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

   1 /*
   2  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   3  * Copyright (C) 2014 Datera Inc.
   4  */
   5
   6 #include "bcache.h"
   7 #include "alloc.h"
   8 #include "bkey_methods.h"
   9 #include "btree_locking.h"
  10 #include "btree_update.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 #include "writeback.h"
  23
  24 #include <linux/slab.h>
  25 #include <linux/bitops.h>
  26 #include <linux/freezer.h>
  27 #include <linux/kthread.h>
  28 #include <linux/rcupdate.h>
  29 #include <trace/events/bcache.h>
  30
  31 struct range_checks {
  32         struct range_level {
  33                 struct bpos     min;
  34                 struct bpos     max;
  35         }                       l[BTREE_MAX_DEPTH];
  36         unsigned                depth;
  37 };
  38
  39 static void btree_node_range_checks_init(struct range_checks *r, unsigned depth)
  40 {
  41         unsigned i;
  42
  43         for (i = 0; i < BTREE_MAX_DEPTH; i++)
  44                 r->l[i].min = r->l[i].max = POS_MIN;
  45         r->depth = depth;
  46 }
  47
  48 static void btree_node_range_checks(struct cache_set *c, struct btree *b,
  49                                     struct range_checks *r)
  50 {
  51         struct range_level *l = &r->l[b->level];
  52
  53         struct bpos expected_min = bkey_cmp(l->min, l->max)
  54                 ? btree_type_successor(b->btree_id, l->max)
  55                 : l->max;
  56
  57         cache_set_inconsistent_on(bkey_cmp(b->data->min_key,
  58                                            expected_min), c,
  59                 "btree node has incorrect min key: %llu:%llu != %llu:%llu",
  60                 b->data->min_key.inode,
  61                 b->data->min_key.offset,
  62                 expected_min.inode,
  63                 expected_min.offset);
  64
  65         l->max = b->data->max_key;
  66
  67         if (b->level > r->depth) {
  68                 l = &r->l[b->level - 1];
  69
  70                 cache_set_inconsistent_on(bkey_cmp(b->data->min_key,
  71                                                    l->min), c,
  72                         "btree node min doesn't match min of child nodes: %llu:%llu != %llu:%llu",
  73                         b->data->min_key.inode,
  74                         b->data->min_key.offset,
  75                         l->min.inode,
  76                         l->min.offset);
  77
  78                 cache_set_inconsistent_on(bkey_cmp(b->data->max_key,
  79                                                    l->max), c,
  80                         "btree node max doesn't match max of child nodes: %llu:%llu != %llu:%llu",
  81                         b->data->max_key.inode,
  82                         b->data->max_key.offset,
  83                         l->max.inode,
  84                         l->max.offset);
  85
  86                 if (bkey_cmp(b->data->max_key, POS_MAX))
  87                         l->min = l->max =
  88                                 btree_type_successor(b->btree_id,
  89                                                      b->data->max_key);
  90         }
  91 }
  92
  93 u8 bch_btree_key_recalc_oldest_gen(struct cache_set *c, struct bkey_s_c k)
  94 {
  95         const struct bch_extent_ptr *ptr;
  96         struct cache *ca;
  97         u8 max_stale = 0;
  98
  99         if (bkey_extent_is_data(k.k)) {
 100                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
 101
 102                 rcu_read_lock();
 103
 104                 extent_for_each_online_device(c, e, ptr, ca) {
 105                         size_t b = PTR_BUCKET_NR(ca, ptr);
 106
 107                         if (__gen_after(ca->oldest_gens[b], ptr->gen))
 108                                 ca->oldest_gens[b] = ptr->gen;
 109
 110                         max_stale = max(max_stale, ptr_stale(ca, ptr));
 111                 }
 112
 113                 rcu_read_unlock();
 114         }
 115
 116         return max_stale;
 117 }
 118
 119 /*
