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 int bch2_btree_mark_key_initial(struct bch_fs *c, enum bkey_type type,
 130                                 struct bkey_s_c k)
 131 {
 132         int ret = 0;
 133
 134         switch (k.k->type) {
 135         case BCH_EXTENT:
 136         case BCH_EXTENT_CACHED: {
 137                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
 138                 const struct bch_extent_ptr *ptr;
 139
 140                 extent_for_each_ptr(e, ptr) {
 141                         struct bch_dev *ca = c->devs[ptr->dev];
 142                         struct bucket *g = PTR_BUCKET(ca, ptr);
 143                         struct bucket_mark new;
 144
 145                         if (fsck_err_on(gen_cmp(ptr->gen, g->mark.gen) > 0, c,
 146                                         "%s ptr gen in the future: %u > %u",
 147                                         type == BKEY_TYPE_BTREE
 148                                         ? "btree" : "data",
 149                                         ptr->gen, g->mark.gen)) {
 150                                 bucket_cmpxchg(g, new, new.gen = ptr->gen);
 151                                 set_bit(BCH_FS_FIXED_GENS, &c->flags);
 152                                 ca->need_prio_write = true;
 153                         }
 154
 155                 }
 156                 break;
 157         }
 158         }
 159
 160         atomic64_set(&c->key_version,
 161                      max_t(u64, k.k->version.lo,
 162                            atomic64_read(&c->key_version)));
 163
 164         bch2_btree_mark_key(c, type, k);
 165 fsck_err:
 166         return ret;
 167 }
 168
 169 static bool btree_gc_mark_node(struct bch_fs *c, struct btree *b)
 170 {
 171         if (btree_node_has_ptrs(b)) {
 172                 struct btree_node_iter iter;
 173                 struct bkey unpacked;
 174                 struct bkey_s_c k;
 175                 u8 stale = 0;
 176
 177                 for_each_btree_node_key_unpack(b, k, &iter,
 178                                                btree_node_is_extents(b),
 179                                                &unpacked) {
 180                         bch2_bkey_debugcheck(c, b, k);
 181                         stale = max(stale, bch2_btree_mark_key(c,
 182                                                         btree_node_type(b), k));
 183                 }
 184
 185                 if (btree_gc_rewrite_disabled(c))
 186                         return false;
 187
 188                 if (stale > 10)
 189                         return true;
 190         }
 191
 192         if (btree_gc_always_rewrite(c))
 193                 return true;
 194
 195         return false;
 196 }
 197
 198 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
 199 {
 200         write_seqcount_begin(&c->gc_pos_lock);
 201         c->gc_pos = new_pos;
 202         write_seqcount_end(&c->gc_pos_lock);
 203 }
 204
 205 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
 206 {
 207         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
 208         __gc_pos_set(c, new_pos);
 209 }
 210
 211 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id)
 212 {
 213         struct btree_iter iter;
 214         struct btree *b;
 215         bool should_rewrite;
 216         struct range_checks r;
 217         unsigned depth = btree_id == BTREE_ID_EXTENTS ? 0 : 1;
 218         int ret;
 219
 220         /*
 221          * if expensive_debug_checks is on, run range_checks on all leaf nodes:
 222          */
 223         if (expensive_debug_checks(c))
 224                 depth = 0;
 225
 226         btree_node_range_checks_init(&r, depth);
 227
 228         __for_each_btree_node(&iter, c, btree_id, POS_MIN,
 229                               0, depth, BTREE_ITER_PREFETCH, b) {
 230                 btree_node_range_checks(c, b, &r);
 231
 232                 bch2_verify_btree_nr_keys(b);
 233
 234                 should_rewrite = btree_gc_mark_node(c, b);
 235
 236                 gc_pos_set(c, gc_pos_btree_node(b));
 237
 238                 if (should_rewrite)
