git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_gc.c

   1 /*
   2  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
   3  * Copyright (C) 2014 Datera Inc.
   4  */
   5
   6 #include "bcachefs.h"
   7 #include "alloc_background.h"
   8 #include "alloc_foreground.h"
   9 #include "bkey_methods.h"
  10 #include "btree_locking.h"
  11 #include "btree_update_interior.h"
  12 #include "btree_io.h"
  13 #include "btree_gc.h"
  14 #include "buckets.h"
  15 #include "clock.h"
  16 #include "debug.h"
  17 #include "error.h"
  18 #include "extents.h"
  19 #include "journal.h"
  20 #include "journal_io.h"
  21 #include "keylist.h"
  22 #include "move.h"
  23 #include "replicas.h"
  24 #include "super-io.h"
  25
  26 #include <linux/slab.h>
  27 #include <linux/bitops.h>
  28 #include <linux/freezer.h>
  29 #include <linux/kthread.h>
  30 #include <linux/preempt.h>
  31 #include <linux/rcupdate.h>
  32 #include <linux/sched/task.h>
  33 #include <trace/events/bcachefs.h>
  34
  35 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  36 {
  37         write_seqcount_begin(&c->gc_pos_lock);
  38         c->gc_pos = new_pos;
  39         write_seqcount_end(&c->gc_pos_lock);
  40 }
  41
  42 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  43 {
  44         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
  45         __gc_pos_set(c, new_pos);
  46 }
  47
  48 /* range_checks - for validating min/max pos of each btree node: */
  49
  50 struct range_checks {
  51         struct range_level {
  52                 struct bpos     min;
  53                 struct bpos     max;
  54         }                       l[BTREE_MAX_DEPTH];
  55         unsigned                depth;
  56 };
  57
  58 static void btree_node_range_checks_init(struct range_checks *r, unsigned depth)
  59 {
  60         unsigned i;
  61
  62         for (i = 0; i < BTREE_MAX_DEPTH; i++)
  63                 r->l[i].min = r->l[i].max = POS_MIN;
  64         r->depth = depth;
  65 }
  66
  67 static void btree_node_range_checks(struct bch_fs *c, struct btree *b,
  68                                     struct range_checks *r)
  69 {
  70         struct range_level *l = &r->l[b->level];
  71
  72         struct bpos expected_min = bkey_cmp(l->min, l->max)
  73                 ? btree_type_successor(b->btree_id, l->max)
  74                 : l->max;
  75
  76         bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, expected_min), c,
  77                 "btree node has incorrect min key: %llu:%llu != %llu:%llu",
  78                 b->data->min_key.inode,
  79                 b->data->min_key.offset,
  80                 expected_min.inode,
  81                 expected_min.offset);
  82
  83         l->max = b->data->max_key;
  84
  85         if (b->level > r->depth) {
  86                 l = &r->l[b->level - 1];
  87
  88                 bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, l->min), c,
  89                         "btree node min doesn't match min of child nodes: %llu:%llu != %llu:%llu",
  90                         b->data->min_key.inode,
  91                         b->data->min_key.offset,
  92                         l->min.inode,
  93                         l->min.offset);
  94
  95                 bch2_fs_inconsistent_on(bkey_cmp(b->data->max_key, l->max), c,
  96                         "btree node max doesn't match max of child nodes: %llu:%llu != %llu:%llu",
  97                         b->data->max_key.inode,
  98                         b->data->max_key.offset,
  99                         l->max.inode,
 100                         l->max.offset);
 101
 102                 if (bkey_cmp(b->data->max_key, POS_MAX))
 103                         l->min = l->max =
 104                                 btree_type_successor(b->btree_id,
 105                                                      b->data->max_key);
 106         }
 107 }
 108
 109 /* marking of btree keys/nodes: */
 110
 111 static bool bkey_type_needs_gc(enum bkey_type type)
 112 {
 113         switch (type) {
 114         case BKEY_TYPE_BTREE:
 115         case BKEY_TYPE_EXTENTS:
 116                 return true;
 117         default:
 118                 return false;
 119         }
 120 }
 121
 122 static void ptr_gen_recalc_oldest(struct bch_fs *c,
 123                                   const struct bch_extent_ptr *ptr,
 124                                   u8 *max_stale)
 125 {
 126         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 127         size_t b = PTR_BUCKET_NR(ca, ptr);
 128
 129         if (gen_after(ca->oldest_gens[b], ptr->gen))
 130                 ca->oldest_gens[b] = ptr->gen;
 131
 132         *max_stale = max(*max_stale, ptr_stale(ca, ptr));
 133 }
 134
 135 static u8 ptr_gens_recalc_oldest(struct bch_fs *c,
 136                                  enum bkey_type type,
 137                                  struct bkey_s_c k)
 138 {
 139         const struct bch_extent_ptr *ptr;
 140         u8 max_stale = 0;
 141
 142         switch (type) {
 143         case BKEY_TYPE_BTREE:
 144         case BKEY_TYPE_EXTENTS:
 145                 switch (k.k->type) {
 146                 case BCH_EXTENT:
 147                 case BCH_EXTENT_CACHED: {
 148                         struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
 149
 150                         extent_for_each_ptr(e, ptr)
 151                                 ptr_gen_recalc_oldest(c, ptr, &max_stale);
 152                         break;
 153                 }
 154                 }
 155                 break;
 156         default:
 157                 break;
 158         }
 159
 160         return max_stale;
 161 }
 162
