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

   1 /*
   2  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   3  * Copyright (C) 2014 Datera Inc.
   4  */
   5
   6 #include "bcachefs.h"
   7 #include "alloc.h"
   8 #include "bkey_methods.h"
   9 #include "btree_locking.h"
  10 #include "btree_update.h"
  11 #include "btree_io.h"
  12 #include "btree_gc.h"
  13 #include "buckets.h"
  14 #include "clock.h"
  15 #include "debug.h"
  16 #include "error.h"
  17 #include "extents.h"
  18 #include "journal.h"
  19 #include "keylist.h"
  20 #include "move.h"
  21 #include "super-io.h"
  22
  23 #include <linux/slab.h>
  24 #include <linux/bitops.h>
  25 #include <linux/freezer.h>
  26 #include <linux/kthread.h>
  27 #include <linux/rcupdate.h>
  28 #include <trace/events/bcachefs.h>
  29
  30 struct range_checks {
  31         struct range_level {
  32                 struct bpos     min;
  33                 struct bpos     max;
  34         }                       l[BTREE_MAX_DEPTH];
  35         unsigned                depth;
  36 };
  37
  38 static void btree_node_range_checks_init(struct range_checks *r, unsigned depth)
  39 {
  40         unsigned i;
  41
  42         for (i = 0; i < BTREE_MAX_DEPTH; i++)
  43                 r->l[i].min = r->l[i].max = POS_MIN;
  44         r->depth = depth;
  45 }
  46
  47 static void btree_node_range_checks(struct bch_fs *c, struct btree *b,
  48                                     struct range_checks *r)
  49 {
  50         struct range_level *l = &r->l[b->level];
  51
  52         struct bpos expected_min = bkey_cmp(l->min, l->max)
  53                 ? btree_type_successor(b->btree_id, l->max)
  54                 : l->max;
  55
  56         bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, expected_min), c,
  57                 "btree node has incorrect min key: %llu:%llu != %llu:%llu",
  58                 b->data->min_key.inode,
  59                 b->data->min_key.offset,
  60                 expected_min.inode,
  61                 expected_min.offset);
  62
  63         l->max = b->data->max_key;
  64
  65         if (b->level > r->depth) {
  66                 l = &r->l[b->level - 1];
  67
  68                 bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, l->min), c,
  69                         "btree node min doesn't match min of child nodes: %llu:%llu != %llu:%llu",
  70                         b->data->min_key.inode,
  71                         b->data->min_key.offset,
  72                         l->min.inode,
  73                         l->min.offset);
  74
  75                 bch2_fs_inconsistent_on(bkey_cmp(b->data->max_key, l->max), c,
  76                         "btree node max doesn't match max of child nodes: %llu:%llu != %llu:%llu",
  77                         b->data->max_key.inode,
  78                         b->data->max_key.offset,
  79                         l->max.inode,
  80                         l->max.offset);
  81
  82                 if (bkey_cmp(b->data->max_key, POS_MAX))
  83                         l->min = l->max =
  84                                 btree_type_successor(b->btree_id,
  85                                                      b->data->max_key);
  86         }
  87 }
  88
  89 u8 bch2_btree_key_recalc_oldest_gen(struct bch_fs *c, struct bkey_s_c k)
  90 {
  91         const struct bch_extent_ptr *ptr;
  92         u8 max_stale = 0;
  93
  94         if (bkey_extent_is_data(k.k)) {
  95                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
  96
  97                 extent_for_each_ptr(e, ptr) {
  98                         struct bch_dev *ca = c->devs[ptr->dev];
  99                         size_t b = PTR_BUCKET_NR(ca, ptr);
 100
 101                         if (__gen_after(ca->oldest_gens[b], ptr->gen))
 102                                 ca->oldest_gens[b] = ptr->gen;
 103
 104                         max_stale = max(max_stale, ptr_stale(ca, ptr));
 105                 }
 106         }
 107
 108         return max_stale;
 109 }
 110
 111 /*
 112  * For runtime mark and sweep:
 113  */
 114 static u8 bch2_btree_mark_key(struct bch_fs *c, enum bkey_type type,
 115                              struct bkey_s_c k)
 116 {
 117         switch (type) {
 118         case BKEY_TYPE_BTREE:
 119                 bch2_gc_mark_key(c, k, c->sb.btree_node_size, true);
 120                 return 0;
 121         case BKEY_TYPE_EXTENTS:
 122                 bch2_gc_mark_key(c, k, k.k->size, false);
 123                 return bch2_btree_key_recalc_oldest_gen(c, k);
 124         default:
 125                 BUG();