 239                         bch2_btree_node_rewrite(&iter, b, NULL);
 240
 241                 bch2_btree_iter_cond_resched(&iter);
 242         }
 243         ret = bch2_btree_iter_unlock(&iter);
 244         if (ret)
 245                 return ret;
 246
 247         mutex_lock(&c->btree_root_lock);
 248
 249         b = c->btree_roots[btree_id].b;
 250         bch2_btree_mark_key(c, BKEY_TYPE_BTREE, bkey_i_to_s_c(&b->key));
 251         gc_pos_set(c, gc_pos_btree_root(b->btree_id));
 252
 253         mutex_unlock(&c->btree_root_lock);
 254         return 0;
 255 }
 256
 257 static void bch2_mark_allocator_buckets(struct bch_fs *c)
 258 {
 259         struct bch_dev *ca;
 260         struct open_bucket *ob;
 261         size_t i, j, iter;
 262         unsigned ci;
 263
 264         for_each_member_device(ca, c, ci) {
 265                 spin_lock(&ca->freelist_lock);
 266
 267                 fifo_for_each_entry(i, &ca->free_inc, iter)
 268                         bch2_mark_alloc_bucket(ca, &ca->buckets[i], true);
 269
 270                 for (j = 0; j < RESERVE_NR; j++)
 271                         fifo_for_each_entry(i, &ca->free[j], iter)
 272                                 bch2_mark_alloc_bucket(ca, &ca->buckets[i], true);
 273
 274                 spin_unlock(&ca->freelist_lock);
 275         }
 276
 277         for (ob = c->open_buckets;
 278              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
 279              ob++) {
 280                 const struct bch_extent_ptr *ptr;
 281
 282                 mutex_lock(&ob->lock);
 283                 open_bucket_for_each_ptr(ob, ptr) {
 284                         ca = c->devs[ptr->dev];
 285                         bch2_mark_alloc_bucket(ca, PTR_BUCKET(ca, ptr), true);
 286                 }
 287                 mutex_unlock(&ob->lock);
 288         }
 289 }
 290
 291 static void mark_metadata_sectors(struct bch_dev *ca, u64 start, u64 end,
 292                                   enum bucket_data_type type)
 293 {
 294         u64 b = start >> ca->bucket_bits;
 295
 296         do {
 297                 bch2_mark_metadata_bucket(ca, ca->buckets + b, type, true);
 298                 b++;
 299         } while (b < end >> ca->bucket_bits);
 300 }
 301
 302 static void bch2_dev_mark_superblocks(struct bch_dev *ca)
 303 {
 304         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
 305         unsigned i;
 306
 307         for (i = 0; i < layout->nr_superblocks; i++) {
 308                 if (layout->sb_offset[i] == BCH_SB_SECTOR)
 309                         mark_metadata_sectors(ca, 0, BCH_SB_SECTOR,
 310                                               BUCKET_SB);
 311
 312                 mark_metadata_sectors(ca,
 313                                       layout->sb_offset[i],
 314                                       layout->sb_offset[i] +
 315                                       (1 << layout->sb_max_size_bits),
 316                                       BUCKET_SB);
 317         }
 318 }
 319
 320 /*
 321  * Mark non btree metadata - prios, journal
 322  */
 323 void bch2_mark_dev_metadata(struct bch_fs *c, struct bch_dev *ca)
 324 {
 325         unsigned i;
 326         u64 b;
 327
 328         lockdep_assert_held(&c->sb_lock);
 329
 330         bch2_dev_mark_superblocks(ca);
 331
 332         spin_lock(&c->journal.lock);
 333
 334         for (i = 0; i < ca->journal.nr; i++) {
 335                 b = ca->journal.buckets[i];
 336                 bch2_mark_metadata_bucket(ca, ca->buckets + b,
 337                                          BUCKET_JOURNAL, true);
 338         }
 339
 340         spin_unlock(&c->journal.lock);
 341
 342         spin_lock(&ca->prio_buckets_lock);
 343
 344         for (i = 0; i < prio_buckets(ca) * 2; i++) {
 345                 b = ca->prio_buckets[i];
 346                 if (b)
 347                         bch2_mark_metadata_bucket(ca, ca->buckets + b,
 348                                                  BUCKET_PRIOS, true);
 349         }
 350
 351         spin_unlock(&ca->prio_buckets_lock);
 352 }
 353
 354 static void bch2_mark_metadata(struct bch_fs *c)
 355 {
 356         struct bch_dev *ca;
