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