 126         }
 127 }
 128
 129 u8 bch2_btree_mark_key_initial(struct bch_fs *c, enum bkey_type type,
 130                                struct bkey_s_c k)
 131 {
 132         atomic64_set(&c->key_version,
 133                      max_t(u64, k.k->version.lo,
 134                            atomic64_read(&c->key_version)));
 135
 136         return bch2_btree_mark_key(c, type, k);
 137 }
 138
 139 static bool btree_gc_mark_node(struct bch_fs *c, struct btree *b)
 140 {
 141         if (btree_node_has_ptrs(b)) {
 142                 struct btree_node_iter iter;
 143                 struct bkey unpacked;
 144                 struct bkey_s_c k;
 145                 u8 stale = 0;
 146
 147                 for_each_btree_node_key_unpack(b, k, &iter,
 148                                                btree_node_is_extents(b),
 149                                                &unpacked) {
 150                         bch2_bkey_debugcheck(c, b, k);
 151                         stale = max(stale, bch2_btree_mark_key(c,
 152                                                         btree_node_type(b), k));
 153                 }
 154
 155                 if (btree_gc_rewrite_disabled(c))
 156                         return false;
 157
 158                 if (stale > 10)
 159                         return true;
 160         }
 161
 162         if (btree_gc_always_rewrite(c))
 163                 return true;
 164
 165         return false;
 166 }
 167
 168 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
 169 {
 170         write_seqcount_begin(&c->gc_pos_lock);
 171         c->gc_pos = new_pos;
 172         write_seqcount_end(&c->gc_pos_lock);
 173 }
 174
 175 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
 176 {
 177         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
 178         __gc_pos_set(c, new_pos);
 179 }
 180
 181 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id)
 182 {
 183         struct btree_iter iter;
 184         struct btree *b;
 185         bool should_rewrite;
 186         struct range_checks r;
 187         unsigned depth = btree_id == BTREE_ID_EXTENTS ? 0 : 1;
 188         int ret;
 189
 190         /*
 191          * if expensive_debug_checks is on, run range_checks on all leaf nodes:
 192          */
 193         if (expensive_debug_checks(c))
 194                 depth = 0;
 195
 196         btree_node_range_checks_init(&r, depth);
 197
 198         for_each_btree_node(&iter, c, btree_id, POS_MIN, depth, b) {
 199                 btree_node_range_checks(c, b, &r);
 200
 201                 bch2_verify_btree_nr_keys(b);
 202
 203                 should_rewrite = btree_gc_mark_node(c, b);
 204
 205                 gc_pos_set(c, gc_pos_btree_node(b));
 206
 207                 if (should_rewrite)
 208                         bch2_btree_node_rewrite(&iter, b, NULL);
 209
 210                 bch2_btree_iter_cond_resched(&iter);
 211         }
 212         ret = bch2_btree_iter_unlock(&iter);
 213         if (ret)
 214                 return ret;
 215
 216         mutex_lock(&c->btree_root_lock);
 217
 218         b = c->btree_roots[btree_id].b;
 219         bch2_btree_mark_key(c, BKEY_TYPE_BTREE, bkey_i_to_s_c(&b->key));
 220         gc_pos_set(c, gc_pos_btree_root(b->btree_id));
 221
 222         mutex_unlock(&c->btree_root_lock);
 223         return 0;
 224 }
 225
 226 static void bch2_mark_allocator_buckets(struct bch_fs *c)
 227 {
 228         struct bch_dev *ca;
 229         struct open_bucket *ob;
 230         size_t i, j, iter;
 231         unsigned ci;
 232
 233         for_each_member_device(ca, c, ci) {
 234                 spin_lock(&ca->freelist_lock);
 235
 236                 fifo_for_each_entry(i, &ca->free_inc, iter)
 237                         bch2_mark_alloc_bucket(ca, &ca->buckets[i], true);
 238
 239                 for (j = 0; j < RESERVE_NR; j++)
 240                         fifo_for_each_entry(i, &ca->free[j], iter)
 241                                 bch2_mark_alloc_bucket(ca, &ca->buckets[i], true);
 242
 243                 spin_unlock(&ca->freelist_lock);
 244         }
 245
 246         for (ob = c->open_buckets;
 247              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
 248              ob++) {
 249                 const struct bch_extent_ptr *ptr;
 250
 251                 mutex_lock(&ob->lock);
 252                 open_bucket_for_each_ptr(ob, ptr) {
 253                         ca = c->devs[ptr->dev];
 254                         bch2_mark_alloc_bucket(ca, PTR_BUCKET(ca, ptr), true);
 255                 }
 256                 mutex_unlock(&ob->lock);
 257         }