 163 static int ptr_gen_check(struct bch_fs *c,
 164                          enum bkey_type type,
 165                          const struct bch_extent_ptr *ptr)
 166 {
 167         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 168         size_t b = PTR_BUCKET_NR(ca, ptr);
 169         struct bucket *g = PTR_BUCKET(ca, ptr);
 170         int ret = 0;
 171
 172         if (mustfix_fsck_err_on(!g->mark.gen_valid, c,
 173                                 "found ptr with missing gen in alloc btree,\n"
 174                                 "type %u gen %u",
 175                                 type, ptr->gen)) {
 176                 g->_mark.gen = ptr->gen;
 177                 g->_mark.gen_valid = 1;
 178                 set_bit(b, ca->buckets_dirty);
 179         }
 180
 181         if (mustfix_fsck_err_on(gen_cmp(ptr->gen, g->mark.gen) > 0, c,
 182                                 "%u ptr gen in the future: %u > %u",
 183                                 type, ptr->gen, g->mark.gen)) {
 184                 g->_mark.gen = ptr->gen;
 185                 g->_mark.gen_valid = 1;
 186                 set_bit(b, ca->buckets_dirty);
 187                 set_bit(BCH_FS_FIXED_GENS, &c->flags);
 188         }
 189 fsck_err:
 190         return ret;
 191 }
 192
 193 static int ptr_gens_check(struct bch_fs *c, enum bkey_type type,
 194                           struct bkey_s_c k)
 195 {
 196         const struct bch_extent_ptr *ptr;
 197         int ret = 0;
 198
 199         switch (type) {
 200         case BKEY_TYPE_BTREE:
 201         case BKEY_TYPE_EXTENTS:
 202                 switch (k.k->type) {
 203                 case BCH_EXTENT:
 204                 case BCH_EXTENT_CACHED: {
 205                         struct bkey_s_c_extent e = bkey_s_c_to_extent(k);
 206
 207                         extent_for_each_ptr(e, ptr) {
 208                                 ret = ptr_gen_check(c, type, ptr);
 209                                 if (ret)
 210                                         return ret;
 211
 212                         }
 213                         break;
 214                 }
 215                 }
 216                 break;
 217         default:
 218                 break;
 219         }
 220
 221         return ret;
 222 }
 223
 224 /*
 225  * For runtime mark and sweep:
 226  */
 227 static int bch2_gc_mark_key(struct bch_fs *c, enum bkey_type type,
 228                             struct bkey_s_c k, bool initial)
 229 {
 230         struct gc_pos pos = { 0 };
 231         unsigned flags =
 232                 BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
 233                 BCH_BUCKET_MARK_GC_LOCK_HELD|
 234                 (initial ? BCH_BUCKET_MARK_NOATOMIC : 0);
 235         int ret = 0;
 236
 237         if (initial) {
 238                 BUG_ON(journal_seq_verify(c) &&
 239                        k.k->version.lo > journal_cur_seq(&c->journal));
 240
 241                 if (k.k->version.lo > atomic64_read(&c->key_version))
 242                         atomic64_set(&c->key_version, k.k->version.lo);
 243
 244                 if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
 245                     fsck_err_on(!bch2_bkey_replicas_marked(c, type, k,
 246                                                            false), c,
 247                                 "superblock not marked as containing replicas (type %u)",
 248                                 type)) {
 249                         ret = bch2_mark_bkey_replicas(c, type, k);
 250                         if (ret)
 251                                 return ret;
 252                 }
 253
 254                 ret = ptr_gens_check(c, type, k);
 255                 if (ret)
 256                         return ret;
 257         }
 258
 259         bch2_mark_key(c, type, k, true, k.k->size, pos, NULL, 0, flags);
 260
 261         ret = ptr_gens_recalc_oldest(c, type, k);
 262 fsck_err:
 263         return ret;
 264 }
 265
 266 static int btree_gc_mark_node(struct bch_fs *c, struct btree *b,
 267                               bool initial)
 268 {
 269         enum bkey_type type = btree_node_type(b);
 270         struct btree_node_iter iter;
 271         struct bkey unpacked;
 272         struct bkey_s_c k;
 273         u8 stale = 0;
 274         int ret;
 275
 276         if (!bkey_type_needs_gc(type))
 277                 return 0;
 278
 279         for_each_btree_node_key_unpack(b, k, &iter,
 280                                        &unpacked) {
 281                 bch2_bkey_debugcheck(c, b, k);
 282
 283                 ret = bch2_gc_mark_key(c, type, k, initial);
 284                 if (ret < 0)
 285                         return ret;
 286
 287                 stale = max_t(u8, stale, ret);
 288         }
 289
 290         return stale;
 291 }
 292
 293 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id,
 294                          bool initial)
 295 {
 296         struct btree_iter iter;
 297         struct btree *b;
 298         struct range_checks r;
 299         unsigned depth = bkey_type_needs_gc(btree_id) ? 0 : 1;
 300         unsigned max_stale;
 301         int ret = 0;
 302
 303         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 304
 305         if (!c->btree_roots[btree_id].b)
 306                 return 0;
 307
 308         /*
 309          * if expensive_debug_checks is on, run range_checks on all leaf nodes:
 310          *
 311          * and on startup, we have to read every btree node (XXX: only if it was
 312          * an unclean shutdown)
 313          */
 314         if (initial || expensive_debug_checks(c))
 315                 depth = 0;
 316
 317         btree_node_range_checks_init(&r, depth);
