fs/bcachefs/btree_gc.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   4  * Copyright (C) 2014 Datera Inc.
   5  */
   6
   7 #include "bcachefs.h"
   8 #include "alloc_background.h"
   9 #include "alloc_foreground.h"
  10 #include "bkey_methods.h"
  11 #include "bkey_on_stack.h"
  12 #include "btree_locking.h"
  13 #include "btree_update_interior.h"
  14 #include "btree_io.h"
  15 #include "btree_gc.h"
  16 #include "buckets.h"
  17 #include "clock.h"
  18 #include "debug.h"
  19 #include "ec.h"
  20 #include "error.h"
  21 #include "extents.h"
  22 #include "journal.h"
  23 #include "keylist.h"
  24 #include "move.h"
  25 #include "recovery.h"
  26 #include "replicas.h"
  27 #include "super-io.h"
  28 #include "trace.h"
  29
  30 #include <linux/slab.h>
  31 #include <linux/bitops.h>
  32 #include <linux/freezer.h>
  33 #include <linux/kthread.h>
  34 #include <linux/preempt.h>
  35 #include <linux/rcupdate.h>
  36 #include <linux/sched/task.h>
  37
  38 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  39 {
  40         preempt_disable();
  41         write_seqcount_begin(&c->gc_pos_lock);
  42         c->gc_pos = new_pos;
  43         write_seqcount_end(&c->gc_pos_lock);
  44         preempt_enable();
  45 }
  46
  47 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  48 {
  49         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
  50         __gc_pos_set(c, new_pos);
  51 }
  52
  53 static int bch2_gc_check_topology(struct bch_fs *c,
  54                                   struct bkey_s_c k,
  55                                   struct bpos *expected_start,
  56                                   struct bpos expected_end,
  57                                   bool is_last)
  58 {
  59         int ret = 0;
  60
  61         if (k.k->type == KEY_TYPE_btree_ptr_v2) {
  62                 struct bkey_s_c_btree_ptr_v2 bp = bkey_s_c_to_btree_ptr_v2(k);
  63
  64                 if (fsck_err_on(bkey_cmp(*expected_start, bp.v->min_key), c,
  65                                 "btree node with incorrect min_key: got %llu:%llu, should be %llu:%llu",
  66                                 bp.v->min_key.inode,
  67                                 bp.v->min_key.offset,
  68                                 expected_start->inode,
  69                                 expected_start->offset)) {
  70                         BUG();
  71                 }
  72         }
  73
  74         *expected_start = bkey_cmp(k.k->p, POS_MAX)
  75                 ? bkey_successor(k.k->p)
  76                 : k.k->p;
  77
  78         if (fsck_err_on(is_last &&
  79                         bkey_cmp(k.k->p, expected_end), c,
  80                         "btree node with incorrect max_key: got %llu:%llu, should be %llu:%llu",
  81                         k.k->p.inode,
  82                         k.k->p.offset,
  83                         expected_end.inode,
  84                         expected_end.offset)) {
  85                 BUG();
  86         }
  87 fsck_err:
  88         return ret;
  89 }
  90
  91 /* marking of btree keys/nodes: */
  92
  93 static int bch2_gc_mark_key(struct bch_fs *c, struct bkey_s_c k,
  94                             u8 *max_stale, bool initial)
  95 {
  96         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
  97         const struct bch_extent_ptr *ptr;
  98         unsigned flags =
  99                 BTREE_TRIGGER_GC|
 100                 (initial ? BTREE_TRIGGER_NOATOMIC : 0);
 101         int ret = 0;
 102
 103         if (initial) {
 104                 BUG_ON(bch2_journal_seq_verify &&
 105                        k.k->version.lo > journal_cur_seq(&c->journal));
 106
 107                 /* XXX change to fsck check */
 108                 if (fsck_err_on(k.k->version.lo > atomic64_read(&c->key_version), c,
 109                                 "key version number higher than recorded: %llu > %llu",
 110                                 k.k->version.lo,
 111                                 atomic64_read(&c->key_version)))
 112                         atomic64_set(&c->key_version, k.k->version.lo);
 113
 114                 if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
 115                     fsck_err_on(!bch2_bkey_replicas_marked(c, k), c,
 116                                 "superblock not marked as containing replicas (type %u)",
 117                                 k.k->type)) {
 118                         ret = bch2_mark_bkey_replicas(c, k);
 119                         if (ret)
 120                                 return ret;
 121                 }
 122
 123                 bkey_for_each_ptr(ptrs, ptr) {
 124                         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 125                         struct bucket *g = PTR_BUCKET(ca, ptr, true);
 126                         struct bucket *g2 = PTR_BUCKET(ca, ptr, false);
 127
 128                         if (mustfix_fsck_err_on(!g->gen_valid, c,
 129                                         "bucket %u:%zu data type %s ptr gen %u missing in alloc btree",
 130                                         ptr->dev, PTR_BUCKET_NR(ca, ptr),
 131                                         bch2_data_types[ptr_data_type(k.k, ptr)],
 132                                         ptr->gen)) {
 133                                 g2->_mark.gen   = g->_mark.gen          = ptr->gen;
 134                                 g2->gen_valid   = g->gen_valid          = true;
 135                         }
 136
 137                         if (mustfix_fsck_err_on(gen_cmp(ptr->gen, g->mark.gen) > 0, c,
 138                                         "bucket %u:%zu data type %s ptr gen in the future: %u > %u",
 139                                         ptr->dev, PTR_BUCKET_NR(ca, ptr),
 140                                         bch2_data_types[ptr_data_type(k.k, ptr)],
 141                                         ptr->gen, g->mark.gen)) {
 142                                 g2->_mark.gen   = g->_mark.gen          = ptr->gen;
 143                                 g2->gen_valid   = g->gen_valid          = true;
 144                                 g2->_mark.data_type             = 0;
 145                                 g2->_mark.dirty_sectors         = 0;
 146                                 g2->_mark.cached_sectors        = 0;
 147                                 set_bit(BCH_FS_FIXED_GENS, &c->flags);
 148                         }
 149                 }
 150         }
 151
 152         bkey_for_each_ptr(ptrs, ptr) {
 153                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 154                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 155
 156                 if (gen_after(g->oldest_gen, ptr->gen))
 157                         g->oldest_gen = ptr->gen;
 158
 159                 *max_stale = max(*max_stale, ptr_stale(ca, ptr));
 160         }
 161
 162         bch2_mark_key(c, k, 0, k.k->size, NULL, 0, flags);
 163 fsck_err:
 164         return ret;
 165 }
 166
 167 static int btree_gc_mark_node(struct bch_fs *c, struct btree *b, u8 *max_stale,
 168                               bool initial)
 169 {
 170         struct bpos next_node_start = b->data->min_key;
 171         struct btree_node_iter iter;
 172         struct bkey unpacked;
 173         struct bkey_s_c k;
 174         int ret = 0;
 175
 176         *max_stale = 0;
 177
 178         if (!btree_node_type_needs_gc(btree_node_type(b)))
 179                 return 0;
 180
 181         bch2_btree_node_iter_init_from_start(&iter, b);
 182
