replicas_endio(struct bio *bio)
1455 {
1456         struct btree_read_bio *rb =
1457                 container_of(bio, struct btree_read_bio, bio);
1458         struct bch_fs *c        = rb->c;
1459         struct btree_node_read_all *ra = rb->ra;
1460
1461         if (rb->have_ioref) {
1462                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1463
1464                 bch2_latency_acct(ca, rb->start_time, READ);
1465         }
1466
1467         ra->err[rb->idx] = bio->bi_status;
1468         closure_put(&ra->cl);
1469 }
1470
1471 /*
1472  * XXX This allocates multiple times from the same mempools, and can deadlock
1473  * under sufficient memory pressure (but is only a debug path)
1474  */
1475 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1476 {
1477         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1478         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1479         const union bch_extent_entry *entry;
1480         struct extent_ptr_decoded pick;
1481         struct btree_node_read_all *ra;
1482         unsigned i;
1483
1484         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1485         if (!ra)
1486                 return -BCH_ERR_ENOMEM_btree_node_read_all_replicas;
1487
1488         closure_init(&ra->cl, NULL);
1489         ra->c   = c;
1490         ra->b   = b;
1491         ra->nr  = bch2_bkey_nr_ptrs(k);
1492
1493         for (i = 0; i < ra->nr; i++) {
1494                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1495                 ra->bio[i] = bio_alloc_bioset(NULL,
1496                                               buf_pages(ra->buf[i], btree_bytes(c)),
1497                                               REQ_OP_READ|REQ_SYNC|REQ_META,
1498                                               GFP_NOFS,
1499                                               &c->btree_bio);
1500         }
1501
1502         i = 0;
1503         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1504                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1505                 struct btree_read_bio *rb =
1506                         container_of(ra->bio[i], struct btree_read_bio, bio);
1507                 rb->c                   = c;
1508                 rb->b                   = b;
1509                 rb->ra                  = ra;
1510                 rb->start_time          = local_clock();
1511                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1512                 rb->idx                 = i;
1513                 rb->pick                = pick;
1514                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1515                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1516                 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1517
1518                 if (rb->have_ioref) {
1519                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1520                                      bio_sectors(&rb->bio));
1521                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1522
1523                         closure_get(&ra->cl);
1524                         submit_bio(&rb->bio);
1525                 } else {
1526                         ra->err[i] = BLK_STS_REMOVED;
1527                 }
1528
1529                 i++;
1530         }
1531
1532         if (sync) {
1533                 closure_sync(&ra->cl);
1534                 btree_node_read_all_replicas_done(&ra->cl);
1535         } else {
1536                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1537                             c->io_complete_wq);
1538         }
1539
1540         return 0;
1541 }
1542
1543 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1544                           bool sync)
1545 {
1546         struct extent_ptr_decoded pick;
1547         struct btree_read_bio *rb;
1548         struct bch_dev *ca;
1549         struct bio *bio;
1550         int ret;
1551
1552         trace_and_count(c, btree_node_read, c, b);
1553
1554         if (bch2_verify_all_btree_replicas &&
1555             !btree_node_read_all_replicas(c, b, sync))
1556                 return;
1557
1558         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1559                                          NULL, &pick);
1560
1561         if (ret <= 0) {
1562                 struct printbuf buf = PRINTBUF;
1563
1564                 prt_str(&buf, "btree node read error: no device to read from\n at ");
1565                 btree_pos_to_text(&buf, c, b);
1566                 bch_err(c, "%s", buf.buf);
1567
1568                 if (test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags))
1569                         bch2_fatal_error(c);
1570
1571                 set_btree_node_read_error(b);