 258 }
 259
 260 static void mark_metadata_sectors(struct bch_dev *ca, u64 start, u64 end,
 261                                   enum bucket_data_type type)
 262 {
 263         u64 b = start >> ca->bucket_bits;
 264
 265         do {
 266                 bch2_mark_metadata_bucket(ca, ca->buckets + b, type, true);
 267                 b++;
 268         } while (b < end >> ca->bucket_bits);
 269 }
 270
 271 static void bch2_dev_mark_superblocks(struct bch_dev *ca)
 272 {
 273         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
 274         unsigned i;
 275
 276         for (i = 0; i < layout->nr_superblocks; i++) {
 277                 if (layout->sb_offset[i] == BCH_SB_SECTOR)
 278                         mark_metadata_sectors(ca, 0, BCH_SB_SECTOR,
 279                                               BUCKET_SB);
 280
 281                 mark_metadata_sectors(ca,
 282                                       layout->sb_offset[i],
 283                                       layout->sb_offset[i] +
 284                                       (1 << layout->sb_max_size_bits),
 285                                       BUCKET_SB);
 286         }
 287 }
 288
 289 /*
 290  * Mark non btree metadata - prios, journal
 291  */
 292 void bch2_mark_dev_metadata(struct bch_fs *c, struct bch_dev *ca)
 293 {
 294         unsigned i;
 295         u64 b;
 296
 297         lockdep_assert_held(&c->sb_lock);
 298
 299         bch2_dev_mark_superblocks(ca);
 300
 301         spin_lock(&c->journal.lock);
 302
 303         for (i = 0; i < ca->journal.nr; i++) {
 304                 b = ca->journal.buckets[i];
 305                 bch2_mark_metadata_bucket(ca, ca->buckets + b,
 306                                          BUCKET_JOURNAL, true);
 307         }
 308
 309         spin_unlock(&c->journal.lock);
 310
 311         spin_lock(&ca->prio_buckets_lock);
 312
 313         for (i = 0; i < prio_buckets(ca) * 2; i++) {
 314                 b = ca->prio_buckets[i];
 315                 if (b)
 316                         bch2_mark_metadata_bucket(ca, ca->buckets + b,
 317                                                  BUCKET_PRIOS, true);
 318         }
 319
 320         spin_unlock(&ca->prio_buckets_lock);
 321 }
 322
 323 static void bch2_mark_metadata(struct bch_fs *c)
 324 {
 325         struct bch_dev *ca;
 326         unsigned i;
 327
 328         mutex_lock(&c->sb_lock);
 329         gc_pos_set(c, gc_phase(GC_PHASE_SB_METADATA));
 330
 331         for_each_online_member(ca, c, i)
 332                 bch2_mark_dev_metadata(c, ca);
 333         mutex_unlock(&c->sb_lock);
 334 }
 335
 336 /* Also see bch2_pending_btree_node_free_insert_done() */
 337 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
 338 {
 339         struct bch_fs_usage stats = { 0 };
 340         struct btree_interior_update *as;
 341         struct pending_btree_node_free *d;
 342
 343         mutex_lock(&c->btree_interior_update_lock);
 344         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
 345
 346         for_each_pending_btree_node_free(c, as, d)
 347                 if (d->index_update_done)
 348                         __bch2_gc_mark_key(c, bkey_i_to_s_c(&d->key),
 349                                           c->sb.btree_node_size, true,
 350                                           &stats);
 351         /*
 352          * Don't apply stats - pending deletes aren't tracked in
 353          * bch_alloc_stats:
 354          */
 355
 356         mutex_unlock(&c->btree_interior_update_lock);
 357 }
 358
 359 /**
 360  * bch_gc - recompute bucket marks and oldest_gen, rewrite btree nodes
 361  */
 362 void bch2_gc(struct bch_fs *c)
 363 {
 364         struct bch_dev *ca;
 365         struct bucket *g;
 366         struct bucket_mark new;
 367         u64 start_time = local_clock();
 368         unsigned i;
 369         int cpu;
 370
 371         /*
 372          * Walk _all_ references to buckets, and recompute them:
 373          *
 374          * Order matters here:
 375          *  - Concurrent GC relies on the fact that we have a total ordering for
 376          *    everything that GC walks - see  gc_will_visit_node(),
 377          *    gc_will_visit_root()
 378          *
 379          *  - also, references move around in the course of index updates and
 380          *    various other crap: everything needs to agree on the ordering
 381          *    references are allowed to move around in - e.g., we're allowed to
 382          *    start with a reference owned by an open_bucket (the allocator) and
 383          *    move it to the btree, but not the reverse.
