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 "bkey_buf.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
  29 #include <linux/slab.h>
  30 #include <linux/bitops.h>
  31 #include <linux/freezer.h>
  32 #include <linux/kthread.h>
  33 #include <linux/preempt.h>
  34 #include <linux/rcupdate.h>
  35 #include <linux/sched/task.h>
  36 #include <trace/events/bcachefs.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                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 136                         }
 137
 138                         if (mustfix_fsck_err_on(gen_cmp(ptr->gen, g->mark.gen) > 0, c,
 139                                         "bucket %u:%zu data type %s ptr gen in the future: %u > %u",
 140                                         ptr->dev, PTR_BUCKET_NR(ca, ptr),
 141                                         bch2_data_types[ptr_data_type(k.k, ptr)],
 142                                         ptr->gen, g->mark.gen)) {
 143                                 g2->_mark.gen   = g->_mark.gen          = ptr->gen;
 144                                 g2->gen_valid   = g->gen_valid          = true;
 145                                 g2->_mark.data_type             = 0;
 146                                 g2->_mark.dirty_sectors         = 0;
 147                                 g2->_mark.cached_sectors        = 0;
 148                                 set_bit(BCH_FS_FIXED_GENS, &c->flags);
 149                                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);
 150                         }
 151                 }
 152         }
 153
 154         bkey_for_each_ptr(ptrs, ptr) {
 155                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 156                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 157
 158                 if (gen_after(g->oldest_gen, ptr->gen))
 159                         g->oldest_gen = ptr->gen;
 160
 161                 *max_stale = max(*max_stale, ptr_stale(ca, ptr));
 162         }
 163
 164         bch2_mark_key(c, k, 0, k.k->size, NULL, 0, flags);
 165 fsck_err:
 166         return ret;
 167 }
 168
 169 static int btree_gc_mark_node(struct bch_fs *c, struct btree *b, u8 *max_stale,
 170                               bool initial)
 171 {
 172         struct bpos next_node_start = b->data->min_key;
 173         struct btree_node_iter iter;
 174         struct bkey unpacked;
 175         struct bkey_s_c k;
 176         int ret = 0;
 177
 178         *max_stale = 0;
 179
 180         if (!btree_node_type_needs_gc(btree_node_type(b)))
 181                 return 0;
 182
 183         bch2_btree_node_iter_init_from_start(&iter, b);
 184
 185         while ((k = bch2_btree_node_iter_peek_unpack(&iter, b, &unpacked)).k) {
 186                 bch2_bkey_debugcheck(c, b, k);
 187
 188                 ret = bch2_gc_mark_key(c, k, max_stale, initial);
 189                 if (ret)
 190                         break;
 191
 192                 bch2_btree_node_iter_advance(&iter, b);
 193
 194                 if (b->c.level) {
 195                         ret = bch2_gc_check_topology(c, k,
 196                                         &next_node_start,
 197                                         b->data->max_key,
 198                                         bch2_btree_node_iter_end(&iter));
 199                         if (ret)
 200                                 break;
 201                 }
 202         }
 203
 204         return ret;
 205 }
 206
 207 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id,
 208                          bool initial)
 209 {
 210         struct btree_trans trans;
 211         struct btree_iter *iter;
 212         struct btree *b;
 213         unsigned depth = bch2_expensive_debug_checks    ? 0
 214                 : !btree_node_type_needs_gc(btree_id)   ? 1
 215                 : 0;
 216         u8 max_stale = 0;
 217         int ret = 0;
 218
 219         bch2_trans_init(&trans, c, 0, 0);
 220
 221         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 222
 223         __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
 224                               0, depth, BTREE_ITER_PREFETCH, b) {
 225                 bch2_verify_btree_nr_keys(b);
 226
 227                 gc_pos_set(c, gc_pos_btree_node(b));
 228
 229                 ret = btree_gc_mark_node(c, b, &max_stale, initial);
 230                 if (ret)
 231                         break;
 232
 233                 if (!initial) {
 234                         if (max_stale > 64)
 235                                 bch2_btree_node_rewrite(c, iter,
 236                                                 b->data->keys.seq,
 237                                                 BTREE_INSERT_NOWAIT|
 238                                                 BTREE_INSERT_GC_LOCK_HELD);
 239                         else if (!bch2_btree_gc_rewrite_disabled &&
 240                                  (bch2_btree_gc_always_rewrite || max_stale > 16))
 241                                 bch2_btree_node_rewrite(c, iter,
 242                                                 b->data->keys.seq,
 243                                                 BTREE_INSERT_NOWAIT|
 244                                                 BTREE_INSERT_GC_LOCK_HELD);
 245                 }
 246
 247                 bch2_trans_cond_resched(&trans);
 248         }
 249         ret = bch2_trans_exit(&trans) ?: ret;
 250         if (ret)
 251                 return ret;
 252
 253         mutex_lock(&c->btree_root_lock);
 254         b = c->btree_roots[btree_id].b;
 255         if (!btree_node_fake(b))
 256                 ret = bch2_gc_mark_key(c, bkey_i_to_s_c(&b->key),
 257                                        &max_stale, initial);
 258         gc_pos_set(c, gc_pos_btree_root(b->c.btree_id));
 259         mutex_unlock(&c->btree_root_lock);
 260
 261         return ret;
 262 }
 263
 264 static int bch2_gc_btree_init_recurse(struct bch_fs *c, struct btree *b,
 265                                       struct journal_keys *journal_keys,
 266                                       unsigned target_depth)
 267 {
 268         struct btree_and_journal_iter iter;
 269         struct bkey_s_c k;
 270         struct bpos next_node_start = b->data->min_key;
 271         struct bkey_buf tmp;
 272         u8 max_stale = 0;
 273         int ret = 0;
 274
 275         bch2_btree_and_journal_iter_init_node_iter(&iter, journal_keys, b);
 276         bch2_bkey_buf_init(&tmp);
 277
 278         while ((k = bch2_btree_and_journal_iter_peek(&iter)).k) {