 183         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 184                 bch2_bkey_debugcheck(c, b, k);
 185
 186                 ret = bch2_gc_mark_key(c, k, max_stale, initial);
 187                 if (ret)
 188                         break;
 189
 190                 bch2_btree_node_iter_advance(&iter, b);
 191
 192                 if (b->c.level) {
 193                         ret = bch2_gc_check_topology(c, k,
 194                                         &next_node_start,
 195                                         b->data->max_key,
 196                                         bch2_btree_node_iter_end(&iter));
 197                         if (ret)
 198                                 break;
 199                 }
 200         }
 201
 202         return ret;
 203 }
 204
 205 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id,
 206                          bool initial, bool metadata_only)
 207 {
 208         struct btree_trans trans;
 209         struct btree_iter *iter;
 210         struct btree *b;
 211         unsigned depth = metadata_only                  ? 1
 212                 : bch2_expensive_debug_checks           ? 0
 213                 : !btree_node_type_needs_gc(btree_id)   ? 1
 214                 : 0;
 215         u8 max_stale = 0;
 216         int ret = 0;
 217
 218         bch2_trans_init(&trans, c, 0, 0);
 219
 220         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 221
 222         __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
 223                               0, depth, BTREE_ITER_PREFETCH, b) {
 224                 bch2_verify_btree_nr_keys(b);
 225
 226                 gc_pos_set(c, gc_pos_btree_node(b));
 227
 228                 ret = btree_gc_mark_node(c, b, &max_stale, initial);
 229                 if (ret)
 230                         break;
 231
 232                 if (!initial) {
 233                         if (max_stale > 64)
 234                                 bch2_btree_node_rewrite(c, iter,
 235                                                 b->data->keys.seq,
 236                                                 BTREE_INSERT_USE_RESERVE|
 237                                                 BTREE_INSERT_NOWAIT|
 238                                                 BTREE_INSERT_GC_LOCK_HELD);
 239                         else if (!bch2_btree_gc_rewrite_disabled &&
 240                                  (bch2_btree_gc_always_rewrite || max_stale > 16))
 241                                 bch2_btree_node_rewrite(c, iter,
 242                                                 b->data->keys.seq,
 243                                                 BTREE_INSERT_NOWAIT|
 244                                                 BTREE_INSERT_GC_LOCK_HELD);
 245                 }
 246
 247                 bch2_trans_cond_resched(&trans);
 248         }
 249         ret = bch2_trans_exit(&trans) ?: ret;
 250         if (ret)
 251                 return ret;
 252
 253         mutex_lock(&c->btree_root_lock);
 254         b = c->btree_roots[btree_id].b;
 255         if (!btree_node_fake(b))
 256                 ret = bch2_gc_mark_key(c, bkey_i_to_s_c(&b->key),
 257                                        &max_stale, initial);
 258         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 259         mutex_unlock(&c->btree_root_lock);
 260
 261         return ret;
 262 }
 263
 264 static int bch2_gc_btree_init_recurse(struct bch_fs *c, struct btree *b,
 265                                       struct journal_keys *journal_keys,
 266                                       unsigned target_depth)
 267 {
 268         struct btree_and_journal_iter iter;
 269         struct bkey_s_c k;
 270         struct bpos next_node_start = b->data->min_key;
 271         u8 max_stale = 0;
 272         int ret = 0;
 273
 274         bch2_btree_and_journal_iter_init_node_iter(&iter, journal_keys, b);
 275
 276         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 277                 bch2_bkey_debugcheck(c, b, k);
 278
 279                 BUG_ON(bkey_cmp(k.k->p, b->data->min_key) < 0);
 280                 BUG_ON(bkey_cmp(k.k->p, b->data->max_key) > 0);
 281
 282                 ret = bch2_gc_mark_key(c, k, &max_stale, true);
 283                 if (ret)
 284                         break;
 285
 286                 if (b->c.level) {
 287                         struct btree *child;
 288                         BKEY_PADDED(k) tmp;
 289
 290                         bkey_reassemble(&tmp.k, k);
 291                         k = bkey_i_to_s_c(&tmp.k);
 292
 293                         bch2_btree_and_journal_iter_advance(&iter);
 294
 295                         ret = bch2_gc_check_topology(c, k,
 296                                         &next_node_start,
 297                                         b->data->max_key,
 298                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 299                         if (ret)
 300                                 break;
 301
 302                         if (b->c.level > target_depth) {
 303                                 child = bch2_btree_node_get_noiter(c, &tmp.k,
 304                                                         b->c.btree_id, b->c.level - 1);
 305                                 ret = PTR_ERR_OR_ZERO(child);
 306                                 if (ret)
 307                                         break;
 308
 309                                 ret = bch2_gc_btree_init_recurse(c, child,
 310                                                 journal_keys, target_depth);
 311                                 six_unlock_read(&child->c.lock);
 312
 313                                 if (ret)
 314                                         break;
 315                         }
 316                 } else {
 317                         bch2_btree_and_journal_iter_advance(&iter);
 318                 }
 319         }
 320
 321         return ret;
 322 }
 323
 324 static int bch2_gc_btree_init(struct bch_fs *c,
 325                               struct journal_keys *journal_keys,
 326                               enum btree_id btree_id,
 327                               bool metadata_only)
 328 {
 329         struct btree *b;
 330         unsigned target_depth = metadata_only           ? 1
 331                 : bch2_expensive_debug_checks           ? 0
 332                 : !btree_node_type_needs_gc(btree_id)   ? 1
 333                 : 0;
 334         u8 max_stale = 0;
 335         int ret = 0;
 336
 337         b = c->btree_roots[btree_id].b;
 338
 339         if (btree_node_fake(b))
 340                 return 0;
 341
 342         six_lock_read(&b->c.lock, NULL, NULL);
 343         if (fsck_err_on(bkey_cmp(b->data->min_key, POS_MIN), c,
 344                         "btree root with incorrect min_key: %llu:%llu",
 345                         b->data->min_key.inode,
 346                         b->data->min_key.offset)) {
 347                 BUG();
 348         }
 349
 350         if (fsck_err_on(bkey_cmp(b->data->max_key, POS_MAX), c,
 351                         "btree root with incorrect min_key: %llu:%llu",
 352                         b->data->max_key.inode,
 353                         b->data->max_key.offset)) {
 354                 BUG();
 355         }
 356
 357         if (b->c.level >= target_depth)
 358                 ret = bch2_gc_btree_init_recurse(c, b,
 359                                         journal_keys, target_depth);
 360
 361         if (!ret)
 362                 ret = bch2_gc_mark_key(c, bkey_i_to_s_c(&b->key),
