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_interior.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/preempt.h>
  28 #include <linux/rcupdate.h>
  29 #include <trace/events/bcachefs.h>
  30
  31 struct range_checks {
  32         struct range_level {
  33                 struct bpos     min;
  34                 struct bpos     max;
  35         }                       l[BTREE_MAX_DEPTH];
  36         unsigned                depth;
  37 };
  38
  39 static void btree_node_range_checks_init(struct range_checks *r, unsigned depth)
  40 {
  41         unsigned i;
  42
  43         for (i = 0; i < BTREE_MAX_DEPTH; i++)
  44                 r->l[i].min = r->l[i].max = POS_MIN;
  45         r->depth = depth;
  46 }
  47
  48 static void btree_node_range_checks(struct bch_fs *c, struct btree *b,
  49                                     struct range_checks *r)
  50 {
  51         struct range_level *l = &r->l[b->level];
  52
  53         struct bpos expected_min = bkey_cmp(l->min, l->max)
  54                 ? btree_type_successor(b->btree_id, l->max)
  55                 : l->max;
  56
  57         bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, expected_min), c,
  58                 "btree node has incorrect min key: %llu:%llu != %llu:%llu",
  59                 b->data->min_key.inode,
  60                 b->data->min_key.offset,
  61                 expected_min.inode,
  62                 expected_min.offset);
  63
  64         l->max = b->data->max_key;
  65
  66         if (b->level > r->depth) {
  67                 l = &r->l[b->level - 1];
  68
  69                 bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, l->min), c,
  70                         "btree node min doesn't match min of child nodes: %llu:%llu != %llu:%llu",
  71                         b->data->min_key.inode,
  72                         b->data->min_key.offset,
  73                         l->min.inode,
  74                         l->min.offset);
  75
  76                 bch2_fs_inconsistent_on(bkey_cmp(b->data->max_key, l->max), c,
  77                         "btree node max doesn't match max of child nodes: %llu:%llu != %llu:%llu",
  78                         b->data->max_key.inode,
  79                         b->data->max_key.offset,
  80                         l->max.inode,
  81                         l->max.offset);
  82
  83                 if (bkey_cmp(b->data->max_key, POS_MAX))
  84                         l->min = l->max =
  85                                 btree_type_successor(b->btree_id,
  86                                                      b->data->max_key);
  87         }
  88 }
  89
  90 u8 bch2_btree_key_recalc_oldest_gen(struct bch_fs *c, struct bkey_s_c k)
  91 {
  92         const struct bch_extent_ptr *ptr;
  93         u8 max_stale = 0;
  94
  95         if (bkey_extent_is_data(k.k)) {
  96                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
  97
  98                 extent_for_each_ptr(e, ptr) {
  99                         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 100                         size_t b = PTR_BUCKET_NR(ca, ptr);
 101
 102                         if (gen_after(ca->oldest_gens[b], ptr->gen))
 103                                 ca->oldest_gens[b] = ptr->gen;
 104
 105                         max_stale = max(max_stale, ptr_stale(ca, ptr));
 106                 }
 107         }
 108
 109         return max_stale;
 110 }
 111
 112 /*
 113  * For runtime mark and sweep:
 114  */
 115 static u8 bch2_gc_mark_key(struct bch_fs *c, enum bkey_type type,
 116                            struct bkey_s_c k, unsigned flags)
 117 {
 118         struct gc_pos pos = { 0 };
 119         struct bch_fs_usage *stats;
 120         u8 ret = 0;
 121
 122         preempt_disable();
 123         stats = this_cpu_ptr(c->usage_percpu);
 124         switch (type) {
 125         case BKEY_TYPE_BTREE:
 126                 bch2_mark_key(c, k, c->opts.btree_node_size, true, pos, stats,
 127                               0, flags|
 128                               BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
 129                               BCH_BUCKET_MARK_GC_LOCK_HELD);
 130                 break;
 131         case BKEY_TYPE_EXTENTS:
 132                 bch2_mark_key(c, k, k.k->size, false, pos, stats,
 133                               0, flags|
 134                               BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
 135                               BCH_BUCKET_MARK_GC_LOCK_HELD);
 136                 ret = bch2_btree_key_recalc_oldest_gen(c, k);
 137                 break;
 138         default:
 139                 BUG();
 140         }
 141         preempt_enable();
 142
 143         return ret;
 144 }
 145
 146 int bch2_btree_mark_key_initial(struct bch_fs *c, enum bkey_type type,
 147                                 struct bkey_s_c k)
 148 {
 149         enum bch_data_type data_type = type == BKEY_TYPE_BTREE
 150                 ? BCH_DATA_BTREE : BCH_DATA_USER;
 151         int ret = 0;
 152
 153         switch (k.k->type) {
 154         case BCH_EXTENT:
 155         case BCH_EXTENT_CACHED: {
