fs/bcachefs/btree_gc.c

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