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 bch_fs *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 bch_fs *c, struct bkey_s_c k)
  91 {
  92         const struct bch_extent_ptr *ptr;
  93         u8 max_stale = 0;
  94
  95         if (bkey_extent_is_data(k.k)) {
  96                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
  97
  98                 extent_for_each_ptr(e, ptr) {
  99                         struct bch_dev *ca = c->devs[ptr->dev];
 100                         size_t b = PTR_BUCKET_NR(ca, ptr);
 101
 102                         if (__gen_after(ca->oldest_gens[b], ptr->gen))
 103                                 ca->oldest_gens[b] = ptr->gen;
 104
 105                         max_stale = max(max_stale, ptr_stale(ca, ptr));
 106                 }
 107         }
 108
 109         return max_stale;
 110 }
 111
 112 /*
 113  * For runtime mark and sweep:
 114  */
 115 static u8 bch_btree_mark_key(struct bch_fs *c, enum bkey_type type,
 116                              struct bkey_s_c k)
 117 {
 118         switch (type) {
 119         case BKEY_TYPE_BTREE:
 120                 bch_gc_mark_key(c, k, c->sb.btree_node_size, true);
 121                 return 0;
 122         case BKEY_TYPE_EXTENTS:
 123                 bch_gc_mark_key(c, k, k.k->size, false);
 124                 return bch_btree_key_recalc_oldest_gen(c, k);
 125         default:
 126                 BUG();
 127         }
 128 }
 129
 130 u8 bch_btree_mark_key_initial(struct bch_fs *c, enum bkey_type type,
 131                           struct bkey_s_c k)
 132 {
 133         atomic64_set(&c->key_version,
 134                      max_t(u64, k.k->version.lo,
 135                            atomic64_read(&c->key_version)));
 136
 137         return bch_btree_mark_key(c, type, k);
 138 }
 139
 140 static bool btree_gc_mark_node(struct bch_fs *c, struct btree *b)
 141 {
 142         if (btree_node_has_ptrs(b)) {
 143                 struct btree_node_iter iter;
 144                 struct bkey unpacked;
 145                 struct bkey_s_c k;
 146                 u8 stale = 0;
 147
 148                 for_each_btree_node_key_unpack(b, k, &iter,
 149                                                btree_node_is_extents(b),
 150                                                &unpacked) {
 151                         bkey_debugcheck(c, b, k);
 152                         stale = max(stale, bch_btree_mark_key(c,
 153                                                         btree_node_type(b), k));
 154                 }
 155
 156                 if (btree_gc_rewrite_disabled(c))
 157                         return false;
 158
 159                 if (stale > 10)
 160                         return true;
 161         }
 162
 163         if (btree_gc_always_rewrite(c))
 164                 return true;
 165
 166         return false;
 167 }
 168
 169 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
 170 {
 171         write_seqcount_begin(&c->gc_pos_lock);
 172         c->gc_pos = new_pos;
 173         write_seqcount_end(&c->gc_pos_lock);
 174 }
 175
 176 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
 177 {
 178         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
 179         __gc_pos_set(c, new_pos);
 180 }
 181
 182 static int bch_gc_btree(struct bch_fs *c, enum btree_id btree_id)
 183 {
 184         struct btree_iter iter;
 185         struct btree *b;
 186         bool should_rewrite;
 187         struct range_checks r;
 188         unsigned depth = btree_id == BTREE_ID_EXTENTS ? 0 : 1;
 189         int ret;
 190
 191         /*
 192          * if expensive_debug_checks is on, run range_checks on all leaf nodes:
 193          */
 194         if (expensive_debug_checks(c))
 195                 depth = 0;
 196
 197         btree_node_range_checks_init(&r, depth);
 198
 199         for_each_btree_node(&iter, c, btree_id, POS_MIN, depth, b) {
 200                 btree_node_range_checks(c, b, &r);
 201
 202                 bch_verify_btree_nr_keys(b);
 203
 204                 should_rewrite = btree_gc_mark_node(c, b);
 205
 206                 gc_pos_set(c, gc_pos_btree_node(b));
 207
 208                 if (should_rewrite)
 209                         bch_btree_node_rewrite(&iter, b, NULL);