 156                 struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
 157                 const struct bch_extent_ptr *ptr;
 158
 159                 if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
 160                     (!c->opts.nofsck &&
 161                      fsck_err_on(!bch2_sb_has_replicas(c, e, data_type), c,
 162                                  "superblock not marked as containing replicas (type %u)",
 163                                  data_type))) {
 164                         ret = bch2_check_mark_super(c, e, data_type);
 165                         if (ret)
 166                                 return ret;
 167                 }
 168
 169                 extent_for_each_ptr(e, ptr) {
 170                         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 171                         struct bucket *g = PTR_BUCKET(ca, ptr);
 172
 173                         if (mustfix_fsck_err_on(!g->mark.gen_valid, c,
 174                                         "found ptr with missing gen in alloc btree,\n"
 175                                         "type %s gen %u",
 176                                         bch2_data_types[data_type],
 177                                         ptr->gen)) {
 178                                 g->_mark.gen = ptr->gen;
 179                                 g->_mark.gen_valid = 1;
 180                                 set_bit(g - ca->buckets, ca->bucket_dirty);
 181                         }
 182
 183                         if (mustfix_fsck_err_on(gen_cmp(ptr->gen, g->mark.gen) > 0, c,
 184                                         "%s ptr gen in the future: %u > %u",
 185                                         bch2_data_types[data_type],
 186                                         ptr->gen, g->mark.gen)) {
 187                                 g->_mark.gen = ptr->gen;
 188                                 g->_mark.gen_valid = 1;
 189                                 set_bit(g - ca->buckets, ca->bucket_dirty);
 190                                 set_bit(BCH_FS_FIXED_GENS, &c->flags);
 191                         }
 192
 193                 }
 194                 break;
 195         }
 196         }
 197
 198
 199         atomic64_set(&c->key_version,
 200                      max_t(u64, k.k->version.lo,
 201                            atomic64_read(&c->key_version)));
 202
 203         bch2_gc_mark_key(c, type, k, BCH_BUCKET_MARK_NOATOMIC);
 204 fsck_err:
 205         return ret;
 206 }
 207
 208 static unsigned btree_gc_mark_node(struct bch_fs *c, struct btree *b)
 209 {
 210         enum bkey_type type = btree_node_type(b);
 211         struct btree_node_iter iter;
 212         struct bkey unpacked;
 213         struct bkey_s_c k;
 214         u8 stale = 0;
 215
 216         if (btree_node_has_ptrs(b))
 217                 for_each_btree_node_key_unpack(b, k, &iter,
 218                                                btree_node_is_extents(b),
 219                                                &unpacked) {
 220                         bch2_bkey_debugcheck(c, b, k);
 221                         stale = max(stale, bch2_gc_mark_key(c, type, k, 0));
 222                 }
 223
 224         return stale;
 225 }
 226
 227 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
 228 {
 229         write_seqcount_begin(&c->gc_pos_lock);
 230         c->gc_pos = new_pos;
 231         write_seqcount_end(&c->gc_pos_lock);
 232 }
 233
 234 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
 235 {
 236         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
 237         __gc_pos_set(c, new_pos);
 238 }
 239
 240 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id)
 241 {
 242         struct btree_iter iter;
 243         struct btree *b;
 244         struct range_checks r;
 245         unsigned depth = btree_id == BTREE_ID_EXTENTS ? 0 : 1;
 246         unsigned max_stale;
 247         int ret = 0;
 248
 249         /*
 250          * if expensive_debug_checks is on, run range_checks on all leaf nodes:
 251          */
 252         if (expensive_debug_checks(c))
 253                 depth = 0;
 254
 255         btree_node_range_checks_init(&r, depth);
 256
 257         __for_each_btree_node(&iter, c, btree_id, POS_MIN,
 258                               0, depth, BTREE_ITER_PREFETCH, b) {
 259                 btree_node_range_checks(c, b, &r);
 260
 261                 bch2_verify_btree_nr_keys(b);
 262
 263                 max_stale = btree_gc_mark_node(c, b);
 264
 265                 gc_pos_set(c, gc_pos_btree_node(b));
 266
 267                 if (max_stale > 32)
 268                         bch2_btree_node_rewrite(c, &iter,
 269                                         b->data->keys.seq,
 270                                         BTREE_INSERT_USE_RESERVE|
 271                                         BTREE_INSERT_GC_LOCK_HELD);
 272                 else if (!btree_gc_rewrite_disabled(c) &&
 273                          (btree_gc_always_rewrite(c) || max_stale > 16))
 274                         bch2_btree_node_rewrite(c, &iter,
 275                                         b->data->keys.seq,
