git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_gc.c

   1 /*
   2  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   3  * Copyright (C) 2014 Datera Inc.
   4  */
   5
   6 #include "bcachefs.h"
   7 #include "alloc.h"
   8 #include "bkey_methods.h"
   9 #include "btree_locking.h"
  10 #include "btree_update_interior.h"
  11 #include "btree_io.h"
  12 #include "btree_gc.h"
  13 #include "buckets.h"
  14 #include "clock.h"
  15 #include "debug.h"
  16 #include "error.h"
  17 #include "extents.h"
  18 #include "journal.h"
  19 #include "keylist.h"
  20 #include "move.h"
  21 #include "super-io.h"
  22
  23 #include <linux/slab.h>
  24 #include <linux/bitops.h>
  25 #include <linux/freezer.h>
  26 #include <linux/kthread.h>
  27 #include <linux/rcupdate.h>
  28 #include <trace/events/bcachefs.h>
  29
  30 struct range_checks {
  31         struct range_level {
  32                 struct bpos     min;
  33                 struct bpos     max;
  34         }                       l[BTREE_MAX_DEPTH];
  35         unsigned                depth;
  36 };
  37
  38 static void btree_node_range_checks_init(struct range_checks *r, unsigned depth)
  39 {
  40         unsigned i;
  41
  42         for (i = 0; i < BTREE_MAX_DEPTH; i++)
  43                 r->l[i].min = r->l[i].max = POS_MIN;
  44         r->depth = depth;
  45 }
  46
  47 static void btree_node_range_checks(struct bch_fs *c, struct btree *b,
  48                                     struct range_checks *r)
  49 {
  50         struct range_level *l = &r->l[b->level];
  51
  52         struct bpos expected_min = bkey_cmp(l->min, l->max)
  53                 ? btree_type_successor(b->btree_id, l->max)
  54                 : l->max;
  55
  56         bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, expected_min), c,
  57                 "btree node has incorrect min key: %llu:%llu != %llu:%llu",
  58                 b->data->min_key.inode,
  59                 b->data->min_key.offset,
  60                 expected_min.inode,
  61                 expected_min.offset);
  62
  63         l->max = b->data->max_key;
  64
  65         if (b->level > r->depth) {
  66                 l = &r->l[b->level - 1];
  67
  68                 bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, l->min), c,
  69                         "btree node min doesn't match min of child nodes: %llu:%llu != %llu:%llu",
  70                         b->data->min_key.inode,
  71                         b->data->min_key.offset,
  72                         l->min.inode,
  73                         l->min.offset);
  74
  75                 bch2_fs_inconsistent_on(bkey_cmp(b->data->max_key, l->max), c,
  76                         "btree node max doesn't match max of child nodes: %llu:%llu != %llu:%llu",
  77                         b->data->max_key.inode,
  78                         b->data->max_key.offset,
  79                         l->max.inode,
  80                         l->max.offset);
  81
  82                 if (bkey_cmp(b->data->max_key, POS_MAX))
  83                         l->min = l->max =
  84                                 btree_type_successor(b->btree_id,
  85                                                      b->data->max_key);
  86         }
  87 }
  88
  89 u8 bch2_btree_key_recalc_oldest_gen(struct bch_fs *c, struct bkey_s_c k)
  90 {
  91         const struct bch_extent_ptr *ptr;
  92         u8 max_stale = 0;
  93
  94         if (bkey_extent_is_data(k.k)) {
  95                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
  96
  97                 extent_for_each_ptr(e, ptr) {
  98                         struct bch_dev *ca = c->devs[ptr->dev];
  99                         size_t b = PTR_BUCKET_NR(ca, ptr);
 100
 101                         if (gen_after(ca->oldest_gens[b], ptr->gen))
 102                                 ca->oldest_gens[b] = ptr->gen;
 103
 104                         max_stale = max(max_stale, ptr_stale(ca, ptr));
 105                 }
 106         }
 107
 108         return max_stale;
 109 }
 110
 111 /*
 112  * For runtime mark and sweep:
 113  */
 114 static u8 bch2_btree_mark_key(struct bch_fs *c, enum bkey_type type,
 115                               struct bkey_s_c k, unsigned flags)
 116 {
 117         switch (type) {
 118         case BKEY_TYPE_BTREE:
 119                 bch2_gc_mark_key(c, k, c->opts.btree_node_size, true, flags);
 120                 return 0;
 121         case BKEY_TYPE_EXTENTS:
 122                 bch2_gc_mark_key(c, k, k.k->size, false, flags);
 123                 return bch2_btree_key_recalc_oldest_gen(c, k);
 124         default:
 125                 BUG();
 126         }
 127 }
 128
 129 int bch2_btree_mark_key_initial(struct bch_fs *c, enum bkey_type type,
 130                                 struct bkey_s_c k)
 131 {
 132         enum bch_data_type data_type = type == BKEY_TYPE_BTREE