 357         unsigned i;
 358
 359         mutex_lock(&c->sb_lock);
 360         gc_pos_set(c, gc_phase(GC_PHASE_SB_METADATA));
 361
 362         for_each_online_member(ca, c, i)
 363                 bch2_mark_dev_metadata(c, ca);
 364         mutex_unlock(&c->sb_lock);
 365 }
 366
 367 /* Also see bch2_pending_btree_node_free_insert_done() */
 368 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
 369 {
 370         struct bch_fs_usage stats = { 0 };
 371         struct btree_interior_update *as;
 372         struct pending_btree_node_free *d;
 373
 374         mutex_lock(&c->btree_interior_update_lock);
 375         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
 376
 377         for_each_pending_btree_node_free(c, as, d)
 378                 if (d->index_update_done)
 379                         __bch2_gc_mark_key(c, bkey_i_to_s_c(&d->key),
 380                                           c->sb.btree_node_size, true,
 381                                           &stats);
 382         /*
 383          * Don't apply stats - pending deletes aren't tracked in
 384          * bch_alloc_stats:
 385          */
 386
 387         mutex_unlock(&c->btree_interior_update_lock);
 388 }
 389
 390 void bch2_gc_start(struct bch_fs *c)
 391 {
 392         struct bch_dev *ca;
 393         struct bucket *g;
 394         struct bucket_mark new;
 395         unsigned i;
 396         int cpu;
 397
 398         lg_global_lock(&c->usage_lock);
 399
 400         /*
 401          * Indicates to buckets code that gc is now in progress - done under
 402          * usage_lock to avoid racing with bch2_mark_key():
 403          */
 404         __gc_pos_set(c, GC_POS_MIN);
 405
 406         /* Save a copy of the existing bucket stats while we recompute them: */
 407         for_each_member_device(ca, c, i) {
 408                 ca->usage_cached = __bch2_dev_usage_read(ca);
 409                 for_each_possible_cpu(cpu) {
 410                         struct bch_dev_usage *p =
 411                                 per_cpu_ptr(ca->usage_percpu, cpu);
 412                         memset(p, 0, sizeof(*p));
 413                 }
 414         }
 415
 416         c->usage_cached = __bch2_fs_usage_read(c);
 417         for_each_possible_cpu(cpu) {
 418                 struct bch_fs_usage *p =
 419                         per_cpu_ptr(c->usage_percpu, cpu);
 420
 421                 memset(p->s, 0, sizeof(p->s));
 422                 p->persistent_reserved = 0;
 423         }
 424
 425         lg_global_unlock(&c->usage_lock);
 426
 427         /* Clear bucket marks: */
 428         for_each_member_device(ca, c, i)
 429                 for_each_bucket(g, ca) {
 430                         bucket_cmpxchg(g, new, ({
 431                                 new.owned_by_allocator  = 0;
 432                                 new.data_type           = 0;
 433                                 new.cached_sectors      = 0;
 434                                 new.dirty_sectors       = 0;
 435                         }));
 436                         ca->oldest_gens[g - ca->buckets] = new.gen;
 437                 }
 438 }
 439
 440 /**
 441  * bch_gc - recompute bucket marks and oldest_gen, rewrite btree nodes
 442  */
 443 void bch2_gc(struct bch_fs *c)
 444 {
 445         struct bch_dev *ca;
 446         u64 start_time = local_clock();
 447         unsigned i;
 448
 449         /*
 450          * Walk _all_ references to buckets, and recompute them:
 451          *
 452          * Order matters here:
 453          *  - Concurrent GC relies on the fact that we have a total ordering for
 454          *    everything that GC walks - see  gc_will_visit_node(),
 455          *    gc_will_visit_root()
 456          *
 457          *  - also, references move around in the course of index updates and
 458          *    various other crap: everything needs to agree on the ordering
 459          *    references are allowed to move around in - e.g., we're allowed to
 460          *    start with a reference owned by an open_bucket (the allocator) and
 461          *    move it to the btree, but not the reverse.