1572                 clear_btree_node_read_in_flight(b);
1573                 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1574                 printbuf_exit(&buf);
1575                 return;
1576         }
1577
1578         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1579
1580         bio = bio_alloc_bioset(NULL,
1581                                buf_pages(b->data, btree_bytes(c)),
1582                                REQ_OP_READ|REQ_SYNC|REQ_META,
1583                                GFP_NOFS,
1584                                &c->btree_bio);
1585         rb = container_of(bio, struct btree_read_bio, bio);
1586         rb->c                   = c;
1587         rb->b                   = b;
1588         rb->ra                  = NULL;
1589         rb->start_time          = local_clock();
1590         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1591         rb->pick                = pick;
1592         INIT_WORK(&rb->work, btree_node_read_work);
1593         bio->bi_iter.bi_sector  = pick.ptr.offset;
1594         bio->bi_end_io          = btree_node_read_endio;
1595         bch2_bio_map(bio, b->data, btree_bytes(c));
1596
1597         if (rb->have_ioref) {
1598                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1599                              bio_sectors(bio));
1600                 bio_set_dev(bio, ca->disk_sb.bdev);
1601
1602                 if (sync) {
1603                         submit_bio_wait(bio);
1604
1605                         btree_node_read_work(&rb->work);
1606                 } else {
1607                         submit_bio(bio);
1608                 }
1609         } else {
1610                 bio->bi_status = BLK_STS_REMOVED;
1611
1612                 if (sync)
1613                         btree_node_read_work(&rb->work);
1614                 else
1615                         queue_work(c->io_complete_wq, &rb->work);
1616         }
1617 }
1618
1619 static int __bch2_btree_root_read(struct btree_trans *trans, enum btree_id id,
1620                                   const struct bkey_i *k, unsigned level)
1621 {
1622         struct bch_fs *c = trans->c;
1623         struct closure cl;
1624         struct btree *b;
1625         int ret;
1626
1627         closure_init_stack(&cl);
1628
1629         do {
1630                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1631                 closure_sync(&cl);
1632         } while (ret);
1633
1634         b = bch2_btree_node_mem_alloc(trans, level != 0);
1635         bch2_btree_cache_cannibalize_unlock(c);
1636
1637         BUG_ON(IS_ERR(b));
1638
1639         bkey_copy(&b->key, k);
1640         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1641
1642         set_btree_node_read_in_flight(b);
1643
1644         bch2_btree_node_read(c, b, true);
1645
1646         if (btree_node_read_error(b)) {
1647                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1648
1649                 mutex_lock(&c->btree_cache.lock);
1650                 list_move(&b->list, &c->btree_cache.freeable);
1651                 mutex_unlock(&c->btree_cache.lock);
1652
1653                 ret = -EIO;
1654                 goto err;
1655         }
1656
1657         bch2_btree_set_root_for_read(c, b);
1658 err:
1659         six_unlock_write(&b->c.lock);
1660         six_unlock_intent(&b->c.lock);
1661
1662         return ret;
1663 }
1664
1665 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1666                         const struct bkey_i *k, unsigned level)
1667 {
1668         return bch2_trans_run(c, __bch2_btree_root_read(&trans, id, k, level));
1669
1670 }
1671
1672 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1673                               struct btree_write *w)
1674 {
1675         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1676
1677         do {
1678                 old = new = v;
1679                 if (!(old & 1))
1680                         break;
1681
1682                 new &= ~1UL;
1683         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1684
1685         if (old & 1)
1686                 closure_put(&((struct btree_update *) new)->cl);
1687
1688         bch2_journal_pin_drop(&c->journal, &w->journal);
1689 }
1690
1691 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1692 {
1693         struct btree_write *w = btree_prev_write(b);
1694         unsigned long old, new, v;
1695         unsigned type = 0;
1696
1697         bch2_btree_complete_write(c, b, w);
1698
1699         v = READ_ONCE(b->flags);
1700         do {
1701                 old = new = v;
1702
1703                 if ((old & (1U << BTREE_NODE_dirty)) &&
1704                     (old & (1U << BTREE_NODE_need_write)) &&
1705                     !(old & (1U << BTREE_NODE_never_write)) &&