 318
 319         __for_each_btree_node(&iter, c, btree_id, POS_MIN,
 320                               0, depth, BTREE_ITER_PREFETCH, b) {
 321                 btree_node_range_checks(c, b, &r);
 322
 323                 bch2_verify_btree_nr_keys(b);
 324
 325                 max_stale = btree_gc_mark_node(c, b, initial);
 326
 327                 gc_pos_set(c, gc_pos_btree_node(b));
 328
 329                 if (!initial) {
 330                         if (max_stale > 64)
 331                                 bch2_btree_node_rewrite(c, &iter,
 332                                                 b->data->keys.seq,
 333                                                 BTREE_INSERT_USE_RESERVE|
 334                                                 BTREE_INSERT_NOWAIT|
 335                                                 BTREE_INSERT_GC_LOCK_HELD);
 336                         else if (!btree_gc_rewrite_disabled(c) &&
 337                                  (btree_gc_always_rewrite(c) || max_stale > 16))
 338                                 bch2_btree_node_rewrite(c, &iter,
 339                                                 b->data->keys.seq,
 340                                                 BTREE_INSERT_NOWAIT|
 341                                                 BTREE_INSERT_GC_LOCK_HELD);
 342                 }
 343
 344                 bch2_btree_iter_cond_resched(&iter);
 345         }
 346         ret = bch2_btree_iter_unlock(&iter);
 347         if (ret)
 348                 return ret;
 349
 350         mutex_lock(&c->btree_root_lock);
 351
 352         b = c->btree_roots[btree_id].b;
 353         if (!btree_node_fake(b))
 354                 bch2_gc_mark_key(c, BKEY_TYPE_BTREE,
 355                                  bkey_i_to_s_c(&b->key), initial);
 356         gc_pos_set(c, gc_pos_btree_root(b->btree_id));
 357
 358         mutex_unlock(&c->btree_root_lock);
 359         return 0;
 360 }
 361
 362 static int bch2_gc_btrees(struct bch_fs *c, struct list_head *journal,
 363                           bool initial)
 364 {
 365         unsigned i;
 366
 367         for (i = 0; i < BTREE_ID_NR; i++) {
 368                 enum bkey_type type = bkey_type(0, i);
 369
 370                 int ret = bch2_gc_btree(c, i, initial);
 371                 if (ret)
 372                         return ret;
 373
 374                 if (journal && bkey_type_needs_gc(type)) {
 375                         struct bkey_i *k, *n;
 376                         struct jset_entry *j;
 377                         struct journal_replay *r;
 378                         int ret;
 379
 380                         list_for_each_entry(r, journal, list)
 381                                 for_each_jset_key(k, n, j, &r->j) {
 382                                         if (type == bkey_type(j->level, j->btree_id)) {
 383                                                 ret = bch2_gc_mark_key(c, type,
 384                                                         bkey_i_to_s_c(k), initial);
 385                                                 if (ret < 0)
 386                                                         return ret;
 387                                         }
 388                                 }
 389                 }
 390         }
 391
 392         return 0;
 393 }
 394
 395 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
 396                                   u64 start, u64 end,
 397                                   enum bch_data_type type,
 398                                   unsigned flags)
 399 {
 400         u64 b = sector_to_bucket(ca, start);
 401
 402         do {
 403                 unsigned sectors =
 404                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
 405
 406                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
 407                                           gc_phase(GC_PHASE_SB), flags);
 408                 b++;
 409                 start += sectors;
 410         } while (start < end);
 411 }
 412
 413 void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
 414                               unsigned flags)
 415 {
 416         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
 417         unsigned i;
 418         u64 b;
 419
 420         /*
 421          * This conditional is kind of gross, but we may be called from the
 422          * device add path, before the new device has actually been added to the
 423          * running filesystem:
 424          */
 425         if (c) {
 426                 lockdep_assert_held(&c->sb_lock);
 427                 percpu_down_read_preempt_disable(&c->usage_lock);
 428         } else {
 429                 preempt_disable();
 430         }
 431
 432         for (i = 0; i < layout->nr_superblocks; i++) {
 433                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
 434
 435                 if (offset == BCH_SB_SECTOR)
 436                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
 437                                               BCH_DATA_SB, flags);
 438
 439                 mark_metadata_sectors(c, ca, offset,
 440                                       offset + (1 << layout->sb_max_size_bits),
 441                                       BCH_DATA_SB, flags);
 442         }
 443
 444         if (c)
 445                 spin_lock(&c->journal.lock);
 446
 447         for (i = 0; i < ca->journal.nr; i++) {
 448                 b = ca->journal.buckets[i];
 449                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_JOURNAL,
 450                                           ca->mi.bucket_size,
 451                                           gc_phase(GC_PHASE_SB), flags);