 120  * For runtime mark and sweep:
 121  */
 122 static u8 bch_btree_mark_key(struct cache_set *c, enum bkey_type type,
 123                              struct bkey_s_c k)
 124 {
 125         switch (type) {
 126         case BKEY_TYPE_BTREE:
 127                 bch_gc_mark_key(c, k, c->sb.btree_node_size, true);
 128                 return 0;
 129         case BKEY_TYPE_EXTENTS:
 130                 bch_gc_mark_key(c, k, k.k->size, false);
 131                 return bch_btree_key_recalc_oldest_gen(c, k);
 132         default:
 133                 BUG();
 134         }
 135 }
 136
 137 u8 bch_btree_mark_key_initial(struct cache_set *c, enum bkey_type type,
 138                           struct bkey_s_c k)
 139 {
 140         atomic64_set(&c->key_version,
 141                      max_t(u64, k.k->version.lo,
 142                            atomic64_read(&c->key_version)));
 143
 144         return bch_btree_mark_key(c, type, k);
 145 }
 146
 147 static bool btree_gc_mark_node(struct cache_set *c, struct btree *b)
 148 {
 149         if (btree_node_has_ptrs(b)) {
 150                 struct btree_node_iter iter;
 151                 struct bkey unpacked;
 152                 struct bkey_s_c k;
 153                 u8 stale = 0;
 154
 155                 for_each_btree_node_key_unpack(b, k, &iter,
 156                                                btree_node_is_extents(b),
 157                                                &unpacked) {
 158                         bkey_debugcheck(c, b, k);
 159                         stale = max(stale, bch_btree_mark_key(c,
 160                                                         btree_node_type(b), k));
 161                 }
 162
 163                 if (btree_gc_rewrite_disabled(c))
 164                         return false;
 165
 166                 if (stale > 10)
 167                         return true;
 168         }
 169
 170         if (btree_gc_always_rewrite(c))
 171                 return true;
 172
 173         return false;
 174 }
 175
 176 static inline void __gc_pos_set(struct cache_set *c, struct gc_pos new_pos)
 177 {
 178         write_seqcount_begin(&c->gc_pos_lock);
 179         c->gc_pos = new_pos;
 180         write_seqcount_end(&c->gc_pos_lock);
 181 }
 182
 183 static inline void gc_pos_set(struct cache_set *c, struct gc_pos new_pos)
 184 {
 185         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
 186         __gc_pos_set(c, new_pos);
 187 }
 188
 189 static int bch_gc_btree(struct cache_set *c, enum btree_id btree_id)
 190 {
 191         struct btree_iter iter;
 192         struct btree *b;
 193         bool should_rewrite;
 194         struct range_checks r;
 195         unsigned depth = btree_id == BTREE_ID_EXTENTS ? 0 : 1;
 196         int ret;
 197
 198         /*
 199          * if expensive_debug_checks is on, run range_checks on all leaf nodes:
 200          */
 201         if (expensive_debug_checks(c))
 202                 depth = 0;
 203
 204         btree_node_range_checks_init(&r, depth);
 205
 206         for_each_btree_node(&iter, c, btree_id, POS_MIN, depth, b) {
 207                 btree_node_range_checks(c, b, &r);
 208
 209                 bch_verify_btree_nr_keys(b);
 210
 211                 should_rewrite = btree_gc_mark_node(c, b);
 212
 213                 gc_pos_set(c, gc_pos_btree_node(b));
 214
 215                 if (should_rewrite)
 216                         bch_btree_node_rewrite(&iter, b, NULL);
 217
 218                 bch_btree_iter_cond_resched(&iter);
 219         }
 220         ret = bch_btree_iter_unlock(&iter);
 221         if (ret)
 222                 return ret;
 223
 224         mutex_lock(&c->btree_root_lock);
 225
 226         b = c->btree_roots[btree_id].b;
 227         bch_btree_mark_key(c, BKEY_TYPE_BTREE, bkey_i_to_s_c(&b->key));
 228         gc_pos_set(c, gc_pos_btree_root(b->btree_id));
 229
 230         mutex_unlock(&c->btree_root_lock);
 231         return 0;
 232 }
 233
 234 static void bch_mark_allocator_buckets(struct cache_set *c)
 235 {
 236         struct cache *ca;