1706                     !(old & (1U << BTREE_NODE_write_blocked)) &&
1707                     !(old & (1U << BTREE_NODE_will_make_reachable))) {
1708                         new &= ~(1U << BTREE_NODE_dirty);
1709                         new &= ~(1U << BTREE_NODE_need_write);
1710                         new |=  (1U << BTREE_NODE_write_in_flight);
1711                         new |=  (1U << BTREE_NODE_write_in_flight_inner);
1712                         new |=  (1U << BTREE_NODE_just_written);
1713                         new ^=  (1U << BTREE_NODE_write_idx);
1714
1715                         type = new & BTREE_WRITE_TYPE_MASK;
1716                         new &= ~BTREE_WRITE_TYPE_MASK;
1717                 } else {
1718                         new &= ~(1U << BTREE_NODE_write_in_flight);
1719                         new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1720                 }
1721         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1722
1723         if (new & (1U << BTREE_NODE_write_in_flight))
1724                 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED|type);
1725         else
1726                 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1727 }
1728
1729 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1730 {
1731         struct btree_trans trans;
1732
1733         bch2_trans_init(&trans, c, 0, 0);
1734
1735         btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
1736         __btree_node_write_done(c, b);
1737         six_unlock_read(&b->c.lock);
1738
1739         bch2_trans_exit(&trans);
1740 }
1741
1742 static void btree_node_write_work(struct work_struct *work)
1743 {
1744         struct btree_write_bio *wbio =
1745                 container_of(work, struct btree_write_bio, work);
1746         struct bch_fs *c        = wbio->wbio.c;
1747         struct btree *b         = wbio->wbio.bio.bi_private;
1748         struct bch_extent_ptr *ptr;
1749         int ret = 0;
1750
1751         btree_bounce_free(c,
1752                 wbio->data_bytes,
1753                 wbio->wbio.used_mempool,
1754                 wbio->data);
1755
1756         bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1757                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1758
1759         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key)))
1760                 goto err;
1761
1762         if (wbio->wbio.first_btree_write) {
1763                 if (wbio->wbio.failed.nr) {
1764
1765                 }
1766         } else {
1767                 ret = bch2_trans_do(c, NULL, NULL, 0,
1768                         bch2_btree_node_update_key_get_iter(&trans, b, &wbio->key,
1769                                         BCH_WATERMARK_reclaim|
1770                                         BTREE_INSERT_JOURNAL_RECLAIM|
1771                                         BTREE_INSERT_NOFAIL|
1772                                         BTREE_INSERT_NOCHECK_RW,
1773                                         !wbio->wbio.failed.nr));
1774                 if (ret)
1775                         goto err;
1776         }
1777 out:
1778         bio_put(&wbio->wbio.bio);
1779         btree_node_write_done(c, b);
1780         return;
1781 err:
1782         set_btree_node_noevict(b);
1783         if (!bch2_err_matches(ret, EROFS))
1784                 bch2_fs_fatal_error(c, "fatal error writing btree node: %s", bch2_err_str(ret));
1785         goto out;
1786 }
1787
1788 static void btree_node_write_endio(struct bio *bio)
1789 {
1790         struct bch_write_bio *wbio      = to_wbio(bio);
1791         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1792         struct bch_write_bio *orig      = parent ?: wbio;
1793         struct btree_write_bio *wb      = container_of(orig, struct btree_write_bio, wbio);
1794         struct bch_fs *c                = wbio->c;
1795         struct btree *b                 = wbio->bio.bi_private;
1796         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1797         unsigned long flags;
1798
1799         if (wbio->have_ioref)
1800                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1801
1802         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1803                                bch2_blk_status_to_str(bio->bi_status)) ||
1804             bch2_meta_write_fault("btree")) {
1805                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1806                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1807                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1808         }
1809
1810         if (wbio->have_ioref)
1811                 percpu_ref_put(&ca->io_ref);
1812
1813         if (parent) {
