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 "ec.h"
  18 #include "error.h"
  19 #include "extents.h"
  20 #include "journal.h"
  21 #include "keylist.h"
  22 #include "move.h"
  23 #include "recovery.h"
  24 #include "replicas.h"
  25 #include "super-io.h"
  26
  27 #include <linux/slab.h>
  28 #include <linux/bitops.h>
  29 #include <linux/freezer.h>
  30 #include <linux/kthread.h>
  31 #include <linux/preempt.h>
  32 #include <linux/rcupdate.h>
  33 #include <linux/sched/task.h>
  34 #include <trace/events/bcachefs.h>
  35
  36 static inline void __gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  37 {
  38         write_seqcount_begin(&c->gc_pos_lock);
  39         c->gc_pos = new_pos;
  40         write_seqcount_end(&c->gc_pos_lock);
  41 }
  42
  43 static inline void gc_pos_set(struct bch_fs *c, struct gc_pos new_pos)
  44 {
  45         BUG_ON(gc_pos_cmp(new_pos, c->gc_pos) <= 0);
  46         __gc_pos_set(c, new_pos);
  47 }
  48
  49 /* range_checks - for validating min/max pos of each btree node: */
  50
  51 struct range_checks {
  52         struct range_level {
  53                 struct bpos     min;
  54                 struct bpos     max;
  55         }                       l[BTREE_MAX_DEPTH];
  56         unsigned                depth;
  57 };
  58
  59 static void btree_node_range_checks_init(struct range_checks *r, unsigned depth)
  60 {
  61         unsigned i;
  62
  63         for (i = 0; i < BTREE_MAX_DEPTH; i++)
  64                 r->l[i].min = r->l[i].max = POS_MIN;
  65         r->depth = depth;
  66 }
  67
  68 static void btree_node_range_checks(struct bch_fs *c, struct btree *b,
  69                                     struct range_checks *r)
  70 {
  71         struct range_level *l = &r->l[b->level];
  72
  73         struct bpos expected_min = bkey_cmp(l->min, l->max)
  74                 ? btree_type_successor(b->btree_id, l->max)
  75                 : l->max;
  76
  77         bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, expected_min), c,
  78                 "btree node has incorrect min key: %llu:%llu != %llu:%llu",
  79                 b->data->min_key.inode,
  80                 b->data->min_key.offset,
  81                 expected_min.inode,
  82                 expected_min.offset);
  83
  84         l->max = b->data->max_key;
  85
  86         if (b->level > r->depth) {
  87                 l = &r->l[b->level - 1];
  88
  89                 bch2_fs_inconsistent_on(bkey_cmp(b->data->min_key, l->min), c,
  90                         "btree node min doesn't match min of child nodes: %llu:%llu != %llu:%llu",
  91                         b->data->min_key.inode,
  92                         b->data->min_key.offset,
  93                         l->min.inode,
  94                         l->min.offset);
  95
  96                 bch2_fs_inconsistent_on(bkey_cmp(b->data->max_key, l->max), c,
  97                         "btree node max doesn't match max of child nodes: %llu:%llu != %llu:%llu",
  98                         b->data->max_key.inode,
  99                         b->data->max_key.offset,
 100                         l->max.inode,
 101                         l->max.offset);
 102
 103                 if (bkey_cmp(b->data->max_key, POS_MAX))
 104                         l->min = l->max =
 105                                 btree_type_successor(b->btree_id,
 106                                                      b->data->max_key);
 107         }
 108 }
 109
 110 /* marking of btree keys/nodes: */
 111
 112 static int bch2_gc_mark_key(struct bch_fs *c, struct bkey_s_c k,
 113                             u8 *max_stale, bool initial)
 114 {
 115         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 116         const struct bch_extent_ptr *ptr;
 117         unsigned flags =
 118                 BCH_BUCKET_MARK_GC|
 119                 (initial ? BCH_BUCKET_MARK_NOATOMIC : 0);
 120         int ret = 0;
 121
 122         if (initial) {
 123                 BUG_ON(journal_seq_verify(c) &&
 124                        k.k->version.lo > journal_cur_seq(&c->journal));
 125
 126                 if (k.k->version.lo > atomic64_read(&c->key_version))
 127                         atomic64_set(&c->key_version, k.k->version.lo);
 128
 129                 if (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
 130                     fsck_err_on(!bch2_bkey_replicas_marked(c, k, false), c,
 131                                 "superblock not marked as containing replicas (type %u)",
 132                                 k.k->type)) {
 133                         ret = bch2_mark_bkey_replicas(c, k);
 134                         if (ret)
 135                                 return ret;
 136                 }
 137
 138                 bkey_for_each_ptr(ptrs, ptr) {
 139                         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 140                         struct bucket *g = PTR_BUCKET(ca, ptr, true);
 141                         struct bucket *g2 = PTR_BUCKET(ca, ptr, false);
 142
 143                         if (mustfix_fsck_err_on(!g->gen_valid, c,
 144                                         "found ptr with missing gen in alloc btree,\n"
 145                                         "type %u gen %u",
 146                                         k.k->type, ptr->gen)) {