 279                 bch2_bkey_debugcheck(c, b, k);
 280
 281                 BUG_ON(bkey_cmp(k.k->p, b->data->min_key) < 0);
 282                 BUG_ON(bkey_cmp(k.k->p, b->data->max_key) > 0);
 283
 284                 ret = bch2_gc_mark_key(c, k, &max_stale, true);
 285                 if (ret)
 286                         break;
 287
 288                 if (b->c.level) {
 289                         struct btree *child;
 290
 291                         bch2_bkey_buf_reassemble(&tmp, c, k);
 292                         k = bkey_i_to_s_c(tmp.k);
 293
 294                         bch2_btree_and_journal_iter_advance(&iter);
 295
 296                         ret = bch2_gc_check_topology(c, k,
 297                                         &next_node_start,
 298                                         b->data->max_key,
 299                                         !bch2_btree_and_journal_iter_peek(&iter).k);
 300                         if (ret)
 301                                 break;
 302
 303                         if (b->c.level > target_depth) {
 304                                 child = bch2_btree_node_get_noiter(c, tmp.k,
 305                                                         b->c.btree_id, b->c.level - 1);
 306                                 ret = PTR_ERR_OR_ZERO(child);
 307                                 if (ret)
 308                                         break;
 309
 310                                 ret = bch2_gc_btree_init_recurse(c, child,
 311                                                 journal_keys, target_depth);
 312                                 six_unlock_read(&child->c.lock);
 313
 314                                 if (ret)
 315                                         break;
 316                         }
 317                 } else {
 318                         bch2_btree_and_journal_iter_advance(&iter);
 319                 }
 320         }
 321
 322         bch2_bkey_buf_exit(&tmp, c);
 323         return ret;
 324 }
 325
 326 static int bch2_gc_btree_init(struct bch_fs *c,
 327                               struct journal_keys *journal_keys,
 328                               enum btree_id btree_id)
 329 {
 330         struct btree *b;
 331         unsigned target_depth = bch2_expensive_debug_checks     ? 0
 332                 : !btree_node_type_needs_gc(btree_id)           ? 1
 333                 : 0;
 334         u8 max_stale = 0;
 335         int ret = 0;
 336
 337         b = c->btree_roots[btree_id].b;
 338
 339         if (btree_node_fake(b))
 340                 return 0;
 341
 342         six_lock_read(&b->c.lock, NULL, NULL);
 343         if (fsck_err_on(bkey_cmp(b->data->min_key, POS_MIN), c,
 344                         "btree root with incorrect min_key: %llu:%llu",
 345                         b->data->min_key.inode,
 346                         b->data->min_key.offset)) {
 347                 BUG();
 348         }
 349
 350         if (fsck_err_on(bkey_cmp(b->data->max_key, POS_MAX), c,
 351                         "btree root with incorrect min_key: %llu:%llu",
 352                         b->data->max_key.inode,
 353                         b->data->max_key.offset)) {
 354                 BUG();
 355         }
 356
 357         if (b->c.level >= target_depth)
 358                 ret = bch2_gc_btree_init_recurse(c, b,
 359                                         journal_keys, target_depth);
 360
 361         if (!ret)
 362                 ret = bch2_gc_mark_key(c, bkey_i_to_s_c(&b->key),
 363                                        &max_stale, true);
 364 fsck_err:
 365         six_unlock_read(&b->c.lock);
 366
 367         return ret;
 368 }
 369
 370 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
 371 {
 372         return  (int) btree_id_to_gc_phase(l) -
 373                 (int) btree_id_to_gc_phase(r);
 374 }
 375
 376 static int bch2_gc_btrees(struct bch_fs *c, struct journal_keys *journal_keys,
 377                           bool initial)
 378 {
 379         enum btree_id ids[BTREE_ID_NR];
 380         unsigned i;
 381
 382         for (i = 0; i < BTREE_ID_NR; i++)
 383                 ids[i] = i;
 384         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
 385
 386         for (i = 0; i < BTREE_ID_NR; i++) {
 387                 enum btree_id id = ids[i];
 388                 int ret = initial
 389                         ? bch2_gc_btree_init(c, journal_keys,
 390                                              id)
 391                         : bch2_gc_btree(c, id, initial);
 392                 if (ret)
 393                         return ret;
 394         }
 395
 396         return 0;
 397 }
 398
 399 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
 400                                   u64 start, u64 end,
 401                                   enum bch_data_type type,
 402                                   unsigned flags)
 403 {
 404         u64 b = sector_to_bucket(ca, start);
 405
 406         do {
 407                 unsigned sectors =
 408                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
 409
 410                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
 411                                           gc_phase(GC_PHASE_SB), flags);
 412                 b++;
 413                 start += sectors;
 414         } while (start < end);
 415 }
 416
 417 void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
 418                               unsigned flags)
 419 {
 420         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
 421         unsigned i;
 422         u64 b;
 423
 424         /*
 425          * This conditional is kind of gross, but we may be called from the
 426          * device add path, before the new device has actually been added to the
 427          * running filesystem:
 428          */
 429         if (c) {
 430                 lockdep_assert_held(&c->sb_lock);
 431                 percpu_down_read(&c->mark_lock);
 432         }
 433
 434         for (i = 0; i < layout->nr_superblocks; i++) {
 435                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
 436
 437                 if (offset == BCH_SB_SECTOR)
 438                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
 439                                               BCH_DATA_sb, flags);
 440
 441                 mark_metadata_sectors(c, ca, offset,
 442                                       offset + (1 << layout->sb_max_size_bits),
 443                                       BCH_DATA_sb, flags);
 444         }
