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->c.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->c.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->c.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->c.level > target_depth) {
 300                                 child = bch2_btree_node_get_noiter(c, &tmp.k,
 301                                                         b->c.btree_id, b->c.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->c.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->c.lock, NULL, NULL);
 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->c.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->c.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         lockdep_assert_held(&c->state_lock);
 802         trace_gc_start(c);
 803
 804         down_write(&c->gc_lock);
 805
 806         /* flush interior btree updates: */
 807         closure_wait_event(&c->btree_interior_update_wait,
 808                            !bch2_btree_interior_updates_nr_pending(c));
 809 again:
 810         ret = bch2_gc_start(c, metadata_only);
 811         if (ret)
 812                 goto out;
 813
 814         bch2_mark_superblocks(c);
 815
 816         ret = bch2_gc_btrees(c, journal_keys, initial, metadata_only);
 817         if (ret)
 818                 goto out;
 819
 820 #if 0
 821         bch2_mark_pending_btree_node_frees(c);
 822 #endif
 823         bch2_mark_allocator_buckets(c);
 824
 825         c->gc_count++;
 826 out:
 827         if (!ret &&
 828             (test_bit(BCH_FS_FIXED_GENS, &c->flags) ||
 829              (!iter && test_restart_gc(c)))) {
 830                 /*
 831                  * XXX: make sure gens we fixed got saved
 832                  */
 833                 if (iter++ <= 2) {
 834                         bch_info(c, "Fixed gens, restarting mark and sweep:");
 835                         clear_bit(BCH_FS_FIXED_GENS, &c->flags);
 836                         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
 837
 838                         percpu_down_write(&c->mark_lock);
 839                         bch2_gc_free(c);
 840                         percpu_up_write(&c->mark_lock);
 841                         /* flush fsck errors, reset counters */
 842                         bch2_flush_fsck_errs(c);
 843
 844                         goto again;
 845                 }
 846
 847                 bch_info(c, "Unable to fix bucket gens, looping");
 848                 ret = -EINVAL;
 849         }
 850
 851         if (!ret) {
 852                 bch2_journal_block(&c->journal);
 853
 854                 percpu_down_write(&c->mark_lock);
 855                 ret = bch2_gc_done(c, initial, metadata_only);
 856
 857                 bch2_journal_unblock(&c->journal);
 858         } else {
 859                 percpu_down_write(&c->mark_lock);
 860         }
 861
 862         /* Indicates that gc is no longer in progress: */
 863         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
 864
 865         bch2_gc_free(c);
 866         percpu_up_write(&c->mark_lock);
 867
 868         up_write(&c->gc_lock);
 869
 870         trace_gc_end(c);
 871         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
 872
 873         /*
 874          * Wake up allocator in case it was waiting for buckets
 875          * because of not being able to inc gens
 876          */
 877         for_each_member_device(ca, c, i)
 878                 bch2_wake_allocator(ca);
 879
 880         /*
 881          * At startup, allocations can happen directly instead of via the
 882          * allocator thread - issue wakeup in case they blocked on gc_lock:
 883          */
 884         closure_wake_up(&c->freelist_wait);
 885         return ret;
 886 }
 887
 888 /*
 889  * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
 890  * node pointers currently never have cached pointers that can become stale:
 891  */
 892 static int bch2_gc_btree_gens(struct bch_fs *c, enum btree_id id)
 893 {
 894         struct btree_trans trans;
 895         struct btree_iter *iter;
 896         struct bkey_s_c k;
 897         int ret;
 898
 899         bch2_trans_init(&trans, c, 0, 0);
 900
 901         for_each_btree_key(&trans, iter, id, POS_MIN, BTREE_ITER_PREFETCH, k, ret) {
 902                 struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 903                 const struct bch_extent_ptr *ptr;
 904
 905                 bkey_for_each_ptr(ptrs, ptr) {
 906                         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 907                         struct bucket *g = PTR_BUCKET(ca, ptr, false);
 908
 909                         if (gen_after(g->gc_gen, ptr->gen))
 910                                 g->gc_gen = ptr->gen;
 911
 912                         if (gen_after(g->mark.gen, ptr->gen) > 32) {
 913                                 /* rewrite btree node */
 914
 915                         }
 916                 }
 917         }
 918
 919         bch2_trans_exit(&trans);
 920         return ret;
 921 }
 922