 210
 211                 bch_btree_iter_cond_resched(&iter);
 212         }
 213         ret = bch_btree_iter_unlock(&iter);
 214         if (ret)
 215                 return ret;
 216
 217         mutex_lock(&c->btree_root_lock);
 218
 219         b = c->btree_roots[btree_id].b;
 220         bch_btree_mark_key(c, BKEY_TYPE_BTREE, bkey_i_to_s_c(&b->key));
 221         gc_pos_set(c, gc_pos_btree_root(b->btree_id));
 222
 223         mutex_unlock(&c->btree_root_lock);
 224         return 0;
 225 }
 226
 227 static void bch_mark_allocator_buckets(struct bch_fs *c)
 228 {
 229         struct bch_dev *ca;
 230         struct open_bucket *ob;
 231         size_t i, j, iter;
 232         unsigned ci;
 233
 234         for_each_member_device(ca, c, ci) {
 235                 spin_lock(&ca->freelist_lock);
 236
 237                 fifo_for_each_entry(i, &ca->free_inc, iter)
 238                         bch_mark_alloc_bucket(ca, &ca->buckets[i], true);
 239
 240                 for (j = 0; j < RESERVE_NR; j++)
 241                         fifo_for_each_entry(i, &ca->free[j], iter)
 242                                 bch_mark_alloc_bucket(ca, &ca->buckets[i], true);
 243
 244                 spin_unlock(&ca->freelist_lock);
 245         }
 246
 247         for (ob = c->open_buckets;
 248              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
 249              ob++) {
 250                 const struct bch_extent_ptr *ptr;
 251
 252                 mutex_lock(&ob->lock);
 253                 open_bucket_for_each_ptr(ob, ptr) {
 254                         ca = c->devs[ptr->dev];
 255                         bch_mark_alloc_bucket(ca, PTR_BUCKET(ca, ptr), true);
 256                 }
 257                 mutex_unlock(&ob->lock);
 258         }
 259 }
 260
 261 static void mark_metadata_sectors(struct bch_dev *ca, u64 start, u64 end,
 262                                   enum bucket_data_type type)
 263 {
 264         u64 b = start >> ca->bucket_bits;
 265
 266         do {
 267                 bch_mark_metadata_bucket(ca, ca->buckets + b, type, true);
 268                 b++;
 269         } while (b < end >> ca->bucket_bits);
 270 }
 271
 272 static void bch_dev_mark_superblocks(struct bch_dev *ca)
 273 {
 274         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
 275         unsigned i;
 276
 277         for (i = 0; i < layout->nr_superblocks; i++) {
 278                 if (layout->sb_offset[i] == BCH_SB_SECTOR)
 279                         mark_metadata_sectors(ca, 0, BCH_SB_SECTOR,
 280                                               BUCKET_SB);
 281
 282                 mark_metadata_sectors(ca,
 283                                       layout->sb_offset[i],
 284                                       layout->sb_offset[i] +
 285                                       (1 << layout->sb_max_size_bits),
 286                                       BUCKET_SB);
 287         }
 288 }
 289
 290 /*
 291  * Mark non btree metadata - prios, journal
 292  */
 293 void bch_mark_dev_metadata(struct bch_fs *c, struct bch_dev *ca)
 294 {
 295         unsigned i;
 296         u64 b;
 297
 298         lockdep_assert_held(&c->sb_lock);
 299
 300         bch_dev_mark_superblocks(ca);
 301
 302         spin_lock(&c->journal.lock);
 303
 304         for (i = 0; i < ca->journal.nr; i++) {
 305                 b = ca->journal.buckets[i];
 306                 bch_mark_metadata_bucket(ca, ca->buckets + b,
 307                                          BUCKET_JOURNAL, true);
 308         }
 309
 310         spin_unlock(&c->journal.lock);
 311
 312         spin_lock(&ca->prio_buckets_lock);
 313
 314         for (i = 0; i < prio_buckets(ca) * 2; i++) {
 315                 b = ca->prio_buckets[i];
 316                 if (b)
 317                         bch_mark_metadata_bucket(ca, ca->buckets + b,
 318                                                  BUCKET_PRIOS, true);
 319         }
 320
 321         spin_unlock(&ca->prio_buckets_lock);
 322 }
 323
 324 static void bch_mark_metadata(struct bch_fs *c)
 325 {
 326         struct bch_dev *ca;
 327         unsigned i;
 328
 329         mutex_lock(&c->sb_lock);
 330         gc_pos_set(c, gc_phase(GC_PHASE_SB_METADATA));
 331
 332         for_each_online_member(ca, c, i)