1814                 bio_put(bio);
1815                 bio_endio(&parent->bio);
1816                 return;
1817         }
1818
1819         clear_btree_node_write_in_flight_inner(b);
1820         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1821         INIT_WORK(&wb->work, btree_node_write_work);
1822         queue_work(c->btree_io_complete_wq, &wb->work);
1823 }
1824
1825 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1826                                    struct bset *i, unsigned sectors)
1827 {
1828         struct printbuf buf = PRINTBUF;
1829         bool saw_error;
1830         int ret;
1831
1832         ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key),
1833                                 BKEY_TYPE_btree, WRITE, &buf);
1834
1835         if (ret)
1836                 bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf);
1837         printbuf_exit(&buf);
1838         if (ret)
1839                 return ret;
1840
1841         ret = validate_bset_keys(c, b, i, WRITE, false, &saw_error) ?:
1842                 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false, &saw_error);
1843         if (ret) {
1844                 bch2_inconsistent_error(c);
1845                 dump_stack();
1846         }
1847
1848         return ret;
1849 }
1850
1851 static void btree_write_submit(struct work_struct *work)
1852 {
1853         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1854         struct bch_extent_ptr *ptr;
1855         BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1856
1857         bkey_copy(&tmp.k, &wbio->key);
1858
1859         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1860                 ptr->offset += wbio->sector_offset;
1861
1862         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree,
1863                                   &tmp.k, false);
1864 }
1865
1866 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1867 {
1868         struct btree_write_bio *wbio;
1869         struct bset_tree *t;
1870         struct bset *i;
1871         struct btree_node *bn = NULL;
1872         struct btree_node_entry *bne = NULL;
1873         struct sort_iter sort_iter;
1874         struct nonce nonce;
1875         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1876         u64 seq = 0;
1877         bool used_mempool;
1878         unsigned long old, new;
1879         bool validate_before_checksum = false;
1880         enum btree_write_type type = flags & BTREE_WRITE_TYPE_MASK;
1881         void *data;
1882         int ret;
1883
1884         if (flags & BTREE_WRITE_ALREADY_STARTED)
1885                 goto do_write;
1886
1887         /*
1888          * We may only have a read lock on the btree node - the dirty bit is our
1889          * "lock" against racing with other threads that may be trying to start
1890          * a write, we do a write iff we clear the dirty bit. Since setting the
1891          * dirty bit requires a write lock, we can't race with other threads
1892          * redirtying it:
1893          */
1894         do {
1895                 old = new = READ_ONCE(b->flags);
1896
1897                 if (!(old & (1 << BTREE_NODE_dirty)))
1898                         return;
1899
1900                 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
1901                     !(old & (1 << BTREE_NODE_need_write)))
1902                         return;
1903
1904                 if (old &
1905                     ((1 << BTREE_NODE_never_write)|
1906                      (1 << BTREE_NODE_write_blocked)))
1907                         return;
1908
1909                 if (b->written &&
1910                     (old & (1 << BTREE_NODE_will_make_reachable)))
1911                         return;
1912
1913                 if (old & (1 << BTREE_NODE_write_in_flight))
1914                         return;
1915
1916                 if (flags & BTREE_WRITE_ONLY_IF_NEED)
1917                         type = new & BTREE_WRITE_TYPE_MASK;
1918                 new &= ~BTREE_WRITE_TYPE_MASK;
1919
1920                 new &= ~(1 << BTREE_NODE_dirty);
1921                 new &= ~(1 << BTREE_NODE_need_write);
1922                 new |=  (1 << BTREE_NODE_write_in_flight);
1923                 new |=  (1 << BTREE_NODE_write_in_flight_inner);
1924                 new |=  (1 << BTREE_NODE_just_written);
1925                 new ^=  (1 << BTREE_NODE_write_idx);
1926         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1927
1928         if (new & (1U << BTREE_NODE_need_write))
1929                 return;
1930 do_write:
1931         BUG_ON((type == BTREE_WRITE_initial) != (b->written == 0));
1932
1933         atomic_dec(&c->btree_cache.dirty);
1934
1935         BUG_ON(btree_node_fake(b));
1936         BUG_ON((b->will_make_reachable != 0) != !b->written);