 237         struct open_bucket *ob;
 238         size_t i, j, iter;
 239         unsigned ci;
 240
 241         for_each_cache(ca, c, ci) {
 242                 spin_lock(&ca->freelist_lock);
 243
 244                 fifo_for_each_entry(i, &ca->free_inc, iter)
 245                         bch_mark_alloc_bucket(ca, &ca->buckets[i], true);
 246
 247                 for (j = 0; j < RESERVE_NR; j++)
 248                         fifo_for_each_entry(i, &ca->free[j], iter)
 249                                 bch_mark_alloc_bucket(ca, &ca->buckets[i], true);
 250
 251                 spin_unlock(&ca->freelist_lock);
 252         }
 253
 254         for (ob = c->open_buckets;
 255              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
 256              ob++) {
 257                 const struct bch_extent_ptr *ptr;
 258
 259                 mutex_lock(&ob->lock);
 260                 rcu_read_lock();
 261                 open_bucket_for_each_online_device(c, ob, ptr, ca)
 262                         bch_mark_alloc_bucket(ca, PTR_BUCKET(ca, ptr), true);
 263                 rcu_read_unlock();
 264                 mutex_unlock(&ob->lock);
 265         }
 266 }
 267
 268 /*
 269  * Mark non btree metadata - prios, journal
 270  */
 271 static void bch_mark_metadata(struct cache_set *c)
 272 {
 273         struct cache *ca;
 274         unsigned i, j;
 275         u64 b;
 276
 277         for_each_cache(ca, c, i) {
 278                 for (j = 0; j < ca->journal.nr; j++) {
 279                         b = ca->journal.buckets[j];
 280                         bch_mark_metadata_bucket(ca, ca->buckets + b, true);
 281                 }
 282
 283                 spin_lock(&ca->prio_buckets_lock);
 284
 285                 for (j = 0; j < prio_buckets(ca) * 2; j++) {
 286                         b = ca->prio_buckets[j];
 287                         bch_mark_metadata_bucket(ca, ca->buckets + b, true);
 288                 }
 289
 290                 spin_unlock(&ca->prio_buckets_lock);
 291         }
 292 }
 293
 294 /* Also see bch_pending_btree_node_free_insert_done() */
 295 static void bch_mark_pending_btree_node_frees(struct cache_set *c)
 296 {
 297         struct bucket_stats_cache_set stats = { 0 };
 298         struct btree_interior_update *as;
 299         struct pending_btree_node_free *d;
 300
 301         mutex_lock(&c->btree_interior_update_lock);
 302         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
 303
 304         for_each_pending_btree_node_free(c, as, d)
 305                 if (d->index_update_done)
 306                         __bch_gc_mark_key(c, bkey_i_to_s_c(&d->key),
 307                                           c->sb.btree_node_size, true,
 308                                           &stats);
 309         /*
 310          * Don't apply stats - pending deletes aren't tracked in
 311          * cache_set_stats:
 312          */
 313
 314         mutex_unlock(&c->btree_interior_update_lock);
 315 }
 316
 317 /**
 318  * bch_gc - recompute bucket marks and oldest_gen, rewrite btree nodes
 319  */
 320 void bch_gc(struct cache_set *c)
 321 {
 322         struct cache *ca;
 323         struct bucket *g;
 324         struct bucket_mark new;
 325         u64 start_time = local_clock();
 326         unsigned i;
 327         int cpu;
 328
 329         /*
 330          * Walk _all_ references to buckets, and recompute them:
 331          *
 332          * Order matters here:
 333          *  - Concurrent GC relies on the fact that we have a total ordering for
 334          *    everything that GC walks - see  gc_will_visit_node(),
 335          *    gc_will_visit_root()
 336          *
 337          *  - also, references move around in the course of index updates and
 338          *    various other crap: everything needs to agree on the ordering
 339          *    references are allowed to move around in - e.g., we're allowed to
 340          *    start with a reference owned by an open_bucket (the allocator) and
 341          *    move it to the btree, but not the reverse.