 276                                         BTREE_INSERT_NOWAIT|
 277                                         BTREE_INSERT_GC_LOCK_HELD);
 278
 279                 bch2_btree_iter_cond_resched(&iter);
 280         }
 281         ret = bch2_btree_iter_unlock(&iter);
 282         if (ret)
 283                 return ret;
 284
 285         mutex_lock(&c->btree_root_lock);
 286
 287         b = c->btree_roots[btree_id].b;
 288         bch2_gc_mark_key(c, BKEY_TYPE_BTREE, bkey_i_to_s_c(&b->key), 0);
 289         gc_pos_set(c, gc_pos_btree_root(b->btree_id));
 290
 291         mutex_unlock(&c->btree_root_lock);
 292         return 0;
 293 }
 294
 295 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
 296                                   u64 start, u64 end,
 297                                   enum bucket_data_type type,
 298                                   unsigned flags)
 299 {
 300         u64 b = sector_to_bucket(ca, start);
 301
 302         do {
 303                 bch2_mark_metadata_bucket(c, ca, ca->buckets + b, type,
 304                                           gc_phase(GC_PHASE_SB), flags);
 305                 b++;
 306         } while (b < sector_to_bucket(ca, end));
 307 }
 308
 309 void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
 310                               unsigned flags)
 311 {
 312         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
 313         unsigned i;
 314         u64 b;
 315
 316         lockdep_assert_held(&c->sb_lock);
 317
 318         for (i = 0; i < layout->nr_superblocks; i++) {
 319                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
 320
 321                 if (offset == BCH_SB_SECTOR)
 322                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
 323                                               BUCKET_SB, flags);
 324
 325                 mark_metadata_sectors(c, ca, offset,
 326                                       offset + (1 << layout->sb_max_size_bits),
 327                                       BUCKET_SB, flags);
 328         }
 329
 330         spin_lock(&c->journal.lock);
 331
 332         for (i = 0; i < ca->journal.nr; i++) {
 333                 b = ca->journal.buckets[i];
 334                 bch2_mark_metadata_bucket(c, ca, ca->buckets + b,
 335                                           BUCKET_JOURNAL,
 336                                           gc_phase(GC_PHASE_SB), flags);
 337         }
 338
 339         spin_unlock(&c->journal.lock);
 340 }
 341
 342 static void bch2_mark_superblocks(struct bch_fs *c)
 343 {
 344         struct bch_dev *ca;
 345         unsigned i;
 346
 347         mutex_lock(&c->sb_lock);
 348         gc_pos_set(c, gc_phase(GC_PHASE_SB));
 349
 350         for_each_online_member(ca, c, i)
 351                 bch2_mark_dev_superblock(c, ca,
 352                                          BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
 353                                          BCH_BUCKET_MARK_GC_LOCK_HELD);
 354         mutex_unlock(&c->sb_lock);
 355 }
 356
 357 /* Also see bch2_pending_btree_node_free_insert_done() */
 358 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
 359 {
 360         struct gc_pos pos = { 0 };
 361         struct bch_fs_usage stats = { 0 };
 362         struct btree_update *as;
 363         struct pending_btree_node_free *d;
 364
 365         mutex_lock(&c->btree_interior_update_lock);
 366         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
 367
 368         for_each_pending_btree_node_free(c, as, d)
 369                 if (d->index_update_done)
 370                         bch2_mark_key(c, bkey_i_to_s_c(&d->key),
 371                                       c->opts.btree_node_size, true, pos,
 372                                       &stats, 0,
 373                                       BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
 374                                       BCH_BUCKET_MARK_GC_LOCK_HELD);
 375         /*
 376          * Don't apply stats - pending deletes aren't tracked in
 377          * bch_alloc_stats:
 378          */
 379
 380         mutex_unlock(&c->btree_interior_update_lock);
 381 }
 382
 383 static void bch2_mark_allocator_buckets(struct bch_fs *c)
 384 {
 385         struct bch_dev *ca;
 386         struct open_bucket *ob;
 387         size_t i, j, iter;
 388         unsigned ci;
 389
 390         spin_lock(&c->freelist_lock);
 391         gc_pos_set(c, gc_pos_alloc(c, NULL));
 392
 393         for_each_member_device(ca, c, ci) {
 394                 fifo_for_each_entry(i, &ca->free_inc, iter)
 395                         bch2_mark_alloc_bucket(c, ca, &ca->buckets[i], true,
 396                                                gc_pos_alloc(c, NULL),
 397                                                BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
 398                                                BCH_BUCKET_MARK_GC_LOCK_HELD);
 399
 400
 401
 402                 for (j = 0; j < RESERVE_NR; j++)
 403                         fifo_for_each_entry(i, &ca->free[j], iter)
 404                                 bch2_mark_alloc_bucket(c, ca, &ca->buckets[i], true,