1937
1938         BUG_ON(b->written >= btree_sectors(c));
1939         BUG_ON(b->written & (block_sectors(c) - 1));
1940         BUG_ON(bset_written(b, btree_bset_last(b)));
1941         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1942         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1943
1944         bch2_sort_whiteouts(c, b);
1945
1946         sort_iter_init(&sort_iter, b);
1947
1948         bytes = !b->written
1949                 ? sizeof(struct btree_node)
1950                 : sizeof(struct btree_node_entry);
1951
1952         bytes += b->whiteout_u64s * sizeof(u64);
1953
1954         for_each_bset(b, t) {
1955                 i = bset(b, t);
1956
1957                 if (bset_written(b, i))
1958                         continue;
1959
1960                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1961                 sort_iter_add(&sort_iter,
1962                               btree_bkey_first(b, t),
1963                               btree_bkey_last(b, t));
1964                 seq = max(seq, le64_to_cpu(i->journal_seq));
1965         }
1966
1967         BUG_ON(b->written && !seq);
1968
1969         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1970         bytes += 8;
1971
1972         /* buffer must be a multiple of the block size */
1973         bytes = round_up(bytes, block_bytes(c));
1974
1975         data = btree_bounce_alloc(c, bytes, &used_mempool);
1976
1977         if (!b->written) {
1978                 bn = data;
1979                 *bn = *b->data;
1980                 i = &bn->keys;
1981         } else {
1982                 bne = data;
1983                 bne->keys = b->data->keys;
1984                 i = &bne->keys;
1985         }
1986
1987         i->journal_seq  = cpu_to_le64(seq);
1988         i->u64s         = 0;
1989
1990         sort_iter_add(&sort_iter,
1991                       unwritten_whiteouts_start(c, b),
1992                       unwritten_whiteouts_end(c, b));
1993         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1994
1995         b->whiteout_u64s = 0;
1996
1997         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1998         le16_add_cpu(&i->u64s, u64s);
1999
2000         BUG_ON(!b->written && i->u64s != b->data->keys.u64s);
2001
2002         set_needs_whiteout(i, false);
2003
2004         /* do we have data to write? */
2005         if (b->written && !i->u64s)
2006                 goto nowrite;
2007
2008         bytes_to_write = vstruct_end(i) - data;
2009         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
2010
2011         if (!b->written &&
2012             b->key.k.type == KEY_TYPE_btree_ptr_v2)
2013                 BUG_ON(btree_ptr_sectors_written(&b->key) != sectors_to_write);
2014
2015         memset(data + bytes_to_write, 0,
2016                (sectors_to_write << 9) - bytes_to_write);
2017
2018         BUG_ON(b->written + sectors_to_write > btree_sectors(c));
2019         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
2020         BUG_ON(i->seq != b->data->keys.seq);
2021
2022         i->version = cpu_to_le16(c->sb.version);
2023         SET_BSET_OFFSET(i, b->written);
2024         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
2025
2026         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
2027                 validate_before_checksum = true;
2028
2029         /* validate_bset will be modifying: */
2030         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
2031                 validate_before_checksum = true;
2032
2033         /* if we're going to be encrypting, check metadata validity first: */
2034         if (validate_before_checksum &&
2035             validate_bset_for_write(c, b, i, sectors_to_write))
2036                 goto err;
2037
2038         ret = bset_encrypt(c, i, b->written << 9);
2039         if (bch2_fs_fatal_err_on(ret, c,
2040                         "error encrypting btree node: %i\n", ret))
2041                 goto err;
2042
2043         nonce = btree_nonce(i, b->written << 9);
2044
2045         if (bn)
2046                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
2047         else
2048                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
2049
2050         /* if we're not encrypting, check metadata after checksumming: */
2051         if (!validate_before_checksum &&
2052             validate_bset_for_write(c, b, i, sectors_to_write))
2053                 goto err;
2054
2055         /*
2056          * We handle btree write errors by immediately halting the journal -
2057          * after we've done that, we can't issue any subsequent btree writes
2058          * because they might have pointers to new nodes that failed to write.