 133                 ? BCH_DATA_BTREE : BCH_DATA_USER;
 134         int ret = 0;
 135
 136         switch (k.k->type) {
 137         case BCH_EXTENT:
 138         case BCH_EXTENT_CACHED: {
 139                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
 140                 const struct bch_extent_ptr *ptr;
 141
 142                 if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
 143                     (!c->opts.nofsck &&
 144                      fsck_err_on(!bch2_sb_has_replicas(c, e, data_type), c,
 145                                  "superblock not marked as containing replicas"))) {
 146                         ret = bch2_check_mark_super(c, e, data_type);
 147                         if (ret)
 148                                 return ret;
 149                 }
 150
 151                 extent_for_each_ptr(e, ptr) {
 152                         struct bch_dev *ca = c->devs[ptr->dev];
 153                         struct bucket *g = PTR_BUCKET(ca, ptr);
 154
 155                         if (mustfix_fsck_err_on(!g->mark.gen_valid, c,
 156                                         "found ptr with missing gen in alloc btree,\n"
 157                                         "type %s gen %u",
 158                                         bch2_data_types[data_type],
 159                                         ptr->gen)) {
 160                                 g->_mark.gen = ptr->gen;
 161                                 g->_mark.gen_valid = 1;
 162                                 set_bit(g - ca->buckets, ca->bucket_dirty);
 163                         }
 164
 165                         if (mustfix_fsck_err_on(gen_cmp(ptr->gen, g->mark.gen) > 0, c,
 166                                         "%s ptr gen in the future: %u > %u",
 167                                         bch2_data_types[data_type],
 168                                         ptr->gen, g->mark.gen)) {
 169                                 g->_mark.gen = ptr->gen;
 170                                 g->_mark.gen_valid = 1;
 171                                 set_bit(g - ca->buckets, ca->bucket_dirty);
 172                                 set_bit(BCH_FS_FIXED_GENS, &c->flags);
 173                         }
 174
 175                 }
 176                 break;
 177         }
 178         }
 179
 180
 181         atomic64_set(&c->key_version,
 182                      max_t(u64, k.k->version.lo,
 183                            atomic64_read(&c->key_version)));
 184
 185         bch2_btree_mark_key(c, type, k, BCH_BUCKET_MARK_NOATOMIC);
 186 fsck_err:
 187         return ret;
 188 }
 189
 190 static unsigned btree_gc_mark_node(struct bch_fs *c, struct btree *b)
 191 {
 192         enum bkey_type type = btree_node_type(b);
 193         struct btree_node_iter iter;
 194         struct bkey unpacked;
 195         struct bkey_s_c k;
 196         u8 stale = 0;
 197
 198         if (btree_node_has_ptrs(b))
 199                 for_each_btree_node_key_unpack(b, k, &iter,
 200                                                btree_node_is_extents(b),
 201                                                &unpacked) {
 202                         bch2_bkey_debugcheck(c, b, k);
 203                         stale = max(stale, bch2_btree_mark_key(c, type, k, 0));
 204                 }
 205
 206         return stale;
 207 }
 208
 209 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
 210 {
 211         write_seqcount_begin(&c->gc_pos_lock);
 212         c->gc_pos = new_pos;
 213         write_seqcount_end(&c->gc_pos_lock);
 214 }
 215
 216 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
 217 {
 218         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
 219         __gc_pos_set(c, new_pos);
 220 }
 221
 222 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id)
 223 {
 224         struct btree_iter iter;
 225         struct btree *b;
 226         struct range_checks r;
 227         unsigned depth = btree_id == BTREE_ID_EXTENTS ? 0 : 1;
 228         unsigned max_stale;
 229         int ret = 0;
 230
 231         /*
 232          * if expensive_debug_checks is on, run range_checks on all leaf nodes:
 233          */
 234         if (expensive_debug_checks(c))
 235                 depth = 0;
 236
 237         btree_node_range_checks_init(&r, depth);
 238
 239         __for_each_btree_node(&iter, c, btree_id, POS_MIN,
 240                               0, depth, BTREE_ITER_PREFETCH, b) {
 241                 btree_node_range_checks(c, b, &r);
 242
 243                 bch2_verify_btree_nr_keys(b);
 244
 245                 max_stale = btree_gc_mark_node(c, b);
 246
 247                 gc_pos_set(c, gc_pos_btree_node(b));
 248
 249                 if (max_stale > 32)
 250                         bch2_btree_node_rewrite(c, &iter,
 251                                         b->data->keys.seq,
 252                                         BTREE_INSERT_USE_RESERVE|