 405                                                        gc_pos_alloc(c, NULL),
 406                                                        BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
 407                                                        BCH_BUCKET_MARK_GC_LOCK_HELD);
 408         }
 409
 410         spin_unlock(&c->freelist_lock);
 411
 412         for (ob = c->open_buckets;
 413              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
 414              ob++) {
 415                 spin_lock(&ob->lock);
 416                 if (ob->valid) {
 417                         gc_pos_set(c, gc_pos_alloc(c, ob));
 418                         ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 419                         bch2_mark_alloc_bucket(c, ca, PTR_BUCKET(ca, &ob->ptr), true,
 420                                                gc_pos_alloc(c, ob),
 421                                                BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
 422                                                BCH_BUCKET_MARK_GC_LOCK_HELD);
 423                 }
 424                 spin_unlock(&ob->lock);
 425         }
 426 }
 427
 428 static void bch2_gc_start(struct bch_fs *c)
 429 {
 430         struct bch_dev *ca;
 431         struct bucket *g;
 432         struct bucket_mark new;
 433         unsigned i;
 434         int cpu;
 435
 436         lg_global_lock(&c->usage_lock);
 437
 438         /*
 439          * Indicates to buckets code that gc is now in progress - done under
 440          * usage_lock to avoid racing with bch2_mark_key():
 441          */
 442         __gc_pos_set(c, GC_POS_MIN);
 443
 444         /* Save a copy of the existing bucket stats while we recompute them: */
 445         for_each_member_device(ca, c, i) {
 446                 ca->usage_cached = __bch2_dev_usage_read(ca);
 447                 for_each_possible_cpu(cpu) {
 448                         struct bch_dev_usage *p =
 449                                 per_cpu_ptr(ca->usage_percpu, cpu);
 450                         memset(p, 0, sizeof(*p));
 451                 }
 452         }
 453
 454         c->usage_cached = __bch2_fs_usage_read(c);
 455         for_each_possible_cpu(cpu) {
 456                 struct bch_fs_usage *p =
 457                         per_cpu_ptr(c->usage_percpu, cpu);
 458
 459                 memset(p->s, 0, sizeof(p->s));
 460         }
 461
 462         lg_global_unlock(&c->usage_lock);
 463
 464         /* Clear bucket marks: */
 465         for_each_member_device(ca, c, i)
 466                 for_each_bucket(g, ca) {
 467                         bucket_cmpxchg(g, new, ({
 468                                 new.owned_by_allocator  = 0;
 469                                 new.data_type           = 0;
 470                                 new.cached_sectors      = 0;
 471                                 new.dirty_sectors       = 0;
 472                         }));
 473                         ca->oldest_gens[g - ca->buckets] = new.gen;
 474                 }
 475 }
 476
 477 /**
 478  * bch_gc - recompute bucket marks and oldest_gen, rewrite btree nodes
 479  */
 480 void bch2_gc(struct bch_fs *c)
 481 {
 482         struct bch_dev *ca;
 483         u64 start_time = local_clock();
 484         unsigned i;
 485
 486         /*
 487          * Walk _all_ references to buckets, and recompute them:
 488          *
 489          * Order matters here:
 490          *  - Concurrent GC relies on the fact that we have a total ordering for
 491          *    everything that GC walks - see  gc_will_visit_node(),
 492          *    gc_will_visit_root()
 493          *
 494          *  - also, references move around in the course of index updates and
 495          *    various other crap: everything needs to agree on the ordering
 496          *    references are allowed to move around in - e.g., we're allowed to
 497          *    start with a reference owned by an open_bucket (the allocator) and
 498          *    move it to the btree, but not the reverse.
 499          *
 500          *    This is necessary to ensure that gc doesn't miss references that
 501          *    move around - if references move backwards in the ordering GC
 502          *    uses, GC could skip past them
 503          */
 504         trace_gc_start(c);
 505
 506         /*
 507          * Do this before taking gc_lock - bch2_disk_reservation_get() blocks on
 508          * gc_lock if sectors_available goes to 0:
 509          */
 510         bch2_recalc_sectors_available(c);
 511
 512         down_write(&c->gc_lock);
 513         if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
 514                 goto out;
 515
 516         bch2_gc_start(c);
 517
 518         /* Walk btree: */
 519         while (c->gc_pos.phase < (int) BTREE_ID_NR) {
 520                 int ret = c->btree_roots[c->gc_pos.phase].b
 521                         ? bch2_gc_btree(c, (int) c->gc_pos.phase)
 522                         : 0;
 523
 524                 if (ret) {
 525                         bch_err(c, "btree gc failed: %d", ret);
 526                         set_bit(BCH_FS_GC_FAILURE, &c->flags);
 527                         goto out;