 452         }
 453
 454         if (c) {
 455                 percpu_up_read_preempt_enable(&c->usage_lock);
 456                 spin_unlock(&c->journal.lock);
 457         } else {
 458                 preempt_enable();
 459         }
 460 }
 461
 462 static void bch2_mark_superblocks(struct bch_fs *c)
 463 {
 464         struct bch_dev *ca;
 465         unsigned i;
 466
 467         mutex_lock(&c->sb_lock);
 468         gc_pos_set(c, gc_phase(GC_PHASE_SB));
 469
 470         for_each_online_member(ca, c, i)
 471                 bch2_mark_dev_superblock(c, ca,
 472                                          BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
 473                                          BCH_BUCKET_MARK_GC_LOCK_HELD);
 474         mutex_unlock(&c->sb_lock);
 475 }
 476
 477 /* Also see bch2_pending_btree_node_free_insert_done() */
 478 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
 479 {
 480         struct gc_pos pos = { 0 };
 481         struct bch_fs_usage stats = { 0 };
 482         struct btree_update *as;
 483         struct pending_btree_node_free *d;
 484
 485         mutex_lock(&c->btree_interior_update_lock);
 486         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
 487
 488         for_each_pending_btree_node_free(c, as, d)
 489                 if (d->index_update_done)
 490                         bch2_mark_key(c, BKEY_TYPE_BTREE,
 491                                       bkey_i_to_s_c(&d->key),
 492                                       true, 0,
 493                                       pos, &stats, 0,
 494                                       BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
 495                                       BCH_BUCKET_MARK_GC_LOCK_HELD);
 496         /*
 497          * Don't apply stats - pending deletes aren't tracked in
 498          * bch_alloc_stats:
 499          */
 500
 501         mutex_unlock(&c->btree_interior_update_lock);
 502 }
 503
 504 static void bch2_mark_allocator_buckets(struct bch_fs *c)
 505 {
 506         struct bch_dev *ca;
 507         struct open_bucket *ob;
 508         size_t i, j, iter;
 509         unsigned ci;
 510
 511         percpu_down_read_preempt_disable(&c->usage_lock);
 512
 513         spin_lock(&c->freelist_lock);
 514         gc_pos_set(c, gc_pos_alloc(c, NULL));
 515
 516         for_each_member_device(ca, c, ci) {
 517                 fifo_for_each_entry(i, &ca->free_inc, iter)
 518                         bch2_mark_alloc_bucket(c, ca, i, true,
 519                                                gc_pos_alloc(c, NULL),
 520                                                BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
 521                                                BCH_BUCKET_MARK_GC_LOCK_HELD);
 522
 523
 524
 525                 for (j = 0; j < RESERVE_NR; j++)
 526                         fifo_for_each_entry(i, &ca->free[j], iter)
 527                                 bch2_mark_alloc_bucket(c, ca, i, true,
 528                                                        gc_pos_alloc(c, NULL),
 529                                                        BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
 530                                                        BCH_BUCKET_MARK_GC_LOCK_HELD);
 531         }
 532
 533         spin_unlock(&c->freelist_lock);
 534
 535         for (ob = c->open_buckets;
 536              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
 537              ob++) {
 538                 spin_lock(&ob->lock);
 539                 if (ob->valid) {
 540                         gc_pos_set(c, gc_pos_alloc(c, ob));
 541                         ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 542                         bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr), true,
 543                                                gc_pos_alloc(c, ob),
 544                                                BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
 545                                                BCH_BUCKET_MARK_GC_LOCK_HELD);
 546                 }
 547                 spin_unlock(&ob->lock);
 548         }
 549
 550         percpu_up_read_preempt_enable(&c->usage_lock);
 551 }
 552
 553 static void bch2_gc_start(struct bch_fs *c)
 554 {
 555         struct bch_dev *ca;
 556         struct bucket_array *buckets;
 557         struct bucket_mark new;
 558         unsigned i;
 559         size_t b;
 560         int cpu;
 561
 562         percpu_down_write(&c->usage_lock);
 563
 564         /*
 565          * Indicates to buckets code that gc is now in progress - done under
 566          * usage_lock to avoid racing with bch2_mark_key():
 567          */
 568         __gc_pos_set(c, gc_phase(GC_PHASE_START));
 569
 570         /* Save a copy of the existing bucket stats while we recompute them: */
 571         for_each_member_device(ca, c, i) {
 572                 ca->usage_cached = __bch2_dev_usage_read(ca);
 573                 for_each_possible_cpu(cpu) {
 574                         struct bch_dev_usage *p =
 575                                 per_cpu_ptr(ca->usage_percpu, cpu);
 576                         memset(p, 0, sizeof(*p));
 577                 }
 578         }
 579
 580         c->usage_cached = __bch2_fs_usage_read(c);
 581         for_each_possible_cpu(cpu) {
 582                 struct bch_fs_usage *p =
 583                         per_cpu_ptr(c->usage_percpu, cpu);
 584
 585                 memset(p->replicas, 0, sizeof(p->replicas));
 586                 memset(p->buckets, 0, sizeof(p->buckets));
 587         }
 588
 589         percpu_up_write(&c->usage_lock);
 590
 591         /* Clear bucket marks: */