 253                                         BTREE_INSERT_GC_LOCK_HELD);
 254                 else if (!btree_gc_rewrite_disabled(c) &&
 255                          (btree_gc_always_rewrite(c) || max_stale > 16))
 256                         bch2_btree_node_rewrite(c, &iter,
 257                                         b->data->keys.seq,
 258                                         BTREE_INSERT_NOWAIT|
 259                                         BTREE_INSERT_GC_LOCK_HELD);
 260
 261                 bch2_btree_iter_cond_resched(&iter);
 262         }
 263         ret = bch2_btree_iter_unlock(&iter);
 264         if (ret)
 265                 return ret;
 266
 267         mutex_lock(&c->btree_root_lock);
 268
 269         b = c->btree_roots[btree_id].b;
 270         bch2_btree_mark_key(c, BKEY_TYPE_BTREE, bkey_i_to_s_c(&b->key), 0);
 271         gc_pos_set(c, gc_pos_btree_root(b->btree_id));
 272
 273         mutex_unlock(&c->btree_root_lock);
 274         return 0;
 275 }
 276
 277 static void bch2_mark_allocator_buckets(struct bch_fs *c)
 278 {
 279         struct bch_dev *ca;
 280         struct open_bucket *ob;
 281         size_t i, j, iter;
 282         unsigned ci;
 283
 284         for_each_member_device(ca, c, ci) {
 285                 spin_lock(&ca->freelist_lock);
 286
 287                 fifo_for_each_entry(i, &ca->free_inc, iter)
 288                         bch2_mark_alloc_bucket(ca, &ca->buckets[i], true);
 289
 290                 for (j = 0; j < RESERVE_NR; j++)
 291                         fifo_for_each_entry(i, &ca->free[j], iter)
 292                                 bch2_mark_alloc_bucket(ca, &ca->buckets[i], true);
 293
 294                 spin_unlock(&ca->freelist_lock);
 295         }
 296
 297         for (ob = c->open_buckets;
 298              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
 299              ob++) {
 300                 const struct bch_extent_ptr *ptr;
 301
 302                 mutex_lock(&ob->lock);
 303                 open_bucket_for_each_ptr(ob, ptr) {
 304                         ca = c->devs[ptr->dev];
 305                         bch2_mark_alloc_bucket(ca, PTR_BUCKET(ca, ptr), true);
 306                 }
 307                 mutex_unlock(&ob->lock);
 308         }
 309 }
 310
 311 static void mark_metadata_sectors(struct bch_dev *ca, u64 start, u64 end,
 312                                   enum bucket_data_type type)
 313 {
 314         u64 b = sector_to_bucket(ca, start);
 315
 316         do {
 317                 bch2_mark_metadata_bucket(ca, ca->buckets + b, type, true);
 318                 b++;
 319         } while (b < sector_to_bucket(ca, end));
 320 }
 321
 322 static void bch2_dev_mark_superblocks(struct bch_dev *ca)
 323 {
 324         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
 325         unsigned i;
 326
 327         for (i = 0; i < layout->nr_superblocks; i++) {
 328                 if (layout->sb_offset[i] == BCH_SB_SECTOR)
 329                         mark_metadata_sectors(ca, 0, BCH_SB_SECTOR,
 330                                               BUCKET_SB);
 331
 332                 mark_metadata_sectors(ca,
 333                                       layout->sb_offset[i],
 334                                       layout->sb_offset[i] +
 335                                       (1 << layout->sb_max_size_bits),
 336                                       BUCKET_SB);
 337         }
 338 }
 339
 340 /*
 341  * Mark non btree metadata - prios, journal
 342  */
 343 void bch2_mark_dev_metadata(struct bch_fs *c, struct bch_dev *ca)
 344 {
 345         unsigned i;
 346         u64 b;
 347
 348         lockdep_assert_held(&c->sb_lock);
 349
 350         bch2_dev_mark_superblocks(ca);
 351
 352         spin_lock(&c->journal.lock);
 353
 354         for (i = 0; i < ca->journal.nr; i++) {
 355                 b = ca->journal.buckets[i];
 356                 bch2_mark_metadata_bucket(ca, ca->buckets + b,
 357                                          BUCKET_JOURNAL, true);
 358         }
 359
 360         spin_unlock(&c->journal.lock);
 361 }
 362
 363 static void bch2_mark_metadata(struct bch_fs *c)
 364 {
 365         struct bch_dev *ca;
 366         unsigned i;
 367
 368         mutex_lock(&c->sb_lock);
 369         gc_pos_set(c, gc_phase(GC_PHASE_SB_METADATA));
 370
 371         for_each_online_member(ca, c, i)
 372                 bch2_mark_dev_metadata(c, ca);
 373         mutex_unlock(&c->sb_lock);
 374 }
 375
 376 /* Also see bch2_pending_btree_node_free_insert_done() */
 377 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
 378 {
 379         struct bch_fs_usage stats = { 0 };
 380         struct btree_update *as;
 381         struct pending_btree_node_free *d;
 382
 383         mutex_lock(&c->btree_interior_update_lock);
 384         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
 385
 386         for_each_pending_btree_node_free(c, as, d)