 445
 446         for (i = 0; i < ca->journal.nr; i++) {
 447                 b = ca->journal.buckets[i];
 448                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_journal,
 449                                           ca->mi.bucket_size,
 450                                           gc_phase(GC_PHASE_SB), flags);
 451         }
 452
 453         if (c)
 454                 percpu_up_read(&c->mark_lock);
 455 }
 456
 457 static void bch2_mark_superblocks(struct bch_fs *c)
 458 {
 459         struct bch_dev *ca;
 460         unsigned i;
 461
 462         mutex_lock(&c->sb_lock);
 463         gc_pos_set(c, gc_phase(GC_PHASE_SB));
 464
 465         for_each_online_member(ca, c, i)
 466                 bch2_mark_dev_superblock(c, ca, BTREE_TRIGGER_GC);
 467         mutex_unlock(&c->sb_lock);
 468 }
 469
 470 #if 0
 471 /* Also see bch2_pending_btree_node_free_insert_done() */
 472 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
 473 {
 474         struct btree_update *as;
 475         struct pending_btree_node_free *d;
 476
 477         mutex_lock(&c->btree_interior_update_lock);
 478         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
 479
 480         for_each_pending_btree_node_free(c, as, d)
 481                 if (d->index_update_done)
 482                         bch2_mark_key(c, bkey_i_to_s_c(&d->key),
 483                                       0, 0, NULL, 0,
 484                                       BTREE_TRIGGER_GC);
 485
 486         mutex_unlock(&c->btree_interior_update_lock);
 487 }
 488 #endif
 489
 490 static void bch2_mark_allocator_buckets(struct bch_fs *c)
 491 {
 492         struct bch_dev *ca;
 493         struct open_bucket *ob;
 494         size_t i, j, iter;
 495         unsigned ci;
 496
 497         percpu_down_read(&c->mark_lock);
 498
 499         spin_lock(&c->freelist_lock);
 500         gc_pos_set(c, gc_pos_alloc(c, NULL));
 501
 502         for_each_member_device(ca, c, ci) {
 503                 fifo_for_each_entry(i, &ca->free_inc, iter)
 504                         bch2_mark_alloc_bucket(c, ca, i, true,
 505                                                gc_pos_alloc(c, NULL),
 506                                                BTREE_TRIGGER_GC);
 507
 508
 509
 510                 for (j = 0; j < RESERVE_NR; j++)
 511                         fifo_for_each_entry(i, &ca->free[j], iter)
 512                                 bch2_mark_alloc_bucket(c, ca, i, true,
 513                                                        gc_pos_alloc(c, NULL),
 514                                                        BTREE_TRIGGER_GC);
 515         }
 516
 517         spin_unlock(&c->freelist_lock);
 518
 519         for (ob = c->open_buckets;
 520              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
 521              ob++) {
 522                 spin_lock(&ob->lock);
 523                 if (ob->valid) {
 524                         gc_pos_set(c, gc_pos_alloc(c, ob));
 525                         ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 526                         bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr), true,
 527                                                gc_pos_alloc(c, ob),
 528                                                BTREE_TRIGGER_GC);
 529                 }
 530                 spin_unlock(&ob->lock);
 531         }
 532
 533         percpu_up_read(&c->mark_lock);
 534 }
 535
 536 static void bch2_gc_free(struct bch_fs *c)
 537 {
 538         struct bch_dev *ca;
 539         unsigned i;
 540
 541         genradix_free(&c->stripes[1]);
 542
 543         for_each_member_device(ca, c, i) {
 544                 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
 545                         sizeof(struct bucket_array) +
 546                         ca->mi.nbuckets * sizeof(struct bucket));
 547                 ca->buckets[1] = NULL;
 548
 549                 free_percpu(ca->usage[1]);
 550                 ca->usage[1] = NULL;
 551         }
 552
 553         free_percpu(c->usage_gc);
 554         c->usage_gc = NULL;
 555 }
 556
 557 static int bch2_gc_done(struct bch_fs *c,
 558                         bool initial)
 559 {
 560         struct bch_dev *ca;
 561         bool verify = (!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                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
 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                                 iter.pos, ##__VA_ARGS__,                \
 580                                 dst->_f, src->_f);                      \
 581                 dst->_f = src->_f;                                      \
 582                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
 583         }
 584 #define copy_bucket_field(_f)                                           \
 585         if (dst->b[b].mark._f != src->b[b].mark._f) {                   \
 586                 if (verify)                                             \
 587                         fsck_err(c, "bucket %u:%zu gen %u data type %s has wrong " #_f  \
 588                                 ": got %u, should be %u", i, b,         \
 589                                 dst->b[b].mark.gen,                     \
 590                                 bch2_data_types[dst->b[b].mark.data_type],\
 591                                 dst->b[b].mark._f, src->b[b].mark._f);  \
 592                 dst->b[b]._mark._f = src->b[b].mark._f;                 \
 593                 set_bit(BCH_FS_NEED_ALLOC_WRITE, &c->flags);            \
 594         }
 595 #define copy_dev_field(_f, _msg, ...)                                   \
 596         copy_field(_f, "dev %u has wrong " _msg, i, ##__VA_ARGS__)
 597 #define copy_fs_field(_f, _msg, ...)                                    \
 598         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
 599
 600         {
 601                 struct genradix_iter iter = genradix_iter_init(&c->stripes[1], 0);
 602                 struct stripe *dst, *src;
 603
 604                 while ((src = genradix_iter_peek(&iter, &c->stripes[1]))) {
 605                         dst = genradix_ptr_alloc(&c->stripes[0], iter.pos, GFP_KERNEL);
 606
 607                         if (dst->alive          != src->alive ||
 608                             dst->sectors        != src->sectors ||
 609                             dst->algorithm      != src->algorithm ||
 610                             dst->nr_blocks      != src->nr_blocks ||
 611                             dst->nr_redundant   != src->nr_redundant) {
