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