 342          *
 343          *    This is necessary to ensure that gc doesn't miss references that
 344          *    move around - if references move backwards in the ordering GC
 345          *    uses, GC could skip past them
 346          */
 347
 348         if (test_bit(CACHE_SET_GC_FAILURE, &c->flags))
 349                 return;
 350
 351         trace_bcache_gc_start(c);
 352
 353         /*
 354          * Do this before taking gc_lock - bch_disk_reservation_get() blocks on
 355          * gc_lock if sectors_available goes to 0:
 356          */
 357         bch_recalc_sectors_available(c);
 358
 359         down_write(&c->gc_lock);
 360
 361         lg_global_lock(&c->bucket_stats_lock);
 362
 363         /*
 364          * Indicates to buckets code that gc is now in progress - done under
 365          * bucket_stats_lock to avoid racing with bch_mark_key():
 366          */
 367         __gc_pos_set(c, GC_POS_MIN);
 368
 369         /* Save a copy of the existing bucket stats while we recompute them: */
 370         for_each_cache(ca, c, i) {
 371                 ca->bucket_stats_cached = __bch_bucket_stats_read_cache(ca);
 372                 for_each_possible_cpu(cpu) {
 373                         struct bucket_stats_cache *p =
 374                                 per_cpu_ptr(ca->bucket_stats_percpu, cpu);
 375                         memset(p, 0, sizeof(*p));
 376                 }
 377         }
 378
 379         c->bucket_stats_cached = __bch_bucket_stats_read_cache_set(c);
 380         for_each_possible_cpu(cpu) {
 381                 struct bucket_stats_cache_set *p =
 382                         per_cpu_ptr(c->bucket_stats_percpu, cpu);
 383
 384                 memset(p->s, 0, sizeof(p->s));
 385                 p->persistent_reserved = 0;
 386         }
 387
 388         lg_global_unlock(&c->bucket_stats_lock);
 389
 390         /* Clear bucket marks: */
 391         for_each_cache(ca, c, i)
 392                 for_each_bucket(g, ca) {
 393                         bucket_cmpxchg(g, new, ({
 394                                 new.owned_by_allocator  = 0;
 395                                 new.is_metadata         = 0;
 396                                 new.cached_sectors      = 0;
 397                                 new.dirty_sectors       = 0;
 398                         }));
 399                         ca->oldest_gens[g - ca->buckets] = new.gen;
 400                 }
 401
 402         /* Walk allocator's references: */
 403         bch_mark_allocator_buckets(c);
 404
 405         /* Walk btree: */
 406         while (c->gc_pos.phase < (int) BTREE_ID_NR) {
 407                 int ret = c->btree_roots[c->gc_pos.phase].b
 408                         ? bch_gc_btree(c, (int) c->gc_pos.phase)
 409                         : 0;
 410
 411                 if (ret) {
 412                         bch_err(c, "btree gc failed: %d", ret);
 413                         set_bit(CACHE_SET_GC_FAILURE, &c->flags);
 414                         up_write(&c->gc_lock);
 415                         return;
 416                 }
 417
 418                 gc_pos_set(c, gc_phase(c->gc_pos.phase + 1));
 419         }
 420
 421         bch_mark_metadata(c);
 422         bch_mark_pending_btree_node_frees(c);
 423         bch_writeback_recalc_oldest_gens(c);
 424
 425         for_each_cache(ca, c, i)
 426                 atomic_long_set(&ca->saturated_count, 0);
 427
 428         /* Indicates that gc is no longer in progress: */
 429         gc_pos_set(c, gc_phase(GC_PHASE_DONE));
 430
 431         up_write(&c->gc_lock);
 432         trace_bcache_gc_end(c);
 433         bch_time_stats_update(&c->btree_gc_time, start_time);
 434
 435         /*
 436          * Wake up allocator in case it was waiting for buckets
 437          * because of not being able to inc gens
 438          */
 439         for_each_cache(ca, c, i)
 440                 bch_wake_allocator(ca);
 441 }
 442
 443 /* Btree coalescing */
 444
 445 static void recalc_packed_keys(struct btree *b)
 446 {
 447         struct bkey_packed *k;
 448
 449         memset(&b->nr, 0, sizeof(b->nr));
 450
 451         BUG_ON(b->nsets != 1);
 452
 453         for (k =  btree_bkey_first(b, b->set);
 454              k != btree_bkey_last(b, b->set);
 455              k = bkey_next(k))
 456                 btree_keys_account_key_add(&b->nr, 0, k);
 457 }
 458
 459 static void bch_coalesce_nodes(struct btree *old_nodes[GC_MERGE_NODES],
 460                                struct btree_iter *iter)
 461 {
 462         struct btree *parent = iter->nodes[old_nodes[0]->level + 1];
 463         struct cache_set *c = iter->c;
 464         unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
 465         unsigned blocks = btree_blocks(c) * 2 / 3;
 466         struct btree *new_nodes[GC_MERGE_NODES];
 467         struct btree_interior_update *as;
 468         struct btree_reserve *res;
 469         struct keylist keylist;
 470         struct bkey_format_state format_state;