 612                                 bch_err(c, "unexpected stripe inconsistency at bch2_gc_done, confused");
 613                                 ret = -EINVAL;
 614                                 goto fsck_err;
 615                         }
 616
 617                         for (i = 0; i < ARRAY_SIZE(dst->block_sectors); i++)
 618                                 copy_stripe_field(block_sectors[i],
 619                                                   "block_sectors[%u]", i);
 620
 621                         dst->blocks_nonempty = 0;
 622                         for (i = 0; i < dst->nr_blocks; i++)
 623                                 dst->blocks_nonempty += dst->block_sectors[i] != 0;
 624
 625                         genradix_iter_advance(&iter, &c->stripes[1]);
 626                 }
 627         }
 628
 629         for_each_member_device(ca, c, i) {
 630                 struct bucket_array *dst = __bucket_array(ca, 0);
 631                 struct bucket_array *src = __bucket_array(ca, 1);
 632                 size_t b;
 633
 634                 for (b = 0; b < src->nbuckets; b++) {
 635                         copy_bucket_field(gen);
 636                         copy_bucket_field(data_type);
 637                         copy_bucket_field(owned_by_allocator);
 638                         copy_bucket_field(stripe);
 639                         copy_bucket_field(dirty_sectors);
 640                         copy_bucket_field(cached_sectors);
 641
 642                         dst->b[b].oldest_gen = src->b[b].oldest_gen;
 643                 }
 644         };
 645
 646         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 647                 bch2_fs_usage_acc_to_base(c, i);
 648
 649         bch2_dev_usage_from_buckets(c);
 650
 651         {
 652                 unsigned nr = fs_usage_u64s(c);
 653                 struct bch_fs_usage *dst = c->usage_base;
 654                 struct bch_fs_usage *src = (void *)
 655                         bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
 656
 657                 copy_fs_field(hidden,           "hidden");
 658                 copy_fs_field(btree,            "btree");
 659                 copy_fs_field(data,     "data");
 660                 copy_fs_field(cached,   "cached");
 661                 copy_fs_field(reserved, "reserved");
 662                 copy_fs_field(nr_inodes,"nr_inodes");
 663
 664                 for (i = 0; i < BCH_REPLICAS_MAX; i++)
 665                         copy_fs_field(persistent_reserved[i],
 666                                       "persistent_reserved[%i]", i);
 667
 668                 for (i = 0; i < c->replicas.nr; i++) {
 669                         struct bch_replicas_entry *e =
 670                                 cpu_replicas_entry(&c->replicas, i);
 671                         char buf[80];
 672
 673                         bch2_replicas_entry_to_text(&PBUF(buf), e);
 674
 675                         copy_fs_field(replicas[i], "%s", buf);
 676                 }
 677         }
 678
 679 #undef copy_fs_field
 680 #undef copy_dev_field
 681 #undef copy_bucket_field
 682 #undef copy_stripe_field
 683 #undef copy_field
 684 fsck_err:
 685         return ret;
 686 }
 687
 688 static int bch2_gc_start(struct bch_fs *c)
 689 {
 690         struct bch_dev *ca;
 691         unsigned i;
 692         int ret;
 693
 694         BUG_ON(c->usage_gc);
 695
 696         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
 697                                          sizeof(u64), GFP_KERNEL);
 698         if (!c->usage_gc) {
 699                 bch_err(c, "error allocating c->usage_gc");
 700                 return -ENOMEM;
 701         }
 702
 703         for_each_member_device(ca, c, i) {
 704                 BUG_ON(ca->buckets[1]);
 705                 BUG_ON(ca->usage[1]);
 706
 707                 ca->buckets[1] = kvpmalloc(sizeof(struct bucket_array) +
 708                                 ca->mi.nbuckets * sizeof(struct bucket),
 709                                 GFP_KERNEL|__GFP_ZERO);
 710                 if (!ca->buckets[1]) {
 711                         percpu_ref_put(&ca->ref);
 712                         bch_err(c, "error allocating ca->buckets[gc]");
 713                         return -ENOMEM;
 714                 }
 715
 716                 ca->usage[1] = alloc_percpu(struct bch_dev_usage);
 717                 if (!ca->usage[1]) {
 718                         bch_err(c, "error allocating ca->usage[gc]");
 719                         percpu_ref_put(&ca->ref);
 720                         return -ENOMEM;
 721                 }
 722         }
 723
 724         ret = bch2_ec_mem_alloc(c, true);
 725         if (ret) {
 726                 bch_err(c, "error allocating ec gc mem");
 727                 return ret;
 728         }
 729
 730         percpu_down_write(&c->mark_lock);
 731
 732         /*
 733          * indicate to stripe code that we need to allocate for the gc stripes
 734          * radix tree, too
 735          */
 736         gc_pos_set(c, gc_phase(GC_PHASE_START));
 737
 738         for_each_member_device(ca, c, i) {
 739                 struct bucket_array *dst = __bucket_array(ca, 1);
 740                 struct bucket_array *src = __bucket_array(ca, 0);
 741                 size_t b;
 742
 743                 dst->first_bucket       = src->first_bucket;
 744                 dst->nbuckets           = src->nbuckets;
 745
 746                 for (b = 0; b < src->nbuckets; b++) {
 747                         struct bucket *d = &dst->b[b];
 748                         struct bucket *s = &src->b[b];
 749
 750                         d->_mark.gen = dst->b[b].oldest_gen = s->mark.gen;
 751                         d->gen_valid = s->gen_valid;
 752                 }
 753         };
 754
 755         percpu_up_write(&c->mark_lock);
 756
 757         return 0;
 758 }
 759
 760 /**
 761  * bch2_gc - walk _all_ references to buckets, and recompute them:
 762  *
 763  * Order matters here:
 764  *  - Concurrent GC relies on the fact that we have a total ordering for
 765  *    everything that GC walks - see  gc_will_visit_node(),
 766  *    gc_will_visit_root()
 767  *
 768  *  - also, references move around in the course of index updates and
 769  *    various other crap: everything needs to agree on the ordering
 770  *    references are allowed to move around in - e.g., we're allowed to
 771  *    start with a reference owned by an open_bucket (the allocator) and
 772  *    move it to the btree, but not the reverse.