 923 int bch2_gc_gens(struct bch_fs *c)
 924 {
 925         struct bch_dev *ca;
 926         unsigned i;
 927         int ret;
 928
 929         down_read(&c->state_lock);
 930
 931         for_each_member_device(ca, c, i) {
 932                 struct bucket_array *buckets = bucket_array(ca);
 933                 struct bucket *g;
 934
 935                 for_each_bucket(g, buckets)
 936                         g->gc_gen = g->mark.gen;
 937         }
 938
 939         for (i = 0; i < BTREE_ID_NR; i++)
 940                 if (btree_node_type_needs_gc(i)) {
 941                         ret = bch2_gc_btree_gens(c, i);
 942                         if (ret)
 943                                 goto err;
 944                 }
 945
 946         for_each_member_device(ca, c, i) {
 947                 struct bucket_array *buckets = bucket_array(ca);
 948                 struct bucket *g;
 949
 950                 for_each_bucket(g, buckets)
 951                         g->oldest_gen = g->gc_gen;
 952         }
 953 err:
 954         up_read(&c->state_lock);
 955         return ret;
 956 }
 957
 958 /* Btree coalescing */
 959
 960 static void recalc_packed_keys(struct btree *b)
 961 {
 962         struct bset *i = btree_bset_first(b);
 963         struct bkey_packed *k;
 964
 965         memset(&b->nr, 0, sizeof(b->nr));
 966
 967         BUG_ON(b->nsets != 1);
 968
 969         vstruct_for_each(i, k)
 970                 btree_keys_account_key_add(&b->nr, 0, k);
 971 }
 972
 973 static void bch2_coalesce_nodes(struct bch_fs *c, struct btree_iter *iter,
 974                                 struct btree *old_nodes[GC_MERGE_NODES])
 975 {
 976         struct btree *parent = btree_node_parent(iter, old_nodes[0]);
 977         unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
 978         unsigned blocks = btree_blocks(c) * 2 / 3;
 979         struct btree *new_nodes[GC_MERGE_NODES];
 980         struct btree_update *as;
 981         struct keylist keylist;
 982         struct bkey_format_state format_state;
 983         struct bkey_format new_format;
 984
 985         memset(new_nodes, 0, sizeof(new_nodes));
 986         bch2_keylist_init(&keylist, NULL);
 987
 988         /* Count keys that are not deleted */
 989         for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
 990                 u64s += old_nodes[i]->nr.live_u64s;
 991
 992         nr_old_nodes = nr_new_nodes = i;
 993
 994         /* Check if all keys in @old_nodes could fit in one fewer node */
 995         if (nr_old_nodes <= 1 ||
 996             __vstruct_blocks(struct btree_node, c->block_bits,
 997                              DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
 998                 return;
 999
1000         /* Find a format that all keys in @old_nodes can pack into */
1001         bch2_bkey_format_init(&format_state);
1002
1003         for (i = 0; i < nr_old_nodes; i++)
1004                 __bch2_btree_calc_format(&format_state, old_nodes[i]);
1005
1006         new_format = bch2_bkey_format_done(&format_state);
1007
1008         /* Check if repacking would make any nodes too big to fit */
1009         for (i = 0; i < nr_old_nodes; i++)
1010                 if (!bch2_btree_node_format_fits(c, old_nodes[i], &new_format)) {
1011                         trace_btree_gc_coalesce_fail(c,
1012                                         BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
1013                         return;
1014                 }
1015
1016         if (bch2_keylist_realloc(&keylist, NULL, 0,
1017                         (BKEY_U64s + BKEY_EXTENT_U64s_MAX) * nr_old_nodes)) {
1018                 trace_btree_gc_coalesce_fail(c,
1019                                 BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
1020                 return;
1021         }
1022
1023         as = bch2_btree_update_start(iter->trans, iter->btree_id,
1024                         btree_update_reserve_required(c, parent) + nr_old_nodes,
1025                         BTREE_INSERT_NOFAIL|
1026                         BTREE_INSERT_USE_RESERVE,
1027                         NULL);
1028         if (IS_ERR(as)) {
1029                 trace_btree_gc_coalesce_fail(c,
1030                                 BTREE_GC_COALESCE_FAIL_RESERVE_GET);
1031                 bch2_keylist_free(&keylist, NULL);
1032                 return;
1033         }
1034
1035         trace_btree_gc_coalesce(c, old_nodes[0]);
1036
1037         for (i = 0; i < nr_old_nodes; i++)
1038                 bch2_btree_interior_update_will_free_node(as, old_nodes[i]);
1039
1040         /* Repack everything with @new_format and sort down to one bset */
1041         for (i = 0; i < nr_old_nodes; i++)
1042                 new_nodes[i] =
1043                         __bch2_btree_node_alloc_replacement(as, old_nodes[i],
1044                                                             new_format);
1045
1046         /*
1047          * Conceptually we concatenate the nodes together and slice them
1048          * up at different boundaries.