 462          *
 463          *    This is necessary to ensure that gc doesn't miss references that
 464          *    move around - if references move backwards in the ordering GC
 465          *    uses, GC could skip past them
 466          */
 467
 468         if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
 469                 return;
 470
 471         trace_gc_start(c);
 472
 473         /*
 474          * Do this before taking gc_lock - bch2_disk_reservation_get() blocks on
 475          * gc_lock if sectors_available goes to 0:
 476          */
 477         bch2_recalc_sectors_available(c);
 478
 479         down_write(&c->gc_lock);
 480
 481         bch2_gc_start(c);
 482
 483         /* Walk allocator's references: */
 484         bch2_mark_allocator_buckets(c);
 485
 486         /* Walk btree: */
 487         while (c->gc_pos.phase < (int) BTREE_ID_NR) {
 488                 int ret = c->btree_roots[c->gc_pos.phase].b
 489                         ? bch2_gc_btree(c, (int) c->gc_pos.phase)
 490                         : 0;
 491
 492                 if (ret) {
 493                         bch_err(c, "btree gc failed: %d", ret);
 494                         set_bit(BCH_FS_GC_FAILURE, &c->flags);
 495                         up_write(&c->gc_lock);
 496                         return;
 497                 }
 498
 499                 gc_pos_set(c, gc_phase(c->gc_pos.phase + 1));
 500         }
 501
 502         bch2_mark_metadata(c);
 503         bch2_mark_pending_btree_node_frees(c);
 504
 505         for_each_member_device(ca, c, i)
 506                 atomic_long_set(&ca->saturated_count, 0);
 507
 508         /* Indicates that gc is no longer in progress: */
 509         gc_pos_set(c, gc_phase(GC_PHASE_DONE));
 510
 511         up_write(&c->gc_lock);
 512         trace_gc_end(c);
 513         bch2_time_stats_update(&c->btree_gc_time, start_time);
 514
 515         /*
 516          * Wake up allocator in case it was waiting for buckets
 517          * because of not being able to inc gens
 518          */
 519         for_each_member_device(ca, c, i)
 520                 bch2_wake_allocator(ca);
 521 }
 522
 523 /* Btree coalescing */
 524
 525 static void recalc_packed_keys(struct btree *b)
 526 {
 527         struct bkey_packed *k;
 528
 529         memset(&b->nr, 0, sizeof(b->nr));
 530
 531         BUG_ON(b->nsets != 1);
 532
 533         for (k =  btree_bkey_first(b, b->set);
 534              k != btree_bkey_last(b, b->set);
 535              k = bkey_next(k))
 536                 btree_keys_account_key_add(&b->nr, 0, k);
 537 }
 538
 539 static void bch2_coalesce_nodes(struct btree *old_nodes[GC_MERGE_NODES],
 540                                 struct btree_iter *iter)
 541 {
 542         struct btree *parent = iter->nodes[old_nodes[0]->level + 1];
 543         struct bch_fs *c = iter->c;
 544         unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
 545         unsigned blocks = btree_blocks(c) * 2 / 3;
 546         struct btree *new_nodes[GC_MERGE_NODES];
 547         struct btree_interior_update *as;
 548         struct btree_reserve *res;
 549         struct keylist keylist;
 550         struct bkey_format_state format_state;
 551         struct bkey_format new_format;
 552
 553         memset(new_nodes, 0, sizeof(new_nodes));
 554         bch2_keylist_init(&keylist, NULL, 0);
 555
 556         /* Count keys that are not deleted */
 557         for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
 558                 u64s += old_nodes[i]->nr.live_u64s;
 559
 560         nr_old_nodes = nr_new_nodes = i;
 561
 562         /* Check if all keys in @old_nodes could fit in one fewer node */
 563         if (nr_old_nodes <= 1 ||
 564             __vstruct_blocks(struct btree_node, c->block_bits,
 565                              DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
 566                 return;
 567
 568         res = bch2_btree_reserve_get(c, parent, nr_old_nodes,
 569                                     BTREE_INSERT_NOFAIL|
 570                                     BTREE_INSERT_USE_RESERVE,
 571                                     NULL);
 572         if (IS_ERR(res)) {
 573                 trace_btree_gc_coalesce_fail(c,
 574                                 BTREE_GC_COALESCE_FAIL_RESERVE_GET);
 575                 return;
 576         }
 577
 578         if (bch2_keylist_realloc(&keylist, NULL, 0,
 579                         (BKEY_U64s + BKEY_EXTENT_U64s_MAX) * nr_old_nodes)) {
 580                 trace_btree_gc_coalesce_fail(c,
 581                                 BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
 582                 goto out;
 583         }
 584
 585         /* Find a format that all keys in @old_nodes can pack into */
 586         bch2_bkey_format_init(&format_state);
 587
 588         for (i = 0; i < nr_old_nodes; i++)
 589                 __bch2_btree_calc_format(&format_state, old_nodes[i]);
 590
 591         new_format = bch2_bkey_format_done(&format_state);
 592
 593         /* Check if repacking would make any nodes too big to fit */