 773  *
 774  *    This is necessary to ensure that gc doesn't miss references that
 775  *    move around - if references move backwards in the ordering GC
 776  *    uses, GC could skip past them
 777  */
 778 int bch2_gc(struct bch_fs *c, struct journal_keys *journal_keys,
 779             bool initial)
 780 {
 781         struct bch_dev *ca;
 782         u64 start_time = local_clock();
 783         unsigned i, iter = 0;
 784         int ret;
 785
 786         lockdep_assert_held(&c->state_lock);
 787         trace_gc_start(c);
 788
 789         down_write(&c->gc_lock);
 790
 791         /* flush interior btree updates: */
 792         closure_wait_event(&c->btree_interior_update_wait,
 793                            !bch2_btree_interior_updates_nr_pending(c));
 794 again:
 795         ret = bch2_gc_start(c);
 796         if (ret)
 797                 goto out;
 798
 799         bch2_mark_superblocks(c);
 800
 801         ret = bch2_gc_btrees(c, journal_keys, initial);
 802         if (ret)
 803                 goto out;
 804
 805 #if 0
 806         bch2_mark_pending_btree_node_frees(c);
 807 #endif
 808         bch2_mark_allocator_buckets(c);
 809
 810         c->gc_count++;
 811 out:
 812         if (!ret &&
 813             (test_bit(BCH_FS_FIXED_GENS, &c->flags) ||
 814              (!iter && bch2_test_restart_gc))) {
 815                 /*
 816                  * XXX: make sure gens we fixed got saved
 817                  */
 818                 if (iter++ <= 2) {
 819                         bch_info(c, "Fixed gens, restarting mark and sweep:");
 820                         clear_bit(BCH_FS_FIXED_GENS, &c->flags);
 821                         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
 822
 823                         percpu_down_write(&c->mark_lock);
 824                         bch2_gc_free(c);
 825                         percpu_up_write(&c->mark_lock);
 826                         /* flush fsck errors, reset counters */
 827                         bch2_flush_fsck_errs(c);
 828
 829                         goto again;
 830                 }
 831
 832                 bch_info(c, "Unable to fix bucket gens, looping");
 833                 ret = -EINVAL;
 834         }
 835
 836         if (!ret) {
 837                 bch2_journal_block(&c->journal);
 838
 839                 percpu_down_write(&c->mark_lock);
 840                 ret = bch2_gc_done(c, initial);
 841
 842                 bch2_journal_unblock(&c->journal);
 843         } else {
 844                 percpu_down_write(&c->mark_lock);
 845         }
 846
 847         /* Indicates that gc is no longer in progress: */
 848         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
 849
 850         bch2_gc_free(c);
 851         percpu_up_write(&c->mark_lock);
 852
 853         up_write(&c->gc_lock);
 854
 855         trace_gc_end(c);
 856         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
 857
 858         /*
 859          * Wake up allocator in case it was waiting for buckets
 860          * because of not being able to inc gens
 861          */
 862         for_each_member_device(ca, c, i)
 863                 bch2_wake_allocator(ca);
 864
 865         /*
 866          * At startup, allocations can happen directly instead of via the
 867          * allocator thread - issue wakeup in case they blocked on gc_lock:
 868          */
 869         closure_wake_up(&c->freelist_wait);
 870         return ret;
 871 }
 872
 873 static bool gc_btree_gens_key(struct bch_fs *c, struct bkey_s_c k)
 874 {
 875         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 876         const struct bch_extent_ptr *ptr;
 877
 878         percpu_down_read(&c->mark_lock);
 879         bkey_for_each_ptr(ptrs, ptr) {
 880                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 881                 struct bucket *g = PTR_BUCKET(ca, ptr, false);
 882
 883                 if (gen_after(g->mark.gen, ptr->gen) > 16) {
 884                         percpu_up_read(&c->mark_lock);
 885                         return true;
 886                 }
 887         }
 888
 889         bkey_for_each_ptr(ptrs, ptr) {
 890                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 891                 struct bucket *g = PTR_BUCKET(ca, ptr, false);
 892
 893                 if (gen_after(g->gc_gen, ptr->gen))
 894                         g->gc_gen = ptr->gen;
 895         }
 896         percpu_up_read(&c->mark_lock);
 897
 898         return false;
 899 }
 900
 901 /*
 902  * For recalculating oldest gen, we only need to walk keys in leaf nodes; btree
 903  * node pointers currently never have cached pointers that can become stale:
 904  */
 905 static int bch2_gc_btree_gens(struct bch_fs *c, enum btree_id btree_id)
 906 {
 907         struct btree_trans trans;
 908         struct btree_iter *iter;
 909         struct bkey_s_c k;
 910         struct bkey_buf sk;
 911         int ret = 0;
 912
 913         bch2_bkey_buf_init(&sk);
 914         bch2_trans_init(&trans, c, 0, 0);
 915
 916         iter = bch2_trans_get_iter(&trans, btree_id, POS_MIN,
 917                                    BTREE_ITER_PREFETCH);
 918
 919         while ((k = bch2_btree_iter_peek(iter)).k &&
 920                !(ret = bkey_err(k))) {
 921                 if (gc_btree_gens_key(c, k)) {
 922                         bch2_bkey_buf_reassemble(&sk, c, k);
 923                         bch2_extent_normalize(c, bkey_i_to_s(sk.k));
 924
 925                         bch2_btree_iter_set_pos(iter, bkey_start_pos(&sk.k->k));
 926
 927                         bch2_trans_update(&trans, iter, sk.k, 0);
 928
 929                         ret = bch2_trans_commit(&trans, NULL, NULL,
 930                                                 BTREE_INSERT_NOFAIL);
 931                         if (ret == -EINTR)
 932                                 continue;
 933                         if (ret) {