 387                 if (d->index_update_done)
 388                         __bch2_mark_key(c, bkey_i_to_s_c(&d->key),
 389                                         c->opts.btree_node_size, true,
 390                                         &stats, 0,
 391                                         BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE);
 392         /*
 393          * Don't apply stats - pending deletes aren't tracked in
 394          * bch_alloc_stats:
 395          */
 396
 397         mutex_unlock(&c->btree_interior_update_lock);
 398 }
 399
 400 void bch2_gc_start(struct bch_fs *c)
 401 {
 402         struct bch_dev *ca;
 403         struct bucket *g;
 404         struct bucket_mark new;
 405         unsigned i;
 406         int cpu;
 407
 408         lg_global_lock(&c->usage_lock);
 409
 410         /*
 411          * Indicates to buckets code that gc is now in progress - done under
 412          * usage_lock to avoid racing with bch2_mark_key():
 413          */
 414         __gc_pos_set(c, GC_POS_MIN);
 415
 416         /* Save a copy of the existing bucket stats while we recompute them: */
 417         for_each_member_device(ca, c, i) {
 418                 ca->usage_cached = __bch2_dev_usage_read(ca);
 419                 for_each_possible_cpu(cpu) {
 420                         struct bch_dev_usage *p =
 421                                 per_cpu_ptr(ca->usage_percpu, cpu);
 422                         memset(p, 0, sizeof(*p));
 423                 }
 424         }
 425
 426         c->usage_cached = __bch2_fs_usage_read(c);
 427         for_each_possible_cpu(cpu) {
 428                 struct bch_fs_usage *p =
 429                         per_cpu_ptr(c->usage_percpu, cpu);
 430
 431                 memset(p->s, 0, sizeof(p->s));
 432         }
 433
 434         lg_global_unlock(&c->usage_lock);
 435
 436         /* Clear bucket marks: */
 437         for_each_member_device(ca, c, i)
 438                 for_each_bucket(g, ca) {
 439                         bucket_cmpxchg(g, new, ({
 440                                 new.owned_by_allocator  = 0;
 441                                 new.data_type           = 0;
 442                                 new.cached_sectors      = 0;
 443                                 new.dirty_sectors       = 0;
 444                         }));
 445                         ca->oldest_gens[g - ca->buckets] = new.gen;
 446                 }
 447 }
 448
 449 /**
 450  * bch_gc - recompute bucket marks and oldest_gen, rewrite btree nodes
 451  */
 452 void bch2_gc(struct bch_fs *c)
 453 {
 454         struct bch_dev *ca;
 455         u64 start_time = local_clock();
 456         unsigned i;
 457
 458         /*
 459          * Walk _all_ references to buckets, and recompute them:
 460          *
 461          * Order matters here:
 462          *  - Concurrent GC relies on the fact that we have a total ordering for
 463          *    everything that GC walks - see  gc_will_visit_node(),
 464          *    gc_will_visit_root()
 465          *
 466          *  - also, references move around in the course of index updates and
 467          *    various other crap: everything needs to agree on the ordering
 468          *    references are allowed to move around in - e.g., we're allowed to
 469          *    start with a reference owned by an open_bucket (the allocator) and
 470          *    move it to the btree, but not the reverse.
 471          *
 472          *    This is necessary to ensure that gc doesn't miss references that
 473          *    move around - if references move backwards in the ordering GC
 474          *    uses, GC could skip past them
 475          */
 476         trace_gc_start(c);
 477
 478         /*
 479          * Do this before taking gc_lock - bch2_disk_reservation_get() blocks on
 480          * gc_lock if sectors_available goes to 0:
 481          */
 482         bch2_recalc_sectors_available(c);
 483
 484         down_write(&c->gc_lock);
 485         if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
 486                 goto out;
 487
 488         bch2_gc_start(c);
 489
 490         /* Walk allocator's references: */
 491         bch2_mark_allocator_buckets(c);
 492
 493         /* Walk btree: */
 494         while (c->gc_pos.phase < (int) BTREE_ID_NR) {
 495                 int ret = c->btree_roots[c->gc_pos.phase].b
 496                         ? bch2_gc_btree(c, (int) c->gc_pos.phase)
 497                         : 0;
 498
 499                 if (ret) {
 500                         bch_err(c, "btree gc failed: %d", ret);
 501                         set_bit(BCH_FS_GC_FAILURE, &c->flags);
 502                         goto out;
 503                 }
 504
 505                 gc_pos_set(c, gc_phase(c->gc_pos.phase + 1));
 506         }
 507
 508         bch2_mark_metadata(c);
 509         bch2_mark_pending_btree_node_frees(c);