1049          */
1050         for (i = nr_new_nodes - 1; i > 0; --i) {
1051                 struct btree *n1 = new_nodes[i];
1052                 struct btree *n2 = new_nodes[i - 1];
1053
1054                 struct bset *s1 = btree_bset_first(n1);
1055                 struct bset *s2 = btree_bset_first(n2);
1056                 struct bkey_packed *k, *last = NULL;
1057
1058                 /* Calculate how many keys from @n2 we could fit inside @n1 */
1059                 u64s = 0;
1060
1061                 for (k = s2->start;
1062                      k < vstruct_last(s2) &&
1063                      vstruct_blocks_plus(n1->data, c->block_bits,
1064                                          u64s + k->u64s) <= blocks;
1065                      k = bkey_next_skip_noops(k, vstruct_last(s2))) {
1066                         last = k;
1067                         u64s += k->u64s;
1068                 }
1069
1070                 if (u64s == le16_to_cpu(s2->u64s)) {
1071                         /* n2 fits entirely in n1 */
1072                         n1->key.k.p = n1->data->max_key = n2->data->max_key;
1073
1074                         memcpy_u64s(vstruct_last(s1),
1075                                     s2->start,
1076                                     le16_to_cpu(s2->u64s));
1077                         le16_add_cpu(&s1->u64s, le16_to_cpu(s2->u64s));
1078
1079                         set_btree_bset_end(n1, n1->set);
1080
1081                         six_unlock_write(&n2->c.lock);
1082                         bch2_btree_node_free_never_inserted(c, n2);
1083                         six_unlock_intent(&n2->c.lock);
1084
1085                         memmove(new_nodes + i - 1,
1086                                 new_nodes + i,
1087                                 sizeof(new_nodes[0]) * (nr_new_nodes - i));
1088                         new_nodes[--nr_new_nodes] = NULL;
1089                 } else if (u64s) {
1090                         /* move part of n2 into n1 */
1091                         n1->key.k.p = n1->data->max_key =
1092                                 bkey_unpack_pos(n1, last);
1093
1094                         n2->data->min_key = bkey_successor(n1->data->max_key);
1095
1096                         memcpy_u64s(vstruct_last(s1),
1097                                     s2->start, u64s);
1098                         le16_add_cpu(&s1->u64s, u64s);
1099
1100                         memmove(s2->start,
1101                                 vstruct_idx(s2, u64s),
1102                                 (le16_to_cpu(s2->u64s) - u64s) * sizeof(u64));
1103                         s2->u64s = cpu_to_le16(le16_to_cpu(s2->u64s) - u64s);
1104
1105                         set_btree_bset_end(n1, n1->set);
1106                         set_btree_bset_end(n2, n2->set);
1107                 }
1108         }
1109
1110         for (i = 0; i < nr_new_nodes; i++) {
1111                 struct btree *n = new_nodes[i];
1112
1113                 recalc_packed_keys(n);
1114                 btree_node_reset_sib_u64s(n);
1115
1116                 bch2_btree_build_aux_trees(n);
1117
1118                 bch2_btree_update_add_new_node(as, n);
1119                 six_unlock_write(&n->c.lock);
1120
1121                 bch2_btree_node_write(c, n, SIX_LOCK_intent);
1122         }
1123
1124         /*
1125          * The keys for the old nodes get deleted. We don't want to insert keys
1126          * that compare equal to the keys for the new nodes we'll also be
1127          * inserting - we can't because keys on a keylist must be strictly
1128          * greater than the previous keys, and we also don't need to since the
1129          * key for the new node will serve the same purpose (overwriting the key
1130          * for the old node).