 333                 bch_mark_dev_metadata(c, ca);
 334         mutex_unlock(&c->sb_lock);
 335 }
 336
 337 /* Also see bch_pending_btree_node_free_insert_done() */
 338 static void bch_mark_pending_btree_node_frees(struct bch_fs *c)
 339 {
 340         struct bch_fs_usage stats = { 0 };
 341         struct btree_interior_update *as;
 342         struct pending_btree_node_free *d;
 343
 344         mutex_lock(&c->btree_interior_update_lock);
 345         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
 346
 347         for_each_pending_btree_node_free(c, as, d)
 348                 if (d->index_update_done)
 349                         __bch_gc_mark_key(c, bkey_i_to_s_c(&d->key),
 350                                           c->sb.btree_node_size, true,
 351                                           &stats);
 352         /*
 353          * Don't apply stats - pending deletes aren't tracked in
 354          * bch_alloc_stats:
 355          */
 356
 357         mutex_unlock(&c->btree_interior_update_lock);
 358 }
 359
 360 /**
 361  * bch_gc - recompute bucket marks and oldest_gen, rewrite btree nodes
 362  */
 363 void bch_gc(struct bch_fs *c)
 364 {
 365         struct bch_dev *ca;
 366         struct bucket *g;
 367         struct bucket_mark new;
 368         u64 start_time = local_clock();
 369         unsigned i;
 370         int cpu;
 371
 372         /*
 373          * Walk _all_ references to buckets, and recompute them:
 374          *
 375          * Order matters here:
 376          *  - Concurrent GC relies on the fact that we have a total ordering for
 377          *    everything that GC walks - see  gc_will_visit_node(),
 378          *    gc_will_visit_root()
 379          *
 380          *  - also, references move around in the course of index updates and
 381          *    various other crap: everything needs to agree on the ordering
 382          *    references are allowed to move around in - e.g., we're allowed to
 383          *    start with a reference owned by an open_bucket (the allocator) and
 384          *    move it to the btree, but not the reverse.
 385          *
 386          *    This is necessary to ensure that gc doesn't miss references that
 387          *    move around - if references move backwards in the ordering GC
 388          *    uses, GC could skip past them
 389          */
 390
 391         if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
 392                 return;
 393
 394         trace_bcache_gc_start(c);
 395
 396         /*
 397          * Do this before taking gc_lock - bch_disk_reservation_get() blocks on
 398          * gc_lock if sectors_available goes to 0:
 399          */
 400         bch_recalc_sectors_available(c);
 401
 402         down_write(&c->gc_lock);
 403
 404         lg_global_lock(&c->usage_lock);
 405
 406         /*
 407          * Indicates to buckets code that gc is now in progress - done under
 408          * usage_lock to avoid racing with bch_mark_key():
 409          */
 410         __gc_pos_set(c, GC_POS_MIN);
 411
 412         /* Save a copy of the existing bucket stats while we recompute them: */
 413         for_each_member_device(ca, c, i) {
 414                 ca->usage_cached = __bch_dev_usage_read(ca);
 415                 for_each_possible_cpu(cpu) {
 416                         struct bch_dev_usage *p =
 417                                 per_cpu_ptr(ca->usage_percpu, cpu);
 418                         memset(p, 0, sizeof(*p));
 419                 }
 420         }
 421
 422         c->usage_cached = __bch_fs_usage_read(c);
 423         for_each_possible_cpu(cpu) {
 424                 struct bch_fs_usage *p =
 425                         per_cpu_ptr(c->usage_percpu, cpu);
 426
 427                 memset(p->s, 0, sizeof(p->s));
 428                 p->persistent_reserved = 0;
 429         }
 430
 431         lg_global_unlock(&c->usage_lock);
 432
 433         /* Clear bucket marks: */
 434         for_each_member_device(ca, c, i)
 435                 for_each_bucket(g, ca) {
 436                         bucket_cmpxchg(g, new, ({
 437                                 new.owned_by_allocator  = 0;
 438                                 new.data_type           = 0;