 471         struct bkey_format new_format;
 472
 473         memset(new_nodes, 0, sizeof(new_nodes));
 474         bch_keylist_init(&keylist, NULL, 0);
 475
 476         /* Count keys that are not deleted */
 477         for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
 478                 u64s += old_nodes[i]->nr.live_u64s;
 479
 480         nr_old_nodes = nr_new_nodes = i;
 481
 482         /* Check if all keys in @old_nodes could fit in one fewer node */
 483         if (nr_old_nodes <= 1 ||
 484             __vstruct_blocks(struct btree_node, c->block_bits,
 485                              DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
 486                 return;
 487
 488         res = bch_btree_reserve_get(c, parent, nr_old_nodes,
 489                                     BTREE_INSERT_NOFAIL|
 490                                     BTREE_INSERT_USE_RESERVE,
 491                                     NULL);
 492         if (IS_ERR(res)) {
 493                 trace_bcache_btree_gc_coalesce_fail(c,
 494                                 BTREE_GC_COALESCE_FAIL_RESERVE_GET);
 495                 return;
 496         }
 497
 498         if (bch_keylist_realloc(&keylist, NULL, 0,
 499                         (BKEY_U64s + BKEY_EXTENT_U64s_MAX) * nr_old_nodes)) {
 500                 trace_bcache_btree_gc_coalesce_fail(c,
 501                                 BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
 502                 goto out;
 503         }
 504
 505         /* Find a format that all keys in @old_nodes can pack into */
 506         bch_bkey_format_init(&format_state);
 507
 508         for (i = 0; i < nr_old_nodes; i++)
 509                 __bch_btree_calc_format(&format_state, old_nodes[i]);
 510
 511         new_format = bch_bkey_format_done(&format_state);
 512
 513         /* Check if repacking would make any nodes too big to fit */
 514         for (i = 0; i < nr_old_nodes; i++)
 515                 if (!bch_btree_node_format_fits(c, old_nodes[i], &new_format)) {
 516                         trace_bcache_btree_gc_coalesce_fail(c,
 517                                         BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
 518                         goto out;
 519                 }
 520
 521         trace_bcache_btree_gc_coalesce(c, parent, nr_old_nodes);
 522
 523         as = bch_btree_interior_update_alloc(c);
 524
 525         for (i = 0; i < nr_old_nodes; i++)
 526                 bch_btree_interior_update_will_free_node(c, as, old_nodes[i]);
 527
 528         /* Repack everything with @new_format and sort down to one bset */
 529         for (i = 0; i < nr_old_nodes; i++)
 530                 new_nodes[i] = __btree_node_alloc_replacement(c, old_nodes[i],
 531                                                               new_format, res);
 532
 533         /*
 534          * Conceptually we concatenate the nodes together and slice them
 535          * up at different boundaries.
 536          */
 537         for (i = nr_new_nodes - 1; i > 0; --i) {
 538                 struct btree *n1 = new_nodes[i];
 539                 struct btree *n2 = new_nodes[i - 1];
 540
 541                 struct bset *s1 = btree_bset_first(n1);
 542                 struct bset *s2 = btree_bset_first(n2);
 543                 struct bkey_packed *k, *last = NULL;
 544
 545                 /* Calculate how many keys from @n2 we could fit inside @n1 */
 546                 u64s = 0;
 547
 548                 for (k = s2->start;
 549                      k < vstruct_last(s2) &&
 550                      vstruct_blocks_plus(n1->data, c->block_bits,
 551                                          u64s + k->u64s) <= blocks;
 552                      k = bkey_next(k)) {
 553                         last = k;
 554                         u64s += k->u64s;
 555                 }
 556
 557                 if (u64s == le16_to_cpu(s2->u64s)) {
 558                         /* n2 fits entirely in n1 */
 559                         n1->key.k.p = n1->data->max_key = n2->data->max_key;
 560
 561                         memcpy_u64s(vstruct_last(s1),
 562                                     s2->start,
 563                                     le16_to_cpu(s2->u64s));
 564                         le16_add_cpu(&s1->u64s, le16_to_cpu(s2->u64s));
 565
 566                         set_btree_bset_end(n1, n1->set);
 567
 568                         six_unlock_write(&n2->lock);
 569                         bch_btree_node_free_never_inserted(c, n2);
 570                         six_unlock_intent(&n2->lock);
 571
 572                         memmove(new_nodes + i - 1,
 573                                 new_nodes + i,
 574                                 sizeof(new_nodes[0]) * (nr_new_nodes - i));
 575                         new_nodes[--nr_new_nodes] = NULL;
 576                 } else if (u64s) {
 577                         /* move part of n2 into n1 */
 578                         n1->key.k.p = n1->data->max_key =