 528                 }
 529
 530                 gc_pos_set(c, gc_phase(c->gc_pos.phase + 1));
 531         }
 532
 533         bch2_mark_superblocks(c);
 534         bch2_mark_pending_btree_node_frees(c);
 535         bch2_mark_allocator_buckets(c);
 536
 537         for_each_member_device(ca, c, i)
 538                 atomic_long_set(&ca->saturated_count, 0);
 539
 540         /* Indicates that gc is no longer in progress: */
 541         gc_pos_set(c, gc_phase(GC_PHASE_DONE));
 542         c->gc_count++;
 543 out:
 544         up_write(&c->gc_lock);
 545         trace_gc_end(c);
 546         bch2_time_stats_update(&c->btree_gc_time, start_time);
 547
 548         /*
 549          * Wake up allocator in case it was waiting for buckets
 550          * because of not being able to inc gens
 551          */
 552         for_each_member_device(ca, c, i)
 553                 bch2_wake_allocator(ca);
 554
 555         /*
 556          * At startup, allocations can happen directly instead of via the
 557          * allocator thread - issue wakeup in case they blocked on gc_lock:
 558          */
 559         closure_wake_up(&c->freelist_wait);
 560 }
 561
 562 /* Btree coalescing */
 563
 564 static void recalc_packed_keys(struct btree *b)
 565 {
 566         struct bkey_packed *k;
 567
 568         memset(&b->nr, 0, sizeof(b->nr));
 569
 570         BUG_ON(b->nsets != 1);
 571
 572         for (k =  btree_bkey_first(b, b->set);
 573              k != btree_bkey_last(b, b->set);
 574              k = bkey_next(k))
 575                 btree_keys_account_key_add(&b->nr, 0, k);
 576 }
 577
 578 static void bch2_coalesce_nodes(struct bch_fs *c, struct btree_iter *iter,
 579                                 struct btree *old_nodes[GC_MERGE_NODES])
 580 {
 581         struct btree *parent = iter->nodes[old_nodes[0]->level + 1];
 582         unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
 583         unsigned blocks = btree_blocks(c) * 2 / 3;
 584         struct btree *new_nodes[GC_MERGE_NODES];
 585         struct btree_update *as;
 586         struct keylist keylist;
 587         struct bkey_format_state format_state;
 588         struct bkey_format new_format;
 589
 590         memset(new_nodes, 0, sizeof(new_nodes));
 591         bch2_keylist_init(&keylist, NULL);
 592
 593         /* Count keys that are not deleted */
 594         for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
 595                 u64s += old_nodes[i]->nr.live_u64s;
 596
 597         nr_old_nodes = nr_new_nodes = i;
 598
 599         /* Check if all keys in @old_nodes could fit in one fewer node */
 600         if (nr_old_nodes <= 1 ||
 601             __vstruct_blocks(struct btree_node, c->block_bits,
 602                              DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
 603                 return;
 604
 605         /* Find a format that all keys in @old_nodes can pack into */
 606         bch2_bkey_format_init(&format_state);
 607
 608         for (i = 0; i < nr_old_nodes; i++)
 609                 __bch2_btree_calc_format(&format_state, old_nodes[i]);
 610
 611         new_format = bch2_bkey_format_done(&format_state);
 612
 613         /* Check if repacking would make any nodes too big to fit */
 614         for (i = 0; i < nr_old_nodes; i++)
 615                 if (!bch2_btree_node_format_fits(c, old_nodes[i], &new_format)) {
 616                         trace_btree_gc_coalesce_fail(c,
 617                                         BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
 618                         return;
 619                 }
 620
 621         if (bch2_keylist_realloc(&keylist, NULL, 0,
 622                         (BKEY_U64s + BKEY_EXTENT_U64s_MAX) * nr_old_nodes)) {
 623                 trace_btree_gc_coalesce_fail(c,
 624                                 BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
 625                 return;
 626         }
 627
 628         as = bch2_btree_update_start(c, iter->btree_id,
 629                         btree_update_reserve_required(c, parent) + nr_old_nodes,
 630                         BTREE_INSERT_NOFAIL|
 631                         BTREE_INSERT_USE_RESERVE,
 632                         NULL);
 633         if (IS_ERR(as)) {
 634                 trace_btree_gc_coalesce_fail(c,
 635                                 BTREE_GC_COALESCE_FAIL_RESERVE_GET);
 636                 bch2_keylist_free(&keylist, NULL);
 637                 return;
 638         }
 639
 640         trace_btree_gc_coalesce(c, old_nodes[0]);
 641
 642         for (i = 0; i < nr_old_nodes; i++)
 643                 bch2_btree_interior_update_will_free_node(as, old_nodes[i]);
 644
 645         /* Repack everything with @new_format and sort down to one bset */
 646         for (i = 0; i < nr_old_nodes; i++)
 647                 new_nodes[i] =
 648                         __bch2_btree_node_alloc_replacement(as, old_nodes[i],
 649                                                             new_format);
 650
 651         /*
 652          * Conceptually we concatenate the nodes together and slice them
 653          * up at different boundaries.
