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(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_NOFS);
 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_p_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_p_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_p_next(k);
 260
 261                         if (!bkey_deleted(k)) {
 262                                 bkey_copy(out, k);
 263                                 out = bkey_p_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(!six_lock_counts(&b->c.lock).n[SIX_LOCK_write]);
 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                bch2_btree_id_root(c, 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_p_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_p_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(!bch2_version_compatible(version),
 703                      BTREE_ERR_INCOMPATIBLE, c, ca, b, i,
 704                      "unsupported bset version %u", version);
 705
 706         if (btree_err_on(version < c->sb.version_min,
 707                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 708                          "bset version %u older than superblock version_min %u",
 709                          version, c->sb.version_min)) {
 710                 mutex_lock(&c->sb_lock);
 711                 c->disk_sb.sb->version_min = cpu_to_le16(version);
 712                 bch2_write_super(c);
 713                 mutex_unlock(&c->sb_lock);
 714         }
 715
 716         if (btree_err_on(version > c->sb.version,
 717                          BTREE_ERR_FIXABLE, c, NULL, b, i,
 718                          "bset version %u newer than superblock version %u",
 719                          version, c->sb.version)) {
 720                 mutex_lock(&c->sb_lock);
 721                 c->disk_sb.sb->version = cpu_to_le16(version);
 722                 bch2_write_super(c);
 723                 mutex_unlock(&c->sb_lock);
 724         }
 725
 726         btree_err_on(BSET_SEPARATE_WHITEOUTS(i),
 727                      BTREE_ERR_INCOMPATIBLE, c, ca, b, i,
 728                      "BSET_SEPARATE_WHITEOUTS no longer supported");
 729
 730         if (btree_err_on(offset + sectors > btree_sectors(c),
 731                          BTREE_ERR_FIXABLE, c, ca, b, i,
 732                          "bset past end of btree node")) {
 733                 i->u64s = 0;
 734                 ret = 0;
 735                 goto out;
 736         }
 737
 738         btree_err_on(offset && !i->u64s,
 739                      BTREE_ERR_FIXABLE, c, ca, b, i,
 740                      "empty bset");
 741
 742         btree_err_on(BSET_OFFSET(i) &&
 743                      BSET_OFFSET(i) != offset,
 744                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 745                      "bset at wrong sector offset");
 746
 747         if (!offset) {
 748                 struct btree_node *bn =
 749                         container_of(i, struct btree_node, keys);
 750                 /* These indicate that we read the wrong btree node: */
 751
 752                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 753                         struct bch_btree_ptr_v2 *bp =
 754                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 755
 756                         /* XXX endianness */
 757                         btree_err_on(bp->seq != bn->keys.seq,
 758                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 759                                      "incorrect sequence number (wrong btree node)");
 760                 }
 761
 762                 btree_err_on(BTREE_NODE_ID(bn) != b->c.btree_id,
 763                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 764                              "incorrect btree id");
 765
 766                 btree_err_on(BTREE_NODE_LEVEL(bn) != b->c.level,
 767                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 768                              "incorrect level");
 769
 770                 if (!write)
 771                         compat_btree_node(b->c.level, b->c.btree_id, version,
 772                                           BSET_BIG_ENDIAN(i), write, bn);
 773
 774                 if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 775                         struct bch_btree_ptr_v2 *bp =
 776                                 &bkey_i_to_btree_ptr_v2(&b->key)->v;
 777
 778                         if (BTREE_PTR_RANGE_UPDATED(bp)) {
 779                                 b->data->min_key = bp->min_key;
 780                                 b->data->max_key = b->key.k.p;
 781                         }
 782
 783                         btree_err_on(!bpos_eq(b->data->min_key, bp->min_key),
 784                                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 785                                      "incorrect min_key: got %s should be %s",
 786                                      (printbuf_reset(&buf1),
 787                                       bch2_bpos_to_text(&buf1, bn->min_key), buf1.buf),
 788                                      (printbuf_reset(&buf2),
 789                                       bch2_bpos_to_text(&buf2, bp->min_key), buf2.buf));
 790                 }
 791
 792                 btree_err_on(!bpos_eq(bn->max_key, b->key.k.p),
 793                              BTREE_ERR_MUST_RETRY, c, ca, b, i,
 794                              "incorrect max key %s",
 795                              (printbuf_reset(&buf1),
 796                               bch2_bpos_to_text(&buf1, bn->max_key), buf1.buf));
 797
 798                 if (write)
 799                         compat_btree_node(b->c.level, b->c.btree_id, version,
 800                                           BSET_BIG_ENDIAN(i), write, bn);
 801
 802                 err = bch2_bkey_format_validate(&bn->format);
 803                 btree_err_on(err,
 804                              BTREE_ERR_BAD_NODE, c, ca, b, i,
 805                              "invalid bkey format: %s", err);
 806
 807                 compat_bformat(b->c.level, b->c.btree_id, version,
 808                                BSET_BIG_ENDIAN(i), write,
 809                                &bn->format);
 810         }
 811 out:
 812 fsck_err:
 813         printbuf_exit(&buf2);
 814         printbuf_exit(&buf1);
 815         return ret;
 816 }
 817
 818 static int bset_key_invalid(struct bch_fs *c, struct btree *b,
 819                             struct bkey_s_c k,
 820                             bool updated_range, int rw,
 821                             struct printbuf *err)
 822 {
 823         return __bch2_bkey_invalid(c, k, btree_node_type(b), READ, err) ?:
 824                 (!updated_range ? bch2_bkey_in_btree_node(b, k, err) : 0) ?:
 825                 (rw == WRITE ? bch2_bkey_val_invalid(c, k, READ, err) : 0);
 826 }
 827
 828 static int validate_bset_keys(struct bch_fs *c, struct btree *b,
 829                          struct bset *i, int write,
 830                          bool have_retry, bool *saw_error)
 831 {
 832         unsigned version = le16_to_cpu(i->version);
 833         struct bkey_packed *k, *prev = NULL;
 834         struct printbuf buf = PRINTBUF;
 835         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 836                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 837         int ret = 0;
 838
 839         for (k = i->start;
 840              k != vstruct_last(i);) {
 841                 struct bkey_s u;
 842                 struct bkey tmp;
 843
 844                 if (btree_err_on(bkey_p_next(k) > vstruct_last(i),
 845                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 846                                  "key extends past end of bset")) {
 847                         i->u64s = cpu_to_le16((u64 *) k - i->_data);
 848                         break;
 849                 }
 850
 851                 if (btree_err_on(k->format > KEY_FORMAT_CURRENT,
 852                                  BTREE_ERR_FIXABLE, c, NULL, b, i,
 853                                  "invalid bkey format %u", k->format)) {
 854                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 855                         memmove_u64s_down(k, bkey_p_next(k),
 856                                           (u64 *) vstruct_end(i) - (u64 *) k);
 857                         continue;
 858                 }