 934                                 break;
 935                         }
 936                 }
 937
 938                 bch2_btree_iter_next(iter);
 939         }
 940
 941         bch2_trans_exit(&trans);
 942         bch2_bkey_buf_exit(&sk, c);
 943
 944         return ret;
 945 }
 946
 947 int bch2_gc_gens(struct bch_fs *c)
 948 {
 949         struct bch_dev *ca;
 950         struct bucket_array *buckets;
 951         struct bucket *g;
 952         unsigned i;
 953         int ret;
 954
 955         /*
 956          * Ideally we would be using state_lock and not gc_lock here, but that
 957          * introduces a deadlock in the RO path - we currently take the state
 958          * lock at the start of going RO, thus the gc thread may get stuck:
 959          */
 960         down_read(&c->gc_lock);
 961
 962         for_each_member_device(ca, c, i) {
 963                 down_read(&ca->bucket_lock);
 964                 buckets = bucket_array(ca);
 965
 966                 for_each_bucket(g, buckets)
 967                         g->gc_gen = g->mark.gen;
 968                 up_read(&ca->bucket_lock);
 969         }
 970
 971         for (i = 0; i < BTREE_ID_NR; i++)
 972                 if (btree_node_type_needs_gc(i)) {
 973                         ret = bch2_gc_btree_gens(c, i);
 974                         if (ret) {
 975                                 bch_err(c, "error recalculating oldest_gen: %i", ret);
 976                                 goto err;
 977                         }
 978                 }
 979
 980         for_each_member_device(ca, c, i) {
 981                 down_read(&ca->bucket_lock);
 982                 buckets = bucket_array(ca);
 983
 984                 for_each_bucket(g, buckets)
 985                         g->oldest_gen = g->gc_gen;
 986                 up_read(&ca->bucket_lock);
 987         }
 988
 989         c->gc_count++;
 990 err:
 991         up_read(&c->gc_lock);
 992         return ret;
 993 }
 994
 995 /* Btree coalescing */
 996
 997 static void recalc_packed_keys(struct btree *b)
 998 {
 999         struct bset *i = btree_bset_first(b);
1000         struct bkey_packed *k;
1001
1002         memset(&b->nr, 0, sizeof(b->nr));
1003
1004         BUG_ON(b->nsets != 1);
1005
1006         vstruct_for_each(i, k)
1007                 btree_keys_account_key_add(&b->nr, 0, k);
1008 }
1009
1010 static void bch2_coalesce_nodes(struct bch_fs *c, struct btree_iter *iter,
1011                                 struct btree *old_nodes[GC_MERGE_NODES])
1012 {
1013         struct btree *parent = btree_node_parent(iter, old_nodes[0]);
1014         unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
1015         unsigned blocks = btree_blocks(c) * 2 / 3;
1016         struct btree *new_nodes[GC_MERGE_NODES];
1017         struct btree_update *as;
1018         struct keylist keylist;
1019         struct bkey_format_state format_state;
1020         struct bkey_format new_format;
1021
1022         memset(new_nodes, 0, sizeof(new_nodes));
1023         bch2_keylist_init(&keylist, NULL);
1024
1025         /* Count keys that are not deleted */
1026         for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
1027                 u64s += old_nodes[i]->nr.live_u64s;
1028
1029         nr_old_nodes = nr_new_nodes = i;
1030
1031         /* Check if all keys in @old_nodes could fit in one fewer node */
1032         if (nr_old_nodes <= 1 ||
1033             __vstruct_blocks(struct btree_node, c->block_bits,
1034                              DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
1035                 return;
1036
1037         /* Find a format that all keys in @old_nodes can pack into */
1038         bch2_bkey_format_init(&format_state);
1039
1040         for (i = 0; i < nr_old_nodes; i++)
1041                 __bch2_btree_calc_format(&format_state, old_nodes[i]);
1042
1043         new_format = bch2_bkey_format_done(&format_state);
1044
1045         /* Check if repacking would make any nodes too big to fit */
1046         for (i = 0; i < nr_old_nodes; i++)
1047                 if (!bch2_btree_node_format_fits(c, old_nodes[i], &new_format)) {
1048                         trace_btree_gc_coalesce_fail(c,
1049                                         BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
1050                         return;
1051                 }
1052
1053         if (bch2_keylist_realloc(&keylist, NULL, 0,
1054                         BKEY_BTREE_PTR_U64s_MAX * nr_old_nodes)) {
1055                 trace_btree_gc_coalesce_fail(c,
1056                                 BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
1057                 return;
1058         }
1059
1060         as = bch2_btree_update_start(iter->trans, iter->btree_id,
1061                         btree_update_reserve_required(c, parent) + nr_old_nodes,
1062                         BTREE_INSERT_NOFAIL|
1063                         BTREE_INSERT_USE_RESERVE,
1064                         NULL);
1065         if (IS_ERR(as)) {
1066                 trace_btree_gc_coalesce_fail(c,
1067                                 BTREE_GC_COALESCE_FAIL_RESERVE_GET);
1068                 bch2_keylist_free(&keylist, NULL);
1069                 return;
1070         }
1071
1072         trace_btree_gc_coalesce(c, old_nodes[0]);
1073
1074         for (i = 0; i < nr_old_nodes; i++)
1075                 bch2_btree_interior_update_will_free_node(as, old_nodes[i]);
1076
1077         /* Repack everything with @new_format and sort down to one bset */
1078         for (i = 0; i < nr_old_nodes; i++)
1079                 new_nodes[i] =
1080                         __bch2_btree_node_alloc_replacement(as, old_nodes[i],
1081                                                             new_format);
1082
1083         /*
1084          * Conceptually we concatenate the nodes together and slice them
1085          * up at different boundaries.
