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