 594         for (i = 0; i < nr_old_nodes; i++)
 595                 if (!bch2_btree_node_format_fits(c, old_nodes[i], &new_format)) {
 596                         trace_btree_gc_coalesce_fail(c,
 597                                         BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
 598                         goto out;
 599                 }
 600
 601         trace_btree_gc_coalesce(c, parent, nr_old_nodes);
 602
 603         as = bch2_btree_interior_update_alloc(c);
 604
 605         for (i = 0; i < nr_old_nodes; i++)
 606                 bch2_btree_interior_update_will_free_node(c, as, old_nodes[i]);
 607
 608         /* Repack everything with @new_format and sort down to one bset */
 609         for (i = 0; i < nr_old_nodes; i++) {
 610                 new_nodes[i] =
 611                         __bch2_btree_node_alloc_replacement(c, old_nodes[i],
 612                                                             new_format, res);
 613                 list_add(&new_nodes[i]->reachable, &as->reachable_list);
 614         }
 615
 616         /*
 617          * Conceptually we concatenate the nodes together and slice them
 618          * up at different boundaries.
 619          */
 620         for (i = nr_new_nodes - 1; i > 0; --i) {
 621                 struct btree *n1 = new_nodes[i];
 622                 struct btree *n2 = new_nodes[i - 1];
 623
 624                 struct bset *s1 = btree_bset_first(n1);
 625                 struct bset *s2 = btree_bset_first(n2);
 626                 struct bkey_packed *k, *last = NULL;
 627
 628                 /* Calculate how many keys from @n2 we could fit inside @n1 */
 629                 u64s = 0;
 630
 631                 for (k = s2->start;
 632                      k < vstruct_last(s2) &&
 633                      vstruct_blocks_plus(n1->data, c->block_bits,
 634                                          u64s + k->u64s) <= blocks;
 635                      k = bkey_next(k)) {
 636                         last = k;
 637                         u64s += k->u64s;
 638                 }
 639
 640                 if (u64s == le16_to_cpu(s2->u64s)) {
 641                         /* n2 fits entirely in n1 */
 642                         n1->key.k.p = n1->data->max_key = n2->data->max_key;
 643
 644                         memcpy_u64s(vstruct_last(s1),
 645                                     s2->start,
 646                                     le16_to_cpu(s2->u64s));
 647                         le16_add_cpu(&s1->u64s, le16_to_cpu(s2->u64s));
 648
 649                         set_btree_bset_end(n1, n1->set);
 650
 651                         list_del_init(&n2->reachable);
 652                         six_unlock_write(&n2->lock);
 653                         bch2_btree_node_free_never_inserted(c, n2);
 654                         six_unlock_intent(&n2->lock);
 655
 656                         memmove(new_nodes + i - 1,
 657                                 new_nodes + i,
 658                                 sizeof(new_nodes[0]) * (nr_new_nodes - i));
 659                         new_nodes[--nr_new_nodes] = NULL;
 660                 } else if (u64s) {
 661                         /* move part of n2 into n1 */
 662                         n1->key.k.p = n1->data->max_key =
 663                                 bkey_unpack_pos(n1, last);
 664
 665                         n2->data->min_key =
 666                                 btree_type_successor(iter->btree_id,
 667                                                      n1->data->max_key);
 668
 669                         memcpy_u64s(vstruct_last(s1),
 670                                     s2->start, u64s);
 671                         le16_add_cpu(&s1->u64s, u64s);
 672
 673                         memmove(s2->start,
 674                                 vstruct_idx(s2, u64s),
 675                                 (le16_to_cpu(s2->u64s) - u64s) * sizeof(u64));
 676                         s2->u64s = cpu_to_le16(le16_to_cpu(s2->u64s) - u64s);
 677
 678                         set_btree_bset_end(n1, n1->set);
 679                         set_btree_bset_end(n2, n2->set);
 680                 }
 681         }
 682
 683         for (i = 0; i < nr_new_nodes; i++) {
 684                 struct btree *n = new_nodes[i];
 685
 686                 recalc_packed_keys(n);
 687                 btree_node_reset_sib_u64s(n);
 688
 689                 bch2_btree_build_aux_trees(n);
 690                 six_unlock_write(&n->lock);
 691
 692                 bch2_btree_node_write(c, n, &as->cl, SIX_LOCK_intent);
 693         }
 694
 695         /*
 696          * The keys for the old nodes get deleted. We don't want to insert keys
 697          * that compare equal to the keys for the new nodes we'll also be
 698          * inserting - we can't because keys on a keylist must be strictly
 699          * greater than the previous keys, and we also don't need to since the
 700          * key for the new node will serve the same purpose (overwriting the key
 701          * for the old node).