1086          */
1087         for (i = nr_new_nodes - 1; i > 0; --i) {
1088                 struct btree *n1 = new_nodes[i];
1089                 struct btree *n2 = new_nodes[i - 1];
1090
1091                 struct bset *s1 = btree_bset_first(n1);
1092                 struct bset *s2 = btree_bset_first(n2);
1093                 struct bkey_packed *k, *last = NULL;
1094
1095                 /* Calculate how many keys from @n2 we could fit inside @n1 */
1096                 u64s = 0;
1097
1098                 for (k = s2->start;
1099                      k < vstruct_last(s2) &&
1100                      vstruct_blocks_plus(n1->data, c->block_bits,
1101                                          u64s + k->u64s) <= blocks;
1102                      k = bkey_next_skip_noops(k, vstruct_last(s2))) {
1103                         last = k;
1104                         u64s += k->u64s;
1105                 }
1106
1107                 if (u64s == le16_to_cpu(s2->u64s)) {
1108                         /* n2 fits entirely in n1 */
1109                         n1->key.k.p = n1->data->max_key = n2->data->max_key;
1110
1111                         memcpy_u64s(vstruct_last(s1),
1112                                     s2->start,
1113                                     le16_to_cpu(s2->u64s));
1114                         le16_add_cpu(&s1->u64s, le16_to_cpu(s2->u64s));
1115
1116                         set_btree_bset_end(n1, n1->set);
1117
1118                         six_unlock_write(&n2->c.lock);
1119                         bch2_btree_node_free_never_inserted(c, n2);
1120                         six_unlock_intent(&n2->c.lock);
1121
1122                         memmove(new_nodes + i - 1,
1123                                 new_nodes + i,
1124                                 sizeof(new_nodes[0]) * (nr_new_nodes - i));
1125                         new_nodes[--nr_new_nodes] = NULL;
1126                 } else if (u64s) {
1127                         /* move part of n2 into n1 */
1128                         n1->key.k.p = n1->data->max_key =
1129                                 bkey_unpack_pos(n1, last);
1130
1131                         n2->data->min_key = bkey_successor(n1->data->max_key);
1132
1133                         memcpy_u64s(vstruct_last(s1),
1134                                     s2->start, u64s);
1135                         le16_add_cpu(&s1->u64s, u64s);
1136
1137                         memmove(s2->start,
1138                                 vstruct_idx(s2, u64s),
1139                                 (le16_to_cpu(s2->u64s) - u64s) * sizeof(u64));
1140                         s2->u64s = cpu_to_le16(le16_to_cpu(s2->u64s) - u64s);
1141
1142                         set_btree_bset_end(n1, n1->set);
1143                         set_btree_bset_end(n2, n2->set);
1144                 }
1145         }
1146
1147         for (i = 0; i < nr_new_nodes; i++) {
1148                 struct btree *n = new_nodes[i];
1149
1150                 recalc_packed_keys(n);
1151                 btree_node_reset_sib_u64s(n);
1152
1153                 bch2_btree_build_aux_trees(n);
1154
1155                 bch2_btree_update_add_new_node(as, n);
1156                 six_unlock_write(&n->c.lock);
1157
1158                 bch2_btree_node_write(c, n, SIX_LOCK_intent);
1159         }
1160
1161         /*
1162          * The keys for the old nodes get deleted. We don't want to insert keys
1163          * that compare equal to the keys for the new nodes we'll also be
1164          * inserting - we can't because keys on a keylist must be strictly
1165          * greater than the previous keys, and we also don't need to since the
1166          * key for the new node will serve the same purpose (overwriting the key
1167          * for the old node).
1168          */
1169         for (i = 0; i < nr_old_nodes; i++) {
1170                 struct bkey_i delete;
1171                 unsigned j;
1172
1173                 for (j = 0; j < nr_new_nodes; j++)
1174                         if (!bkey_cmp(old_nodes[i]->key.k.p,
1175                                       new_nodes[j]->key.k.p))
1176                                 goto next;
1177
1178                 bkey_init(&delete.k);
1179                 delete.k.p = old_nodes[i]->key.k.p;
1180                 bch2_keylist_add_in_order(&keylist, &delete);
1181 next:
1182                 i = i;
1183         }
1184
1185         /*
1186          * Keys for the new nodes get inserted: bch2_btree_insert_keys() only
1187          * does the lookup once and thus expects the keys to be in sorted order
1188          * so we have to make sure the new keys are correctly ordered with
1189          * respect to the deleted keys added in the previous loop
1190          */
1191         for (i = 0; i < nr_new_nodes; i++)
1192                 bch2_keylist_add_in_order(&keylist, &new_nodes[i]->key);
1193
1194         /* Insert the newly coalesced nodes */
1195         bch2_btree_insert_node(as, parent, iter, &keylist, 0);
1196
1197         BUG_ON(!bch2_keylist_empty(&keylist));
1198
1199         BUG_ON(iter->l[old_nodes[0]->c.level].b != old_nodes[0]);
1200
1201         bch2_btree_iter_node_replace(iter, new_nodes[0]);
1202
1203         for (i = 0; i < nr_new_nodes; i++)
1204                 bch2_btree_update_get_open_buckets(as, new_nodes[i]);
1205
1206         /* Free the old nodes and update our sliding window */
1207         for (i = 0; i < nr_old_nodes; i++) {
1208                 bch2_btree_node_free_inmem(c, old_nodes[i], iter);
1209
1210                 /*
1211                  * the index update might have triggered a split, in which case
1212                  * the nodes we coalesced - the new nodes we just created -
1213                  * might not be sibling nodes anymore - don't add them to the
1214                  * sliding window (except the first):
1215                  */
1216                 if (!i) {
1217                         old_nodes[i] = new_nodes[i];
1218                 } else {
1219                         old_nodes[i] = NULL;
1220                 }
1221         }
1222
1223         for (i = 0; i < nr_new_nodes; i++)
1224                 six_unlock_intent(&new_nodes[i]->c.lock);
1225
1226         bch2_btree_update_done(as);
1227         bch2_keylist_free(&keylist, NULL);
1228 }
1229
1230 static int bch2_coalesce_btree(struct bch_fs *c, enum btree_id btree_id)
1231 {
1232         struct btree_trans trans;
1233         struct btree_iter *iter;
1234         struct btree *b;
1235         bool kthread = (current->flags & PF_KTHREAD) != 0;
1236         unsigned i;
1237
1238         /* Sliding window of adjacent btree nodes */
1239         struct btree *merge[GC_MERGE_NODES];
1240         u32 lock_seq[GC_MERGE_NODES];
1241
1242         bch2_trans_init(&trans, c, 0, 0);
1243
1244         /*
1245          * XXX: We don't have a good way of positively matching on sibling nodes
1246          * that have the same parent - this code works by handling the cases
1247          * where they might not have the same parent, and is thus fragile. Ugh.