 859
 860                 /* XXX: validate k->u64s */
 861                 if (!write)
 862                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 863                                     BSET_BIG_ENDIAN(i), write,
 864                                     &b->format, k);
 865
 866                 u = __bkey_disassemble(b, k, &tmp);
 867
 868                 printbuf_reset(&buf);
 869                 if (bset_key_invalid(c, b, u.s_c, updated_range, write, &buf)) {
 870                         printbuf_reset(&buf);
 871                         prt_printf(&buf, "invalid bkey:  ");
 872                         bset_key_invalid(c, b, u.s_c, updated_range, write, &buf);
 873                         prt_printf(&buf, "\n  ");
 874                         bch2_bkey_val_to_text(&buf, c, u.s_c);
 875
 876                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf);
 877
 878                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 879                         memmove_u64s_down(k, bkey_p_next(k),
 880                                           (u64 *) vstruct_end(i) - (u64 *) k);
 881                         continue;
 882                 }
 883
 884                 if (write)
 885                         bch2_bkey_compat(b->c.level, b->c.btree_id, version,
 886                                     BSET_BIG_ENDIAN(i), write,
 887                                     &b->format, k);
 888
 889                 if (prev && bkey_iter_cmp(b, prev, k) > 0) {
 890                         struct bkey up = bkey_unpack_key(b, prev);
 891
 892                         printbuf_reset(&buf);
 893                         prt_printf(&buf, "keys out of order: ");
 894                         bch2_bkey_to_text(&buf, &up);
 895                         prt_printf(&buf, " > ");
 896                         bch2_bkey_to_text(&buf, u.k);
 897
 898                         bch2_dump_bset(c, b, i, 0);
 899
 900                         if (btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf)) {
 901                                 i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
 902                                 memmove_u64s_down(k, bkey_p_next(k),
 903                                                   (u64 *) vstruct_end(i) - (u64 *) k);
 904                                 continue;
 905                         }
 906                 }
 907
 908                 prev = k;
 909                 k = bkey_p_next(k);
 910         }
 911 fsck_err:
 912         printbuf_exit(&buf);
 913         return ret;
 914 }
 915
 916 int bch2_btree_node_read_done(struct bch_fs *c, struct bch_dev *ca,
 917                               struct btree *b, bool have_retry, bool *saw_error)
 918 {
 919         struct btree_node_entry *bne;
 920         struct sort_iter *iter;
 921         struct btree_node *sorted;
 922         struct bkey_packed *k;
 923         struct bch_extent_ptr *ptr;
 924         struct bset *i;
 925         bool used_mempool, blacklisted;
 926         bool updated_range = b->key.k.type == KEY_TYPE_btree_ptr_v2 &&
 927                 BTREE_PTR_RANGE_UPDATED(&bkey_i_to_btree_ptr_v2(&b->key)->v);
 928         unsigned u64s;
 929         unsigned blacklisted_written, nonblacklisted_written = 0;
 930         unsigned ptr_written = btree_ptr_sectors_written(&b->key);
 931         struct printbuf buf = PRINTBUF;
 932         int ret = 0, retry_read = 0, write = READ;
 933
 934         b->version_ondisk = U16_MAX;
 935         /* We might get called multiple times on read retry: */
 936         b->written = 0;
 937
 938         iter = mempool_alloc(&c->fill_iter, GFP_NOFS);
 939         sort_iter_init(iter, b);
 940         iter->size = (btree_blocks(c) + 1) * 2;
 941
 942         if (bch2_meta_read_fault("btree"))
 943                 btree_err(BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 944                           "dynamic fault");
 945
 946         btree_err_on(le64_to_cpu(b->data->magic) != bset_magic(c),
 947                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 948                      "bad magic: want %llx, got %llx",
 949                      bset_magic(c), le64_to_cpu(b->data->magic));
 950
 951         btree_err_on(!b->data->keys.seq,
 952                      BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 953                      "bad btree header: seq 0");
 954
 955         if (b->key.k.type == KEY_TYPE_btree_ptr_v2) {
 956                 struct bch_btree_ptr_v2 *bp =
 957                         &bkey_i_to_btree_ptr_v2(&b->key)->v;
 958
 959                 btree_err_on(b->data->keys.seq != bp->seq,
 960                              BTREE_ERR_MUST_RETRY, c, ca, b, NULL,
 961                              "got wrong btree node (seq %llx want %llx)",
 962                              b->data->keys.seq, bp->seq);
 963         }
 964
 965         while (b->written < (ptr_written ?: btree_sectors(c))) {
 966                 unsigned sectors;
 967                 struct nonce nonce;
 968                 struct bch_csum csum;
 969                 bool first = !b->written;
 970
 971                 if (!b->written) {
 972                         i = &b->data->keys;
 973
 974                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
 975                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 976                                      "unknown checksum type %llu",
 977                                      BSET_CSUM_TYPE(i));
 978
 979                         nonce = btree_nonce(i, b->written << 9);
 980                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, b->data);
 981
 982                         btree_err_on(bch2_crc_cmp(csum, b->data->csum),
 983                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
 984                                      "invalid checksum");
 985
 986                         ret = bset_encrypt(c, i, b->written << 9);
 987                         if (bch2_fs_fatal_err_on(ret, c,
 988                                         "error decrypting btree node: %i", ret))
 989                                 goto fsck_err;
 990
 991                         btree_err_on(btree_node_type_is_extents(btree_node_type(b)) &&
 992                                      !BTREE_NODE_NEW_EXTENT_OVERWRITE(b->data),
 993                                      BTREE_ERR_INCOMPATIBLE, c, NULL, b, NULL,
 994                                      "btree node does not have NEW_EXTENT_OVERWRITE set");
 995
 996                         sectors = vstruct_sectors(b->data, c->block_bits);
 997                 } else {
 998                         bne = write_block(b);
 999                         i = &bne->keys;
1000
1001                         if (i->seq != b->data->keys.seq)
1002                                 break;
1003
1004                         btree_err_on(!bch2_checksum_type_valid(c, BSET_CSUM_TYPE(i)),
1005                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
1006                                      "unknown checksum type %llu",
1007                                      BSET_CSUM_TYPE(i));
1008
1009                         nonce = btree_nonce(i, b->written << 9);
1010                         csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
1011
1012                         btree_err_on(bch2_crc_cmp(csum, bne->csum),
1013                                      BTREE_ERR_WANT_RETRY, c, ca, b, i,
1014                                      "invalid checksum");
1015
1016                         ret = bset_encrypt(c, i, b->written << 9);
1017                         if (bch2_fs_fatal_err_on(ret, c,
1018                                         "error decrypting btree node: %i\n", ret))
1019                                 goto fsck_err;
1020
1021                         sectors = vstruct_sectors(bne, c->block_bits);
1022                 }
1023
1024                 b->version_ondisk = min(b->version_ondisk,
1025                                         le16_to_cpu(i->version));
1026
1027                 ret = validate_bset(c, ca, b, i, b->written, sectors,
1028                                     READ, have_retry, saw_error);
1029                 if (ret)
1030                         goto fsck_err;
1031
1032                 if (!b->written)
1033                         btree_node_set_format(b, b->data->format);
1034