 702          */
 703         for (i = 0; i < nr_old_nodes; i++) {
 704                 struct bkey_i delete;
 705                 unsigned j;
 706
 707                 for (j = 0; j < nr_new_nodes; j++)
 708                         if (!bkey_cmp(old_nodes[i]->key.k.p,
 709                                       new_nodes[j]->key.k.p))
 710                                 goto next;
 711
 712                 bkey_init(&delete.k);
 713                 delete.k.p = old_nodes[i]->key.k.p;
 714                 bch2_keylist_add_in_order(&keylist, &delete);
 715 next:
 716                 i = i;
 717         }
 718
 719         /*
 720          * Keys for the new nodes get inserted: bch2_btree_insert_keys() only
 721          * does the lookup once and thus expects the keys to be in sorted order
 722          * so we have to make sure the new keys are correctly ordered with
 723          * respect to the deleted keys added in the previous loop
 724          */
 725         for (i = 0; i < nr_new_nodes; i++)
 726                 bch2_keylist_add_in_order(&keylist, &new_nodes[i]->key);
 727
 728         /* Insert the newly coalesced nodes */
 729         bch2_btree_insert_node(parent, iter, &keylist, res, as);
 730
 731         BUG_ON(!bch2_keylist_empty(&keylist));
 732
 733         BUG_ON(iter->nodes[old_nodes[0]->level] != old_nodes[0]);
 734
 735         BUG_ON(!bch2_btree_iter_node_replace(iter, new_nodes[0]));
 736
 737         for (i = 0; i < nr_new_nodes; i++)
 738                 bch2_btree_open_bucket_put(c, new_nodes[i]);
 739
 740         /* Free the old nodes and update our sliding window */
 741         for (i = 0; i < nr_old_nodes; i++) {
 742                 bch2_btree_node_free_inmem(iter, old_nodes[i]);
 743                 six_unlock_intent(&old_nodes[i]->lock);
 744
 745                 /*
 746                  * the index update might have triggered a split, in which case
 747                  * the nodes we coalesced - the new nodes we just created -
 748                  * might not be sibling nodes anymore - don't add them to the
 749                  * sliding window (except the first):
 750                  */
 751                 if (!i) {
 752                         old_nodes[i] = new_nodes[i];
 753                 } else {
 754                         old_nodes[i] = NULL;
 755                         if (new_nodes[i])
 756                                 six_unlock_intent(&new_nodes[i]->lock);
 757                 }
 758         }
 759 out:
 760         bch2_keylist_free(&keylist, NULL);
 761         bch2_btree_reserve_put(c, res);
 762 }
 763
 764 static int bch2_coalesce_btree(struct bch_fs *c, enum btree_id btree_id)
 765 {
 766         struct btree_iter iter;
 767         struct btree *b;
 768         unsigned i;
 769
 770         /* Sliding window of adjacent btree nodes */
 771         struct btree *merge[GC_MERGE_NODES];
 772         u32 lock_seq[GC_MERGE_NODES];
 773
 774         /*
 775          * XXX: We don't have a good way of positively matching on sibling nodes
 776          * that have the same parent - this code works by handling the cases
 777          * where they might not have the same parent, and is thus fragile. Ugh.
 778          *
 779          * Perhaps redo this to use multiple linked iterators?