 363                                        &max_stale, true);
 364 fsck_err:
 365         six_unlock_read(&b->c.lock);
 366
 367         return ret;
 368 }
 369
 370 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
 371 {
 372         return  (int) btree_id_to_gc_phase(l) -
 373                 (int) btree_id_to_gc_phase(r);
 374 }
 375
 376 static int bch2_gc_btrees(struct bch_fs *c, struct journal_keys *journal_keys,
 377                           bool initial, bool metadata_only)
 378 {
 379         enum btree_id ids[BTREE_ID_NR];
 380         unsigned i;
 381
 382         for (i = 0; i < BTREE_ID_NR; i++)
 383                 ids[i] = i;
 384         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
 385
 386         for (i = 0; i < BTREE_ID_NR; i++) {
 387                 enum btree_id id = ids[i];
 388                 int ret = initial
 389                         ? bch2_gc_btree_init(c, journal_keys,
 390                                              id, metadata_only)
 391                         : bch2_gc_btree(c, id, initial, metadata_only);
 392                 if (ret)
 393                         return ret;
 394         }
 395
 396         return 0;
 397 }
 398
 399 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
 400                                   u64 start, u64 end,
 401                                   enum bch_data_type type,
 402                                   unsigned flags)
 403 {
 404         u64 b = sector_to_bucket(ca, start);
 405
 406         do {
 407                 unsigned sectors =
 408                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
 409
 410                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
 411                                           gc_phase(GC_PHASE_SB), flags);
 412                 b++;
 413                 start += sectors;
 414         } while (start < end);
 415 }
 416
 417 void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
 418                               unsigned flags)
 419 {
 420         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
 421         unsigned i;
 422         u64 b;
 423
 424         /*
 425          * This conditional is kind of gross, but we may be called from the
 426          * device add path, before the new device has actually been added to the
 427          * running filesystem:
 428          */
 429         if (c) {
 430                 lockdep_assert_held(&c->sb_lock);
 431                 percpu_down_read(&c->mark_lock);
 432         }
 433
 434         for (i = 0; i < layout->nr_superblocks; i++) {
 435                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
 436
 437                 if (offset == BCH_SB_SECTOR)
 438                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
 439                                               BCH_DATA_sb, flags);
 440
 441                 mark_metadata_sectors(c, ca, offset,
 442                                       offset + (1 << layout->sb_max_size_bits),
 443                                       BCH_DATA_sb, flags);
 444         }
 445
 446         for (i = 0; i < ca->journal.nr; i++) {
 447                 b = ca->journal.buckets[i];
 448                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
 449                                           ca->mi.bucket_size,
 450                                           gc_phase(GC_PHASE_SB), flags);
 451         }
 452
 453         if (c)
 454                 percpu_up_read(&c->mark_lock);
 455 }
 456
 457 static void bch2_mark_superblocks(struct bch_fs *c)
 458 {
 459         struct bch_dev *ca;
 460         unsigned i;
 461
 462         mutex_lock(&c->sb_lock);
 463         gc_pos_set(c, gc_phase(GC_PHASE_SB));
 464
 465         for_each_online_member(ca, c, i)
 466                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
 467         mutex_unlock(&c->sb_lock);
 468 }
 469
 470 #if 0
 471 /* Also see bch2_pending_btree_node_free_insert_done() */
 472 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
 473 {
 474         struct btree_update *as;
 475         struct pending_btree_node_free *d;
 476
 477         mutex_lock(&c->btree_interior_update_lock);
 478         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
 479
 480         for_each_pending_btree_node_free(c, as, d)
 481                 if (d->index_update_done)
 482                         bch2_mark_key(c, bkey_i_to_s_c(&d->key),
 483                                       0, 0, NULL, 0,
 484                                       BTREE_TRIGGER_GC);
 485
 486         mutex_unlock(&c->btree_interior_update_lock);
 487 }
 488 #endif
 489
 490 static void bch2_mark_allocator_buckets(struct bch_fs *c)
 491 {
 492         struct bch_dev *ca;
 493         struct open_bucket *ob;
 494         size_t i, j, iter;
 495         unsigned ci;
 496
 497         percpu_down_read(&c->mark_lock);
 498
 499         spin_lock(&c->freelist_lock);
 500         gc_pos_set(c, gc_pos_alloc(c, NULL));
 501
 502         for_each_member_device(ca, c, ci) {
 503                 fifo_for_each_entry(i, &ca->free_inc, iter)
 504                         bch2_mark_alloc_bucket(c, ca, i, true,
 505                                                gc_pos_alloc(c, NULL),
 506                                                BTREE_TRIGGER_GC);
 507
 508
 509
 510                 for (j = 0; j < RESERVE_NR; j++)
 511                         fifo_for_each_entry(i, &ca->free[j], iter)
 512                                 bch2_mark_alloc_bucket(c, ca, i, true,
 513                                                        gc_pos_alloc(c, NULL),
 514                                                        BTREE_TRIGGER_GC);
 515         }
 516
 517         spin_unlock(&c->freelist_lock);
 518
 519         for (ob = c->open_buckets;
 520              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
 521              ob++) {
 522                 spin_lock(&ob->lock);
 523                 if (ob->valid) {
 524                         gc_pos_set(c, gc_pos_alloc(c, ob));
 525                         ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 526                         bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr), true,
 527                                                gc_pos_alloc(c, ob),
 528                                                BTREE_TRIGGER_GC);
 529                 }
 530                 spin_unlock(&ob->lock);
 531         }
 532
 533         percpu_up_read(&c->mark_lock);
 534 }
 535
 536 static void bch2_gc_free(struct bch_fs *c)
 537 {
 538         struct bch_dev *ca;
 539         unsigned i;
 540
 541         genradix_free(&c->stripes[1]);
 542
 543         for_each_member_device(ca, c, i) {
 544                 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
 545                         sizeof(struct bucket_array) +
 546                         ca->mi.nbuckets * sizeof(struct bucket));
 547                 ca->buckets[1] = NULL;
 548
 549                 free_percpu(ca->usage[1]);
 550                 ca->usage[1] = NULL;
 551         }
 552
 553         free_percpu(c->usage_gc);
 554         c->usage_gc = NULL;
 555 }
 556
 557 static int bch2_gc_done(struct bch_fs *c,
 558                         bool initial, bool metadata_only)
 559 {
 560         struct bch_dev *ca;
 561         bool verify = !metadata_only &&
 562                 (!initial ||
 563                  (c->sb.compat & (1ULL << BCH_COMPAT_FEAT_ALLOC_INFO)));
 564         unsigned i;
 565         int ret = 0;
 566