 654          */
 655         for (i = nr_new_nodes - 1; i > 0; --i) {
 656                 struct btree *n1 = new_nodes[i];
 657                 struct btree *n2 = new_nodes[i - 1];
 658
 659                 struct bset *s1 = btree_bset_first(n1);
 660                 struct bset *s2 = btree_bset_first(n2);
 661                 struct bkey_packed *k, *last = NULL;
 662
 663                 /* Calculate how many keys from @n2 we could fit inside @n1 */
 664                 u64s = 0;
 665
 666                 for (k = s2->start;
 667                      k < vstruct_last(s2) &&
 668                      vstruct_blocks_plus(n1->data, c->block_bits,
 669                                          u64s + k->u64s) <= blocks;
 670                      k = bkey_next(k)) {
 671                         last = k;
 672                         u64s += k->u64s;
 673                 }
 674
 675                 if (u64s == le16_to_cpu(s2->u64s)) {
 676                         /* n2 fits entirely in n1 */
 677                         n1->key.k.p = n1->data->max_key = n2->data->max_key;
 678
 679                         memcpy_u64s(vstruct_last(s1),
 680                                     s2->start,
 681                                     le16_to_cpu(s2->u64s));
 682                         le16_add_cpu(&s1->u64s, le16_to_cpu(s2->u64s));
 683
 684                         set_btree_bset_end(n1, n1->set);
 685
 686                         six_unlock_write(&n2->lock);
 687                         bch2_btree_node_free_never_inserted(c, n2);
 688                         six_unlock_intent(&n2->lock);
 689
 690                         memmove(new_nodes + i - 1,
 691                                 new_nodes + i,
 692                                 sizeof(new_nodes[0]) * (nr_new_nodes - i));
 693                         new_nodes[--nr_new_nodes] = NULL;
 694                 } else if (u64s) {
 695                         /* move part of n2 into n1 */
 696                         n1->key.k.p = n1->data->max_key =
 697                                 bkey_unpack_pos(n1, last);
 698
 699                         n2->data->min_key =
 700                                 btree_type_successor(iter->btree_id,
 701                                                      n1->data->max_key);
 702
 703                         memcpy_u64s(vstruct_last(s1),
 704                                     s2->start, u64s);
 705                         le16_add_cpu(&s1->u64s, u64s);
 706
 707                         memmove(s2->start,
 708                                 vstruct_idx(s2, u64s),
 709                                 (le16_to_cpu(s2->u64s) - u64s) * sizeof(u64));
 710                         s2->u64s = cpu_to_le16(le16_to_cpu(s2->u64s) - u64s);
 711
 712                         set_btree_bset_end(n1, n1->set);
 713                         set_btree_bset_end(n2, n2->set);
 714                 }
 715         }
 716
 717         for (i = 0; i < nr_new_nodes; i++) {
 718                 struct btree *n = new_nodes[i];
 719
 720                 recalc_packed_keys(n);
 721                 btree_node_reset_sib_u64s(n);
 722
 723                 bch2_btree_build_aux_trees(n);
 724                 six_unlock_write(&n->lock);
 725
 726                 bch2_btree_node_write(c, n, &as->cl, SIX_LOCK_intent);
 727         }
 728
 729         /*
 730          * The keys for the old nodes get deleted. We don't want to insert keys
 731          * that compare equal to the keys for the new nodes we'll also be
 732          * inserting - we can't because keys on a keylist must be strictly
 733          * greater than the previous keys, and we also don't need to since the
 734          * key for the new node will serve the same purpose (overwriting the key
 735          * for the old node).
 736          */
 737         for (i = 0; i < nr_old_nodes; i++) {
 738                 struct bkey_i delete;
 739                 unsigned j;
 740
 741                 for (j = 0; j < nr_new_nodes; j++)
 742                         if (!bkey_cmp(old_nodes[i]->key.k.p,
 743                                       new_nodes[j]->key.k.p))
 744                                 goto next;
 745
 746                 bkey_init(&delete.k);
 747                 delete.k.p = old_nodes[i]->key.k.p;
 748                 bch2_keylist_add_in_order(&keylist, &delete);
 749 next:
 750                 i = i;
 751         }
 752
 753         /*
 754          * Keys for the new nodes get inserted: bch2_btree_insert_keys() only
 755          * does the lookup once and thus expects the keys to be in sorted order
 756          * so we have to make sure the new keys are correctly ordered with
 757          * respect to the deleted keys added in the previous loop
 758          */
 759         for (i = 0; i < nr_new_nodes; i++)
 760                 bch2_keylist_add_in_order(&keylist, &new_nodes[i]->key);
 761
 762         /* Insert the newly coalesced nodes */
 763         bch2_btree_insert_node(as, parent, iter, &keylist);
 764
 765         BUG_ON(!bch2_keylist_empty(&keylist));
 766
 767         BUG_ON(iter->nodes[old_nodes[0]->level] != old_nodes[0]);
 768
 769         BUG_ON(!bch2_btree_iter_node_replace(iter, new_nodes[0]));
 770
 771         for (i = 0; i < nr_new_nodes; i++)
 772                 bch2_btree_open_bucket_put(c, new_nodes[i]);
 773
 774         /* Free the old nodes and update our sliding window */
 775         for (i = 0; i < nr_old_nodes; i++) {
 776                 bch2_btree_node_free_inmem(c, old_nodes[i], iter);
 777                 six_unlock_intent(&old_nodes[i]->lock);
 778
 779                 /*
 780                  * the index update might have triggered a split, in which case
 781                  * the nodes we coalesced - the new nodes we just created -
 782                  * might not be sibling nodes anymore - don't add them to the
 783                  * sliding window (except the first):
 784                  */
 785                 if (!i) {
 786                         old_nodes[i] = new_nodes[i];
 787                 } else {
 788                         old_nodes[i] = NULL;
 789                         if (new_nodes[i])
 790                                 six_unlock_intent(&new_nodes[i]->lock);
 791                 }
 792         }
 793
 794         bch2_btree_update_done(as);
 795         bch2_keylist_free(&keylist, NULL);
 796 }
 797
 798 static int bch2_coalesce_btree(struct bch_fs *c, enum btree_id btree_id)
 799 {
 800         struct btree_iter iter;
 801         struct btree *b;
 802         unsigned i;
 803
 804         /* Sliding window of adjacent btree nodes */
 805         struct btree *merge[GC_MERGE_NODES];
 806         u32 lock_seq[GC_MERGE_NODES];
 807
 808         /*
 809          * XXX: We don't have a good way of positively matching on sibling nodes
 810          * that have the same parent - this code works by handling the cases
 811          * where they might not have the same parent, and is thus fragile. Ugh.