 780          */
 781         memset(merge, 0, sizeof(merge));
 782
 783         __for_each_btree_node(&iter, c, btree_id, POS_MIN,
 784                               U8_MAX, 0, BTREE_ITER_PREFETCH, b) {
 785                 memmove(merge + 1, merge,
 786                         sizeof(merge) - sizeof(merge[0]));
 787                 memmove(lock_seq + 1, lock_seq,
 788                         sizeof(lock_seq) - sizeof(lock_seq[0]));
 789
 790                 merge[0] = b;
 791
 792                 for (i = 1; i < GC_MERGE_NODES; i++) {
 793                         if (!merge[i] ||
 794                             !six_relock_intent(&merge[i]->lock, lock_seq[i]))
 795                                 break;
 796
 797                         if (merge[i]->level != merge[0]->level) {
 798                                 six_unlock_intent(&merge[i]->lock);
 799                                 break;
 800                         }
 801                 }
 802                 memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));
 803
 804                 bch2_coalesce_nodes(merge, &iter);
 805
 806                 for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
 807                         lock_seq[i] = merge[i]->lock.state.seq;
 808                         six_unlock_intent(&merge[i]->lock);
 809                 }
 810
 811                 lock_seq[0] = merge[0]->lock.state.seq;
 812
 813                 if (test_bit(BCH_FS_GC_STOPPING, &c->flags)) {
 814                         bch2_btree_iter_unlock(&iter);
 815                         return -ESHUTDOWN;
 816                 }
 817
 818                 bch2_btree_iter_cond_resched(&iter);
 819
 820                 /*
 821                  * If the parent node wasn't relocked, it might have been split
 822                  * and the nodes in our sliding window might not have the same
 823                  * parent anymore - blow away the sliding window:
 824                  */
 825                 if (iter.nodes[iter.level + 1] &&
 826                     !btree_node_intent_locked(&iter, iter.level + 1))
 827                         memset(merge + 1, 0,
 828                                (GC_MERGE_NODES - 1) * sizeof(merge[0]));
 829         }
 830         return bch2_btree_iter_unlock(&iter);
 831 }
 832
 833 /**
 834  * bch_coalesce - coalesce adjacent nodes with low occupancy
 835  */
 836 void bch2_coalesce(struct bch_fs *c)
 837 {
 838         u64 start_time;
 839         enum btree_id id;
 840
 841         if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
 842                 return;
 843
 844         down_read(&c->gc_lock);
 845         trace_gc_coalesce_start(c);
 846         start_time = local_clock();
 847
 848         for (id = 0; id < BTREE_ID_NR; id++) {
 849                 int ret = c->btree_roots[id].b
 850                         ? bch2_coalesce_btree(c, id)
 851                         : 0;
 852
 853                 if (ret) {
 854                         if (ret != -ESHUTDOWN)
 855                                 bch_err(c, "btree coalescing failed: %d", ret);
 856                         set_bit(BCH_FS_GC_FAILURE, &c->flags);
 857                         return;
 858                 }
 859         }
 860
 861         bch2_time_stats_update(&c->btree_coalesce_time, start_time);
 862         trace_gc_coalesce_end(c);
 863         up_read(&c->gc_lock);
 864 }
 865
 866 static int bch2_gc_thread(void *arg)
 867 {
 868         struct bch_fs *c = arg;
 869         struct io_clock *clock = &c->io_clock[WRITE];
 870         unsigned long last = atomic_long_read(&clock->now);
 871         unsigned last_kick = atomic_read(&c->kick_gc);
 872
 873         set_freezable();
 874
 875         while (1) {
 876                 unsigned long next = last + c->capacity / 16;
 877
 878                 while (atomic_long_read(&clock->now) < next) {
 879                         set_current_state(TASK_INTERRUPTIBLE);
 880
 881                         if (kthread_should_stop()) {
 882                                 __set_current_state(TASK_RUNNING);
 883                                 return 0;
 884                         }
 885
 886                         if (atomic_read(&c->kick_gc) != last_kick) {
 887                                 __set_current_state(TASK_RUNNING);
 888                                 break;
 889                         }
 890
 891                         bch2_io_clock_schedule_timeout(clock, next);