 567 #define copy_field(_f, _msg, ...)                                       \
 568         if (dst->_f != src->_f) {                                       \
 569                 if (verify)                                             \
 570                         fsck_err(c, _msg ": got %llu, should be %llu"   \
 571                                 , ##__VA_ARGS__, dst->_f, src->_f);     \
 572                 dst->_f = src->_f;                                      \
 573                 ret = 1;                                                \
 574         }
 575 #define copy_stripe_field(_f, _msg, ...)                                \
 576         if (dst->_f != src->_f) {                                       \
 577                 if (verify)                                             \
 578                         fsck_err(c, "stripe %zu has wrong "_msg         \
 579                                 ": got %u, should be %u",               \
 580                                 dst_iter.pos, ##__VA_ARGS__,            \
 581                                 dst->_f, src->_f);                      \
 582                 dst->_f = src->_f;                                      \
 583                 dst->dirty = true;                                      \
 584                 ret = 1;                                                \
 585         }
 586 #define copy_bucket_field(_f)                                           \
 587         if (dst->b[b].mark._f != src->b[b].mark._f) {                   \
 588                 if (verify)                                             \
 589                         fsck_err(c, "bucket %u:%zu gen %u data type %s has wrong " #_f  \
 590                                 ": got %u, should be %u", i, b,         \
 591                                 dst->b[b].mark.gen,                     \
 592                                 bch2_data_types[dst->b[b].mark.data_type],\
 593                                 dst->b[b].mark._f, src->b[b].mark._f);  \
 594                 dst->b[b]._mark._f = src->b[b].mark._f;                 \
 595                 ret = 1;                                                \
 596         }
 597 #define copy_dev_field(_f, _msg, ...)                                   \
 598         copy_field(_f, "dev %u has wrong " _msg, i, ##__VA_ARGS__)
 599 #define copy_fs_field(_f, _msg, ...)                                    \
 600         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
 601
 602         if (!metadata_only) {
 603                 struct genradix_iter dst_iter = genradix_iter_init(&c->stripes[0], 0);
 604                 struct genradix_iter src_iter = genradix_iter_init(&c->stripes[1], 0);
 605                 struct stripe *dst, *src;
 606
 607                 c->ec_stripes_heap.used = 0;
 608
 609                 while ((dst = genradix_iter_peek(&dst_iter, &c->stripes[0])) &&
 610                        (src = genradix_iter_peek(&src_iter, &c->stripes[1]))) {
 611                         BUG_ON(src_iter.pos != dst_iter.pos);
 612
 613                         copy_stripe_field(alive,        "alive");
 614                         copy_stripe_field(sectors,      "sectors");
 615                         copy_stripe_field(algorithm,    "algorithm");
 616                         copy_stripe_field(nr_blocks,    "nr_blocks");
 617                         copy_stripe_field(nr_redundant, "nr_redundant");
 618                         copy_stripe_field(blocks_nonempty,
 619                                           "blocks_nonempty");
 620
 621                         for (i = 0; i < ARRAY_SIZE(dst->block_sectors); i++)
 622                                 copy_stripe_field(block_sectors[i],
 623                                                   "block_sectors[%u]", i);
 624
 625                         if (dst->alive) {
 626                                 spin_lock(&c->ec_stripes_heap_lock);
 627                                 bch2_stripes_heap_insert(c, dst, dst_iter.pos);
 628                                 spin_unlock(&c->ec_stripes_heap_lock);
 629                         }
 630
 631                         genradix_iter_advance(&dst_iter, &c->stripes[0]);
 632                         genradix_iter_advance(&src_iter, &c->stripes[1]);
 633                 }
 634         }
 635
 636         for_each_member_device(ca, c, i) {
 637                 struct bucket_array *dst = __bucket_array(ca, 0);
 638                 struct bucket_array *src = __bucket_array(ca, 1);
 639                 size_t b;
 640
 641                 for (b = 0; b < src->nbuckets; b++) {
 642                         copy_bucket_field(gen);
 643                         copy_bucket_field(data_type);
 644                         copy_bucket_field(owned_by_allocator);
 645                         copy_bucket_field(stripe);
 646                         copy_bucket_field(dirty_sectors);
 647                         copy_bucket_field(cached_sectors);
 648
 649                         dst->b[b].oldest_gen = src->b[b].oldest_gen;
 650                 }
 651         };
 652
 653         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 654                 bch2_fs_usage_acc_to_base(c, i);
 655
 656         bch2_dev_usage_from_buckets(c);
 657
 658         {
 659                 unsigned nr = fs_usage_u64s(c);
 660                 struct bch_fs_usage *dst = c->usage_base;
 661                 struct bch_fs_usage *src = (void *)
 662                         bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
 663
 664                 copy_fs_field(hidden,           "hidden");
 665                 copy_fs_field(btree,            "btree");
 666
 667                 if (!metadata_only) {
 668                         copy_fs_field(data,     "data");
 669                         copy_fs_field(cached,   "cached");
 670                         copy_fs_field(reserved, "reserved");
 671                         copy_fs_field(nr_inodes,"nr_inodes");
 672
 673                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
 674                                 copy_fs_field(persistent_reserved[i],
 675                                               "persistent_reserved[%i]", i);
 676                 }
 677
 678                 for (i = 0; i < c->replicas.nr; i++) {
 679                         struct bch_replicas_entry *e =
 680                                 cpu_replicas_entry(&c->replicas, i);
 681                         char buf[80];
 682
 683                         if (metadata_only &&
 684                             (e->data_type == BCH_DATA_user ||
 685                              e->data_type == BCH_DATA_cached))
 686                                 continue;
 687
 688                         bch2_replicas_entry_to_text(&PBUF(buf), e);
 689
 690                         copy_fs_field(replicas[i], "%s", buf);
 691                 }
 692         }
 693
 694 #undef copy_fs_field
 695 #undef copy_dev_field
 696 #undef copy_bucket_field
 697 #undef copy_stripe_field
 698 #undef copy_field
 699 fsck_err:
 700         return ret;
 701 }
 702
 703 static int bch2_gc_start(struct bch_fs *c,
 704                          bool metadata_only)
 705 {
 706         struct bch_dev *ca;
 707         unsigned i;
 708         int ret;
 709
 710         BUG_ON(c->usage_gc);
 711
 712         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
 713                                          sizeof(u64), GFP_KERNEL);
 714         if (!c->usage_gc) {
 715                 bch_err(c, "error allocating c->usage_gc");
 716                 return -ENOMEM;