 439                                 new.cached_sectors      = 0;
 440                                 new.dirty_sectors       = 0;
 441                         }));
 442                         ca->oldest_gens[g - ca->buckets] = new.gen;
 443                 }
 444
 445         /* Walk allocator's references: */
 446         bch_mark_allocator_buckets(c);
 447
 448         /* Walk btree: */
 449         while (c->gc_pos.phase < (int) BTREE_ID_NR) {
 450                 int ret = c->btree_roots[c->gc_pos.phase].b
 451                         ? bch_gc_btree(c, (int) c->gc_pos.phase)
 452                         : 0;
 453
 454                 if (ret) {
 455                         bch_err(c, "btree gc failed: %d", ret);
 456                         set_bit(BCH_FS_GC_FAILURE, &c->flags);
 457                         up_write(&c->gc_lock);
 458                         return;
 459                 }
 460
 461                 gc_pos_set(c, gc_phase(c->gc_pos.phase + 1));
 462         }
 463
 464         bch_mark_metadata(c);
 465         bch_mark_pending_btree_node_frees(c);
 466         bch_writeback_recalc_oldest_gens(c);
 467
 468         for_each_member_device(ca, c, i)
 469                 atomic_long_set(&ca->saturated_count, 0);
 470
 471         /* Indicates that gc is no longer in progress: */
 472         gc_pos_set(c, gc_phase(GC_PHASE_DONE));
 473
 474         up_write(&c->gc_lock);
 475         trace_bcache_gc_end(c);
 476         bch_time_stats_update(&c->btree_gc_time, start_time);
 477
 478         /*
 479          * Wake up allocator in case it was waiting for buckets
 480          * because of not being able to inc gens
 481          */
 482         for_each_member_device(ca, c, i)
 483                 bch_wake_allocator(ca);
 484 }
 485
 486 /* Btree coalescing */
 487
 488 static void recalc_packed_keys(struct btree *b)
 489 {
 490         struct bkey_packed *k;
 491
 492         memset(&b->nr, 0, sizeof(b->nr));
 493
 494         BUG_ON(b->nsets != 1);
 495
 496         for (k =  btree_bkey_first(b, b->set);
 497              k != btree_bkey_last(b, b->set);
 498              k = bkey_next(k))
 499                 btree_keys_account_key_add(&b->nr, 0, k);
 500 }
 501
 502 static void bch_coalesce_nodes(struct btree *old_nodes[GC_MERGE_NODES],
 503                                struct btree_iter *iter)
 504 {
 505         struct btree *parent = iter->nodes[old_nodes[0]->level + 1];
 506         struct bch_fs *c = iter->c;
 507         unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
 508         unsigned blocks = btree_blocks(c) * 2 / 3;
 509         struct btree *new_nodes[GC_MERGE_NODES];
 510         struct btree_interior_update *as;
 511         struct btree_reserve *res;
 512         struct keylist keylist;
 513         struct bkey_format_state format_state;
 514         struct bkey_format new_format;
 515
 516         memset(new_nodes, 0, sizeof(new_nodes));
 517         bch_keylist_init(&keylist, NULL, 0);
 518
 519         /* Count keys that are not deleted */
 520         for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
 521                 u64s += old_nodes[i]->nr.live_u64s;
 522
 523         nr_old_nodes = nr_new_nodes = i;
 524
 525         /* Check if all keys in @old_nodes could fit in one fewer node */
 526         if (nr_old_nodes <= 1 ||
 527             __vstruct_blocks(struct btree_node, c->block_bits,
 528                              DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
 529                 return;
 530
 531         res = bch_btree_reserve_get(c, parent, nr_old_nodes,
 532                                     BTREE_INSERT_NOFAIL|
 533                                     BTREE_INSERT_USE_RESERVE,
 534                                     NULL);
 535         if (IS_ERR(res)) {
 536                 trace_bcache_btree_gc_coalesce_fail(c,
 537                                 BTREE_GC_COALESCE_FAIL_RESERVE_GET);
 538                 return;
 539         }
 540
 541         if (bch_keylist_realloc(&keylist, NULL, 0,
 542                         (BKEY_U64s + BKEY_EXTENT_U64s_MAX) * nr_old_nodes)) {
 543                 trace_bcache_btree_gc_coalesce_fail(c,
 544                                 BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