 510
 511         for_each_member_device(ca, c, i)
 512                 atomic_long_set(&ca->saturated_count, 0);
 513
 514         /* Indicates that gc is no longer in progress: */
 515         gc_pos_set(c, gc_phase(GC_PHASE_DONE));
 516         c->gc_count++;
 517 out:
 518         up_write(&c->gc_lock);
 519         trace_gc_end(c);
 520         bch2_time_stats_update(&c->btree_gc_time, start_time);
 521
 522         /*
 523          * Wake up allocator in case it was waiting for buckets
 524          * because of not being able to inc gens
 525          */
 526         for_each_member_device(ca, c, i)
 527                 bch2_wake_allocator(ca);
 528
 529         /*
 530          * At startup, allocations can happen directly instead of via the
 531          * allocator thread - issue wakeup in case they blocked on gc_lock:
 532          */
 533         closure_wake_up(&c->freelist_wait);
 534 }
 535
 536 /* Btree coalescing */
 537
 538 static void recalc_packed_keys(struct btree *b)
 539 {
 540         struct bkey_packed *k;
 541
 542         memset(&b->nr, 0, sizeof(b->nr));
 543
 544         BUG_ON(b->nsets != 1);
 545
 546         for (k =  btree_bkey_first(b, b->set);
 547              k != btree_bkey_last(b, b->set);
 548              k = bkey_next(k))
 549                 btree_keys_account_key_add(&b->nr, 0, k);
 550 }
 551
 552 static void bch2_coalesce_nodes(struct bch_fs *c, struct btree_iter *iter,
 553                                 struct btree *old_nodes[GC_MERGE_NODES])
 554 {
 555         struct btree *parent = iter->nodes[old_nodes[0]->level + 1];
 556         unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
 557         unsigned blocks = btree_blocks(c) * 2 / 3;
 558         struct btree *new_nodes[GC_MERGE_NODES];
 559         struct btree_update *as;
 560         struct keylist keylist;
 561         struct bkey_format_state format_state;
 562         struct bkey_format new_format;
 563
 564         memset(new_nodes, 0, sizeof(new_nodes));
 565         bch2_keylist_init(&keylist, NULL, 0);
 566
 567         /* Count keys that are not deleted */
 568         for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
 569                 u64s += old_nodes[i]->nr.live_u64s;
 570
 571         nr_old_nodes = nr_new_nodes = i;
 572
 573         /* Check if all keys in @old_nodes could fit in one fewer node */
 574         if (nr_old_nodes <= 1 ||
 575             __vstruct_blocks(struct btree_node, c->block_bits,
 576                              DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
 577                 return;
 578
 579         /* Find a format that all keys in @old_nodes can pack into */
 580         bch2_bkey_format_init(&format_state);
 581
 582         for (i = 0; i < nr_old_nodes; i++)
 583                 __bch2_btree_calc_format(&format_state, old_nodes[i]);
 584
 585         new_format = bch2_bkey_format_done(&format_state);
 586
 587         /* Check if repacking would make any nodes too big to fit */
 588         for (i = 0; i < nr_old_nodes; i++)
 589                 if (!bch2_btree_node_format_fits(c, old_nodes[i], &new_format)) {
 590                         trace_btree_gc_coalesce_fail(c,
 591                                         BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
 592                         return;
 593                 }
 594
 595         if (bch2_keylist_realloc(&keylist, NULL, 0,
 596                         (BKEY_U64s + BKEY_EXTENT_U64s_MAX) * nr_old_nodes)) {
 597                 trace_btree_gc_coalesce_fail(c,
 598                                 BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
 599                 return;
 600         }
 601
 602         as = bch2_btree_update_start(c, iter->btree_id,
 603                         btree_update_reserve_required(c, parent) + nr_old_nodes,
 604                         BTREE_INSERT_NOFAIL|
 605                         BTREE_INSERT_USE_RESERVE,
 606                         NULL);
 607         if (IS_ERR(as)) {
 608                 trace_btree_gc_coalesce_fail(c,
 609                                 BTREE_GC_COALESCE_FAIL_RESERVE_GET);
 610                 bch2_keylist_free(&keylist, NULL);
 611                 return;
 612         }
 613
 614         trace_btree_gc_coalesce(c, old_nodes[0]);
 615
 616         for (i = 0; i < nr_old_nodes; i++)
 617                 bch2_btree_interior_update_will_free_node(as, old_nodes[i]);
 618
 619         /* Repack everything with @new_format and sort down to one bset */
 620         for (i = 0; i < nr_old_nodes; i++)
 621                 new_nodes[i] =
 622                         __bch2_btree_node_alloc_replacement(as, old_nodes[i],
 623                                                             new_format);
 624
 625         /*
 626          * Conceptually we concatenate the nodes together and slice them
 627          * up at different boundaries.