 592         for_each_member_device(ca, c, i) {
 593                 down_read(&ca->bucket_lock);
 594                 buckets = bucket_array(ca);
 595
 596                 for (b = buckets->first_bucket; b < buckets->nbuckets; b++) {
 597                         bucket_cmpxchg(buckets->b + b, new, ({
 598                                 new.owned_by_allocator  = 0;
 599                                 new.data_type           = 0;
 600                                 new.cached_sectors      = 0;
 601                                 new.dirty_sectors       = 0;
 602                         }));
 603                         ca->oldest_gens[b] = new.gen;
 604                 }
 605                 up_read(&ca->bucket_lock);
 606         }
 607 }
 608
 609 /**
 610  * bch_gc - recompute bucket marks and oldest_gen, rewrite btree nodes
 611  */
 612 void bch2_gc(struct bch_fs *c)
 613 {
 614         struct bch_dev *ca;
 615         u64 start_time = local_clock();
 616         unsigned i;
 617         int ret;
 618
 619         /*
 620          * Walk _all_ references to buckets, and recompute them:
 621          *
 622          * Order matters here:
 623          *  - Concurrent GC relies on the fact that we have a total ordering for
 624          *    everything that GC walks - see  gc_will_visit_node(),
 625          *    gc_will_visit_root()
 626          *
 627          *  - also, references move around in the course of index updates and
 628          *    various other crap: everything needs to agree on the ordering
 629          *    references are allowed to move around in - e.g., we're allowed to
 630          *    start with a reference owned by an open_bucket (the allocator) and
 631          *    move it to the btree, but not the reverse.
 632          *
 633          *    This is necessary to ensure that gc doesn't miss references that
 634          *    move around - if references move backwards in the ordering GC
 635          *    uses, GC could skip past them
 636          */
 637         trace_gc_start(c);
 638
 639         /*
 640          * Do this before taking gc_lock - bch2_disk_reservation_get() blocks on
 641          * gc_lock if sectors_available goes to 0:
 642          */
 643         bch2_recalc_sectors_available(c);
 644
 645         down_write(&c->gc_lock);
 646         if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
 647                 goto out;
 648
 649         bch2_gc_start(c);
 650
 651         bch2_mark_superblocks(c);
 652
 653         ret = bch2_gc_btrees(c, NULL, false);
 654         if (ret) {
 655                 bch_err(c, "btree gc failed: %d", ret);
 656                 set_bit(BCH_FS_GC_FAILURE, &c->flags);
 657                 goto out;
 658         }
 659
 660         bch2_mark_pending_btree_node_frees(c);
 661         bch2_mark_allocator_buckets(c);
 662
 663         /* Indicates that gc is no longer in progress: */
 664         gc_pos_set(c, gc_phase(GC_PHASE_DONE));
 665         c->gc_count++;
 666 out:
 667         up_write(&c->gc_lock);
 668         trace_gc_end(c);
 669         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
 670
 671         /*
 672          * Wake up allocator in case it was waiting for buckets
 673          * because of not being able to inc gens
 674          */
 675         for_each_member_device(ca, c, i)
 676                 bch2_wake_allocator(ca);
 677
 678         /*
 679          * At startup, allocations can happen directly instead of via the
 680          * allocator thread - issue wakeup in case they blocked on gc_lock:
 681          */
 682         closure_wake_up(&c->freelist_wait);
 683 }
 684
 685 /* Btree coalescing */
 686
 687 static void recalc_packed_keys(struct btree *b)
 688 {
 689         struct bset *i = btree_bset_first(b);
 690         struct bkey_packed *k;
 691
 692         memset(&b->nr, 0, sizeof(b->nr));
 693
 694         BUG_ON(b->nsets != 1);
 695
 696         vstruct_for_each(i, k)
 697                 btree_keys_account_key_add(&b->nr, 0, k);
 698 }
 699
 700 static void bch2_coalesce_nodes(struct bch_fs *c, struct btree_iter *iter,
 701                                 struct btree *old_nodes[GC_MERGE_NODES])
 702 {
 703         struct btree *parent = btree_node_parent(iter, old_nodes[0]);
 704         unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
 705         unsigned blocks = btree_blocks(c) * 2 / 3;
 706         struct btree *new_nodes[GC_MERGE_NODES];
 707         struct btree_update *as;
 708         struct keylist keylist;
 709         struct bkey_format_state format_state;
 710         struct bkey_format new_format;
 711
 712         memset(new_nodes, 0, sizeof(new_nodes));
 713         bch2_keylist_init(&keylist, NULL);
 714
 715         /* Count keys that are not deleted */
 716         for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
 717                 u64s += old_nodes[i]->nr.live_u64s;
 718
 719         nr_old_nodes = nr_new_nodes = i;
 720
 721         /* Check if all keys in @old_nodes could fit in one fewer node */
 722         if (nr_old_nodes <= 1 ||
 723             __vstruct_blocks(struct btree_node, c->block_bits,
 724                              DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
 725                 return;
 726
 727         /* Find a format that all keys in @old_nodes can pack into */
 728         bch2_bkey_format_init(&format_state);
 729
 730         for (i = 0; i < nr_old_nodes; i++)
 731                 __bch2_btree_calc_format(&format_state, old_nodes[i]);
 732
 733         new_format = bch2_bkey_format_done(&format_state);
 734
 735         /* Check if repacking would make any nodes too big to fit */