 545                 goto out;
 546         }
 547
 548         /* Find a format that all keys in @old_nodes can pack into */
 549         bch_bkey_format_init(&format_state);
 550
 551         for (i = 0; i < nr_old_nodes; i++)
 552                 __bch_btree_calc_format(&format_state, old_nodes[i]);
 553
 554         new_format = bch_bkey_format_done(&format_state);
 555
 556         /* Check if repacking would make any nodes too big to fit */
 557         for (i = 0; i < nr_old_nodes; i++)
 558                 if (!bch_btree_node_format_fits(c, old_nodes[i], &new_format)) {
 559                         trace_bcache_btree_gc_coalesce_fail(c,
 560                                         BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
 561                         goto out;
 562                 }
 563
 564         trace_bcache_btree_gc_coalesce(c, parent, nr_old_nodes);
 565
 566         as = bch_btree_interior_update_alloc(c);
 567
 568         for (i = 0; i < nr_old_nodes; i++)
 569                 bch_btree_interior_update_will_free_node(c, as, old_nodes[i]);
 570
 571         /* Repack everything with @new_format and sort down to one bset */
 572         for (i = 0; i < nr_old_nodes; i++)
 573                 new_nodes[i] = __btree_node_alloc_replacement(c, old_nodes[i],
 574                                                               new_format, res);
 575
 576         /*
 577          * Conceptually we concatenate the nodes together and slice them
 578          * up at different boundaries.
 579          */
 580         for (i = nr_new_nodes - 1; i > 0; --i) {
 581                 struct btree *n1 = new_nodes[i];
 582                 struct btree *n2 = new_nodes[i - 1];
 583
 584                 struct bset *s1 = btree_bset_first(n1);
 585                 struct bset *s2 = btree_bset_first(n2);
 586                 struct bkey_packed *k, *last = NULL;
 587
 588                 /* Calculate how many keys from @n2 we could fit inside @n1 */
 589                 u64s = 0;
 590
 591                 for (k = s2->start;
 592                      k < vstruct_last(s2) &&
 593                      vstruct_blocks_plus(n1->data, c->block_bits,
 594                                          u64s + k->u64s) <= blocks;
 595                      k = bkey_next(k)) {
 596                         last = k;
 597                         u64s += k->u64s;
 598                 }
 599
 600                 if (u64s == le16_to_cpu(s2->u64s)) {
 601                         /* n2 fits entirely in n1 */
 602                         n1->key.k.p = n1->data->max_key = n2->data->max_key;
 603
 604                         memcpy_u64s(vstruct_last(s1),
 605                                     s2->start,
 606                                     le16_to_cpu(s2->u64s));
 607                         le16_add_cpu(&s1->u64s, le16_to_cpu(s2->u64s));
 608
 609                         set_btree_bset_end(n1, n1->set);
 610
 611                         six_unlock_write(&n2->lock);
 612                         bch_btree_node_free_never_inserted(c, n2);
 613                         six_unlock_intent(&n2->lock);
 614
 615                         memmove(new_nodes + i - 1,
 616                                 new_nodes + i,
 617                                 sizeof(new_nodes[0]) * (nr_new_nodes - i));
 618                         new_nodes[--nr_new_nodes] = NULL;
 619                 } else if (u64s) {
 620                         /* move part of n2 into n1 */
 621                         n1->key.k.p = n1->data->max_key =
 622                                 bkey_unpack_pos(n1, last);
 623
 624                         n2->data->min_key =
 625                                 btree_type_successor(iter->btree_id,
 626                                                      n1->data->max_key);
 627
 628                         memcpy_u64s(vstruct_last(s1),
 629                                     s2->start, u64s);
 630                         le16_add_cpu(&s1->u64s, u64s);
 631
 632                         memmove(s2->start,
 633                                 vstruct_idx(s2, u64s),
 634                                 (le16_to_cpu(s2->u64s) - u64s) * sizeof(u64));
 635                         s2->u64s = cpu_to_le16(le16_to_cpu(s2->u64s) - u64s);