2059          *
2060          * Furthermore, there's no point in doing any more btree writes because
2061          * with the journal stopped, we're never going to update the journal to
2062          * reflect that those writes were done and the data flushed from the
2063          * journal:
2064          *
2065          * Also on journal error, the pending write may have updates that were
2066          * never journalled (interior nodes, see btree_update_nodes_written()) -
2067          * it's critical that we don't do the write in that case otherwise we
2068          * will have updates visible that weren't in the journal:
2069          *
2070          * Make sure to update b->written so bch2_btree_init_next() doesn't
2071          * break:
2072          */
2073         if (bch2_journal_error(&c->journal) ||
2074             c->opts.nochanges)
2075                 goto err;
2076
2077         trace_and_count(c, btree_node_write, b, bytes_to_write, sectors_to_write);
2078
2079         wbio = container_of(bio_alloc_bioset(NULL,
2080                                 buf_pages(data, sectors_to_write << 9),
2081                                 REQ_OP_WRITE|REQ_META,
2082                                 GFP_NOFS,
2083                                 &c->btree_bio),
2084                             struct btree_write_bio, wbio.bio);
2085         wbio_init(&wbio->wbio.bio);
2086         wbio->data                      = data;
2087         wbio->data_bytes                = bytes;
2088         wbio->sector_offset             = b->written;
2089         wbio->wbio.c                    = c;
2090         wbio->wbio.used_mempool         = used_mempool;
2091         wbio->wbio.first_btree_write    = !b->written;
2092         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
2093         wbio->wbio.bio.bi_private       = b;
2094
2095         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
2096
2097         bkey_copy(&wbio->key, &b->key);
2098
2099         b->written += sectors_to_write;
2100
2101         if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
2102                 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2103                         cpu_to_le16(b->written);
2104
2105         atomic64_inc(&c->btree_write_stats[type].nr);
2106         atomic64_add(bytes_to_write, &c->btree_write_stats[type].bytes);
2107
2108         INIT_WORK(&wbio->work, btree_write_submit);
2109         queue_work(c->io_complete_wq, &wbio->work);
2110         return;
2111 err:
2112         set_btree_node_noevict(b);
2113         b->written += sectors_to_write;
2114 nowrite:
2115         btree_bounce_free(c, bytes, used_mempool, data);
2116         __btree_node_write_done(c, b);
2117 }
2118
2119 /*
2120  * Work that must be done with write lock held:
2121  */
2122 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2123 {
2124         bool invalidated_iter = false;
2125         struct btree_node_entry *bne;
2126         struct bset_tree *t;
2127
2128         if (!btree_node_just_written(b))
2129                 return false;
2130
2131         BUG_ON(b->whiteout_u64s);
2132
2133         clear_btree_node_just_written(b);
2134
2135         /*
2136          * Note: immediately after write, bset_written() doesn't work - the
2137          * amount of data we had to write after compaction might have been
2138          * smaller than the offset of the last bset.