 579                                 bkey_unpack_pos(n1, last);
 580
 581                         n2->data->min_key =
 582                                 btree_type_successor(iter->btree_id,
 583                                                      n1->data->max_key);
 584
 585                         memcpy_u64s(vstruct_last(s1),
 586                                     s2->start, u64s);
 587                         le16_add_cpu(&s1->u64s, u64s);
 588
 589                         memmove(s2->start,
 590                                 vstruct_idx(s2, u64s),
 591                                 (le16_to_cpu(s2->u64s) - u64s) * sizeof(u64));
 592                         s2->u64s = cpu_to_le16(le16_to_cpu(s2->u64s) - u64s);
 593
 594                         set_btree_bset_end(n1, n1->set);
 595                         set_btree_bset_end(n2, n2->set);
 596                 }
 597         }
 598
 599         for (i = 0; i < nr_new_nodes; i++) {
 600                 struct btree *n = new_nodes[i];
 601
 602                 recalc_packed_keys(n);
 603                 btree_node_reset_sib_u64s(n);
 604
 605                 bch_btree_build_aux_trees(n);
 606                 six_unlock_write(&n->lock);
 607
 608                 bch_btree_node_write(c, n, &as->cl, SIX_LOCK_intent, -1);
 609         }
 610
 611         /*
 612          * The keys for the old nodes get deleted. We don't want to insert keys
 613          * that compare equal to the keys for the new nodes we'll also be
 614          * inserting - we can't because keys on a keylist must be strictly
 615          * greater than the previous keys, and we also don't need to since the
 616          * key for the new node will serve the same purpose (overwriting the key
 617          * for the old node).
 618          */
 619         for (i = 0; i < nr_old_nodes; i++) {
 620                 struct bkey_i delete;
 621                 unsigned j;
 622
 623                 for (j = 0; j < nr_new_nodes; j++)
 624                         if (!bkey_cmp(old_nodes[i]->key.k.p,
 625                                       new_nodes[j]->key.k.p))
 626                                 goto next;
 627
 628                 bkey_init(&delete.k);
 629                 delete.k.p = old_nodes[i]->key.k.p;
 630                 bch_keylist_add_in_order(&keylist, &delete);
 631 next:
 632                 i = i;
 633         }
 634
 635         /*
 636          * Keys for the new nodes get inserted: bch_btree_insert_keys() only
 637          * does the lookup once and thus expects the keys to be in sorted order
 638          * so we have to make sure the new keys are correctly ordered with
 639          * respect to the deleted keys added in the previous loop
 640          */
 641         for (i = 0; i < nr_new_nodes; i++)
 642                 bch_keylist_add_in_order(&keylist, &new_nodes[i]->key);
 643
 644         /* Insert the newly coalesced nodes */
 645         bch_btree_insert_node(parent, iter, &keylist, res, as);
 646
 647         BUG_ON(!bch_keylist_empty(&keylist));
 648
 649         BUG_ON(iter->nodes[old_nodes[0]->level] != old_nodes[0]);
 650
 651         BUG_ON(!bch_btree_iter_node_replace(iter, new_nodes[0]));
 652
 653         for (i = 0; i < nr_new_nodes; i++)
 654                 btree_open_bucket_put(c, new_nodes[i]);
 655
 656         /* Free the old nodes and update our sliding window */
 657         for (i = 0; i < nr_old_nodes; i++) {
 658                 bch_btree_node_free_inmem(iter, old_nodes[i]);
 659                 six_unlock_intent(&old_nodes[i]->lock);
 660
 661                 /*
 662                  * the index update might have triggered a split, in which case
 663                  * the nodes we coalesced - the new nodes we just created -
 664                  * might not be sibling nodes anymore - don't add them to the
 665                  * sliding window (except the first):
 666                  */
 667                 if (!i) {
 668                         old_nodes[i] = new_nodes[i];
 669                 } else {
 670                         old_nodes[i] = NULL;
 671                         if (new_nodes[i])
 672                                 six_unlock_intent(&new_nodes[i]->lock);
 673                 }
 674         }
 675 out:
 676         bch_keylist_free(&keylist, NULL);
 677         bch_btree_reserve_put(c, res);
 678 }
 679
 680 static int bch_coalesce_btree(struct cache_set *c, enum btree_id btree_id)
 681 {
 682         struct btree_iter iter;
 683         struct btree *b;
 684         unsigned i;
 685
 686         /* Sliding window of adjacent btree nodes */
 687         struct btree *merge[GC_MERGE_NODES];
 688         u32 lock_seq[GC_MERGE_NODES];
 689
 690         /*
 691          * XXX: We don't have a good way of positively matching on sibling nodes
 692          * that have the same parent - this code works by handling the cases
 693          * where they might not have the same parent, and is thus fragile. Ugh.