1035                 ret = validate_bset_keys(c, b, i, READ, have_retry, saw_error);
1036                 if (ret)
1037                         goto fsck_err;
1038
1039                 SET_BSET_BIG_ENDIAN(i, CPU_BIG_ENDIAN);
1040
1041                 blacklisted = bch2_journal_seq_is_blacklisted(c,
1042                                         le64_to_cpu(i->journal_seq),
1043                                         true);
1044
1045                 btree_err_on(blacklisted && first,
1046                              BTREE_ERR_FIXABLE, c, ca, b, i,
1047                              "first btree node bset has blacklisted journal seq (%llu)",
1048                              le64_to_cpu(i->journal_seq));
1049
1050                 btree_err_on(blacklisted && ptr_written,
1051                              BTREE_ERR_FIXABLE, c, ca, b, i,
1052                              "found blacklisted bset (journal seq %llu) in btree node at offset %u-%u/%u",
1053                              le64_to_cpu(i->journal_seq),
1054                              b->written, b->written + sectors, ptr_written);
1055
1056                 b->written += sectors;
1057
1058                 if (blacklisted && !first)
1059                         continue;
1060
1061                 sort_iter_add(iter,
1062                               vstruct_idx(i, 0),
1063                               vstruct_last(i));
1064
1065                 nonblacklisted_written = b->written;
1066         }
1067
1068         if (ptr_written) {
1069                 btree_err_on(b->written < ptr_written,
1070                              BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1071                              "btree node data missing: expected %u sectors, found %u",
1072                              ptr_written, b->written);
1073         } else {
1074                 for (bne = write_block(b);
1075                      bset_byte_offset(b, bne) < btree_bytes(c);
1076                      bne = (void *) bne + block_bytes(c))
1077                         btree_err_on(bne->keys.seq == b->data->keys.seq &&
1078                                      !bch2_journal_seq_is_blacklisted(c,
1079                                                                       le64_to_cpu(bne->keys.journal_seq),
1080                                                                       true),
1081                                      BTREE_ERR_WANT_RETRY, c, ca, b, NULL,
1082                                      "found bset signature after last bset");
1083
1084                 /*
1085                  * Blacklisted bsets are those that were written after the most recent
1086                  * (flush) journal write. Since there wasn't a flush, they may not have
1087                  * made it to all devices - which means we shouldn't write new bsets
1088                  * after them, as that could leave a gap and then reads from that device
1089                  * wouldn't find all the bsets in that btree node - which means it's
1090                  * important that we start writing new bsets after the most recent _non_
1091                  * blacklisted bset:
1092                  */
1093                 blacklisted_written = b->written;
1094                 b->written = nonblacklisted_written;
1095         }
1096
1097         sorted = btree_bounce_alloc(c, btree_bytes(c), &used_mempool);
1098         sorted->keys.u64s = 0;
1099
1100         set_btree_bset(b, b->set, &b->data->keys);
1101
1102         b->nr = bch2_key_sort_fix_overlapping(c, &sorted->keys, iter);
1103
1104         u64s = le16_to_cpu(sorted->keys.u64s);
1105         *sorted = *b->data;
1106         sorted->keys.u64s = cpu_to_le16(u64s);
1107         swap(sorted, b->data);
1108         set_btree_bset(b, b->set, &b->data->keys);
1109         b->nsets = 1;
1110
1111         BUG_ON(b->nr.live_u64s != u64s);
1112
1113         btree_bounce_free(c, btree_bytes(c), used_mempool, sorted);
1114
1115         if (updated_range)
1116                 bch2_btree_node_drop_keys_outside_node(b);
1117
1118         i = &b->data->keys;
1119         for (k = i->start; k != vstruct_last(i);) {
1120                 struct bkey tmp;
1121                 struct bkey_s u = __bkey_disassemble(b, k, &tmp);
1122
1123                 printbuf_reset(&buf);
1124
1125                 if (bch2_bkey_val_invalid(c, u.s_c, READ, &buf) ||
1126                     (bch2_inject_invalid_keys &&
1127                      !bversion_cmp(u.k->version, MAX_VERSION))) {
1128                         printbuf_reset(&buf);
1129
1130                         prt_printf(&buf, "invalid bkey: ");
1131                         bch2_bkey_val_invalid(c, u.s_c, READ, &buf);
1132                         prt_printf(&buf, "\n  ");
1133                         bch2_bkey_val_to_text(&buf, c, u.s_c);
1134
1135                         btree_err(BTREE_ERR_FIXABLE, c, NULL, b, i, "%s", buf.buf);
1136
1137                         btree_keys_account_key_drop(&b->nr, 0, k);
1138
1139                         i->u64s = cpu_to_le16(le16_to_cpu(i->u64s) - k->u64s);
1140                         memmove_u64s_down(k, bkey_p_next(k),
1141                                           (u64 *) vstruct_end(i) - (u64 *) k);
1142                         set_btree_bset_end(b, b->set);
1143                         continue;
1144                 }
1145
1146                 if (u.k->type == KEY_TYPE_btree_ptr_v2) {
1147                         struct bkey_s_btree_ptr_v2 bp = bkey_s_to_btree_ptr_v2(u);
1148
1149                         bp.v->mem_ptr = 0;
1150                 }
1151
1152                 k = bkey_p_next(k);
1153         }
1154
1155         bch2_bset_build_aux_tree(b, b->set, false);
1156
1157         set_needs_whiteout(btree_bset_first(b), true);
1158
1159         btree_node_reset_sib_u64s(b);
1160
1161         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&b->key)), ptr) {
1162                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1163
1164                 if (ca->mi.state != BCH_MEMBER_STATE_rw)
1165                         set_btree_node_need_rewrite(b);
1166         }
1167
1168         if (!ptr_written)
1169                 set_btree_node_need_rewrite(b);
1170 out:
1171         mempool_free(iter, &c->fill_iter);
1172         printbuf_exit(&buf);
1173         return retry_read;
1174 fsck_err:
1175         if (ret == BTREE_RETRY_READ)
1176                 retry_read = 1;
1177         else
1178                 set_btree_node_read_error(b);
1179         goto out;
1180 }
1181
1182 static void btree_node_read_work(struct work_struct *work)
1183 {
1184         struct btree_read_bio *rb =
1185                 container_of(work, struct btree_read_bio, work);
1186         struct bch_fs *c        = rb->c;
1187         struct btree *b         = rb->b;
1188         struct bch_dev *ca      = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1189         struct bio *bio         = &rb->bio;
1190         struct bch_io_failures failed = { .nr = 0 };
1191         struct printbuf buf = PRINTBUF;
1192         bool saw_error = false;
1193         bool retry = false;
1194         bool can_retry;
1195
1196         goto start;
1197         while (1) {
1198                 retry = true;
1199                 bch_info(c, "retrying read");
1200                 ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1201                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1202                 bio_reset(bio, NULL, REQ_OP_READ|REQ_SYNC|REQ_META);
1203                 bio->bi_iter.bi_sector  = rb->pick.ptr.offset;
1204                 bio->bi_iter.bi_size    = btree_bytes(c);
1205
1206                 if (rb->have_ioref) {
1207                         bio_set_dev(bio, ca->disk_sb.bdev);
1208                         submit_bio_wait(bio);
1209                 } else {
1210                         bio->bi_status = BLK_STS_REMOVED;
1211                 }
1212 start:
1213                 printbuf_reset(&buf);
1214                 btree_pos_to_text(&buf, c, b);
1215                 bch2_dev_io_err_on(bio->bi_status, ca, "btree read error %s for %s",
1216                                    bch2_blk_status_to_str(bio->bi_status), buf.buf);
1217                 if (rb->have_ioref)
1218                         percpu_ref_put(&ca->io_ref);
1219                 rb->have_ioref = false;
1220
1221                 bch2_mark_io_failure(&failed, &rb->pick);
1222
1223                 can_retry = bch2_bkey_pick_read_device(c,