 384          *
 385          *    This is necessary to ensure that gc doesn't miss references that
 386          *    move around - if references move backwards in the ordering GC
 387          *    uses, GC could skip past them
 388          */
 389
 390         if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
 391                 return;
 392
 393         trace_gc_start(c);
 394
 395         /*
 396          * Do this before taking gc_lock - bch2_disk_reservation_get() blocks on
 397          * gc_lock if sectors_available goes to 0:
 398          */
 399         bch2_recalc_sectors_available(c);
 400
 401         down_write(&c->gc_lock);
 402
 403         lg_global_lock(&c->usage_lock);
 404
 405         /*
 406          * Indicates to buckets code that gc is now in progress - done under
 407          * usage_lock to avoid racing with bch2_mark_key():
 408          */
 409         __gc_pos_set(c, GC_POS_MIN);
 410
 411         /* Save a copy of the existing bucket stats while we recompute them: */
 412         for_each_member_device(ca, c, i) {
 413                 ca->usage_cached = __bch2_dev_usage_read(ca);
 414                 for_each_possible_cpu(cpu) {
 415                         struct bch_dev_usage *p =
 416                                 per_cpu_ptr(ca->usage_percpu, cpu);
 417                         memset(p, 0, sizeof(*p));
 418                 }
 419         }
 420
 421         c->usage_cached = __bch2_fs_usage_read(c);
 422         for_each_possible_cpu(cpu) {
 423                 struct bch_fs_usage *p =
 424                         per_cpu_ptr(c->usage_percpu, cpu);
 425
 426                 memset(p->s, 0, sizeof(p->s));
 427                 p->persistent_reserved = 0;
 428         }
 429
 430         lg_global_unlock(&c->usage_lock);
 431
 432         /* Clear bucket marks: */
 433         for_each_member_device(ca, c, i)
 434                 for_each_bucket(g, ca) {
 435                         bucket_cmpxchg(g, new, ({
 436                                 new.owned_by_allocator  = 0;
 437                                 new.data_type           = 0;
 438                                 new.cached_sectors      = 0;
 439                                 new.dirty_sectors       = 0;
 440                         }));
 441                         ca->oldest_gens[g - ca->buckets] = new.gen;
 442                 }
 443
 444         /* Walk allocator's references: */
 445         bch2_mark_allocator_buckets(c);
 446
 447         /* Walk btree: */
 448         while (c->gc_pos.phase < (int) BTREE_ID_NR) {
 449                 int ret = c->btree_roots[c->gc_pos.phase].b
 450                         ? bch2_gc_btree(c, (int) c->gc_pos.phase)
 451                         : 0;
 452
 453                 if (ret) {
 454                         bch_err(c, "btree gc failed: %d", ret);
 455                         set_bit(BCH_FS_GC_FAILURE, &c->flags);
 456                         up_write(&c->gc_lock);
 457                         return;
 458                 }
 459
 460                 gc_pos_set(c, gc_phase(c->gc_pos.phase + 1));
 461         }
 462
 463         bch2_mark_metadata(c);
 464         bch2_mark_pending_btree_node_frees(c);
 465
 466         for_each_member_device(ca, c, i)
 467                 atomic_long_set(&ca->saturated_count, 0);
 468
 469         /* Indicates that gc is no longer in progress: */
 470         gc_pos_set(c, gc_phase(GC_PHASE_DONE));
 471
 472         up_write(&c->gc_lock);
 473         trace_gc_end(c);
 474         bch2_time_stats_update(&c->btree_gc_time, start_time);
 475
 476         /*
 477          * Wake up allocator in case it was waiting for buckets
 478          * because of not being able to inc gens
 479          */
 480         for_each_member_device(ca, c, i)
 481                 bch2_wake_allocator(ca);
 482 }
 483
 484 /* Btree coalescing */
 485
 486 static void recalc_packed_keys(struct btree *b)
 487 {