1248          *
1249          * Perhaps redo this to use multiple linked iterators?
1250          */
1251         memset(merge, 0, sizeof(merge));
1252
1253         __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
1254                               BTREE_MAX_DEPTH, 0,
1255                               BTREE_ITER_PREFETCH, b) {
1256                 memmove(merge + 1, merge,
1257                         sizeof(merge) - sizeof(merge[0]));
1258                 memmove(lock_seq + 1, lock_seq,
1259                         sizeof(lock_seq) - sizeof(lock_seq[0]));
1260
1261                 merge[0] = b;
1262
1263                 for (i = 1; i < GC_MERGE_NODES; i++) {
1264                         if (!merge[i] ||
1265                             !six_relock_intent(&merge[i]->c.lock, lock_seq[i]))
1266                                 break;
1267
1268                         if (merge[i]->c.level != merge[0]->c.level) {
1269                                 six_unlock_intent(&merge[i]->c.lock);
1270                                 break;
1271                         }
1272                 }
1273                 memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));
1274
1275                 bch2_coalesce_nodes(c, iter, merge);
1276
1277                 for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
1278                         lock_seq[i] = merge[i]->c.lock.state.seq;
1279                         six_unlock_intent(&merge[i]->c.lock);
1280                 }
1281
1282                 lock_seq[0] = merge[0]->c.lock.state.seq;
1283
1284                 if (kthread && kthread_should_stop()) {
1285                         bch2_trans_exit(&trans);
1286                         return -ESHUTDOWN;
1287                 }
1288
1289                 bch2_trans_cond_resched(&trans);
1290
1291                 /*
1292                  * If the parent node wasn't relocked, it might have been split
1293                  * and the nodes in our sliding window might not have the same
1294                  * parent anymore - blow away the sliding window:
1295                  */
1296                 if (btree_iter_node(iter, iter->level + 1) &&
1297                     !btree_node_intent_locked(iter, iter->level + 1))
1298                         memset(merge + 1, 0,
1299                                (GC_MERGE_NODES - 1) * sizeof(merge[0]));
1300         }
1301         return bch2_trans_exit(&trans);
1302 }
1303
1304 /**
1305  * bch_coalesce - coalesce adjacent nodes with low occupancy
1306  */
1307 void bch2_coalesce(struct bch_fs *c)
1308 {
1309         enum btree_id id;
1310
1311         down_read(&c->gc_lock);
1312         trace_gc_coalesce_start(c);
1313
1314         for (id = 0; id < BTREE_ID_NR; id++) {
1315                 int ret = c->btree_roots[id].b
1316                         ? bch2_coalesce_btree(c, id)
1317                         : 0;
1318
1319                 if (ret) {
1320                         if (ret != -ESHUTDOWN)
1321                                 bch_err(c, "btree coalescing failed: %d", ret);
1322                         return;
1323                 }
1324         }
1325
1326         trace_gc_coalesce_end(c);
1327         up_read(&c->gc_lock);
1328 }
1329
1330 static int bch2_gc_thread(void *arg)
1331 {
1332         struct bch_fs *c = arg;
1333         struct io_clock *clock = &c->io_clock[WRITE];
1334         unsigned long last = atomic_long_read(&clock->now);
1335         unsigned last_kick = atomic_read(&c->kick_gc);
1336         int ret;
1337
1338         set_freezable();
1339
1340         while (1) {
1341                 while (1) {
1342                         set_current_state(TASK_INTERRUPTIBLE);
1343
1344                         if (kthread_should_stop()) {
1345                                 __set_current_state(TASK_RUNNING);
1346                                 return 0;
1347                         }
1348
1349                         if (atomic_read(&c->kick_gc) != last_kick)
1350                                 break;
1351
1352                         if (c->btree_gc_periodic) {
1353                                 unsigned long next = last + c->capacity / 16;
1354
1355                                 if (atomic_long_read(&clock->now) >= next)
1356                                         break;
1357
1358                                 bch2_io_clock_schedule_timeout(clock, next);
1359                         } else {
1360                                 schedule();
1361                         }
1362
1363                         try_to_freeze();
1364                 }
1365                 __set_current_state(TASK_RUNNING);
1366
1367                 last = atomic_long_read(&clock->now);
1368                 last_kick = atomic_read(&c->kick_gc);
1369
1370                 /*
1371                  * Full gc is currently incompatible with btree key cache:
1372                  */
1373 #if 0
1374                 ret = bch2_gc(c, NULL, false, false);
1375 #else
1376                 ret = bch2_gc_gens(c);
1377 #endif
1378                 if (ret < 0)
1379                         bch_err(c, "btree gc failed: %i", ret);
1380
1381                 debug_check_no_locks_held();
1382         }
1383
1384         return 0;
1385 }
1386
1387 void bch2_gc_thread_stop(struct bch_fs *c)
1388 {
1389         struct task_struct *p;
1390
1391         p = c->gc_thread;
1392         c->gc_thread = NULL;
1393
1394         if (p) {
1395                 kthread_stop(p);
1396                 put_task_struct(p);
1397         }
1398 }
1399
1400 int bch2_gc_thread_start(struct bch_fs *c)
1401 {
1402         struct task_struct *p;
1403
1404         BUG_ON(c->gc_thread);
1405
1406         p = kthread_create(bch2_gc_thread, c, "bch-gc/%s", c->name);
1407         if (IS_ERR(p))
1408                 return PTR_ERR(p);
1409
1410         get_task_struct(p);
1411         c->gc_thread = p;
1412         wake_up_process(p);
1413         return 0;
1414 }