1224                                 bkey_i_to_s_c(&b->key),
1225                                 &failed, &rb->pick) > 0;
1226
1227                 if (!bio->bi_status &&
1228                     !bch2_btree_node_read_done(c, ca, b, can_retry, &saw_error)) {
1229                         if (retry)
1230                                 bch_info(c, "retry success");
1231                         break;
1232                 }
1233
1234                 saw_error = true;
1235
1236                 if (!can_retry) {
1237                         set_btree_node_read_error(b);
1238                         break;
1239                 }
1240         }
1241
1242         bch2_time_stats_update(&c->times[BCH_TIME_btree_node_read],
1243                                rb->start_time);
1244         bio_put(&rb->bio);
1245         printbuf_exit(&buf);
1246
1247         if (saw_error && !btree_node_read_error(b)) {
1248                 struct printbuf buf = PRINTBUF;
1249
1250                 bch2_bpos_to_text(&buf, b->key.k.p);
1251                 bch_info(c, "%s: rewriting btree node at btree=%s level=%u %s due to error",
1252                          __func__, bch2_btree_ids[b->c.btree_id], b->c.level, buf.buf);
1253                 printbuf_exit(&buf);
1254
1255                 bch2_btree_node_rewrite_async(c, b);
1256         }
1257
1258         clear_btree_node_read_in_flight(b);
1259         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1260 }
1261
1262 static void btree_node_read_endio(struct bio *bio)
1263 {
1264         struct btree_read_bio *rb =
1265                 container_of(bio, struct btree_read_bio, bio);
1266         struct bch_fs *c        = rb->c;
1267
1268         if (rb->have_ioref) {
1269                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1270
1271                 bch2_latency_acct(ca, rb->start_time, READ);
1272         }
1273
1274         queue_work(c->io_complete_wq, &rb->work);
1275 }
1276
1277 struct btree_node_read_all {
1278         struct closure          cl;
1279         struct bch_fs           *c;
1280         struct btree            *b;
1281         unsigned                nr;
1282         void                    *buf[BCH_REPLICAS_MAX];
1283         struct bio              *bio[BCH_REPLICAS_MAX];
1284         blk_status_t            err[BCH_REPLICAS_MAX];
1285 };
1286
1287 static unsigned btree_node_sectors_written(struct bch_fs *c, void *data)
1288 {
1289         struct btree_node *bn = data;
1290         struct btree_node_entry *bne;
1291         unsigned offset = 0;
1292
1293         if (le64_to_cpu(bn->magic) !=  bset_magic(c))
1294                 return 0;
1295
1296         while (offset < btree_sectors(c)) {
1297                 if (!offset) {
1298                         offset += vstruct_sectors(bn, c->block_bits);
1299                 } else {
1300                         bne = data + (offset << 9);
1301                         if (bne->keys.seq != bn->keys.seq)
1302                                 break;
1303                         offset += vstruct_sectors(bne, c->block_bits);
1304                 }
1305         }
1306
1307         return offset;
1308 }
1309
1310 static bool btree_node_has_extra_bsets(struct bch_fs *c, unsigned offset, void *data)
1311 {
1312         struct btree_node *bn = data;
1313         struct btree_node_entry *bne;
1314
1315         if (!offset)
1316                 return false;
1317
1318         while (offset < btree_sectors(c)) {
1319                 bne = data + (offset << 9);
1320                 if (bne->keys.seq == bn->keys.seq)
1321                         return true;
1322                 offset++;
1323         }
1324
1325         return false;
1326         return offset;
1327 }
1328
1329 static void btree_node_read_all_replicas_done(struct closure *cl)
1330 {
1331         struct btree_node_read_all *ra =
1332                 container_of(cl, struct btree_node_read_all, cl);
1333         struct bch_fs *c = ra->c;
1334         struct btree *b = ra->b;
1335         struct printbuf buf = PRINTBUF;
1336         bool dump_bset_maps = false;
1337         bool have_retry = false;
1338         int ret = 0, best = -1, write = READ;
1339         unsigned i, written = 0, written2 = 0;
1340         __le64 seq = b->key.k.type == KEY_TYPE_btree_ptr_v2
1341                 ? bkey_i_to_btree_ptr_v2(&b->key)->v.seq : 0;
1342         bool _saw_error = false, *saw_error = &_saw_error;
1343
1344         for (i = 0; i < ra->nr; i++) {
1345                 struct btree_node *bn = ra->buf[i];
1346
1347                 if (ra->err[i])
1348                         continue;
1349
1350                 if (le64_to_cpu(bn->magic) != bset_magic(c) ||
1351                     (seq && seq != bn->keys.seq))
1352                         continue;
1353
1354                 if (best < 0) {
1355                         best = i;
1356                         written = btree_node_sectors_written(c, bn);
1357                         continue;
1358                 }
1359
1360                 written2 = btree_node_sectors_written(c, ra->buf[i]);
1361                 if (btree_err_on(written2 != written, BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1362                                  "btree node sectors written mismatch: %u != %u",
1363                                  written, written2) ||
1364                     btree_err_on(btree_node_has_extra_bsets(c, written2, ra->buf[i]),
1365                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1366                                  "found bset signature after last bset") ||
1367                     btree_err_on(memcmp(ra->buf[best], ra->buf[i], written << 9),
1368                                  BTREE_ERR_FIXABLE, c, NULL, b, NULL,
1369                                  "btree node replicas content mismatch"))
1370                         dump_bset_maps = true;
1371
1372                 if (written2 > written) {
1373                         written = written2;
1374                         best = i;
1375                 }
1376         }
1377 fsck_err:
1378         if (dump_bset_maps) {
1379                 for (i = 0; i < ra->nr; i++) {
1380                         struct btree_node *bn = ra->buf[i];
1381                         struct btree_node_entry *bne = NULL;
1382                         unsigned offset = 0, sectors;
1383                         bool gap = false;
1384
1385                         if (ra->err[i])
1386                                 continue;
1387
1388                         printbuf_reset(&buf);
1389
1390                         while (offset < btree_sectors(c)) {
1391                                 if (!offset) {
1392                                         sectors = vstruct_sectors(bn, c->block_bits);
1393                                 } else {
1394                                         bne = ra->buf[i] + (offset << 9);
1395                                         if (bne->keys.seq != bn->keys.seq)
1396                                                 break;
1397                                         sectors = vstruct_sectors(bne, c->block_bits);
1398                                 }
1399
1400                                 prt_printf(&buf, " %u-%u", offset, offset + sectors);
1401                                 if (bne && bch2_journal_seq_is_blacklisted(c,
1402                                                         le64_to_cpu(bne->keys.journal_seq), false))
1403                                         prt_printf(&buf, "*");
1404                                 offset += sectors;
1405                         }
1406
1407                         while (offset < btree_sectors(c)) {
1408                                 bne = ra->buf[i] + (offset << 9);
1409                                 if (bne->keys.seq == bn->keys.seq) {
1410                                         if (!gap)
1411                                                 prt_printf(&buf, " GAP");
1412                                         gap = true;
1413
1414                                         sectors = vstruct_sectors(bne, c->block_bits);
1415                                         prt_printf(&buf, " %u-%u", offset, offset + sectors);
1416                                         if (bch2_journal_seq_is_blacklisted(c,
1417                                                         le64_to_cpu(bne->keys.journal_seq), false))