 488         struct bkey_packed *k;
 489
 490         memset(&b->nr, 0, sizeof(b->nr));
 491
 492         BUG_ON(b->nsets != 1);
 493
 494         for (k =  btree_bkey_first(b, b->set);
 495              k != btree_bkey_last(b, b->set);
 496              k = bkey_next(k))
 497                 btree_keys_account_key_add(&b->nr, 0, k);
 498 }
 499
 500 static void bch2_coalesce_nodes(struct btree *old_nodes[GC_MERGE_NODES],
 501                                 struct btree_iter *iter)
 502 {
 503         struct btree *parent = iter->nodes[old_nodes[0]->level + 1];
 504         struct bch_fs *c = iter->c;
 505         unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
 506         unsigned blocks = btree_blocks(c) * 2 / 3;
 507         struct btree *new_nodes[GC_MERGE_NODES];
 508         struct btree_interior_update *as;
 509         struct btree_reserve *res;
 510         struct keylist keylist;
 511         struct bkey_format_state format_state;
 512         struct bkey_format new_format;
 513
 514         memset(new_nodes, 0, sizeof(new_nodes));
 515         bch2_keylist_init(&keylist, NULL, 0);
 516
 517         /* Count keys that are not deleted */
 518         for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
 519                 u64s += old_nodes[i]->nr.live_u64s;
 520
 521         nr_old_nodes = nr_new_nodes = i;
 522
 523         /* Check if all keys in @old_nodes could fit in one fewer node */
 524         if (nr_old_nodes <= 1 ||
 525             __vstruct_blocks(struct btree_node, c->block_bits,
 526                              DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
 527                 return;
 528
 529         res = bch2_btree_reserve_get(c, parent, nr_old_nodes,
 530                                     BTREE_INSERT_NOFAIL|
 531                                     BTREE_INSERT_USE_RESERVE,
 532                                     NULL);
 533         if (IS_ERR(res)) {
 534                 trace_btree_gc_coalesce_fail(c,
 535                                 BTREE_GC_COALESCE_FAIL_RESERVE_GET);
 536                 return;
 537         }
 538
 539         if (bch2_keylist_realloc(&keylist, NULL, 0,
 540                         (BKEY_U64s + BKEY_EXTENT_U64s_MAX) * nr_old_nodes)) {
 541                 trace_btree_gc_coalesce_fail(c,
 542                                 BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
 543                 goto out;
 544         }
 545
 546         /* Find a format that all keys in @old_nodes can pack into */
 547         bch2_bkey_format_init(&format_state);
 548
 549         for (i = 0; i < nr_old_nodes; i++)
 550                 __bch2_btree_calc_format(&format_state, old_nodes[i]);
 551
 552         new_format = bch2_bkey_format_done(&format_state);
 553
 554         /* Check if repacking would make any nodes too big to fit */
 555         for (i = 0; i < nr_old_nodes; i++)
 556                 if (!bch2_btree_node_format_fits(c, old_nodes[i], &new_format)) {
 557                         trace_btree_gc_coalesce_fail(c,
 558                                         BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
 559                         goto out;
 560                 }
 561
 562         trace_btree_gc_coalesce(c, parent, nr_old_nodes);
 563
 564         as = bch2_btree_interior_update_alloc(c);
 565
 566         for (i = 0; i < nr_old_nodes; i++)
 567                 bch2_btree_interior_update_will_free_node(c, as, old_nodes[i]);
 568
 569         /* Repack everything with @new_format and sort down to one bset */
 570         for (i = 0; i < nr_old_nodes; i++)
 571                 new_nodes[i] =
 572                         __bch2_btree_node_alloc_replacement(c, old_nodes[i],
 573                                                             new_format, res);
 574
 575         /*
 576          * Conceptually we concatenate the nodes together and slice them
 577          * up at different boundaries.