 736         for (i = 0; i < nr_old_nodes; i++)
 737                 if (!bch2_btree_node_format_fits(c, old_nodes[i], &new_format)) {
 738                         trace_btree_gc_coalesce_fail(c,
 739                                         BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
 740                         return;
 741                 }
 742
 743         if (bch2_keylist_realloc(&keylist, NULL, 0,
 744                         (BKEY_U64s + BKEY_EXTENT_U64s_MAX) * nr_old_nodes)) {
 745                 trace_btree_gc_coalesce_fail(c,
 746                                 BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
 747                 return;
 748         }
 749
 750         as = bch2_btree_update_start(c, iter->btree_id,
 751                         btree_update_reserve_required(c, parent) + nr_old_nodes,
 752                         BTREE_INSERT_NOFAIL|
 753                         BTREE_INSERT_USE_RESERVE,
 754                         NULL);
 755         if (IS_ERR(as)) {
 756                 trace_btree_gc_coalesce_fail(c,
 757                                 BTREE_GC_COALESCE_FAIL_RESERVE_GET);
 758                 bch2_keylist_free(&keylist, NULL);
 759                 return;
 760         }
 761
 762         trace_btree_gc_coalesce(c, old_nodes[0]);
 763
 764         for (i = 0; i < nr_old_nodes; i++)
 765                 bch2_btree_interior_update_will_free_node(as, old_nodes[i]);
 766
 767         /* Repack everything with @new_format and sort down to one bset */
 768         for (i = 0; i < nr_old_nodes; i++)
 769                 new_nodes[i] =
 770                         __bch2_btree_node_alloc_replacement(as, old_nodes[i],
 771                                                             new_format);
 772
 773         /*
 774          * Conceptually we concatenate the nodes together and slice them
 775          * up at different boundaries.
 776          */
 777         for (i = nr_new_nodes - 1; i > 0; --i) {
 778                 struct btree *n1 = new_nodes[i];
 779                 struct btree *n2 = new_nodes[i - 1];
 780
 781                 struct bset *s1 = btree_bset_first(n1);
 782                 struct bset *s2 = btree_bset_first(n2);
 783                 struct bkey_packed *k, *last = NULL;
 784
 785                 /* Calculate how many keys from @n2 we could fit inside @n1 */
 786                 u64s = 0;
 787
 788                 for (k = s2->start;
 789                      k < vstruct_last(s2) &&
 790                      vstruct_blocks_plus(n1->data, c->block_bits,
 791                                          u64s + k->u64s) <= blocks;
 792                      k = bkey_next(k)) {
 793                         last = k;
 794                         u64s += k->u64s;
 795                 }
 796
 797                 if (u64s == le16_to_cpu(s2->u64s)) {
 798                         /* n2 fits entirely in n1 */
 799                         n1->key.k.p = n1->data->max_key = n2->data->max_key;
 800
 801                         memcpy_u64s(vstruct_last(s1),
 802                                     s2->start,
 803                                     le16_to_cpu(s2->u64s));
 804                         le16_add_cpu(&s1->u64s, le16_to_cpu(s2->u64s));
 805
 806                         set_btree_bset_end(n1, n1->set);
 807
 808                         six_unlock_write(&n2->lock);
 809                         bch2_btree_node_free_never_inserted(c, n2);
 810                         six_unlock_intent(&n2->lock);
 811
 812                         memmove(new_nodes + i - 1,
 813                                 new_nodes + i,
 814                                 sizeof(new_nodes[0]) * (nr_new_nodes - i));
 815                         new_nodes[--nr_new_nodes] = NULL;
 816                 } else if (u64s) {
 817                         /* move part of n2 into n1 */
 818                         n1->key.k.p = n1->data->max_key =
 819                                 bkey_unpack_pos(n1, last);
 820
 821                         n2->data->min_key =
 822                                 btree_type_successor(iter->btree_id,
 823                                                      n1->data->max_key);
 824
 825                         memcpy_u64s(vstruct_last(s1),
 826                                     s2->start, u64s);
 827                         le16_add_cpu(&s1->u64s, u64s);
 828
 829                         memmove(s2->start,
 830                                 vstruct_idx(s2, u64s),
 831                                 (le16_to_cpu(s2->u64s) - u64s) * sizeof(u64));
 832                         s2->u64s = cpu_to_le16(le16_to_cpu(s2->u64s) - u64s);
 833
 834                         set_btree_bset_end(n1, n1->set);
 835                         set_btree_bset_end(n2, n2->set);
 836                 }
 837         }
 838
 839         for (i = 0; i < nr_new_nodes; i++) {
 840                 struct btree *n = new_nodes[i];
 841
 842                 recalc_packed_keys(n);
 843                 btree_node_reset_sib_u64s(n);
 844
 845                 bch2_btree_build_aux_trees(n);
 846                 six_unlock_write(&n->lock);
 847
 848                 bch2_btree_node_write(c, n, SIX_LOCK_intent);
 849         }
 850
 851         /*
 852          * The keys for the old nodes get deleted. We don't want to insert keys
 853          * that compare equal to the keys for the new nodes we'll also be
 854          * inserting - we can't because keys on a keylist must be strictly
 855          * greater than the previous keys, and we also don't need to since the
 856          * key for the new node will serve the same purpose (overwriting the key
 857          * for the old node).