 628          */
 629         for (i = nr_new_nodes - 1; i > 0; --i) {
 630                 struct btree *n1 = new_nodes[i];
 631                 struct btree *n2 = new_nodes[i - 1];
 632
 633                 struct bset *s1 = btree_bset_first(n1);
 634                 struct bset *s2 = btree_bset_first(n2);
 635                 struct bkey_packed *k, *last = NULL;
 636
 637                 /* Calculate how many keys from @n2 we could fit inside @n1 */
 638                 u64s = 0;
 639
 640                 for (k = s2->start;
 641                      k < vstruct_last(s2) &&
 642                      vstruct_blocks_plus(n1->data, c->block_bits,
 643                                          u64s + k->u64s) <= blocks;
 644                      k = bkey_next(k)) {
 645                         last = k;
 646                         u64s += k->u64s;
 647                 }
 648
 649                 if (u64s == le16_to_cpu(s2->u64s)) {
 650                         /* n2 fits entirely in n1 */
 651                         n1->key.k.p = n1->data->max_key = n2->data->max_key;
 652
 653                         memcpy_u64s(vstruct_last(s1),
 654                                     s2->start,
 655                                     le16_to_cpu(s2->u64s));
 656                         le16_add_cpu(&s1->u64s, le16_to_cpu(s2->u64s));
 657
 658                         set_btree_bset_end(n1, n1->set);
 659
 660                         six_unlock_write(&n2->lock);
 661                         bch2_btree_node_free_never_inserted(c, n2);
 662                         six_unlock_intent(&n2->lock);
 663
 664                         memmove(new_nodes + i - 1,
 665                                 new_nodes + i,
 666                                 sizeof(new_nodes[0]) * (nr_new_nodes - i));
 667                         new_nodes[--nr_new_nodes] = NULL;
 668                 } else if (u64s) {
 669                         /* move part of n2 into n1 */
 670                         n1->key.k.p = n1->data->max_key =
 671                                 bkey_unpack_pos(n1, last);
 672
 673                         n2->data->min_key =
 674                                 btree_type_successor(iter->btree_id,
 675                                                      n1->data->max_key);
 676
 677                         memcpy_u64s(vstruct_last(s1),
 678                                     s2->start, u64s);
 679                         le16_add_cpu(&s1->u64s, u64s);
 680
 681                         memmove(s2->start,
 682                                 vstruct_idx(s2, u64s),
 683                                 (le16_to_cpu(s2->u64s) - u64s) * sizeof(u64));
 684                         s2->u64s = cpu_to_le16(le16_to_cpu(s2->u64s) - u64s);
 685
 686                         set_btree_bset_end(n1, n1->set);
 687                         set_btree_bset_end(n2, n2->set);
 688                 }
 689         }
 690
 691         for (i = 0; i < nr_new_nodes; i++) {
 692                 struct btree *n = new_nodes[i];
 693
 694                 recalc_packed_keys(n);
 695                 btree_node_reset_sib_u64s(n);
 696
 697                 bch2_btree_build_aux_trees(n);
 698                 six_unlock_write(&n->lock);
 699
 700                 bch2_btree_node_write(c, n, &as->cl, SIX_LOCK_intent);
 701         }
 702
 703         /*
 704          * The keys for the old nodes get deleted. We don't want to insert keys
 705          * that compare equal to the keys for the new nodes we'll also be
 706          * inserting - we can't because keys on a keylist must be strictly
 707          * greater than the previous keys, and we also don't need to since the
 708          * key for the new node will serve the same purpose (overwriting the key
 709          * for the old node).
 710          */
 711         for (i = 0; i < nr_old_nodes; i++) {
 712                 struct bkey_i delete;
 713                 unsigned j;
 714
 715                 for (j = 0; j < nr_new_nodes; j++)
 716                         if (!bkey_cmp(old_nodes[i]->key.k.p,
 717                                       new_nodes[j]->key.k.p))
 718                                 goto next;
 719
 720                 bkey_init(&delete.k);
 721                 delete.k.p = old_nodes[i]->key.k.p;
 722                 bch2_keylist_add_in_order(&keylist, &delete);
 723 next:
 724                 i = i;
 725         }
 726
 727         /*
 728          * Keys for the new nodes get inserted: bch2_btree_insert_keys() only
 729          * does the lookup once and thus expects the keys to be in sorted order
 730          * so we have to make sure the new keys are correctly ordered with
 731          * respect to the deleted keys added in the previous loop
 732          */
 733         for (i = 0; i < nr_new_nodes; i++)
 734                 bch2_keylist_add_in_order(&keylist, &new_nodes[i]->key);
 735
 736         /* Insert the newly coalesced nodes */
 737         bch2_btree_insert_node(as, parent, iter, &keylist);
 738
 739         BUG_ON(!bch2_keylist_empty(&keylist));
 740
 741         BUG_ON(iter->nodes[old_nodes[0]->level] != old_nodes[0]);
 742
 743         BUG_ON(!bch2_btree_iter_node_replace(iter, new_nodes[0]));
 744
 745         for (i = 0; i < nr_new_nodes; i++)
 746                 bch2_btree_open_bucket_put(c, new_nodes[i]);
 747
 748         /* Free the old nodes and update our sliding window */
 749         for (i = 0; i < nr_old_nodes; i++) {
 750                 bch2_btree_node_free_inmem(c, old_nodes[i], iter);
 751                 six_unlock_intent(&old_nodes[i]->lock);
 752
 753                 /*
 754                  * the index update might have triggered a split, in which case
 755                  * the nodes we coalesced - the new nodes we just created -
 756                  * might not be sibling nodes anymore - don't add them to the
 757                  * sliding window (except the first):
 758                  */
 759                 if (!i) {
 760                         old_nodes[i] = new_nodes[i];
 761                 } else {
 762                         old_nodes[i] = NULL;
 763                         if (new_nodes[i])
 764                                 six_unlock_intent(&new_nodes[i]->lock);
 765                 }
 766         }
 767
 768         bch2_btree_update_done(as);
 769         bch2_keylist_free(&keylist, NULL);
 770 }
 771
 772 static int bch2_coalesce_btree(struct bch_fs *c, enum btree_id btree_id)
 773 {
 774         struct btree_iter iter;
 775         struct btree *b;
 776         unsigned i;
 777
 778         /* Sliding window of adjacent btree nodes */
 779         struct btree *merge[GC_MERGE_NODES];
 780         u32 lock_seq[GC_MERGE_NODES];
 781
 782         /*
 783          * XXX: We don't have a good way of positively matching on sibling nodes
 784          * that have the same parent - this code works by handling the cases
 785          * where they might not have the same parent, and is thus fragile. Ugh.