 717         }
 718
 719         for_each_member_device(ca, c, i) {
 720                 BUG_ON(ca->buckets[1]);
 721                 BUG_ON(ca->usage[1]);
 722
 723                 ca->buckets[1] = kvpmalloc(sizeof(struct bucket_array) +
 724                                 ca->mi.nbuckets * sizeof(struct bucket),
 725                                 GFP_KERNEL|__GFP_ZERO);
 726                 if (!ca->buckets[1]) {
 727                         percpu_ref_put(&ca->ref);
 728                         bch_err(c, "error allocating ca->buckets[gc]");
 729                         return -ENOMEM;
 730                 }
 731
 732                 ca->usage[1] = alloc_percpu(struct bch_dev_usage);
 733                 if (!ca->usage[1]) {
 734                         bch_err(c, "error allocating ca->usage[gc]");
 735                         percpu_ref_put(&ca->ref);
 736                         return -ENOMEM;
 737                 }
 738         }
 739
 740         ret = bch2_ec_mem_alloc(c, true);
 741         if (ret) {
 742                 bch_err(c, "error allocating ec gc mem");
 743                 return ret;
 744         }
 745
 746         percpu_down_write(&c->mark_lock);
 747
 748         /*
 749          * indicate to stripe code that we need to allocate for the gc stripes
 750          * radix tree, too
 751          */
 752         gc_pos_set(c, gc_phase(GC_PHASE_START));
 753
 754         for_each_member_device(ca, c, i) {
 755                 struct bucket_array *dst = __bucket_array(ca, 1);
 756                 struct bucket_array *src = __bucket_array(ca, 0);
 757                 size_t b;
 758
 759                 dst->first_bucket       = src->first_bucket;
 760                 dst->nbuckets           = src->nbuckets;
 761
 762                 for (b = 0; b < src->nbuckets; b++) {
 763                         struct bucket *d = &dst->b[b];
 764                         struct bucket *s = &src->b[b];
 765
 766                         d->_mark.gen = dst->b[b].oldest_gen = s->mark.gen;
 767                         d->gen_valid = s->gen_valid;
 768
 769                         if (metadata_only &&
 770                             (s->mark.data_type == BCH_DATA_user ||
 771                              s->mark.data_type == BCH_DATA_cached)) {
 772                                 d->_mark = s->mark;
 773                                 d->_mark.owned_by_allocator = 0;
 774                         }
 775                 }
 776         };
 777
 778         percpu_up_write(&c->mark_lock);
 779
 780         return 0;
 781 }
 782
 783 /**
 784  * bch2_gc - walk _all_ references to buckets, and recompute them:
 785  *
 786  * Order matters here:
 787  *  - Concurrent GC relies on the fact that we have a total ordering for
 788  *    everything that GC walks - see  gc_will_visit_node(),
 789  *    gc_will_visit_root()
 790  *
 791  *  - also, references move around in the course of index updates and
 792  *    various other crap: everything needs to agree on the ordering
 793  *    references are allowed to move around in - e.g., we're allowed to
 794  *    start with a reference owned by an open_bucket (the allocator) and
 795  *    move it to the btree, but not the reverse.
 796  *
 797  *    This is necessary to ensure that gc doesn't miss references that
 798  *    move around - if references move backwards in the ordering GC
 799  *    uses, GC could skip past them
 800  */
 801 int bch2_gc(struct bch_fs *c, struct journal_keys *journal_keys,
 802             bool initial, bool metadata_only)
 803 {
 804         struct bch_dev *ca;
 805         u64 start_time = local_clock();
 806         unsigned i, iter = 0;
 807         int ret;
 808
 809         lockdep_assert_held(&c->state_lock);
 810         trace_gc_start(c);
 811
 812         down_write(&c->gc_lock);
 813
 814         /* flush interior btree updates: */
 815         closure_wait_event(&c->btree_interior_update_wait,
 816                            !bch2_btree_interior_updates_nr_pending(c));
 817 again:
 818         ret = bch2_gc_start(c, metadata_only);
 819         if (ret)
 820                 goto out;
 821
 822         bch2_mark_superblocks(c);
 823
 824         ret = bch2_gc_btrees(c, journal_keys, initial, metadata_only);
 825         if (ret)
 826                 goto out;
 827
 828 #if 0
 829         bch2_mark_pending_btree_node_frees(c);
 830 #endif
 831         bch2_mark_allocator_buckets(c);
 832
 833         c->gc_count++;
 834 out:
 835         if (!ret &&
 836             (test_bit(BCH_FS_FIXED_GENS, &c->flags) ||
 837              (!iter && bch2_test_restart_gc))) {
 838                 /*
 839                  * XXX: make sure gens we fixed got saved
 840                  */
 841                 if (iter++ <= 2) {
 842                         bch_info(c, "Fixed gens, restarting mark and sweep:");
 843                         clear_bit(BCH_FS_FIXED_GENS, &c->flags);
 844                         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
 845
 846                         percpu_down_write(&c->mark_lock);
 847                         bch2_gc_free(c);
 848                         percpu_up_write(&c->mark_lock);
 849                         /* flush fsck errors, reset counters */
 850                         bch2_flush_fsck_errs(c);
 851
 852                         goto again;
 853                 }
 854
 855                 bch_info(c, "Unable to fix bucket gens, looping");
 856                 ret = -EINVAL;
 857         }
 858
 859         if (!ret) {
 860                 bch2_journal_block(&c->journal);
 861
 862                 percpu_down_write(&c->mark_lock);
 863                 ret = bch2_gc_done(c, initial, metadata_only);
 864
 865                 bch2_journal_unblock(&c->journal);
 866         } else {
 867                 percpu_down_write(&c->mark_lock);
 868         }
 869
 870         /* Indicates that gc is no longer in progress: */
 871         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
 872
 873         bch2_gc_free(c);
 874         percpu_up_write(&c->mark_lock);
 875
 876         up_write(&c->gc_lock);
 877
 878         trace_gc_end(c);
 879         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
 880
 881         /*
 882          * Wake up allocator in case it was waiting for buckets
 883          * because of not being able to inc gens
 884          */
 885         for_each_member_device(ca, c, i)
 886                 bch2_wake_allocator(ca);
 887
 888         /*
 889          * At startup, allocations can happen directly instead of via the
 890          * allocator thread - issue wakeup in case they blocked on gc_lock:
 891          */
 892         closure_wake_up(&c->freelist_wait);
 893         return ret;
 894 }
 895
 896 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
 897 {
 898         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 899         const struct bch_extent_ptr *ptr;
 900
 901         percpu_down_read(&c->mark_lock);
 902         bkey_for_each_ptr(ptrs, ptr) {
 903                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 904                 struct bucket *g = PTR_BUCKET(ca, ptr, false);
 905
 906                 if (gen_after(g->mark.gen, ptr->gen) > 16) {
 907                         percpu_up_read(&c->mark_lock);
 908                         return true;
 909                 }
 910         }
 911
 912         bkey_for_each_ptr(ptrs, ptr) {
 913                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 914                 struct bucket *g = PTR_BUCKET(ca, ptr, false);
