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