 786          *
 787          * Perhaps redo this to use multiple linked iterators?
 788          */
 789         memset(merge, 0, sizeof(merge));
 790
 791         __for_each_btree_node(&iter, c, btree_id, POS_MIN,
 792                               BTREE_MAX_DEPTH, 0,
 793                               BTREE_ITER_PREFETCH, b) {
 794                 memmove(merge + 1, merge,
 795                         sizeof(merge) - sizeof(merge[0]));
 796                 memmove(lock_seq + 1, lock_seq,
 797                         sizeof(lock_seq) - sizeof(lock_seq[0]));
 798
 799                 merge[0] = b;
 800
 801                 for (i = 1; i < GC_MERGE_NODES; i++) {
 802                         if (!merge[i] ||
 803                             !six_relock_intent(&merge[i]->lock, lock_seq[i]))
 804                                 break;
 805
 806                         if (merge[i]->level != merge[0]->level) {
 807                                 six_unlock_intent(&merge[i]->lock);
 808                                 break;
 809                         }
 810                 }
 811                 memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));
 812
 813                 bch2_coalesce_nodes(c, &iter, merge);
 814
 815                 for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
 816                         lock_seq[i] = merge[i]->lock.state.seq;
 817                         six_unlock_intent(&merge[i]->lock);
 818                 }
 819
 820                 lock_seq[0] = merge[0]->lock.state.seq;
 821
 822                 if (test_bit(BCH_FS_GC_STOPPING, &c->flags)) {
 823                         bch2_btree_iter_unlock(&iter);
 824                         return -ESHUTDOWN;
 825                 }
 826
 827                 bch2_btree_iter_cond_resched(&iter);
 828
 829                 /*
 830                  * If the parent node wasn't relocked, it might have been split
 831                  * and the nodes in our sliding window might not have the same
 832                  * parent anymore - blow away the sliding window:
 833                  */
 834                 if (iter.nodes[iter.level + 1] &&
 835                     !btree_node_intent_locked(&iter, iter.level + 1))
 836                         memset(merge + 1, 0,
 837                                (GC_MERGE_NODES - 1) * sizeof(merge[0]));
 838         }
 839         return bch2_btree_iter_unlock(&iter);
 840 }
 841
 842 /**
 843  * bch_coalesce - coalesce adjacent nodes with low occupancy
 844  */
 845 void bch2_coalesce(struct bch_fs *c)
 846 {
 847         enum btree_id id;
 848
 849         if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
 850                 return;
 851
 852         down_read(&c->gc_lock);
 853         trace_gc_coalesce_start(c);
 854
 855         for (id = 0; id < BTREE_ID_NR; id++) {
 856                 int ret = c->btree_roots[id].b
 857                         ? bch2_coalesce_btree(c, id)
 858                         : 0;
 859
 860                 if (ret) {
 861                         if (ret != -ESHUTDOWN)
 862                                 bch_err(c, "btree coalescing failed: %d", ret);
 863                         set_bit(BCH_FS_GC_FAILURE, &c->flags);
 864                         return;
 865                 }
 866         }
 867
 868         trace_gc_coalesce_end(c);
 869         up_read(&c->gc_lock);
 870 }
 871
 872 static int bch2_gc_thread(void *arg)
 873 {
 874         struct bch_fs *c = arg;
 875         struct io_clock *clock = &c->io_clock[WRITE];
 876         unsigned long last = atomic_long_read(&clock->now);
 877         unsigned last_kick = atomic_read(&c->kick_gc);
 878
 879         set_freezable();
 880
 881         while (1) {
 882                 while (1) {
 883                         set_current_state(TASK_INTERRUPTIBLE);
 884
 885                         if (kthread_should_stop()) {
 886                                 __set_current_state(TASK_RUNNING);
 887                                 return 0;
 888                         }
 889
 890                         if (atomic_read(&c->kick_gc) != last_kick)
 891                                 break;
 892
 893                         if (c->btree_gc_periodic) {
 894                                 unsigned long next = last + c->capacity / 16;
 895
 896                                 if (atomic_long_read(&clock->now) >= next)
 897                                         break;
 898
 899                                 bch2_io_clock_schedule_timeout(clock, next);
 900                         } else {
 901                                 schedule();
 902                         }
 903
 904                         try_to_freeze();