1418                                                 prt_printf(&buf, "*");
1419                                 }
1420                                 offset++;
1421                         }
1422
1423                         bch_err(c, "replica %u:%s", i, buf.buf);
1424                 }
1425         }
1426
1427         if (best >= 0) {
1428                 memcpy(b->data, ra->buf[best], btree_bytes(c));
1429                 ret = bch2_btree_node_read_done(c, NULL, b, false, saw_error);
1430         } else {
1431                 ret = -1;
1432         }
1433
1434         if (ret)
1435                 set_btree_node_read_error(b);
1436         else if (*saw_error)
1437                 bch2_btree_node_rewrite_async(c, b);
1438
1439         for (i = 0; i < ra->nr; i++) {
1440                 mempool_free(ra->buf[i], &c->btree_bounce_pool);
1441                 bio_put(ra->bio[i]);
1442         }
1443
1444         closure_debug_destroy(&ra->cl);
1445         kfree(ra);
1446         printbuf_exit(&buf);
1447
1448         clear_btree_node_read_in_flight(b);
1449         wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1450 }
1451
1452 static void btree_node_read_all_replicas_endio(struct bio *bio)
1453 {
1454         struct btree_read_bio *rb =
1455                 container_of(bio, struct btree_read_bio, bio);
1456         struct bch_fs *c        = rb->c;
1457         struct btree_node_read_all *ra = rb->ra;
1458
1459         if (rb->have_ioref) {
1460                 struct bch_dev *ca = bch_dev_bkey_exists(c, rb->pick.ptr.dev);
1461
1462                 bch2_latency_acct(ca, rb->start_time, READ);
1463         }
1464
1465         ra->err[rb->idx] = bio->bi_status;
1466         closure_put(&ra->cl);
1467 }
1468
1469 /*
1470  * XXX This allocates multiple times from the same mempools, and can deadlock
1471  * under sufficient memory pressure (but is only a debug path)
1472  */
1473 static int btree_node_read_all_replicas(struct bch_fs *c, struct btree *b, bool sync)
1474 {
1475         struct bkey_s_c k = bkey_i_to_s_c(&b->key);
1476         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1477         const union bch_extent_entry *entry;
1478         struct extent_ptr_decoded pick;
1479         struct btree_node_read_all *ra;
1480         unsigned i;
1481
1482         ra = kzalloc(sizeof(*ra), GFP_NOFS);
1483         if (!ra)
1484                 return -BCH_ERR_ENOMEM_btree_node_read_all_replicas;
1485
1486         closure_init(&ra->cl, NULL);
1487         ra->c   = c;
1488         ra->b   = b;
1489         ra->nr  = bch2_bkey_nr_ptrs(k);
1490
1491         for (i = 0; i < ra->nr; i++) {
1492                 ra->buf[i] = mempool_alloc(&c->btree_bounce_pool, GFP_NOFS);
1493                 ra->bio[i] = bio_alloc_bioset(NULL,
1494                                               buf_pages(ra->buf[i], btree_bytes(c)),
1495                                               REQ_OP_READ|REQ_SYNC|REQ_META,
1496                                               GFP_NOFS,
1497                                               &c->btree_bio);
1498         }
1499
1500         i = 0;
1501         bkey_for_each_ptr_decode(k.k, ptrs, pick, entry) {
1502                 struct bch_dev *ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1503                 struct btree_read_bio *rb =
1504                         container_of(ra->bio[i], struct btree_read_bio, bio);
1505                 rb->c                   = c;
1506                 rb->b                   = b;
1507                 rb->ra                  = ra;
1508                 rb->start_time          = local_clock();
1509                 rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1510                 rb->idx                 = i;
1511                 rb->pick                = pick;
1512                 rb->bio.bi_iter.bi_sector = pick.ptr.offset;
1513                 rb->bio.bi_end_io       = btree_node_read_all_replicas_endio;
1514                 bch2_bio_map(&rb->bio, ra->buf[i], btree_bytes(c));
1515
1516                 if (rb->have_ioref) {
1517                         this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1518                                      bio_sectors(&rb->bio));
1519                         bio_set_dev(&rb->bio, ca->disk_sb.bdev);
1520
1521                         closure_get(&ra->cl);
1522                         submit_bio(&rb->bio);
1523                 } else {
1524                         ra->err[i] = BLK_STS_REMOVED;
1525                 }
1526
1527                 i++;
1528         }
1529
1530         if (sync) {
1531                 closure_sync(&ra->cl);
1532                 btree_node_read_all_replicas_done(&ra->cl);
1533         } else {
1534                 continue_at(&ra->cl, btree_node_read_all_replicas_done,
1535                             c->io_complete_wq);
1536         }
1537
1538         return 0;
1539 }
1540
1541 void bch2_btree_node_read(struct bch_fs *c, struct btree *b,
1542                           bool sync)
1543 {
1544         struct extent_ptr_decoded pick;
1545         struct btree_read_bio *rb;
1546         struct bch_dev *ca;
1547         struct bio *bio;
1548         int ret;
1549
1550         trace_and_count(c, btree_node_read, c, b);
1551
1552         if (bch2_verify_all_btree_replicas &&
1553             !btree_node_read_all_replicas(c, b, sync))
1554                 return;
1555
1556         ret = bch2_bkey_pick_read_device(c, bkey_i_to_s_c(&b->key),
1557                                          NULL, &pick);
1558
1559         if (ret <= 0) {
1560                 struct printbuf buf = PRINTBUF;
1561
1562                 prt_str(&buf, "btree node read error: no device to read from\n at ");
1563                 btree_pos_to_text(&buf, c, b);
1564                 bch_err(c, "%s", buf.buf);
1565
1566                 if (test_bit(BCH_FS_TOPOLOGY_REPAIR_DONE, &c->flags))
1567                         bch2_fatal_error(c);
1568
1569                 set_btree_node_read_error(b);
1570                 clear_btree_node_read_in_flight(b);
1571                 wake_up_bit(&b->flags, BTREE_NODE_read_in_flight);
1572                 printbuf_exit(&buf);
1573                 return;
1574         }
1575
1576         ca = bch_dev_bkey_exists(c, pick.ptr.dev);
1577
1578         bio = bio_alloc_bioset(NULL,
1579                                buf_pages(b->data, btree_bytes(c)),
1580                                REQ_OP_READ|REQ_SYNC|REQ_META,
1581                                GFP_NOFS,
1582                                &c->btree_bio);
1583         rb = container_of(bio, struct btree_read_bio, bio);
1584         rb->c                   = c;
1585         rb->b                   = b;
1586         rb->ra                  = NULL;
1587         rb->start_time          = local_clock();
1588         rb->have_ioref          = bch2_dev_get_ioref(ca, READ);
1589         rb->pick                = pick;
1590         INIT_WORK(&rb->work, btree_node_read_work);
1591         bio->bi_iter.bi_sector  = pick.ptr.offset;
1592         bio->bi_end_io          = btree_node_read_endio;
1593         bch2_bio_map(bio, b->data, btree_bytes(c));
1594
1595         if (rb->have_ioref) {
1596                 this_cpu_add(ca->io_done->sectors[READ][BCH_DATA_btree],
1597                              bio_sectors(bio));
1598                 bio_set_dev(bio, ca->disk_sb.bdev);
1599
1600                 if (sync) {
1601                         submit_bio_wait(bio);
1602
1603                         btree_node_read_work(&rb->work);
1604                 } else {
1605                         submit_bio(bio);
1606                 }
1607         } else {
1608                 bio->bi_status = BLK_STS_REMOVED;
1609
1610                 if (sync)
1611                         btree_node_read_work(&rb->work);
1612                 else
1613                         queue_work(c->io_complete_wq, &rb->work);
1614         }
1615 }
1616
1617 static int __bch2_btree_root_read(struct btree_trans *trans, enum btree_id id,
1618                                   const struct bkey_i *k, unsigned level)
1619 {
1620         struct bch_fs *c = trans->c;
1621         struct closure cl;
1622         struct btree *b;
1623         int ret;
1624
1625         closure_init_stack(&cl);
1626
1627         do {
1628                 ret = bch2_btree_cache_cannibalize_lock(c, &cl);
1629                 closure_sync(&cl);
1630         } while (ret);