 812          *
 813          * Perhaps redo this to use multiple linked iterators?
 814          */
 815         memset(merge, 0, sizeof(merge));
 816
 817         __for_each_btree_node(&iter, c, btree_id, POS_MIN,
 818                               BTREE_MAX_DEPTH, 0,
 819                               BTREE_ITER_PREFETCH, b) {
 820                 memmove(merge + 1, merge,
 821                         sizeof(merge) - sizeof(merge[0]));
 822                 memmove(lock_seq + 1, lock_seq,
 823                         sizeof(lock_seq) - sizeof(lock_seq[0]));
 824
 825                 merge[0] = b;
 826
 827                 for (i = 1; i < GC_MERGE_NODES; i++) {
 828                         if (!merge[i] ||
 829                             !six_relock_intent(&merge[i]->lock, lock_seq[i]))
 830                                 break;
 831
 832                         if (merge[i]->level != merge[0]->level) {
 833                                 six_unlock_intent(&merge[i]->lock);
 834                                 break;
 835                         }
 836                 }
 837                 memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));
 838
 839                 bch2_coalesce_nodes(c, &iter, merge);
 840
 841                 for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
 842                         lock_seq[i] = merge[i]->lock.state.seq;
 843                         six_unlock_intent(&merge[i]->lock);
 844                 }
 845
 846                 lock_seq[0] = merge[0]->lock.state.seq;
 847
 848                 if (test_bit(BCH_FS_GC_STOPPING, &c->flags)) {
 849                         bch2_btree_iter_unlock(&iter);
 850                         return -ESHUTDOWN;
 851                 }
 852
 853                 bch2_btree_iter_cond_resched(&iter);
 854
 855                 /*
 856                  * If the parent node wasn't relocked, it might have been split
 857                  * and the nodes in our sliding window might not have the same
 858                  * parent anymore - blow away the sliding window:
 859                  */
 860                 if (iter.nodes[iter.level + 1] &&
 861                     !btree_node_intent_locked(&iter, iter.level + 1))
 862                         memset(merge + 1, 0,
 863                                (GC_MERGE_NODES - 1) * sizeof(merge[0]));
 864         }
 865         return bch2_btree_iter_unlock(&iter);
 866 }
 867
 868 /**
 869  * bch_coalesce - coalesce adjacent nodes with low occupancy
 870  */
 871 void bch2_coalesce(struct bch_fs *c)
 872 {
 873         enum btree_id id;
 874
 875         if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
 876                 return;
 877
 878         down_read(&c->gc_lock);
 879         trace_gc_coalesce_start(c);
 880
 881         for (id = 0; id < BTREE_ID_NR; id++) {
 882                 int ret = c->btree_roots[id].b
 883                         ? bch2_coalesce_btree(c, id)
 884                         : 0;
 885
 886                 if (ret) {
 887                         if (ret != -ESHUTDOWN)
 888                                 bch_err(c, "btree coalescing failed: %d", ret);
 889                         set_bit(BCH_FS_GC_FAILURE, &c->flags);
 890                         return;
 891                 }
 892         }
 893
 894         trace_gc_coalesce_end(c);
 895         up_read(&c->gc_lock);
 896 }
 897
 898 static int bch2_gc_thread(void *arg)
 899 {
 900         struct bch_fs *c = arg;
 901         struct io_clock *clock = &c->io_clock[WRITE];
 902         unsigned long last = atomic_long_read(&clock->now);
 903         unsigned last_kick = atomic_read(&c->kick_gc);
 904
 905         set_freezable();
 906
 907         while (1) {
 908                 while (1) {
 909                         set_current_state(TASK_INTERRUPTIBLE);
 910
 911                         if (kthread_should_stop()) {
 912                                 __set_current_state(TASK_RUNNING);
 913                                 return 0;
 914                         }
 915
 916                         if (atomic_read(&c->kick_gc) != last_kick)
 917                                 break;
 918
 919                         if (c->btree_gc_periodic) {
 920                                 unsigned long next = last + c->capacity / 16;
 921
 922                                 if (atomic_long_read(&clock->now) >= next)
 923                                         break;
 924
 925                                 bch2_io_clock_schedule_timeout(clock, next);
 926                         } else {
 927                                 schedule();
 928                         }
 929
 930                         try_to_freeze();