2139          *
2140          * However, we know that all bsets have been written here, as long as
2141          * we're still holding the write lock:
2142          */
2143
2144         /*
2145          * XXX: decide if we really want to unconditionally sort down to a
2146          * single bset:
2147          */
2148         if (b->nsets > 1) {
2149                 btree_node_sort(c, b, 0, b->nsets, true);
2150                 invalidated_iter = true;
2151         } else {
2152                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2153         }
2154
2155         for_each_bset(b, t)
2156                 set_needs_whiteout(bset(b, t), true);
2157
2158         bch2_btree_verify(c, b);
2159
2160         /*
2161          * If later we don't unconditionally sort down to a single bset, we have
2162          * to ensure this is still true:
2163          */
2164         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2165
2166         bne = want_new_bset(c, b);
2167         if (bne)
2168                 bch2_bset_init_next(c, b, bne);
2169
2170         bch2_btree_build_aux_trees(b);
2171
2172         return invalidated_iter;
2173 }
2174
2175 /*
2176  * Use this one if the node is intent locked:
2177  */
2178 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2179                            enum six_lock_type lock_type_held,
2180                            unsigned flags)
2181 {
2182         if (lock_type_held == SIX_LOCK_intent ||
2183             (lock_type_held == SIX_LOCK_read &&
2184              six_lock_tryupgrade(&b->c.lock))) {
2185                 __bch2_btree_node_write(c, b, flags);
2186
2187                 /* don't cycle lock unnecessarily: */
2188                 if (btree_node_just_written(b) &&
2189                     six_trylock_write(&b->c.lock)) {
2190                         bch2_btree_post_write_cleanup(c, b);
2191                         six_unlock_write(&b->c.lock);
2192                 }
2193
2194                 if (lock_type_held == SIX_LOCK_read)
2195                         six_lock_downgrade(&b->c.lock);
2196         } else {
2197                 __bch2_btree_node_write(c, b, flags);
2198                 if (lock_type_held == SIX_LOCK_write &&
2199                     btree_node_just_written(b))
2200                         bch2_btree_post_write_cleanup(c, b);
2201         }
2202 }
2203
2204 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2205 {
2206         struct bucket_table *tbl;
2207         struct rhash_head *pos;
2208         struct btree *b;
2209         unsigned i;
2210         bool ret = false;
2211 restart:
2212         rcu_read_lock();
2213         for_each_cached_btree(b, c, tbl, i, pos)
2214                 if (test_bit(flag, &b->flags)) {
2215                         rcu_read_unlock();
2216                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2217                         ret = true;
2218                         goto restart;
2219                 }
2220         rcu_read_unlock();
2221
2222         return ret;
2223 }
2224
2225 bool bch2_btree_flush_all_reads(struct bch_fs *c)
2226 {
2227         return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2228 }
2229
2230 bool bch2_btree_flush_all_writes(struct bch_fs *c)
2231 {
2232         return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2233 }
2234
2235 static const char * const bch2_btree_write_types[] = {
2236 #define x(t, n) [n] = #t,
2237         BCH_BTREE_WRITE_TYPES()
2238         NULL
2239 };
2240
2241 void bch2_btree_write_stats_to_text(struct printbuf *out, struct bch_fs *c)
2242 {
2243         printbuf_tabstop_push(out, 20);
2244         printbuf_tabstop_push(out, 10);
2245
2246         prt_tab(out);
2247         prt_str(out, "nr");
2248         prt_tab(out);
2249         prt_str(out, "size");
2250         prt_newline(out);
2251
2252         for (unsigned i = 0; i < BTREE_WRITE_TYPE_NR; i++) {
2253                 u64 nr          = atomic64_read(&c->btree_write_stats[i].nr);
2254                 u64 bytes       = atomic64_read(&c->btree_write_stats[i].bytes);
2255
2256                 prt_printf(out, "%s:", bch2_btree_write_types[i]);
2257                 prt_tab(out);
2258                 prt_u64(out, nr);
2259                 prt_tab(out);
2260                 prt_human_readable_u64(out, nr ? div64_u64(bytes, nr) : 0);
2261                 prt_newline(out);
2262         }
2263 }