1131          */
1132         for (i = 0; i < nr_old_nodes; i++) {
1133                 struct bkey_i delete;
1134                 unsigned j;
1135
1136                 for (j = 0; j < nr_new_nodes; j++)
1137                         if (!bkey_cmp(old_nodes[i]->key.k.p,
1138                                       new_nodes[j]->key.k.p))
1139                                 goto next;
1140
1141                 bkey_init(&delete.k);
1142                 delete.k.p = old_nodes[i]->key.k.p;
1143                 bch2_keylist_add_in_order(&keylist, &delete);
1144 next:
1145                 i = i;
1146         }
1147
1148         /*
1149          * Keys for the new nodes get inserted: bch2_btree_insert_keys() only
1150          * does the lookup once and thus expects the keys to be in sorted order
1151          * so we have to make sure the new keys are correctly ordered with
1152          * respect to the deleted keys added in the previous loop
1153          */
1154         for (i = 0; i < nr_new_nodes; i++)
1155                 bch2_keylist_add_in_order(&keylist, &new_nodes[i]->key);
1156
1157         /* Insert the newly coalesced nodes */
1158         bch2_btree_insert_node(as, parent, iter, &keylist, 0);
1159
1160         BUG_ON(!bch2_keylist_empty(&keylist));
1161
1162         BUG_ON(iter->l[old_nodes[0]->c.level].b != old_nodes[0]);
1163
1164         bch2_btree_iter_node_replace(iter, new_nodes[0]);
1165
1166         for (i = 0; i < nr_new_nodes; i++)
1167                 bch2_btree_update_get_open_buckets(as, new_nodes[i]);
1168
1169         /* Free the old nodes and update our sliding window */
1170         for (i = 0; i < nr_old_nodes; i++) {
1171                 bch2_btree_node_free_inmem(c, old_nodes[i], iter);
1172
1173                 /*
1174                  * the index update might have triggered a split, in which case
1175                  * the nodes we coalesced - the new nodes we just created -
1176                  * might not be sibling nodes anymore - don't add them to the
1177                  * sliding window (except the first):
1178                  */
1179                 if (!i) {
1180                         old_nodes[i] = new_nodes[i];
1181                 } else {
1182                         old_nodes[i] = NULL;
1183                 }
1184         }
1185
1186         for (i = 0; i < nr_new_nodes; i++)
1187                 six_unlock_intent(&new_nodes[i]->c.lock);
1188
1189         bch2_btree_update_done(as);
1190         bch2_keylist_free(&keylist, NULL);
1191 }
1192
1193 static int bch2_coalesce_btree(struct bch_fs *c, enum btree_id btree_id)
1194 {
1195         struct btree_trans trans;
1196         struct btree_iter *iter;
1197         struct btree *b;
1198         bool kthread = (current->flags & PF_KTHREAD) != 0;
1199         unsigned i;
1200
1201         /* Sliding window of adjacent btree nodes */
1202         struct btree *merge[GC_MERGE_NODES];
1203         u32 lock_seq[GC_MERGE_NODES];
1204
1205         bch2_trans_init(&trans, c, 0, 0);
1206
1207         /*
1208          * XXX: We don't have a good way of positively matching on sibling nodes
1209          * that have the same parent - this code works by handling the cases
1210          * where they might not have the same parent, and is thus fragile. Ugh.
1211          *
1212          * Perhaps redo this to use multiple linked iterators?