 931                 }
 932                 __set_current_state(TASK_RUNNING);
 933
 934                 last = atomic_long_read(&clock->now);
 935                 last_kick = atomic_read(&c->kick_gc);
 936
 937                 bch2_gc(c);
 938
 939                 debug_check_no_locks_held();
 940         }
 941
 942         return 0;
 943 }
 944
 945 void bch2_gc_thread_stop(struct bch_fs *c)
 946 {
 947         set_bit(BCH_FS_GC_STOPPING, &c->flags);
 948
 949         if (c->gc_thread)
 950                 kthread_stop(c->gc_thread);
 951
 952         c->gc_thread = NULL;
 953         clear_bit(BCH_FS_GC_STOPPING, &c->flags);
 954 }
 955
 956 int bch2_gc_thread_start(struct bch_fs *c)
 957 {
 958         struct task_struct *p;
 959
 960         BUG_ON(c->gc_thread);
 961
 962         p = kthread_create(bch2_gc_thread, c, "bcache_gc");
 963         if (IS_ERR(p))
 964                 return PTR_ERR(p);
 965
 966         c->gc_thread = p;
 967         wake_up_process(c->gc_thread);
 968         return 0;
 969 }
 970
 971 /* Initial GC computes bucket marks during startup */
 972
 973 static int bch2_initial_gc_btree(struct bch_fs *c, enum btree_id id)
 974 {
 975         struct btree_iter iter;
 976         struct btree *b;
 977         struct range_checks r;
 978         int ret = 0;
 979
 980         btree_node_range_checks_init(&r, 0);
 981
 982         if (!c->btree_roots[id].b)
 983                 return 0;
 984
 985         ret = bch2_btree_mark_key_initial(c, BKEY_TYPE_BTREE,
 986                            bkey_i_to_s_c(&c->btree_roots[id].b->key));
 987         if (ret)
 988                 return ret;
 989
 990         /*
 991          * We have to hit every btree node before starting journal replay, in
 992          * order for the journal seq blacklist machinery to work:
 993          */
 994         for_each_btree_node(&iter, c, id, POS_MIN, BTREE_ITER_PREFETCH, b) {
 995                 btree_node_range_checks(c, b, &r);
 996
 997                 if (btree_node_has_ptrs(b)) {
 998                         struct btree_node_iter node_iter;
 999                         struct bkey unpacked;
1000                         struct bkey_s_c k;
1001
1002                         for_each_btree_node_key_unpack(b, k, &node_iter,
1003                                                        btree_node_is_extents(b),
1004                                                        &unpacked) {
1005                                 ret = bch2_btree_mark_key_initial(c,
1006                                                         btree_node_type(b), k);
1007                                 if (ret)
1008                                         goto err;
1009                         }
1010                 }
1011
1012                 bch2_btree_iter_cond_resched(&iter);
1013         }
1014 err:
1015         return bch2_btree_iter_unlock(&iter) ?: ret;
1016 }
1017
1018 int bch2_initial_gc(struct bch_fs *c, struct list_head *journal)
1019 {
1020         unsigned iter = 0;
1021         enum btree_id id;
1022         int ret;
1023
1024         mutex_lock(&c->sb_lock);
1025         if (!bch2_sb_get_replicas(c->disk_sb)) {
1026                 if (BCH_SB_INITIALIZED(c->disk_sb))
1027                         bch_info(c, "building replicas info");
1028                 set_bit(BCH_FS_REBUILD_REPLICAS, &c->flags);
1029         }
1030         mutex_unlock(&c->sb_lock);
1031 again:
1032         bch2_gc_start(c);
1033
1034         for (id = 0; id < BTREE_ID_NR; id++) {
1035                 ret = bch2_initial_gc_btree(c, id);
1036                 if (ret)
1037                         return ret;
1038         }
1039
1040         ret = bch2_journal_mark(c, journal);
1041         if (ret)
1042         return ret;
1043
1044         if (test_bit(BCH_FS_FIXED_GENS, &c->flags)) {
1045                 if (iter++ > 2) {
1046                         bch_info(c, "Unable to fix bucket gens, looping");
1047                         return -EINVAL;
1048                 }
1049
1050                 bch_info(c, "Fixed gens, restarting initial mark and sweep:");
1051                 clear_bit(BCH_FS_FIXED_GENS, &c->flags);
1052                 goto again;
1053         }
1054
1055         /*
1056          * Skip past versions that might have possibly been used (as nonces),
1057          * but hadn't had their pointers written:
1058          */
1059         if (c->sb.encryption_type)
1060                 atomic64_add(1 << 16, &c->key_version);
1061
1062         bch2_mark_superblocks(c);
1063
1064         gc_pos_set(c, gc_phase(GC_PHASE_DONE));
1065         set_bit(BCH_FS_INITIAL_GC_DONE, &c->flags);
1066
1067         return 0;
1068 }