 636
 637                         set_btree_bset_end(n1, n1->set);
 638                         set_btree_bset_end(n2, n2->set);
 639                 }
 640         }
 641
 642         for (i = 0; i < nr_new_nodes; i++) {
 643                 struct btree *n = new_nodes[i];
 644
 645                 recalc_packed_keys(n);
 646                 btree_node_reset_sib_u64s(n);
 647
 648                 bch_btree_build_aux_trees(n);
 649                 six_unlock_write(&n->lock);
 650
 651                 bch_btree_node_write(c, n, &as->cl, SIX_LOCK_intent, -1);
 652         }
 653
 654         /*
 655          * The keys for the old nodes get deleted. We don't want to insert keys
 656          * that compare equal to the keys for the new nodes we'll also be
 657          * inserting - we can't because keys on a keylist must be strictly
 658          * greater than the previous keys, and we also don't need to since the
 659          * key for the new node will serve the same purpose (overwriting the key
 660          * for the old node).
 661          */
 662         for (i = 0; i < nr_old_nodes; i++) {
 663                 struct bkey_i delete;
 664                 unsigned j;
 665
 666                 for (j = 0; j < nr_new_nodes; j++)
 667                         if (!bkey_cmp(old_nodes[i]->key.k.p,
 668                                       new_nodes[j]->key.k.p))
 669                                 goto next;
 670
 671                 bkey_init(&delete.k);
 672                 delete.k.p = old_nodes[i]->key.k.p;
 673                 bch_keylist_add_in_order(&keylist, &delete);
 674 next:
 675                 i = i;
 676         }
 677
 678         /*
 679          * Keys for the new nodes get inserted: bch_btree_insert_keys() only
 680          * does the lookup once and thus expects the keys to be in sorted order
 681          * so we have to make sure the new keys are correctly ordered with
 682          * respect to the deleted keys added in the previous loop
 683          */
 684         for (i = 0; i < nr_new_nodes; i++)
 685                 bch_keylist_add_in_order(&keylist, &new_nodes[i]->key);
 686
 687         /* Insert the newly coalesced nodes */
 688         bch_btree_insert_node(parent, iter, &keylist, res, as);
 689
 690         BUG_ON(!bch_keylist_empty(&keylist));
 691
 692         BUG_ON(iter->nodes[old_nodes[0]->level] != old_nodes[0]);
 693
 694         BUG_ON(!bch_btree_iter_node_replace(iter, new_nodes[0]));
 695
 696         for (i = 0; i < nr_new_nodes; i++)
 697                 btree_open_bucket_put(c, new_nodes[i]);
 698
 699         /* Free the old nodes and update our sliding window */
 700         for (i = 0; i < nr_old_nodes; i++) {
 701                 bch_btree_node_free_inmem(iter, old_nodes[i]);
 702                 six_unlock_intent(&old_nodes[i]->lock);
 703
 704                 /*
 705                  * the index update might have triggered a split, in which case
 706                  * the nodes we coalesced - the new nodes we just created -
 707                  * might not be sibling nodes anymore - don't add them to the
 708                  * sliding window (except the first):
 709                  */
 710                 if (!i) {
 711                         old_nodes[i] = new_nodes[i];
 712                 } else {
 713                         old_nodes[i] = NULL;
 714                         if (new_nodes[i])
 715                                 six_unlock_intent(&new_nodes[i]->lock);
 716                 }
 717         }
 718 out:
 719         bch_keylist_free(&keylist, NULL);
 720         bch_btree_reserve_put(c, res);
 721 }
 722
 723 static int bch_coalesce_btree(struct bch_fs *c, enum btree_id btree_id)
 724 {
 725         struct btree_iter iter;
 726         struct btree *b;
 727         unsigned i;
 728
 729         /* Sliding window of adjacent btree nodes */
 730         struct btree *merge[GC_MERGE_NODES];
 731         u32 lock_seq[GC_MERGE_NODES];
 732
 733         /*
 734          * XXX: We don't have a good way of positively matching on sibling nodes
 735          * that have the same parent - this code works by handling the cases
 736          * where they might not have the same parent, and is thus fragile. Ugh.