 858          */
 859         for (i = 0; i < nr_old_nodes; i++) {
 860                 struct bkey_i delete;
 861                 unsigned j;
 862
 863                 for (j = 0; j < nr_new_nodes; j++)
 864                         if (!bkey_cmp(old_nodes[i]->key.k.p,
 865                                       new_nodes[j]->key.k.p))
 866                                 goto next;
 867
 868                 bkey_init(&delete.k);
 869                 delete.k.p = old_nodes[i]->key.k.p;
 870                 bch2_keylist_add_in_order(&keylist, &delete);
 871 next:
 872                 i = i;
 873         }
 874
 875         /*
 876          * Keys for the new nodes get inserted: bch2_btree_insert_keys() only
 877          * does the lookup once and thus expects the keys to be in sorted order
 878          * so we have to make sure the new keys are correctly ordered with
 879          * respect to the deleted keys added in the previous loop
 880          */
 881         for (i = 0; i < nr_new_nodes; i++)
 882                 bch2_keylist_add_in_order(&keylist, &new_nodes[i]->key);
 883
 884         /* Insert the newly coalesced nodes */
 885         bch2_btree_insert_node(as, parent, iter, &keylist, 0);
 886
 887         BUG_ON(!bch2_keylist_empty(&keylist));
 888
 889         BUG_ON(iter->l[old_nodes[0]->level].b != old_nodes[0]);
 890
 891         bch2_btree_iter_node_replace(iter, new_nodes[0]);
 892
 893         for (i = 0; i < nr_new_nodes; i++)
 894                 bch2_open_buckets_put(c, &new_nodes[i]->ob);
 895
 896         /* Free the old nodes and update our sliding window */
 897         for (i = 0; i < nr_old_nodes; i++) {
 898                 bch2_btree_node_free_inmem(c, old_nodes[i], iter);
 899                 six_unlock_intent(&old_nodes[i]->lock);
 900
 901                 /*
 902                  * the index update might have triggered a split, in which case
 903                  * the nodes we coalesced - the new nodes we just created -
 904                  * might not be sibling nodes anymore - don't add them to the
 905                  * sliding window (except the first):
 906                  */
 907                 if (!i) {
 908                         old_nodes[i] = new_nodes[i];
 909                 } else {
 910                         old_nodes[i] = NULL;
 911                         if (new_nodes[i])
 912                                 six_unlock_intent(&new_nodes[i]->lock);
 913                 }
 914         }
 915
 916         bch2_btree_update_done(as);
 917         bch2_keylist_free(&keylist, NULL);
 918 }
 919
 920 static int bch2_coalesce_btree(struct bch_fs *c, enum btree_id btree_id)
 921 {
 922         struct btree_iter iter;
 923         struct btree *b;
 924         bool kthread = (current->flags & PF_KTHREAD) != 0;
 925         unsigned i;
 926
 927         /* Sliding window of adjacent btree nodes */
 928         struct btree *merge[GC_MERGE_NODES];
 929         u32 lock_seq[GC_MERGE_NODES];
 930
 931         /*
 932          * XXX: We don't have a good way of positively matching on sibling nodes
 933          * that have the same parent - this code works by handling the cases
 934          * where they might not have the same parent, and is thus fragile. Ugh.
 935          *
 936          * Perhaps redo this to use multiple linked iterators?
 937          */
 938         memset(merge, 0, sizeof(merge));
 939
 940         __for_each_btree_node(&iter, c, btree_id, POS_MIN,
 941                               BTREE_MAX_DEPTH, 0,
 942                               BTREE_ITER_PREFETCH, b) {
 943                 memmove(merge + 1, merge,
 944                         sizeof(merge) - sizeof(merge[0]));
 945                 memmove(lock_seq + 1, lock_seq,
 946                         sizeof(lock_seq) - sizeof(lock_seq[0]));
 947
 948                 merge[0] = b;
 949
 950                 for (i = 1; i < GC_MERGE_NODES; i++) {
 951                         if (!merge[i] ||
 952                             !six_relock_intent(&merge[i]->lock, lock_seq[i]))
 953                                 break;
 954
 955                         if (merge[i]->level != merge[0]->level) {
 956                                 six_unlock_intent(&merge[i]->lock);
 957                                 break;
 958                         }
 959                 }
 960                 memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));
 961
 962                 bch2_coalesce_nodes(c, &iter, merge);
 963
 964                 for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
 965                         lock_seq[i] = merge[i]->lock.state.seq;
 966                         six_unlock_intent(&merge[i]->lock);
 967                 }
 968
 969                 lock_seq[0] = merge[0]->lock.state.seq;
 970
 971                 if (kthread && kthread_should_stop()) {
 972                         bch2_btree_iter_unlock(&iter);
 973                         return -ESHUTDOWN;
 974                 }
 975
 976                 bch2_btree_iter_cond_resched(&iter);
 977
 978                 /*
 979                  * If the parent node wasn't relocked, it might have been split