 578          */
 579         for (i = nr_new_nodes - 1; i > 0; --i) {
 580                 struct btree *n1 = new_nodes[i];
 581                 struct btree *n2 = new_nodes[i - 1];
 582
 583                 struct bset *s1 = btree_bset_first(n1);
 584                 struct bset *s2 = btree_bset_first(n2);
 585                 struct bkey_packed *k, *last = NULL;
 586
 587                 /* Calculate how many keys from @n2 we could fit inside @n1 */
 588                 u64s = 0;
 589
 590                 for (k = s2->start;
 591                      k < vstruct_last(s2) &&
 592                      vstruct_blocks_plus(n1->data, c->block_bits,
 593                                          u64s + k->u64s) <= blocks;
 594                      k = bkey_next(k)) {
 595                         last = k;
 596                         u64s += k->u64s;
 597                 }
 598
 599                 if (u64s == le16_to_cpu(s2->u64s)) {
 600                         /* n2 fits entirely in n1 */
 601                         n1->key.k.p = n1->data->max_key = n2->data->max_key;
 602
 603                         memcpy_u64s(vstruct_last(s1),
 604                                     s2->start,
 605                                     le16_to_cpu(s2->u64s));
 606                         le16_add_cpu(&s1->u64s, le16_to_cpu(s2->u64s));
 607
 608                         set_btree_bset_end(n1, n1->set);
 609
 610                         six_unlock_write(&n2->lock);
 611                         bch2_btree_node_free_never_inserted(c, n2);
 612                         six_unlock_intent(&n2->lock);
 613
 614                         memmove(new_nodes + i - 1,
 615                                 new_nodes + i,
 616                                 sizeof(new_nodes[0]) * (nr_new_nodes - i));
 617                         new_nodes[--nr_new_nodes] = NULL;
 618                 } else if (u64s) {
 619                         /* move part of n2 into n1 */
 620                         n1->key.k.p = n1->data->max_key =
 621                                 bkey_unpack_pos(n1, last);
 622
 623                         n2->data->min_key =
 624                                 btree_type_successor(iter->btree_id,
 625                                                      n1->data->max_key);
 626
 627                         memcpy_u64s(vstruct_last(s1),
 628                                     s2->start, u64s);
 629                         le16_add_cpu(&s1->u64s, u64s);
 630
 631                         memmove(s2->start,
 632                                 vstruct_idx(s2, u64s),
 633                                 (le16_to_cpu(s2->u64s) - u64s) * sizeof(u64));
 634                         s2->u64s = cpu_to_le16(le16_to_cpu(s2->u64s) - u64s);
 635
 636                         set_btree_bset_end(n1, n1->set);
 637                         set_btree_bset_end(n2, n2->set);
 638                 }
 639         }
 640
 641         for (i = 0; i < nr_new_nodes; i++) {
 642                 struct btree *n = new_nodes[i];
 643
 644                 recalc_packed_keys(n);
 645                 btree_node_reset_sib_u64s(n);
 646
 647                 bch2_btree_build_aux_trees(n);
 648                 six_unlock_write(&n->lock);
 649
 650                 bch2_btree_node_write(c, n, &as->cl, SIX_LOCK_intent, -1);
 651         }
 652
 653         /*
 654          * The keys for the old nodes get deleted. We don't want to insert keys
 655          * that compare equal to the keys for the new nodes we'll also be
 656          * inserting - we can't because keys on a keylist must be strictly
 657          * greater than the previous keys, and we also don't need to since the
 658          * key for the new node will serve the same purpose (overwriting the key
 659          * for the old node).
 660          */
 661         for (i = 0; i < nr_old_nodes; i++) {
 662                 struct bkey_i delete;
 663                 unsigned j;
 664
 665                 for (j = 0; j < nr_new_nodes; j++)
 666                         if (!bkey_cmp(old_nodes[i]->key.k.p,
 667                                       new_nodes[j]->key.k.p))
 668                                 goto next;
 669
 670                 bkey_init(&delete.k);
 671                 delete.k.p = old_nodes[i]->key.k.p;
 672                 bch2_keylist_add_in_order(&keylist, &delete);
 673 next:
 674                 i = i;
 675         }
 676
 677         /*
 678          * Keys for the new nodes get inserted: bch2_btree_insert_keys() only
 679          * does the lookup once and thus expects the keys to be in sorted order
 680          * so we have to make sure the new keys are correctly ordered with
 681          * respect to the deleted keys added in the previous loop
 682          */
 683         for (i = 0; i < nr_new_nodes; i++)
 684                 bch2_keylist_add_in_order(&keylist, &new_nodes[i]->key);
 685
 686         /* Insert the newly coalesced nodes */
 687         bch2_btree_insert_node(parent, iter, &keylist, res, as);
 688
 689         BUG_ON(!bch2_keylist_empty(&keylist));
 690
 691         BUG_ON(iter->nodes[old_nodes[0]->level] != old_nodes[0]);
 692
 693         BUG_ON(!bch2_btree_iter_node_replace(iter, new_nodes[0]));
 694
 695         for (i = 0; i < nr_new_nodes; i++)
 696                 bch2_btree_open_bucket_put(c, new_nodes[i]);
 697
 698         /* Free the old nodes and update our sliding window */
 699         for (i = 0; i < nr_old_nodes; i++) {
 700                 bch2_btree_node_free_inmem(iter, old_nodes[i]);
 701                 six_unlock_intent(&old_nodes[i]->lock);
 702
 703                 /*
 704                  * the index update might have triggered a split, in which case
 705                  * the nodes we coalesced - the new nodes we just created -
 706                  * might not be sibling nodes anymore - don't add them to the
 707                  * sliding window (except the first):
 708                  */
 709                 if (!i) {
 710                         old_nodes[i] = new_nodes[i];
 711                 } else {
 712                         old_nodes[i] = NULL;
 713                         if (new_nodes[i])
 714                                 six_unlock_intent(&new_nodes[i]->lock);
 715                 }
 716         }
 717 out:
 718         bch2_keylist_free(&keylist, NULL);
 719         bch2_btree_reserve_put(c, res);
 720 }
 721
 722 static int bch2_coalesce_btree(struct bch_fs *c, enum btree_id btree_id)
 723 {
 724         struct btree_iter iter;
 725         struct btree *b;
 726         unsigned i;
 727
 728         /* Sliding window of adjacent btree nodes */
 729         struct btree *merge[GC_MERGE_NODES];
 730         u32 lock_seq[GC_MERGE_NODES];
 731
 732         /*
 733          * XXX: We don't have a good way of positively matching on sibling nodes
 734          * that have the same parent - this code works by handling the cases
 735          * where they might not have the same parent, and is thus fragile. Ugh.