 737          *
 738          * Perhaps redo this to use multiple linked iterators?
 739          */
 740         memset(merge, 0, sizeof(merge));
 741
 742         __for_each_btree_node(&iter, c, btree_id, POS_MIN, 0, b, U8_MAX) {
 743                 memmove(merge + 1, merge,
 744                         sizeof(merge) - sizeof(merge[0]));
 745                 memmove(lock_seq + 1, lock_seq,
 746                         sizeof(lock_seq) - sizeof(lock_seq[0]));
 747
 748                 merge[0] = b;
 749
 750                 for (i = 1; i < GC_MERGE_NODES; i++) {
 751                         if (!merge[i] ||
 752                             !six_relock_intent(&merge[i]->lock, lock_seq[i]))
 753                                 break;
 754
 755                         if (merge[i]->level != merge[0]->level) {
 756                                 six_unlock_intent(&merge[i]->lock);
 757                                 break;
 758                         }
 759                 }
 760                 memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));
 761
 762                 bch_coalesce_nodes(merge, &iter);
 763
 764                 for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
 765                         lock_seq[i] = merge[i]->lock.state.seq;
 766                         six_unlock_intent(&merge[i]->lock);
 767                 }
 768
 769                 lock_seq[0] = merge[0]->lock.state.seq;
 770
 771                 if (test_bit(BCH_FS_GC_STOPPING, &c->flags)) {
 772                         bch_btree_iter_unlock(&iter);
 773                         return -ESHUTDOWN;
 774                 }
 775
 776                 bch_btree_iter_cond_resched(&iter);
 777
 778                 /*
 779                  * If the parent node wasn't relocked, it might have been split
 780                  * and the nodes in our sliding window might not have the same
 781                  * parent anymore - blow away the sliding window:
 782                  */
 783                 if (iter.nodes[iter.level + 1] &&
 784                     !btree_node_intent_locked(&iter, iter.level + 1))
 785                         memset(merge + 1, 0,
 786                                (GC_MERGE_NODES - 1) * sizeof(merge[0]));
 787         }
 788         return bch_btree_iter_unlock(&iter);
 789 }
 790
 791 /**
 792  * bch_coalesce - coalesce adjacent nodes with low occupancy
 793  */
 794 void bch_coalesce(struct bch_fs *c)
 795 {
 796         u64 start_time;
 797         enum btree_id id;
 798
 799         if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
 800                 return;
 801
 802         down_read(&c->gc_lock);
 803         trace_bcache_gc_coalesce_start(c);
 804         start_time = local_clock();
 805
 806         for (id = 0; id < BTREE_ID_NR; id++) {
 807                 int ret = c->btree_roots[id].b
 808                         ? bch_coalesce_btree(c, id)
 809                         : 0;
 810
 811                 if (ret) {
 812                         if (ret != -ESHUTDOWN)
 813                                 bch_err(c, "btree coalescing failed: %d", ret);
 814                         set_bit(BCH_FS_GC_FAILURE, &c->flags);
 815                         return;
 816                 }
 817         }
 818
 819         bch_time_stats_update(&c->btree_coalesce_time, start_time);
 820         trace_bcache_gc_coalesce_end(c);
 821         up_read(&c->gc_lock);
 822 }
 823
 824 static int bch_gc_thread(void *arg)
 825 {
 826         struct bch_fs *c = arg;
 827         struct io_clock *clock = &c->io_clock[WRITE];
 828         unsigned long last = atomic_long_read(&clock->now);
 829         unsigned last_kick = atomic_read(&c->kick_gc);
 830
 831         set_freezable();
 832
 833         while (1) {
 834                 unsigned long next = last + c->capacity / 16;
 835
 836                 while (atomic_long_read(&clock->now) < next) {
 837                         set_current_state(TASK_INTERRUPTIBLE);