1213          */
1214         memset(merge, 0, sizeof(merge));
1215
1216         __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
1217                               BTREE_MAX_DEPTH, 0,
1218                               BTREE_ITER_PREFETCH, b) {
1219                 memmove(merge + 1, merge,
1220                         sizeof(merge) - sizeof(merge[0]));
1221                 memmove(lock_seq + 1, lock_seq,
1222                         sizeof(lock_seq) - sizeof(lock_seq[0]));
1223
1224                 merge[0] = b;
1225
1226                 for (i = 1; i < GC_MERGE_NODES; i++) {
1227                         if (!merge[i] ||
1228                             !six_relock_intent(&merge[i]->c.lock, lock_seq[i]))
1229                                 break;
1230
1231                         if (merge[i]->c.level != merge[0]->c.level) {
1232                                 six_unlock_intent(&merge[i]->c.lock);
1233                                 break;
1234                         }
1235                 }
1236                 memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));
1237
1238                 bch2_coalesce_nodes(c, iter, merge);
1239
1240                 for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
1241                         lock_seq[i] = merge[i]->c.lock.state.seq;
1242                         six_unlock_intent(&merge[i]->c.lock);
1243                 }
1244
1245                 lock_seq[0] = merge[0]->c.lock.state.seq;
1246
1247                 if (kthread && kthread_should_stop()) {
1248                         bch2_trans_exit(&trans);
1249                         return -ESHUTDOWN;
1250                 }
1251
1252                 bch2_trans_cond_resched(&trans);
1253
1254                 /*
1255                  * If the parent node wasn't relocked, it might have been split
1256                  * and the nodes in our sliding window might not have the same
1257                  * parent anymore - blow away the sliding window:
1258                  */
1259                 if (btree_iter_node(iter, iter->level + 1) &&
1260                     !btree_node_intent_locked(iter, iter->level + 1))
1261                         memset(merge + 1, 0,
1262                                (GC_MERGE_NODES - 1) * sizeof(merge[0]));
1263         }
1264         return bch2_trans_exit(&trans);
1265 }
1266
1267 /**
1268  * bch_coalesce - coalesce adjacent nodes with low occupancy
1269  */
1270 void bch2_coalesce(struct bch_fs *c)
1271 {
1272         enum btree_id id;
1273
1274         down_read(&c->gc_lock);
1275         trace_gc_coalesce_start(c);
1276
1277         for (id = 0; id < BTREE_ID_NR; id++) {
1278                 int ret = c->btree_roots[id].b
1279                         ? bch2_coalesce_btree(c, id)
1280                         : 0;
1281
1282                 if (ret) {
1283                         if (ret != -ESHUTDOWN)
1284                                 bch_err(c, "btree coalescing failed: %d", ret);
1285                         return;
1286                 }
1287         }
1288
1289         trace_gc_coalesce_end(c);
1290         up_read(&c->gc_lock);
1291 }
1292
1293 static int bch2_gc_thread(void *arg)
1294 {
1295         struct bch_fs *c = arg;
1296         struct io_clock *clock = &c->io_clock[WRITE];
1297         unsigned long last = atomic_long_read(&clock->now);
1298         unsigned last_kick = atomic_read(&c->kick_gc);
1299         int ret;
1300
1301         set_freezable();
1302
1303         while (1) {
1304                 while (1) {
1305                         set_current_state(TASK_INTERRUPTIBLE);
1306
1307                         if (kthread_should_stop()) {
1308                                 __set_current_state(TASK_RUNNING);
1309                                 return 0;
1310                         }
1311
1312                         if (atomic_read(&c->kick_gc) != last_kick)
1313                                 break;
1314
1315                         if (c->btree_gc_periodic) {
1316                                 unsigned long next = last + c->capacity / 16;
1317
1318                                 if (atomic_long_read(&clock->now) >= next)
1319                                         break;
1320
1321                                 bch2_io_clock_schedule_timeout(clock, next);
1322                         } else {
1323                                 schedule();
1324                         }
1325
1326                         try_to_freeze();
1327                 }
1328                 __set_current_state(TASK_RUNNING);
1329
1330                 last = atomic_long_read(&clock->now);
1331                 last_kick = atomic_read(&c->kick_gc);
1332
1333                 /*
1334                  * Full gc is currently incompatible with btree key cache:
1335                  */
1336 #if 0
1337                 ret = bch2_gc(c, NULL, false, false);
1338 #else
1339                 ret = bch2_gc_gens(c);
1340 #endif
1341                 if (ret)
1342                         bch_err(c, "btree gc failed: %i", ret);
1343
1344                 debug_check_no_locks_held();
1345         }
1346
1347         return 0;
1348 }
1349
1350 void bch2_gc_thread_stop(struct bch_fs *c)
1351 {
1352         struct task_struct *p;
1353
1354         p = c->gc_thread;
1355         c->gc_thread = NULL;
1356
1357         if (p) {
1358                 kthread_stop(p);
1359                 put_task_struct(p);
1360         }
1361 }
1362
1363 int bch2_gc_thread_start(struct bch_fs *c)
1364 {
1365         struct task_struct *p;
1366
1367         BUG_ON(c->gc_thread);
1368
1369         p = kthread_create(bch2_gc_thread, c, "bch_gc");
1370         if (IS_ERR(p))
1371                 return PTR_ERR(p);
1372
1373         get_task_struct(p);
1374         c->gc_thread = p;
1375         wake_up_process(p);
1376         return 0;
1377 }