1631
1632         b = bch2_btree_node_mem_alloc(trans, level != 0);
1633         bch2_btree_cache_cannibalize_unlock(c);
1634
1635         BUG_ON(IS_ERR(b));
1636
1637         bkey_copy(&b->key, k);
1638         BUG_ON(bch2_btree_node_hash_insert(&c->btree_cache, b, level, id));
1639
1640         set_btree_node_read_in_flight(b);
1641
1642         bch2_btree_node_read(c, b, true);
1643
1644         if (btree_node_read_error(b)) {
1645                 bch2_btree_node_hash_remove(&c->btree_cache, b);
1646
1647                 mutex_lock(&c->btree_cache.lock);
1648                 list_move(&b->list, &c->btree_cache.freeable);
1649                 mutex_unlock(&c->btree_cache.lock);
1650
1651                 ret = -EIO;
1652                 goto err;
1653         }
1654
1655         bch2_btree_set_root_for_read(c, b);
1656 err:
1657         six_unlock_write(&b->c.lock);
1658         six_unlock_intent(&b->c.lock);
1659
1660         return ret;
1661 }
1662
1663 int bch2_btree_root_read(struct bch_fs *c, enum btree_id id,
1664                         const struct bkey_i *k, unsigned level)
1665 {
1666         return bch2_trans_run(c, __bch2_btree_root_read(&trans, id, k, level));
1667
1668 }
1669
1670 void bch2_btree_complete_write(struct bch_fs *c, struct btree *b,
1671                               struct btree_write *w)
1672 {
1673         unsigned long old, new, v = READ_ONCE(b->will_make_reachable);
1674
1675         do {
1676                 old = new = v;
1677                 if (!(old & 1))
1678                         break;
1679
1680                 new &= ~1UL;
1681         } while ((v = cmpxchg(&b->will_make_reachable, old, new)) != old);
1682
1683         if (old & 1)
1684                 closure_put(&((struct btree_update *) new)->cl);
1685
1686         bch2_journal_pin_drop(&c->journal, &w->journal);
1687 }
1688
1689 static void __btree_node_write_done(struct bch_fs *c, struct btree *b)
1690 {
1691         struct btree_write *w = btree_prev_write(b);
1692         unsigned long old, new, v;
1693         unsigned type = 0;
1694
1695         bch2_btree_complete_write(c, b, w);
1696
1697         v = READ_ONCE(b->flags);
1698         do {
1699                 old = new = v;
1700
1701                 if ((old & (1U << BTREE_NODE_dirty)) &&
1702                     (old & (1U << BTREE_NODE_need_write)) &&
1703                     !(old & (1U << BTREE_NODE_never_write)) &&
1704                     !(old & (1U << BTREE_NODE_write_blocked)) &&
1705                     !(old & (1U << BTREE_NODE_will_make_reachable))) {
1706                         new &= ~(1U << BTREE_NODE_dirty);
1707                         new &= ~(1U << BTREE_NODE_need_write);
1708                         new |=  (1U << BTREE_NODE_write_in_flight);
1709                         new |=  (1U << BTREE_NODE_write_in_flight_inner);
1710                         new |=  (1U << BTREE_NODE_just_written);
1711                         new ^=  (1U << BTREE_NODE_write_idx);
1712
1713                         type = new & BTREE_WRITE_TYPE_MASK;
1714                         new &= ~BTREE_WRITE_TYPE_MASK;
1715                 } else {
1716                         new &= ~(1U << BTREE_NODE_write_in_flight);
1717                         new &= ~(1U << BTREE_NODE_write_in_flight_inner);
1718                 }
1719         } while ((v = cmpxchg(&b->flags, old, new)) != old);
1720
1721         if (new & (1U << BTREE_NODE_write_in_flight))
1722                 __bch2_btree_node_write(c, b, BTREE_WRITE_ALREADY_STARTED|type);
1723         else
1724                 wake_up_bit(&b->flags, BTREE_NODE_write_in_flight);
1725 }
1726
1727 static void btree_node_write_done(struct bch_fs *c, struct btree *b)
1728 {
1729         struct btree_trans trans;
1730
1731         bch2_trans_init(&trans, c, 0, 0);
1732
1733         btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
1734         __btree_node_write_done(c, b);
1735         six_unlock_read(&b->c.lock);
1736
1737         bch2_trans_exit(&trans);
1738 }
1739
1740 static void btree_node_write_work(struct work_struct *work)
1741 {
1742         struct btree_write_bio *wbio =
1743                 container_of(work, struct btree_write_bio, work);
1744         struct bch_fs *c        = wbio->wbio.c;
1745         struct btree *b         = wbio->wbio.bio.bi_private;
1746         struct bch_extent_ptr *ptr;
1747         int ret = 0;
1748
1749         btree_bounce_free(c,
1750                 wbio->data_bytes,
1751                 wbio->wbio.used_mempool,
1752                 wbio->data);
1753
1754         bch2_bkey_drop_ptrs(bkey_i_to_s(&wbio->key), ptr,
1755                 bch2_dev_list_has_dev(wbio->wbio.failed, ptr->dev));
1756
1757         if (!bch2_bkey_nr_ptrs(bkey_i_to_s_c(&wbio->key)))
1758                 goto err;
1759
1760         if (wbio->wbio.first_btree_write) {
1761                 if (wbio->wbio.failed.nr) {
1762
1763                 }
1764         } else {
1765                 ret = bch2_trans_do(c, NULL, NULL, 0,
1766                         bch2_btree_node_update_key_get_iter(&trans, b, &wbio->key,
1767                                         BCH_WATERMARK_reclaim|
1768                                         BTREE_INSERT_JOURNAL_RECLAIM|
1769                                         BTREE_INSERT_NOFAIL|
1770                                         BTREE_INSERT_NOCHECK_RW,
1771                                         !wbio->wbio.failed.nr));
1772                 if (ret)
1773                         goto err;
1774         }
1775 out:
1776         bio_put(&wbio->wbio.bio);
1777         btree_node_write_done(c, b);
1778         return;
1779 err:
1780         set_btree_node_noevict(b);
1781         if (!bch2_err_matches(ret, EROFS))
1782                 bch2_fs_fatal_error(c, "fatal error writing btree node: %s", bch2_err_str(ret));
1783         goto out;
1784 }
1785
1786 static void btree_node_write_endio(struct bio *bio)
1787 {
1788         struct bch_write_bio *wbio      = to_wbio(bio);
1789         struct bch_write_bio *parent    = wbio->split ? wbio->parent : NULL;
1790         struct bch_write_bio *orig      = parent ?: wbio;
1791         struct btree_write_bio *wb      = container_of(orig, struct btree_write_bio, wbio);
1792         struct bch_fs *c                = wbio->c;
1793         struct btree *b                 = wbio->bio.bi_private;
1794         struct bch_dev *ca              = bch_dev_bkey_exists(c, wbio->dev);
1795         unsigned long flags;
1796
1797         if (wbio->have_ioref)
1798                 bch2_latency_acct(ca, wbio->submit_time, WRITE);
1799
1800         if (bch2_dev_io_err_on(bio->bi_status, ca, "btree write error: %s",
1801                                bch2_blk_status_to_str(bio->bi_status)) ||
1802             bch2_meta_write_fault("btree")) {
1803                 spin_lock_irqsave(&c->btree_write_error_lock, flags);
1804                 bch2_dev_list_add_dev(&orig->failed, wbio->dev);
1805                 spin_unlock_irqrestore(&c->btree_write_error_lock, flags);
1806         }
1807
1808         if (wbio->have_ioref)
1809                 percpu_ref_put(&ca->io_ref);
1810
1811         if (parent) {
1812                 bio_put(bio);
1813                 bio_endio(&parent->bio);
1814                 return;
1815         }
1816
1817         clear_btree_node_write_in_flight_inner(b);
1818         wake_up_bit(&b->flags, BTREE_NODE_write_in_flight_inner);
1819         INIT_WORK(&wb->work, btree_node_write_work);
1820         queue_work(c->btree_io_complete_wq, &wb->work);
1821 }
1822
1823 static int validate_bset_for_write(struct bch_fs *c, struct btree *b,
1824                                    struct bset *i, unsigned sectors)
1825 {
1826         struct printbuf buf = PRINTBUF;
1827         bool saw_error;
1828         int ret;
1829
1830         ret = bch2_bkey_invalid(c, bkey_i_to_s_c(&b->key),
1831                                 BKEY_TYPE_btree, WRITE, &buf);
1832
1833         if (ret)
1834                 bch2_fs_inconsistent(c, "invalid btree node key before write: %s", buf.buf);
1835         printbuf_exit(&buf);
1836         if (ret)
1837                 return ret;
1838
1839         ret = validate_bset_keys(c, b, i, WRITE, false, &saw_error) ?:
1840                 validate_bset(c, NULL, b, i, b->written, sectors, WRITE, false, &saw_error);
1841         if (ret) {
1842                 bch2_inconsistent_error(c);