 905                 }
 906                 __set_current_state(TASK_RUNNING);
 907
 908                 last = atomic_long_read(&clock->now);
 909                 last_kick = atomic_read(&c->kick_gc);
 910
 911                 bch2_gc(c);
 912
 913                 debug_check_no_locks_held();
 914         }
 915
 916         return 0;
 917 }
 918
 919 void bch2_gc_thread_stop(struct bch_fs *c)
 920 {
 921         set_bit(BCH_FS_GC_STOPPING, &c->flags);
 922
 923         if (c->gc_thread)
 924                 kthread_stop(c->gc_thread);
 925
 926         c->gc_thread = NULL;
 927         clear_bit(BCH_FS_GC_STOPPING, &c->flags);
 928 }
 929
 930 int bch2_gc_thread_start(struct bch_fs *c)
 931 {
 932         struct task_struct *p;
 933
 934         BUG_ON(c->gc_thread);
 935
 936         p = kthread_create(bch2_gc_thread, c, "bcache_gc");
 937         if (IS_ERR(p))
 938                 return PTR_ERR(p);
 939
 940         c->gc_thread = p;
 941         wake_up_process(c->gc_thread);
 942         return 0;
 943 }
 944
 945 /* Initial GC computes bucket marks during startup */
 946
 947 static int bch2_initial_gc_btree(struct bch_fs *c, enum btree_id id)
 948 {
 949         struct btree_iter iter;
 950         struct btree *b;
 951         struct range_checks r;
 952         int ret = 0;
 953
 954         btree_node_range_checks_init(&r, 0);
 955
 956         if (!c->btree_roots[id].b)
 957                 return 0;
 958
 959         ret = bch2_btree_mark_key_initial(c, BKEY_TYPE_BTREE,
 960                            bkey_i_to_s_c(&c->btree_roots[id].b->key));
 961         if (ret)
 962                 return ret;
 963
 964         /*
 965          * We have to hit every btree node before starting journal replay, in
 966          * order for the journal seq blacklist machinery to work:
 967          */
 968         for_each_btree_node(&iter, c, id, POS_MIN, BTREE_ITER_PREFETCH, b) {
 969                 btree_node_range_checks(c, b, &r);
 970
 971                 if (btree_node_has_ptrs(b)) {
 972                         struct btree_node_iter node_iter;
 973                         struct bkey unpacked;
 974                         struct bkey_s_c k;
 975
 976                         for_each_btree_node_key_unpack(b, k, &node_iter,
 977                                                        btree_node_is_extents(b),
 978                                                        &unpacked) {
 979                                 ret = bch2_btree_mark_key_initial(c,
 980                                                         btree_node_type(b), k);
 981                                 if (ret)
 982                                         goto err;
 983                         }
 984                 }
 985
 986                 bch2_btree_iter_cond_resched(&iter);
 987         }
 988 err:
 989         return bch2_btree_iter_unlock(&iter) ?: ret;
 990 }
 991
 992 int bch2_initial_gc(struct bch_fs *c, struct list_head *journal)
 993 {
 994         unsigned iter = 0;
 995         enum btree_id id;
 996         int ret;
 997
 998         mutex_lock(&c->sb_lock);
 999         if (!bch2_sb_get_replicas(c->disk_sb)) {
1000                 if (BCH_SB_INITIALIZED(c->disk_sb))
1001                         bch_info(c, "building replicas info");
1002                 set_bit(BCH_FS_REBUILD_REPLICAS, &c->flags);
1003         }
1004         mutex_unlock(&c->sb_lock);
1005 again:
1006         bch2_gc_start(c);
1007
1008         for (id = 0; id < BTREE_ID_NR; id++) {
1009                 ret = bch2_initial_gc_btree(c, id);
1010                 if (ret)
1011                         return ret;
1012         }
1013
1014         ret = bch2_journal_mark(c, journal);
1015         if (ret)
1016         return ret;
1017
1018         bch2_mark_metadata(c);
1019
1020         if (test_bit(BCH_FS_FIXED_GENS, &c->flags)) {
1021                 if (iter++ > 2) {
1022                         bch_info(c, "Unable to fix bucket gens, looping");
1023                         return -EINVAL;
1024                 }
1025
1026                 bch_info(c, "Fixed gens, restarting initial mark and sweep:");
1027                 clear_bit(BCH_FS_FIXED_GENS, &c->flags);
1028                 goto again;
1029         }
1030
1031         /*
1032          * Skip past versions that might have possibly been used (as nonces),
1033          * but hadn't had their pointers written:
1034          */
1035         if (c->sb.encryption_type)
1036                 atomic64_add(1 << 16, &c->key_version);
1037
1038         gc_pos_set(c, gc_phase(GC_PHASE_DONE));
1039         set_bit(BCH_FS_INITIAL_GC_DONE, &c->flags);
1040
1041         return 0;
1042 }