 892                         try_to_freeze();
 893                 }
 894
 895                 last = atomic_long_read(&clock->now);
 896                 last_kick = atomic_read(&c->kick_gc);
 897
 898                 bch2_gc(c);
 899                 if (!btree_gc_coalesce_disabled(c))
 900                         bch2_coalesce(c);
 901
 902                 debug_check_no_locks_held();
 903         }
 904
 905         return 0;
 906 }
 907
 908 void bch2_gc_thread_stop(struct bch_fs *c)
 909 {
 910         set_bit(BCH_FS_GC_STOPPING, &c->flags);
 911
 912         if (c->gc_thread)
 913                 kthread_stop(c->gc_thread);
 914
 915         c->gc_thread = NULL;
 916         clear_bit(BCH_FS_GC_STOPPING, &c->flags);
 917 }
 918
 919 int bch2_gc_thread_start(struct bch_fs *c)
 920 {
 921         struct task_struct *p;
 922
 923         BUG_ON(c->gc_thread);
 924
 925         p = kthread_create(bch2_gc_thread, c, "bcache_gc");
 926         if (IS_ERR(p))
 927                 return PTR_ERR(p);
 928
 929         c->gc_thread = p;
 930         wake_up_process(c->gc_thread);
 931         return 0;
 932 }
 933
 934 /* Initial GC computes bucket marks during startup */
 935
 936 static int bch2_initial_gc_btree(struct bch_fs *c, enum btree_id id)
 937 {
 938         struct btree_iter iter;
 939         struct btree *b;
 940         struct range_checks r;
 941         int ret = 0;
 942
 943         btree_node_range_checks_init(&r, 0);
 944
 945         if (!c->btree_roots[id].b)
 946                 return 0;
 947
 948         ret = bch2_btree_mark_key_initial(c, BKEY_TYPE_BTREE,
 949                            bkey_i_to_s_c(&c->btree_roots[id].b->key));
 950         if (ret)
 951                 return ret;
 952
 953         /*
 954          * We have to hit every btree node before starting journal replay, in
 955          * order for the journal seq blacklist machinery to work:
 956          */
 957         for_each_btree_node(&iter, c, id, POS_MIN, BTREE_ITER_PREFETCH, b) {
 958                 btree_node_range_checks(c, b, &r);
 959
 960                 if (btree_node_has_ptrs(b)) {
 961                         struct btree_node_iter node_iter;
 962                         struct bkey unpacked;
 963                         struct bkey_s_c k;
 964
 965                         for_each_btree_node_key_unpack(b, k, &node_iter,
 966                                                        btree_node_is_extents(b),
 967                                                        &unpacked) {
 968                                 ret = bch2_btree_mark_key_initial(c,
 969                                                         btree_node_type(b), k);
 970                                 if (ret)
 971                                         goto err;
 972                         }
 973                 }
 974
 975                 bch2_btree_iter_cond_resched(&iter);
 976         }
 977 err:
 978         bch2_btree_iter_unlock(&iter);
 979         return ret;
 980 }
 981
 982 int bch2_initial_gc(struct bch_fs *c, struct list_head *journal)
 983 {
 984         unsigned iter = 0;
 985         enum btree_id id;
 986         int ret;
 987 again:
 988         bch2_gc_start(c);
 989
 990         for (id = 0; id < BTREE_ID_NR; id++) {
 991                 ret = bch2_initial_gc_btree(c, id);
 992                 if (ret)
 993                         return ret;
 994         }
 995
 996         if (journal) {
 997                 ret = bch2_journal_mark(c, journal);
 998                 if (ret)
 999                         return ret;
1000         }
1001
1002         bch2_mark_metadata(c);
1003
1004         if (test_bit(BCH_FS_FIXED_GENS, &c->flags)) {
1005                 if (iter++ > 2) {
1006                         bch_info(c, "Unable to fix bucket gens, looping");
1007                         return -EINVAL;
1008                 }
1009
1010                 bch_info(c, "Fixed gens, restarting initial mark and sweep:");
1011                 clear_bit(BCH_FS_FIXED_GENS, &c->flags);
1012                 goto again;
1013         }
1014
1015         /*
1016          * Skip past versions that might have possibly been used (as nonces),
1017          * but hadn't had their pointers written:
1018          */
1019         if (c->sb.encryption_type)
1020                 atomic64_add(1 << 16, &c->key_version);
1021
1022         gc_pos_set(c, gc_phase(GC_PHASE_DONE));
1023         set_bit(BCH_FS_INITIAL_GC_DONE, &c->flags);
1024
1025         return 0;
1026 }