1843                 dump_stack();
1844         }
1845
1846         return ret;
1847 }
1848
1849 static void btree_write_submit(struct work_struct *work)
1850 {
1851         struct btree_write_bio *wbio = container_of(work, struct btree_write_bio, work);
1852         struct bch_extent_ptr *ptr;
1853         BKEY_PADDED_ONSTACK(k, BKEY_BTREE_PTR_VAL_U64s_MAX) tmp;
1854
1855         bkey_copy(&tmp.k, &wbio->key);
1856
1857         bkey_for_each_ptr(bch2_bkey_ptrs(bkey_i_to_s(&tmp.k)), ptr)
1858                 ptr->offset += wbio->sector_offset;
1859
1860         bch2_submit_wbio_replicas(&wbio->wbio, wbio->wbio.c, BCH_DATA_btree,
1861                                   &tmp.k, false);
1862 }
1863
1864 void __bch2_btree_node_write(struct bch_fs *c, struct btree *b, unsigned flags)
1865 {
1866         struct btree_write_bio *wbio;
1867         struct bset_tree *t;
1868         struct bset *i;
1869         struct btree_node *bn = NULL;
1870         struct btree_node_entry *bne = NULL;
1871         struct sort_iter sort_iter;
1872         struct nonce nonce;
1873         unsigned bytes_to_write, sectors_to_write, bytes, u64s;
1874         u64 seq = 0;
1875         bool used_mempool;
1876         unsigned long old, new;
1877         bool validate_before_checksum = false;
1878         enum btree_write_type type = flags & BTREE_WRITE_TYPE_MASK;
1879         void *data;
1880         int ret;
1881
1882         if (flags & BTREE_WRITE_ALREADY_STARTED)
1883                 goto do_write;
1884
1885         /*
1886          * We may only have a read lock on the btree node - the dirty bit is our
1887          * "lock" against racing with other threads that may be trying to start
1888          * a write, we do a write iff we clear the dirty bit. Since setting the
1889          * dirty bit requires a write lock, we can't race with other threads
1890          * redirtying it:
1891          */
1892         do {
1893                 old = new = READ_ONCE(b->flags);
1894
1895                 if (!(old & (1 << BTREE_NODE_dirty)))
1896                         return;
1897
1898                 if ((flags & BTREE_WRITE_ONLY_IF_NEED) &&
1899                     !(old & (1 << BTREE_NODE_need_write)))
1900                         return;
1901
1902                 if (old &
1903                     ((1 << BTREE_NODE_never_write)|
1904                      (1 << BTREE_NODE_write_blocked)))
1905                         return;
1906
1907                 if (b->written &&
1908                     (old & (1 << BTREE_NODE_will_make_reachable)))
1909                         return;
1910
1911                 if (old & (1 << BTREE_NODE_write_in_flight))
1912                         return;
1913
1914                 if (flags & BTREE_WRITE_ONLY_IF_NEED)
1915                         type = new & BTREE_WRITE_TYPE_MASK;
1916                 new &= ~BTREE_WRITE_TYPE_MASK;
1917
1918                 new &= ~(1 << BTREE_NODE_dirty);
1919                 new &= ~(1 << BTREE_NODE_need_write);
1920                 new |=  (1 << BTREE_NODE_write_in_flight);
1921                 new |=  (1 << BTREE_NODE_write_in_flight_inner);
1922                 new |=  (1 << BTREE_NODE_just_written);
1923                 new ^=  (1 << BTREE_NODE_write_idx);
1924         } while (cmpxchg_acquire(&b->flags, old, new) != old);
1925
1926         if (new & (1U << BTREE_NODE_need_write))
1927                 return;
1928 do_write:
1929         BUG_ON((type == BTREE_WRITE_initial) != (b->written == 0));
1930
1931         atomic_dec(&c->btree_cache.dirty);
1932
1933         BUG_ON(btree_node_fake(b));
1934         BUG_ON((b->will_make_reachable != 0) != !b->written);
1935
1936         BUG_ON(b->written >= btree_sectors(c));
1937         BUG_ON(b->written & (block_sectors(c) - 1));
1938         BUG_ON(bset_written(b, btree_bset_last(b)));
1939         BUG_ON(le64_to_cpu(b->data->magic) != bset_magic(c));
1940         BUG_ON(memcmp(&b->data->format, &b->format, sizeof(b->format)));
1941
1942         bch2_sort_whiteouts(c, b);
1943
1944         sort_iter_init(&sort_iter, b);
1945
1946         bytes = !b->written
1947                 ? sizeof(struct btree_node)
1948                 : sizeof(struct btree_node_entry);
1949
1950         bytes += b->whiteout_u64s * sizeof(u64);
1951
1952         for_each_bset(b, t) {
1953                 i = bset(b, t);
1954
1955                 if (bset_written(b, i))
1956                         continue;
1957
1958                 bytes += le16_to_cpu(i->u64s) * sizeof(u64);
1959                 sort_iter_add(&sort_iter,
1960                               btree_bkey_first(b, t),
1961                               btree_bkey_last(b, t));
1962                 seq = max(seq, le64_to_cpu(i->journal_seq));
1963         }
1964
1965         BUG_ON(b->written && !seq);
1966
1967         /* bch2_varint_decode may read up to 7 bytes past the end of the buffer: */
1968         bytes += 8;
1969
1970         /* buffer must be a multiple of the block size */
1971         bytes = round_up(bytes, block_bytes(c));
1972
1973         data = btree_bounce_alloc(c, bytes, &used_mempool);
1974
1975         if (!b->written) {
1976                 bn = data;
1977                 *bn = *b->data;
1978                 i = &bn->keys;
1979         } else {
1980                 bne = data;
1981                 bne->keys = b->data->keys;
1982                 i = &bne->keys;
1983         }
1984
1985         i->journal_seq  = cpu_to_le64(seq);
1986         i->u64s         = 0;
1987
1988         sort_iter_add(&sort_iter,
1989                       unwritten_whiteouts_start(c, b),
1990                       unwritten_whiteouts_end(c, b));
1991         SET_BSET_SEPARATE_WHITEOUTS(i, false);
1992
1993         b->whiteout_u64s = 0;
1994
1995         u64s = bch2_sort_keys(i->start, &sort_iter, false);
1996         le16_add_cpu(&i->u64s, u64s);
1997
1998         BUG_ON(!b->written && i->u64s != b->data->keys.u64s);
1999
2000         set_needs_whiteout(i, false);
2001
2002         /* do we have data to write? */
2003         if (b->written && !i->u64s)
2004                 goto nowrite;
2005
2006         bytes_to_write = vstruct_end(i) - data;
2007         sectors_to_write = round_up(bytes_to_write, block_bytes(c)) >> 9;
2008
2009         if (!b->written &&
2010             b->key.k.type == KEY_TYPE_btree_ptr_v2)
2011                 BUG_ON(btree_ptr_sectors_written(&b->key) != sectors_to_write);
2012
2013         memset(data + bytes_to_write, 0,
2014                (sectors_to_write << 9) - bytes_to_write);
2015
2016         BUG_ON(b->written + sectors_to_write > btree_sectors(c));
2017         BUG_ON(BSET_BIG_ENDIAN(i) != CPU_BIG_ENDIAN);
2018         BUG_ON(i->seq != b->data->keys.seq);
2019
2020         i->version = cpu_to_le16(c->sb.version);
2021         SET_BSET_OFFSET(i, b->written);
2022         SET_BSET_CSUM_TYPE(i, bch2_meta_checksum_type(c));
2023
2024         if (bch2_csum_type_is_encryption(BSET_CSUM_TYPE(i)))
2025                 validate_before_checksum = true;
2026
2027         /* validate_bset will be modifying: */
2028         if (le16_to_cpu(i->version) < bcachefs_metadata_version_current)
2029                 validate_before_checksum = true;
2030
2031         /* if we're going to be encrypting, check metadata validity first: */
2032         if (validate_before_checksum &&
2033             validate_bset_for_write(c, b, i, sectors_to_write))
2034                 goto err;
2035
2036         ret = bset_encrypt(c, i, b->written << 9);
2037         if (bch2_fs_fatal_err_on(ret, c,
2038                         "error encrypting btree node: %i\n", ret))
2039                 goto err;
2040
2041         nonce = btree_nonce(i, b->written << 9);
2042
2043         if (bn)
2044                 bn->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bn);
2045         else
2046                 bne->csum = csum_vstruct(c, BSET_CSUM_TYPE(i), nonce, bne);
2047
2048         /* if we're not encrypting, check metadata after checksumming: */
2049         if (!validate_before_checksum &&
2050             validate_bset_for_write(c, b, i, sectors_to_write))
2051                 goto err;
2052
2053         /*
2054          * We handle btree write errors by immediately halting the journal -
2055          * after we've done that, we can't issue any subsequent btree writes
2056          * because they might have pointers to new nodes that failed to write.