 838
 839                         if (kthread_should_stop()) {
 840                                 __set_current_state(TASK_RUNNING);
 841                                 return 0;
 842                         }
 843
 844                         if (atomic_read(&c->kick_gc) != last_kick) {
 845                                 __set_current_state(TASK_RUNNING);
 846                                 break;
 847                         }
 848
 849                         bch_io_clock_schedule_timeout(clock, next);
 850                         try_to_freeze();
 851                 }
 852
 853                 last = atomic_long_read(&clock->now);
 854                 last_kick = atomic_read(&c->kick_gc);
 855
 856                 bch_gc(c);
 857                 if (!btree_gc_coalesce_disabled(c))
 858                         bch_coalesce(c);
 859
 860                 debug_check_no_locks_held();
 861         }
 862
 863         return 0;
 864 }
 865
 866 void bch_gc_thread_stop(struct bch_fs *c)
 867 {
 868         set_bit(BCH_FS_GC_STOPPING, &c->flags);
 869
 870         if (c->gc_thread)
 871                 kthread_stop(c->gc_thread);
 872
 873         c->gc_thread = NULL;
 874         clear_bit(BCH_FS_GC_STOPPING, &c->flags);
 875 }
 876
 877 int bch_gc_thread_start(struct bch_fs *c)
 878 {
 879         struct task_struct *p;
 880
 881         BUG_ON(c->gc_thread);
 882
 883         p = kthread_create(bch_gc_thread, c, "bcache_gc");
 884         if (IS_ERR(p))
 885                 return PTR_ERR(p);
 886
 887         c->gc_thread = p;
 888         wake_up_process(c->gc_thread);
 889         return 0;
 890 }
 891
 892 /* Initial GC computes bucket marks during startup */
 893
 894 static void bch_initial_gc_btree(struct bch_fs *c, enum btree_id id)
 895 {
 896         struct btree_iter iter;
 897         struct btree *b;
 898         struct range_checks r;
 899
 900         btree_node_range_checks_init(&r, 0);
 901
 902         if (!c->btree_roots[id].b)
 903                 return;
 904
 905         /*
 906          * We have to hit every btree node before starting journal replay, in
 907          * order for the journal seq blacklist machinery to work:
 908          */
 909         for_each_btree_node(&iter, c, id, POS_MIN, 0, b) {
 910                 btree_node_range_checks(c, b, &r);
 911
 912                 if (btree_node_has_ptrs(b)) {
 913                         struct btree_node_iter node_iter;
 914                         struct bkey unpacked;
 915                         struct bkey_s_c k;
 916
 917                         for_each_btree_node_key_unpack(b, k, &node_iter,
 918                                                        btree_node_is_extents(b),
 919                                                        &unpacked)
 920                                 bch_btree_mark_key_initial(c, btree_node_type(b), k);
 921                 }
 922
 923                 bch_btree_iter_cond_resched(&iter);
 924         }
 925
 926         bch_btree_iter_unlock(&iter);
 927
 928         bch_btree_mark_key(c, BKEY_TYPE_BTREE,
 929                            bkey_i_to_s_c(&c->btree_roots[id].b->key));
 930 }
 931
 932 int bch_initial_gc(struct bch_fs *c, struct list_head *journal)
 933 {
 934         enum btree_id id;
 935
 936         for (id = 0; id < BTREE_ID_NR; id++)
 937                 bch_initial_gc_btree(c, id);
 938
 939         if (journal)
 940                 bch_journal_mark(c, journal);
 941
 942         bch_mark_metadata(c);
 943
 944         /*
 945          * Skip past versions that might have possibly been used (as nonces),
 946          * but hadn't had their pointers written:
 947          */
 948         if (c->sb.encryption_type)
 949                 atomic64_add(1 << 16, &c->key_version);
 950
 951         gc_pos_set(c, gc_phase(GC_PHASE_DONE));
 952         set_bit(BCH_FS_INITIAL_GC_DONE, &c->flags);
 953
 954         return 0;
 955 }