 915
 916                 if (gen_after(g->gc_gen, ptr->gen))
 917                         g->gc_gen = ptr->gen;
 918         }
 919         percpu_up_read(&c->mark_lock);
 920
 921         return false;
 922 }
 923
 924 /*
 925  * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
 926  * node pointers currently never have cached pointers that can become stale:
 927  */
 928 static int bch2_gc_btree_gens(struct bch_fs *c, enum btree_id btree_id)
 929 {
 930         struct btree_trans trans;
 931         struct btree_iter *iter;
 932         struct bkey_s_c k;
 933         struct bkey_on_stack sk;
 934         int ret = 0;
 935
 936         bkey_on_stack_init(&sk);
 937         bch2_trans_init(&trans, c, 0, 0);
 938
 939         iter = bch2_trans_get_iter(&trans, btree_id, POS_MIN,
 940                                    BTREE_ITER_PREFETCH);
 941
 942         while ((k = bch2_btree_iter_peek(iter)).k &&
 943                !(ret = bkey_err(k))) {
 944                 if (gc_btree_gens_key(c, k)) {
 945                         bkey_on_stack_reassemble(&sk, c, k);
 946                         bch2_extent_normalize(c, bkey_i_to_s(sk.k));
 947
 948                         bch2_btree_iter_set_pos(iter, bkey_start_pos(&sk.k->k));
 949
 950                         bch2_trans_update(&trans, iter, sk.k, 0);
 951
 952                         ret = bch2_trans_commit(&trans, NULL, NULL,
 953                                                 BTREE_INSERT_NOFAIL);
 954                         if (ret == -EINTR)
 955                                 continue;
 956                         if (ret) {
 957                                 break;
 958                         }
 959                 }
 960
 961                 bch2_btree_iter_next(iter);
 962         }
 963
 964         bch2_trans_exit(&trans);
 965         bkey_on_stack_exit(&sk, c);
 966
 967         return ret;
 968 }
 969
 970 int bch2_gc_gens(struct bch_fs *c)
 971 {
 972         struct bch_dev *ca;
 973         struct bucket_array *buckets;
 974         struct bucket *g;
 975         unsigned i;
 976         int ret;
 977
 978         /*
 979          * Ideally we would be using state_lock and not gc_lock here, but that
 980          * introduces a deadlock in the RO path - we currently take the state
 981          * lock at the start of going RO, thus the gc thread may get stuck:
 982          */
 983         down_read(&c->gc_lock);
 984
 985         for_each_member_device(ca, c, i) {
 986                 down_read(&ca->bucket_lock);
 987                 buckets = bucket_array(ca);
 988
 989                 for_each_bucket(g, buckets)
 990                         g->gc_gen = g->mark.gen;
 991                 up_read(&ca->bucket_lock);
 992         }
 993
 994         for (i = 0; i < BTREE_ID_NR; i++)
 995                 if (btree_node_type_needs_gc(i)) {
 996                         ret = bch2_gc_btree_gens(c, i);
 997                         if (ret) {
 998                                 bch_err(c, "error recalculating oldest_gen: %i", ret);
 999                                 goto err;
1000                         }
1001                 }
1002
1003         for_each_member_device(ca, c, i) {
1004                 down_read(&ca->bucket_lock);
1005                 buckets = bucket_array(ca);
1006
1007                 for_each_bucket(g, buckets)
1008                         g->oldest_gen = g->gc_gen;
1009                 up_read(&ca->bucket_lock);
1010         }
1011
1012         c->gc_count++;
1013 err:
1014         up_read(&c->gc_lock);
1015         return ret;
1016 }
1017
1018 /* Btree coalescing */
1019
1020 static void recalc_packed_keys(struct btree *b)
1021 {
1022         struct bset *i = btree_bset_first(b);
1023         struct bkey_packed *k;
1024
1025         memset(&b->nr, 0, sizeof(b->nr));
1026
1027         BUG_ON(b->nsets != 1);
1028
1029         vstruct_for_each(i, k)
1030                 btree_keys_account_key_add(&b->nr, 0, k);
1031 }
1032
1033 static void bch2_coalesce_nodes(struct bch_fs *c, struct btree_iter *iter,
1034                                 struct btree *old_nodes[GC_MERGE_NODES])
1035 {
1036         struct btree *parent = btree_node_parent(iter, old_nodes[0]);
1037         unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
1038         unsigned blocks = btree_blocks(c) * 2 / 3;
1039         struct btree *new_nodes[GC_MERGE_NODES];
1040         struct btree_update *as;
1041         struct keylist keylist;
1042         struct bkey_format_state format_state;
1043         struct bkey_format new_format;
1044
1045         memset(new_nodes, 0, sizeof(new_nodes));
1046         bch2_keylist_init(&keylist, NULL);
1047
1048         /* Count keys that are not deleted */
1049         for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
1050                 u64s += old_nodes[i]->nr.live_u64s;
1051
1052         nr_old_nodes = nr_new_nodes = i;
1053
1054         /* Check if all keys in @old_nodes could fit in one fewer node */
1055         if (nr_old_nodes <= 1 ||
1056             __vstruct_blocks(struct btree_node, c->block_bits,
1057                              DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
1058                 return;
1059
1060         /* Find a format that all keys in @old_nodes can pack into */
1061         bch2_bkey_format_init(&format_state);
1062
1063         for (i = 0; i < nr_old_nodes; i++)
1064                 __bch2_btree_calc_format(&format_state, old_nodes[i]);
1065
1066         new_format = bch2_bkey_format_done(&format_state);
1067
1068         /* Check if repacking would make any nodes too big to fit */
1069         for (i = 0; i < nr_old_nodes; i++)
1070                 if (!bch2_btree_node_format_fits(c, old_nodes[i], &new_format)) {
1071                         trace_btree_gc_coalesce_fail(c,
1072                                         BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
1073                         return;
1074                 }
1075
1076         if (bch2_keylist_realloc(&keylist, NULL, 0,
1077                         (BKEY_U64s + BKEY_EXTENT_U64s_MAX) * nr_old_nodes)) {
1078                 trace_btree_gc_coalesce_fail(c,
1079                                 BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
1080                 return;
1081         }
1082
1083         as = bch2_btree_update_start(iter->trans, iter->btree_id,
1084                         btree_update_reserve_required(c, parent) + nr_old_nodes,
1085                         BTREE_INSERT_NOFAIL|
1086                         BTREE_INSERT_USE_RESERVE,
1087                         NULL);
1088         if (IS_ERR(as)) {
1089                 trace_btree_gc_coalesce_fail(c,
1090                                 BTREE_GC_COALESCE_FAIL_RESERVE_GET);
1091                 bch2_keylist_free(&keylist, NULL);
1092                 return;
1093         }
1094
1095         trace_btree_gc_coalesce(c, old_nodes[0]);
1096
1097         for (i = 0; i < nr_old_nodes; i++)
1098                 bch2_btree_interior_update_will_free_node(as, old_nodes[i]);
1099
1100         /* Repack everything with @new_format and sort down to one bset */
1101         for (i = 0; i < nr_old_nodes; i++)
1102                 new_nodes[i] =
1103                         __bch2_btree_node_alloc_replacement(as, old_nodes[i],
1104                                                             new_format);
1105
1106         /*
1107          * Conceptually we concatenate the nodes together and slice them
1108          * up at different boundaries.