 736          *
 737          * Perhaps redo this to use multiple linked iterators?
 738          */
 739         memset(merge, 0, sizeof(merge));
 740
 741         __for_each_btree_node(&iter, c, btree_id, POS_MIN, 0, b, U8_MAX) {
 742                 memmove(merge + 1, merge,
 743                         sizeof(merge) - sizeof(merge[0]));
 744                 memmove(lock_seq + 1, lock_seq,
 745                         sizeof(lock_seq) - sizeof(lock_seq[0]));
 746
 747                 merge[0] = b;
 748
 749                 for (i = 1; i < GC_MERGE_NODES; i++) {
 750                         if (!merge[i] ||
 751                             !six_relock_intent(&merge[i]->lock, lock_seq[i]))
 752                                 break;
 753
 754                         if (merge[i]->level != merge[0]->level) {
 755                                 six_unlock_intent(&merge[i]->lock);
 756                                 break;
 757                         }
 758                 }
 759                 memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));
 760
 761                 bch2_coalesce_nodes(merge, &iter);
 762
 763                 for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
 764                         lock_seq[i] = merge[i]->lock.state.seq;
 765                         six_unlock_intent(&merge[i]->lock);
 766                 }
 767
 768                 lock_seq[0] = merge[0]->lock.state.seq;
 769
 770                 if (test_bit(BCH_FS_GC_STOPPING, &c->flags)) {
 771                         bch2_btree_iter_unlock(&iter);
 772                         return -ESHUTDOWN;
 773                 }
 774
 775                 bch2_btree_iter_cond_resched(&iter);
 776
 777                 /*
 778                  * If the parent node wasn't relocked, it might have been split
 779                  * and the nodes in our sliding window might not have the same
 780                  * parent anymore - blow away the sliding window:
 781                  */
 782                 if (iter.nodes[iter.level + 1] &&
 783                     !btree_node_intent_locked(&iter, iter.level + 1))
 784                         memset(merge + 1, 0,
 785                                (GC_MERGE_NODES - 1) * sizeof(merge[0]));
 786         }
 787         return bch2_btree_iter_unlock(&iter);
 788 }
 789
 790 /**
 791  * bch_coalesce - coalesce adjacent nodes with low occupancy
 792  */
 793 void bch2_coalesce(struct bch_fs *c)
 794 {
 795         u64 start_time;
 796         enum btree_id id;
 797
 798         if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
 799                 return;
 800
 801         down_read(&c->gc_lock);
 802         trace_gc_coalesce_start(c);
 803         start_time = local_clock();
 804
 805         for (id = 0; id < BTREE_ID_NR; id++) {
 806                 int ret = c->btree_roots[id].b
 807                         ? bch2_coalesce_btree(c, id)
 808                         : 0;
 809
 810                 if (ret) {
 811                         if (ret != -ESHUTDOWN)
 812                                 bch_err(c, "btree coalescing failed: %d", ret);
 813                         set_bit(BCH_FS_GC_FAILURE, &c->flags);
 814                         return;
 815                 }
 816         }
 817
 818         bch2_time_stats_update(&c->btree_coalesce_time, start_time);
 819         trace_gc_coalesce_end(c);
 820         up_read(&c->gc_lock);
 821 }
 822
 823 static int bch2_gc_thread(void *arg)
 824 {
 825         struct bch_fs *c = arg;
 826         struct io_clock *clock = &c->io_clock[WRITE];
 827         unsigned long last = atomic_long_read(&clock->now);
 828         unsigned last_kick = atomic_read(&c->kick_gc);
 829
 830         set_freezable();
 831
 832         while (1) {
 833                 unsigned long next = last + c->capacity / 16;
 834
 835                 while (atomic_long_read(&clock->now) < next) {
 836                         set_current_state(TASK_INTERRUPTIBLE);