 980                  * and the nodes in our sliding window might not have the same
 981                  * parent anymore - blow away the sliding window:
 982                  */
 983                 if (btree_iter_node(&iter, iter.level + 1) &&
 984                     !btree_node_intent_locked(&iter, iter.level + 1))
 985                         memset(merge + 1, 0,
 986                                (GC_MERGE_NODES - 1) * sizeof(merge[0]));
 987         }
 988         return bch2_btree_iter_unlock(&iter);
 989 }
 990
 991 /**
 992  * bch_coalesce - coalesce adjacent nodes with low occupancy
 993  */
 994 void bch2_coalesce(struct bch_fs *c)
 995 {
 996         enum btree_id id;
 997
 998         if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
 999                 return;
1000
1001         down_read(&c->gc_lock);
1002         trace_gc_coalesce_start(c);
1003
1004         for (id = 0; id < BTREE_ID_NR; id++) {
1005                 int ret = c->btree_roots[id].b
1006                         ? bch2_coalesce_btree(c, id)
1007                         : 0;
1008
1009                 if (ret) {
1010                         if (ret != -ESHUTDOWN)
1011                                 bch_err(c, "btree coalescing failed: %d", ret);
1012                         set_bit(BCH_FS_GC_FAILURE, &c->flags);
1013                         return;
1014                 }
1015         }
1016
1017         trace_gc_coalesce_end(c);
1018         up_read(&c->gc_lock);
1019 }
1020
1021 static int bch2_gc_thread(void *arg)
1022 {
1023         struct bch_fs *c = arg;
1024         struct io_clock *clock = &c->io_clock[WRITE];
1025         unsigned long last = atomic_long_read(&clock->now);
1026         unsigned last_kick = atomic_read(&c->kick_gc);
1027
1028         set_freezable();
1029
1030         while (1) {
1031                 while (1) {
1032                         set_current_state(TASK_INTERRUPTIBLE);
1033
1034                         if (kthread_should_stop()) {
1035                                 __set_current_state(TASK_RUNNING);
1036                                 return 0;
1037                         }
1038
1039                         if (atomic_read(&c->kick_gc) != last_kick)
1040                                 break;
1041
1042                         if (c->btree_gc_periodic) {
1043                                 unsigned long next = last + c->capacity / 16;
1044
1045                                 if (atomic_long_read(&clock->now) >= next)
1046                                         break;
1047
1048                                 bch2_io_clock_schedule_timeout(clock, next);
1049                         } else {
1050                                 schedule();
1051                         }
1052
1053                         try_to_freeze();
1054                 }
1055                 __set_current_state(TASK_RUNNING);
1056
1057                 last = atomic_long_read(&clock->now);
1058                 last_kick = atomic_read(&c->kick_gc);
1059
1060                 bch2_gc(c);
1061
1062                 debug_check_no_locks_held();
1063         }
1064
1065         return 0;
1066 }
1067
1068 void bch2_gc_thread_stop(struct bch_fs *c)
1069 {
1070         struct task_struct *p;
1071
1072         p = c->gc_thread;
1073         c->gc_thread = NULL;
1074
1075         if (p) {
1076                 kthread_stop(p);
1077                 put_task_struct(p);
1078         }
1079 }
1080
1081 int bch2_gc_thread_start(struct bch_fs *c)
1082 {
1083         struct task_struct *p;
1084
1085         BUG_ON(c->gc_thread);
1086
1087         p = kthread_create(bch2_gc_thread, c, "bch_gc");
1088         if (IS_ERR(p))
1089                 return PTR_ERR(p);
1090
1091         get_task_struct(p);
1092         c->gc_thread = p;
1093         wake_up_process(p);
1094         return 0;
1095 }
1096
1097 /* Initial GC computes bucket marks during startup */
1098
1099 int bch2_initial_gc(struct bch_fs *c, struct list_head *journal)
1100 {
1101         unsigned iter = 0;
1102         int ret = 0;
1103
1104         down_write(&c->gc_lock);
1105 again:
1106         bch2_gc_start(c);
1107
1108         bch2_mark_superblocks(c);
1109
1110         ret = bch2_gc_btrees(c, journal, true);
1111         if (ret)
1112                 goto err;
1113
1114         if (test_bit(BCH_FS_FIXED_GENS, &c->flags)) {
1115                 if (iter++ > 2) {
1116                         bch_info(c, "Unable to fix bucket gens, looping");
1117                         ret = -EINVAL;
1118                         goto err;
1119                 }
1120
1121                 bch_info(c, "Fixed gens, restarting initial mark and sweep:");
1122                 clear_bit(BCH_FS_FIXED_GENS, &c->flags);
1123                 goto again;
1124         }
1125
1126         /*
1127          * Skip past versions that might have possibly been used (as nonces),
1128          * but hadn't had their pointers written:
1129          */
1130         if (c->sb.encryption_type)
1131                 atomic64_add(1 << 16, &c->key_version);
1132
1133         gc_pos_set(c, gc_phase(GC_PHASE_DONE));
1134         set_bit(BCH_FS_INITIAL_GC_DONE, &c->flags);
1135 err:
1136         up_write(&c->gc_lock);
1137         return ret;
1138 }