1109          */
1110         for (i = nr_new_nodes - 1; i > 0; --i) {
1111                 struct btree *n1 = new_nodes[i];
1112                 struct btree *n2 = new_nodes[i - 1];
1113
1114                 struct bset *s1 = btree_bset_first(n1);
1115                 struct bset *s2 = btree_bset_first(n2);
1116                 struct bkey_packed *k, *last = NULL;
1117
1118                 /* Calculate how many keys from @n2 we could fit inside @n1 */
1119                 u64s = 0;
1120
1121                 for (k = s2->start;
1122                      k < vstruct_last(s2) &&
1123                      vstruct_blocks_plus(n1->data, c->block_bits,
1124                                          u64s + k->u64s) <= blocks;
1125                      k = bkey_next_skip_noops(k, vstruct_last(s2))) {
1126                         last = k;
1127                         u64s += k->u64s;
1128                 }
1129
1130                 if (u64s == le16_to_cpu(s2->u64s)) {
1131                         /* n2 fits entirely in n1 */
1132                         n1->key.k.p = n1->data->max_key = n2->data->max_key;
1133
1134                         memcpy_u64s(vstruct_last(s1),
1135                                     s2->start,
1136                                     le16_to_cpu(s2->u64s));
1137                         le16_add_cpu(&s1->u64s, le16_to_cpu(s2->u64s));
1138
1139                         set_btree_bset_end(n1, n1->set);
1140
1141                         six_unlock_write(&n2->c.lock);
1142                         bch2_btree_node_free_never_inserted(c, n2);
1143                         six_unlock_intent(&n2->c.lock);
1144
1145                         memmove(new_nodes + i - 1,
1146                                 new_nodes + i,
1147                                 sizeof(new_nodes[0]) * (nr_new_nodes - i));
1148                         new_nodes[--nr_new_nodes] = NULL;
1149                 } else if (u64s) {
1150                         /* move part of n2 into n1 */
1151                         n1->key.k.p = n1->data->max_key =
1152                                 bkey_unpack_pos(n1, last);
1153
1154                         n2->data->min_key = bkey_successor(n1->data->max_key);
1155
1156                         memcpy_u64s(vstruct_last(s1),
1157                                     s2->start, u64s);
1158                         le16_add_cpu(&s1->u64s, u64s);
1159
1160                         memmove(s2->start,
1161                                 vstruct_idx(s2, u64s),
1162                                 (le16_to_cpu(s2->u64s) - u64s) * sizeof(u64));
1163                         s2->u64s = cpu_to_le16(le16_to_cpu(s2->u64s) - u64s);
1164
1165                         set_btree_bset_end(n1, n1->set);
1166                         set_btree_bset_end(n2, n2->set);
1167                 }
1168         }
1169
1170         for (i = 0; i < nr_new_nodes; i++) {
1171                 struct btree *n = new_nodes[i];
1172
1173                 recalc_packed_keys(n);
1174                 btree_node_reset_sib_u64s(n);
1175
1176                 bch2_btree_build_aux_trees(n);
1177
1178                 bch2_btree_update_add_new_node(as, n);
1179                 six_unlock_write(&n->c.lock);
1180
1181                 bch2_btree_node_write(c, n, SIX_LOCK_intent);
1182         }
1183
1184         /*
1185          * The keys for the old nodes get deleted. We don't want to insert keys
1186          * that compare equal to the keys for the new nodes we'll also be
1187          * inserting - we can't because keys on a keylist must be strictly
1188          * greater than the previous keys, and we also don't need to since the
1189          * key for the new node will serve the same purpose (overwriting the key
1190          * for the old node).
1191          */
1192         for (i = 0; i < nr_old_nodes; i++) {
1193                 struct bkey_i delete;
1194                 unsigned j;
1195
1196                 for (j = 0; j < nr_new_nodes; j++)
1197                         if (!bkey_cmp(old_nodes[i]->key.k.p,
1198                                       new_nodes[j]->key.k.p))
1199                                 goto next;
1200
1201                 bkey_init(&delete.k);
1202                 delete.k.p = old_nodes[i]->key.k.p;
1203                 bch2_keylist_add_in_order(&keylist, &delete);
1204 next:
1205                 i = i;
1206         }
1207
1208         /*
1209          * Keys for the new nodes get inserted: bch2_btree_insert_keys() only
1210          * does the lookup once and thus expects the keys to be in sorted order
1211          * so we have to make sure the new keys are correctly ordered with
1212          * respect to the deleted keys added in the previous loop
1213          */
1214         for (i = 0; i < nr_new_nodes; i++)
1215                 bch2_keylist_add_in_order(&keylist, &new_nodes[i]->key);
1216
1217         /* Insert the newly coalesced nodes */
1218         bch2_btree_insert_node(as, parent, iter, &keylist, 0);
1219
1220         BUG_ON(!bch2_keylist_empty(&keylist));
1221
1222         BUG_ON(iter->l[old_nodes[0]->c.level].b != old_nodes[0]);
1223
1224         bch2_btree_iter_node_replace(iter, new_nodes[0]);
1225
1226         for (i = 0; i < nr_new_nodes; i++)
1227                 bch2_btree_update_get_open_buckets(as, new_nodes[i]);
1228
1229         /* Free the old nodes and update our sliding window */
1230         for (i = 0; i < nr_old_nodes; i++) {
1231                 bch2_btree_node_free_inmem(c, old_nodes[i], iter);
1232
1233                 /*
1234                  * the index update might have triggered a split, in which case
1235                  * the nodes we coalesced - the new nodes we just created -
1236                  * might not be sibling nodes anymore - don't add them to the
1237                  * sliding window (except the first):
1238                  */
1239                 if (!i) {
1240                         old_nodes[i] = new_nodes[i];
1241                 } else {
1242                         old_nodes[i] = NULL;
1243                 }
1244         }
1245
1246         for (i = 0; i < nr_new_nodes; i++)
1247                 six_unlock_intent(&new_nodes[i]->c.lock);
1248
1249         bch2_btree_update_done(as);
1250         bch2_keylist_free(&keylist, NULL);
1251 }
1252
1253 static int bch2_coalesce_btree(struct bch_fs *c, enum btree_id btree_id)
1254 {
1255         struct btree_trans trans;
1256         struct btree_iter *iter;
1257         struct btree *b;
1258         bool kthread = (current->flags & PF_KTHREAD) != 0;
1259         unsigned i;
1260
1261         /* Sliding window of adjacent btree nodes */
1262         struct btree *merge[GC_MERGE_NODES];
1263         u32 lock_seq[GC_MERGE_NODES];
1264
1265         bch2_trans_init(&trans, c, 0, 0);
1266
1267         /*
1268          * XXX: We don't have a good way of positively matching on sibling nodes
1269          * that have the same parent - this code works by handling the cases
1270          * where they might not have the same parent, and is thus fragile. Ugh.