 837
 838                         if (kthread_should_stop()) {
 839                                 __set_current_state(TASK_RUNNING);
 840                                 return 0;
 841                         }
 842
 843                         if (atomic_read(&c->kick_gc) != last_kick) {
 844                                 __set_current_state(TASK_RUNNING);
 845                                 break;
 846                         }
 847
 848                         bch2_io_clock_schedule_timeout(clock, next);
 849                         try_to_freeze();
 850                 }
 851
 852                 last = atomic_long_read(&clock->now);
 853                 last_kick = atomic_read(&c->kick_gc);
 854
 855                 bch2_gc(c);
 856                 if (!btree_gc_coalesce_disabled(c))
 857                         bch2_coalesce(c);
 858
 859                 debug_check_no_locks_held();
 860         }
 861
 862         return 0;
 863 }
 864
 865 void bch2_gc_thread_stop(struct bch_fs *c)
 866 {
 867         set_bit(BCH_FS_GC_STOPPING, &c->flags);
 868
 869         if (c->gc_thread)
 870                 kthread_stop(c->gc_thread);
 871
 872         c->gc_thread = NULL;
 873         clear_bit(BCH_FS_GC_STOPPING, &c->flags);
 874 }
 875
 876 int bch2_gc_thread_start(struct bch_fs *c)
 877 {
 878         struct task_struct *p;
 879
 880         BUG_ON(c->gc_thread);
 881
 882         p = kthread_create(bch2_gc_thread, c, "bcache_gc");
 883         if (IS_ERR(p))
 884                 return PTR_ERR(p);
 885
 886         c->gc_thread = p;
 887         wake_up_process(c->gc_thread);
 888         return 0;
 889 }
 890
 891 /* Initial GC computes bucket marks during startup */
 892
 893 static void bch2_initial_gc_btree(struct bch_fs *c, enum btree_id id)
 894 {
 895         struct btree_iter iter;
 896         struct btree *b;
 897         struct range_checks r;
 898
 899         btree_node_range_checks_init(&r, 0);
 900
 901         if (!c->btree_roots[id].b)
 902                 return;
 903
 904         /*
 905          * We have to hit every btree node before starting journal replay, in
 906          * order for the journal seq blacklist machinery to work:
 907          */
 908         for_each_btree_node(&iter, c, id, POS_MIN, 0, b) {
 909                 btree_node_range_checks(c, b, &r);
 910
 911                 if (btree_node_has_ptrs(b)) {
 912                         struct btree_node_iter node_iter;
 913                         struct bkey unpacked;
 914                         struct bkey_s_c k;
 915
 916                         for_each_btree_node_key_unpack(b, k, &node_iter,
 917                                                        btree_node_is_extents(b),
 918                                                        &unpacked)
 919                                 bch2_btree_mark_key_initial(c, btree_node_type(b), k);
 920                 }
 921
 922                 bch2_btree_iter_cond_resched(&iter);
 923         }
 924
 925         bch2_btree_iter_unlock(&iter);
 926
 927         bch2_btree_mark_key(c, BKEY_TYPE_BTREE,
 928                            bkey_i_to_s_c(&c->btree_roots[id].b->key));
 929 }
 930
 931 int bch2_initial_gc(struct bch_fs *c, struct list_head *journal)
 932 {
 933         enum btree_id id;
 934
 935         for (id = 0; id < BTREE_ID_NR; id++)
 936                 bch2_initial_gc_btree(c, id);
 937
 938         if (journal)
 939                 bch2_journal_mark(c, journal);
 940
 941         bch2_mark_metadata(c);
 942
 943         /*
 944          * Skip past versions that might have possibly been used (as nonces),
 945          * but hadn't had their pointers written:
 946          */
 947         if (c->sb.encryption_type)
 948                 atomic64_add(1 << 16, &c->key_version);
 949
 950         gc_pos_set(c, gc_phase(GC_PHASE_DONE));
 951         set_bit(BCH_FS_INITIAL_GC_DONE, &c->flags);
 952
 953         return 0;
 954 }