 694          *
 695          * Perhaps redo this to use multiple linked iterators?
 696          */
 697         memset(merge, 0, sizeof(merge));
 698
 699         __for_each_btree_node(&iter, c, btree_id, POS_MIN, 0, b, U8_MAX) {
 700                 memmove(merge + 1, merge,
 701                         sizeof(merge) - sizeof(merge[0]));
 702                 memmove(lock_seq + 1, lock_seq,
 703                         sizeof(lock_seq) - sizeof(lock_seq[0]));
 704
 705                 merge[0] = b;
 706
 707                 for (i = 1; i < GC_MERGE_NODES; i++) {
 708                         if (!merge[i] ||
 709                             !six_relock_intent(&merge[i]->lock, lock_seq[i]))
 710                                 break;
 711
 712                         if (merge[i]->level != merge[0]->level) {
 713                                 six_unlock_intent(&merge[i]->lock);
 714                                 break;
 715                         }
 716                 }
 717                 memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));
 718
 719                 bch_coalesce_nodes(merge, &iter);
 720
 721                 for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
 722                         lock_seq[i] = merge[i]->lock.state.seq;
 723                         six_unlock_intent(&merge[i]->lock);
 724                 }
 725
 726                 lock_seq[0] = merge[0]->lock.state.seq;
 727
 728                 if (test_bit(CACHE_SET_GC_STOPPING, &c->flags)) {
 729                         bch_btree_iter_unlock(&iter);
 730                         return -ESHUTDOWN;
 731                 }
 732
 733                 bch_btree_iter_cond_resched(&iter);
 734
 735                 /*
 736                  * If the parent node wasn't relocked, it might have been split
 737                  * and the nodes in our sliding window might not have the same
 738                  * parent anymore - blow away the sliding window:
 739                  */
 740                 if (iter.nodes[iter.level + 1] &&
 741                     !btree_node_intent_locked(&iter, iter.level + 1))
 742                         memset(merge + 1, 0,
 743                                (GC_MERGE_NODES - 1) * sizeof(merge[0]));
 744         }
 745         return bch_btree_iter_unlock(&iter);
 746 }
 747
 748 /**
 749  * bch_coalesce - coalesce adjacent nodes with low occupancy
 750  */
 751 void bch_coalesce(struct cache_set *c)
 752 {
 753         u64 start_time;
 754         enum btree_id id;
 755
 756         if (btree_gc_coalesce_disabled(c))
 757                 return;
 758
 759         if (test_bit(CACHE_SET_GC_FAILURE, &c->flags))
 760                 return;
 761
 762         down_read(&c->gc_lock);
 763         trace_bcache_gc_coalesce_start(c);
 764         start_time = local_clock();
 765
 766         for (id = 0; id < BTREE_ID_NR; id++) {
 767                 int ret = c->btree_roots[id].b
 768                         ? bch_coalesce_btree(c, id)
 769                         : 0;
 770
 771                 if (ret) {
 772                         if (ret != -ESHUTDOWN)
 773                                 bch_err(c, "btree coalescing failed: %d", ret);
 774                         set_bit(CACHE_SET_GC_FAILURE, &c->flags);
 775                         return;
 776                 }
 777         }
 778
 779         bch_time_stats_update(&c->btree_coalesce_time, start_time);
 780         trace_bcache_gc_coalesce_end(c);
 781         up_read(&c->gc_lock);
 782 }
 783
 784 static int bch_gc_thread(void *arg)
 785 {
 786         struct cache_set *c = arg;
 787         struct io_clock *clock = &c->io_clock[WRITE];
 788         unsigned long last = atomic_long_read(&clock->now);
 789         unsigned last_kick = atomic_read(&c->kick_gc);
 790
 791         set_freezable();
 792
 793         while (1) {
 794                 unsigned long next = last + c->capacity / 16;
 795