1271          *
1272          * Perhaps redo this to use multiple linked iterators?
1273          */
1274         memset(merge, 0, sizeof(merge));
1275
1276         __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
1277                               BTREE_MAX_DEPTH, 0,
1278                               BTREE_ITER_PREFETCH, b) {
1279                 memmove(merge + 1, merge,
1280                         sizeof(merge) - sizeof(merge[0]));
1281                 memmove(lock_seq + 1, lock_seq,
1282                         sizeof(lock_seq) - sizeof(lock_seq[0]));
1283
1284                 merge[0] = b;
1285
1286                 for (i = 1; i < GC_MERGE_NODES; i++) {
1287                         if (!merge[i] ||
1288                             !six_relock_intent(&merge[i]->c.lock, lock_seq[i]))
1289                                 break;
1290
1291                         if (merge[i]->c.level != merge[0]->c.level) {
1292                                 six_unlock_intent(&merge[i]->c.lock);
1293                                 break;
1294                         }
1295                 }
1296                 memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));
1297
1298                 bch2_coalesce_nodes(c, iter, merge);
1299
1300                 for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
1301                         lock_seq[i] = merge[i]->c.lock.state.seq;
1302                         six_unlock_intent(&merge[i]->c.lock);
1303                 }
1304
1305                 lock_seq[0] = merge[0]->c.lock.state.seq;
1306
1307                 if (kthread && kthread_should_stop()) {
1308                         bch2_trans_exit(&trans);
1309                         return -ESHUTDOWN;
1310                 }
1311
1312                 bch2_trans_cond_resched(&trans);
1313
1314                 /*
1315                  * If the parent node wasn't relocked, it might have been split
1316                  * and the nodes in our sliding window might not have the same
1317                  * parent anymore - blow away the sliding window:
1318                  */
1319                 if (btree_iter_node(iter, iter->level + 1) &&
1320                     !btree_node_intent_locked(iter, iter->level + 1))
1321                         memset(merge + 1, 0,
1322                                (GC_MERGE_NODES - 1) * sizeof(merge[0]));
1323         }
1324         return bch2_trans_exit(&trans);
1325 }
1326
1327 /**
1328  * bch_coalesce - coalesce adjacent nodes with low occupancy
1329  */
1330 void bch2_coalesce(struct bch_fs *c)
1331 {
1332         enum btree_id id;
1333
1334         down_read(&c->gc_lock);
1335         trace_gc_coalesce_start(c);
1336
1337         for (id = 0; id < BTREE_ID_NR; id++) {
1338                 int ret = c->btree_roots[id].b
1339                         ? bch2_coalesce_btree(c, id)
1340                         : 0;
1341
1342                 if (ret) {
1343                         if (ret != -ESHUTDOWN)
1344                                 bch_err(c, "btree coalescing failed: %d", ret);
1345                         return;
1346                 }
1347         }
1348
1349         trace_gc_coalesce_end(c);
1350         up_read(&c->gc_lock);
1351 }
1352
1353 static int bch2_gc_thread(void *arg)
1354 {
1355         struct bch_fs *c = arg;
1356         struct io_clock *clock = &c->io_clock[WRITE];
1357         unsigned long last = atomic_long_read(&clock->now);
1358         unsigned last_kick = atomic_read(&c->kick_gc);
1359         int ret;
1360
1361         set_freezable();
1362
1363         while (1) {
1364                 while (1) {
1365                         set_current_state(TASK_INTERRUPTIBLE);
1366
1367                         if (kthread_should_stop()) {
1368                                 __set_current_state(TASK_RUNNING);
1369                                 return 0;
1370                         }
1371
1372                         if (atomic_read(&c->kick_gc) != last_kick)
1373                                 break;
1374
1375                         if (c->btree_gc_periodic) {
1376                                 unsigned long next = last + c->capacity / 16;
1377
1378                                 if (atomic_long_read(&clock->now) >= next)
1379                                         break;
1380
1381                                 bch2_io_clock_schedule_timeout(clock, next);
1382                         } else {
1383                                 schedule();
1384                         }
1385
1386                         try_to_freeze();
1387                 }
1388                 __set_current_state(TASK_RUNNING);
1389
1390                 last = atomic_long_read(&clock->now);
1391                 last_kick = atomic_read(&c->kick_gc);
1392
1393                 /*
1394                  * Full gc is currently incompatible with btree key cache:
1395                  */
1396 #if 0
1397                 ret = bch2_gc(c, NULL, false, false);
1398 #else
1399                 ret = bch2_gc_gens(c);
1400 #endif
1401                 if (ret < 0)
1402                         bch_err(c, "btree gc failed: %i", ret);
1403
1404                 debug_check_no_locks_held();
1405         }
1406
1407         return 0;
1408 }
1409
1410 void bch2_gc_thread_stop(struct bch_fs *c)
1411 {
1412         struct task_struct *p;
1413
1414         p = c->gc_thread;
1415         c->gc_thread = NULL;
1416
1417         if (p) {
1418                 kthread_stop(p);
1419                 put_task_struct(p);
1420         }
1421 }
1422
1423 int bch2_gc_thread_start(struct bch_fs *c)
1424 {
1425         struct task_struct *p;
1426
1427         BUG_ON(c->gc_thread);
1428
1429         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
1430         if (IS_ERR(p))
1431                 return PTR_ERR(p);
1432
1433         get_task_struct(p);
1434         c->gc_thread = p;
1435         wake_up_process(p);
1436         return 0;
1437 }