2057          *
2058          * Furthermore, there's no point in doing any more btree writes because
2059          * with the journal stopped, we're never going to update the journal to
2060          * reflect that those writes were done and the data flushed from the
2061          * journal:
2062          *
2063          * Also on journal error, the pending write may have updates that were
2064          * never journalled (interior nodes, see btree_update_nodes_written()) -
2065          * it's critical that we don't do the write in that case otherwise we
2066          * will have updates visible that weren't in the journal:
2067          *
2068          * Make sure to update b->written so bch2_btree_init_next() doesn't
2069          * break:
2070          */
2071         if (bch2_journal_error(&c->journal) ||
2072             c->opts.nochanges)
2073                 goto err;
2074
2075         trace_and_count(c, btree_node_write, b, bytes_to_write, sectors_to_write);
2076
2077         wbio = container_of(bio_alloc_bioset(NULL,
2078                                 buf_pages(data, sectors_to_write << 9),
2079                                 REQ_OP_WRITE|REQ_META,
2080                                 GFP_NOFS,
2081                                 &c->btree_bio),
2082                             struct btree_write_bio, wbio.bio);
2083         wbio_init(&wbio->wbio.bio);
2084         wbio->data                      = data;
2085         wbio->data_bytes                = bytes;
2086         wbio->sector_offset             = b->written;
2087         wbio->wbio.c                    = c;
2088         wbio->wbio.used_mempool         = used_mempool;
2089         wbio->wbio.first_btree_write    = !b->written;
2090         wbio->wbio.bio.bi_end_io        = btree_node_write_endio;
2091         wbio->wbio.bio.bi_private       = b;
2092
2093         bch2_bio_map(&wbio->wbio.bio, data, sectors_to_write << 9);
2094
2095         bkey_copy(&wbio->key, &b->key);
2096
2097         b->written += sectors_to_write;
2098
2099         if (wbio->key.k.type == KEY_TYPE_btree_ptr_v2)
2100                 bkey_i_to_btree_ptr_v2(&wbio->key)->v.sectors_written =
2101                         cpu_to_le16(b->written);
2102
2103         atomic64_inc(&c->btree_write_stats[type].nr);
2104         atomic64_add(bytes_to_write, &c->btree_write_stats[type].bytes);
2105
2106         INIT_WORK(&wbio->work, btree_write_submit);
2107         queue_work(c->io_complete_wq, &wbio->work);
2108         return;
2109 err:
2110         set_btree_node_noevict(b);
2111         b->written += sectors_to_write;
2112 nowrite:
2113         btree_bounce_free(c, bytes, used_mempool, data);
2114         __btree_node_write_done(c, b);
2115 }
2116
2117 /*
2118  * Work that must be done with write lock held:
2119  */
2120 bool bch2_btree_post_write_cleanup(struct bch_fs *c, struct btree *b)
2121 {
2122         bool invalidated_iter = false;
2123         struct btree_node_entry *bne;
2124         struct bset_tree *t;
2125
2126         if (!btree_node_just_written(b))
2127                 return false;
2128
2129         BUG_ON(b->whiteout_u64s);
2130
2131         clear_btree_node_just_written(b);
2132
2133         /*
2134          * Note: immediately after write, bset_written() doesn't work - the
2135          * amount of data we had to write after compaction might have been
2136          * smaller than the offset of the last bset.
2137          *
2138          * However, we know that all bsets have been written here, as long as
2139          * we're still holding the write lock:
2140          */
2141
2142         /*
2143          * XXX: decide if we really want to unconditionally sort down to a
2144          * single bset:
2145          */
2146         if (b->nsets > 1) {
2147                 btree_node_sort(c, b, 0, b->nsets, true);
2148                 invalidated_iter = true;
2149         } else {
2150                 invalidated_iter = bch2_drop_whiteouts(b, COMPACT_ALL);
2151         }
2152
2153         for_each_bset(b, t)
2154                 set_needs_whiteout(bset(b, t), true);
2155
2156         bch2_btree_verify(c, b);
2157
2158         /*
2159          * If later we don't unconditionally sort down to a single bset, we have
2160          * to ensure this is still true:
2161          */
2162         BUG_ON((void *) btree_bkey_last(b, bset_tree_last(b)) > write_block(b));
2163
2164         bne = want_new_bset(c, b);
2165         if (bne)
2166                 bch2_bset_init_next(c, b, bne);
2167
2168         bch2_btree_build_aux_trees(b);
2169
2170         return invalidated_iter;
2171 }
2172
2173 /*
2174  * Use this one if the node is intent locked:
2175  */
2176 void bch2_btree_node_write(struct bch_fs *c, struct btree *b,
2177                            enum six_lock_type lock_type_held,
2178                            unsigned flags)
2179 {
2180         if (lock_type_held == SIX_LOCK_intent ||
2181             (lock_type_held == SIX_LOCK_read &&
2182              six_lock_tryupgrade(&b->c.lock))) {
2183                 __bch2_btree_node_write(c, b, flags);
2184
2185                 /* don't cycle lock unnecessarily: */
2186                 if (btree_node_just_written(b) &&
2187                     six_trylock_write(&b->c.lock)) {
2188                         bch2_btree_post_write_cleanup(c, b);
2189                         six_unlock_write(&b->c.lock);
2190                 }
2191
2192                 if (lock_type_held == SIX_LOCK_read)
2193                         six_lock_downgrade(&b->c.lock);
2194         } else {
2195                 __bch2_btree_node_write(c, b, flags);
2196                 if (lock_type_held == SIX_LOCK_write &&
2197                     btree_node_just_written(b))
2198                         bch2_btree_post_write_cleanup(c, b);
2199         }
2200 }
2201
2202 static bool __bch2_btree_flush_all(struct bch_fs *c, unsigned flag)
2203 {
2204         struct bucket_table *tbl;
2205         struct rhash_head *pos;
2206         struct btree *b;
2207         unsigned i;
2208         bool ret = false;
2209 restart:
2210         rcu_read_lock();
2211         for_each_cached_btree(b, c, tbl, i, pos)
2212                 if (test_bit(flag, &b->flags)) {
2213                         rcu_read_unlock();
2214                         wait_on_bit_io(&b->flags, flag, TASK_UNINTERRUPTIBLE);
2215                         ret = true;
2216                         goto restart;
2217                 }
2218         rcu_read_unlock();
2219
2220         return ret;
2221 }
2222
2223 bool bch2_btree_flush_all_reads(struct bch_fs *c)
2224 {
2225         return __bch2_btree_flush_all(c, BTREE_NODE_read_in_flight);
2226 }
2227
2228 bool bch2_btree_flush_all_writes(struct bch_fs *c)
2229 {
2230         return __bch2_btree_flush_all(c, BTREE_NODE_write_in_flight);
2231 }
2232
2233 static const char * const bch2_btree_write_types[] = {
2234 #define x(t, n) [n] = #t,
2235         BCH_BTREE_WRITE_TYPES()
2236         NULL
2237 };
2238
2239 void bch2_btree_write_stats_to_text(struct printbuf *out, struct bch_fs *c)
2240 {
2241         printbuf_tabstop_push(out, 20);
2242         printbuf_tabstop_push(out, 10);
2243
2244         prt_tab(out);
2245         prt_str(out, "nr");
2246         prt_tab(out);
2247         prt_str(out, "size");
2248         prt_newline(out);
2249
2250         for (unsigned i = 0; i < BTREE_WRITE_TYPE_NR; i++) {
2251                 u64 nr          = atomic64_read(&c->btree_write_stats[i].nr);
2252                 u64 bytes       = atomic64_read(&c->btree_write_stats[i].bytes);
2253
2254                 prt_printf(out, "%s:", bch2_btree_write_types[i]);
2255                 prt_tab(out);
2256                 prt_u64(out, nr);
2257                 prt_tab(out);
2258                 prt_human_readable_u64(out, nr ? div64_u64(bytes, nr) : 0);
2259                 prt_newline(out);
2260         }
2261 }