 796                 while (atomic_long_read(&clock->now) < next) {
 797                         set_current_state(TASK_INTERRUPTIBLE);
 798
 799                         if (kthread_should_stop()) {
 800                                 __set_current_state(TASK_RUNNING);
 801                                 return 0;
 802                         }
 803
 804                         if (atomic_read(&c->kick_gc) != last_kick) {
 805                                 __set_current_state(TASK_RUNNING);
 806                                 break;
 807                         }
 808
 809                         bch_io_clock_schedule_timeout(clock, next);
 810                         try_to_freeze();
 811                 }
 812
 813                 last = atomic_long_read(&clock->now);
 814                 last_kick = atomic_read(&c->kick_gc);
 815
 816                 bch_gc(c);
 817                 bch_coalesce(c);
 818
 819                 debug_check_no_locks_held();
 820         }
 821
 822         return 0;
 823 }
 824
 825 void bch_gc_thread_stop(struct cache_set *c)
 826 {
 827         set_bit(CACHE_SET_GC_STOPPING, &c->flags);
 828
 829         if (!IS_ERR_OR_NULL(c->gc_thread))
 830                 kthread_stop(c->gc_thread);
 831 }
 832
 833 int bch_gc_thread_start(struct cache_set *c)
 834 {
 835         clear_bit(CACHE_SET_GC_STOPPING, &c->flags);
 836
 837         c->gc_thread = kthread_create(bch_gc_thread, c, "bcache_gc");
 838         if (IS_ERR(c->gc_thread))
 839                 return PTR_ERR(c->gc_thread);
 840
 841         wake_up_process(c->gc_thread);
 842         return 0;
 843 }
 844
 845 /* Initial GC computes bucket marks during startup */
 846
 847 static void bch_initial_gc_btree(struct cache_set *c, enum btree_id id)
 848 {
 849         struct btree_iter iter;
 850         struct btree *b;
 851         struct range_checks r;
 852
 853         btree_node_range_checks_init(&r, 0);
 854
 855         if (!c->btree_roots[id].b)
 856                 return;
 857
 858         /*
 859          * We have to hit every btree node before starting journal replay, in
 860          * order for the journal seq blacklist machinery to work:
 861          */
 862         for_each_btree_node(&iter, c, id, POS_MIN, 0, b) {
 863                 btree_node_range_checks(c, b, &r);
 864
 865                 if (btree_node_has_ptrs(b)) {
 866                         struct btree_node_iter node_iter;
 867                         struct bkey unpacked;
 868                         struct bkey_s_c k;
 869
 870                         for_each_btree_node_key_unpack(b, k, &node_iter,
 871                                                        btree_node_is_extents(b),
 872                                                        &unpacked)
 873                                 bch_btree_mark_key_initial(c, btree_node_type(b), k);
 874                 }
 875
 876                 bch_btree_iter_cond_resched(&iter);
 877         }
 878
 879         bch_btree_iter_unlock(&iter);
 880
 881         bch_btree_mark_key(c, BKEY_TYPE_BTREE,
 882                            bkey_i_to_s_c(&c->btree_roots[id].b->key));
 883 }
 884
 885 int bch_initial_gc(struct cache_set *c, struct list_head *journal)
 886 {
 887         enum btree_id id;
 888
 889         if (journal) {
 890                 for (id = 0; id < BTREE_ID_NR; id++)
 891                         bch_initial_gc_btree(c, id);
 892
 893                 bch_journal_mark(c, journal);
 894         }
 895
 896         /*
 897          * Skip past versions that might have possibly been used (as nonces),
 898          * but hadn't had their pointers written:
 899          */
 900         if (c->sb.encryption_type)
 901                 atomic64_add(1 << 16, &c->key_version);
 902
 903         bch_mark_metadata(c);
 904
 905         gc_pos_set(c, gc_phase(GC_PHASE_DONE));
 906         set_bit(CACHE_SET_INITIAL_GC_DONE, &c->flags);
 907
 908         return 0;
 909 }