 147                                 g2->_mark.gen   = g->_mark.gen          = ptr->gen;
 148                                 g2->_mark.dirty = g->_mark.dirty        = true;
 149                                 g2->gen_valid   = g->gen_valid          = true;
 150                         }
 151
 152                         if (mustfix_fsck_err_on(gen_cmp(ptr->gen, g->mark.gen) > 0, c,
 153                                         "%u ptr gen in the future: %u > %u",
 154                                         k.k->type, ptr->gen, g->mark.gen)) {
 155                                 g2->_mark.gen   = g->_mark.gen          = ptr->gen;
 156                                 g2->_mark.dirty = g->_mark.dirty        = true;
 157                                 g2->gen_valid   = g->gen_valid          = true;
 158                                 set_bit(BCH_FS_FIXED_GENS, &c->flags);
 159                         }
 160                 }
 161         }
 162
 163         bkey_for_each_ptr(ptrs, ptr) {
 164                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 165                 struct bucket *g = PTR_BUCKET(ca, ptr, true);
 166
 167                 if (gen_after(g->oldest_gen, ptr->gen))
 168                         g->oldest_gen = ptr->gen;
 169
 170                 *max_stale = max(*max_stale, ptr_stale(ca, ptr));
 171         }
 172
 173         bch2_mark_key(c, k, k.k->size, NULL, 0, flags);
 174 fsck_err:
 175         return ret;
 176 }
 177
 178 static int btree_gc_mark_node(struct bch_fs *c, struct btree *b,
 179                               u8 *max_stale, bool initial)
 180 {
 181         struct btree_node_iter iter;
 182         struct bkey unpacked;
 183         struct bkey_s_c k;
 184         int ret = 0;
 185
 186         *max_stale = 0;
 187
 188         if (!btree_node_type_needs_gc(btree_node_type(b)))
 189                 return 0;
 190
 191         for_each_btree_node_key_unpack(b, k, &iter,
 192                                        &unpacked) {
 193                 bch2_bkey_debugcheck(c, b, k);
 194
 195                 ret = bch2_gc_mark_key(c, k, max_stale, initial);
 196                 if (ret)
 197                         break;
 198         }
 199
 200         return ret;
 201 }
 202
 203 static int bch2_gc_btree(struct bch_fs *c, enum btree_id btree_id,
 204                          bool initial, bool metadata_only)
 205 {
 206         struct btree_trans trans;
 207         struct btree_iter *iter;
 208         struct btree *b;
 209         struct range_checks r;
 210         unsigned depth = metadata_only                  ? 1
 211                 : expensive_debug_checks(c)             ? 0
 212                 : !btree_node_type_needs_gc(btree_id)   ? 1
 213                 : 0;
 214         u8 max_stale;
 215         int ret = 0;
 216
 217         bch2_trans_init(&trans, c, 0, 0);
 218
 219         gc_pos_set(c, gc_pos_btree(btree_id, POS_MIN, 0));
 220
 221         btree_node_range_checks_init(&r, depth);
 222
 223         __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
 224                               0, depth, BTREE_ITER_PREFETCH, b) {
 225                 btree_node_range_checks(c, b, &r);
 226
 227                 bch2_verify_btree_nr_keys(b);
 228
 229                 gc_pos_set(c, gc_pos_btree_node(b));
 230
 231                 ret = btree_gc_mark_node(c, b, &max_stale, initial);
 232                 if (ret)
 233                         break;
 234
 235                 if (!initial) {
 236                         if (max_stale > 64)
 237                                 bch2_btree_node_rewrite(c, iter,
 238                                                 b->data->keys.seq,
 239                                                 BTREE_INSERT_USE_RESERVE|
 240                                                 BTREE_INSERT_NOWAIT|
 241                                                 BTREE_INSERT_GC_LOCK_HELD);
 242                         else if (!btree_gc_rewrite_disabled(c) &&
 243                                  (btree_gc_always_rewrite(c) || max_stale > 16))
 244                                 bch2_btree_node_rewrite(c, iter,
 245                                                 b->data->keys.seq,
 246                                                 BTREE_INSERT_NOWAIT|
 247                                                 BTREE_INSERT_GC_LOCK_HELD);
 248                 }
 249
 250                 bch2_trans_cond_resched(&trans);
 251         }
 252         ret = bch2_trans_exit(&trans) ?: ret;
 253         if (ret)
 254                 return ret;
 255
 256         mutex_lock(&c->btree_root_lock);
 257         b = c->btree_roots[btree_id].b;
 258         if (!btree_node_fake(b))
 259                 ret = bch2_gc_mark_key(c, bkey_i_to_s_c(&b->key),
 260                                        &max_stale, initial);
 261         gc_pos_set(c, gc_pos_btree_root(b->btree_id));
 262         mutex_unlock(&c->btree_root_lock);
 263
 264         return ret;
 265 }
 266
 267 static inline int btree_id_gc_phase_cmp(enum btree_id l, enum btree_id r)
 268 {
 269         return  (int) btree_id_to_gc_phase(l) -
 270                 (int) btree_id_to_gc_phase(r);
 271 }
 272
 273 static int mark_journal_key(struct bch_fs *c, enum btree_id id,
 274                             struct bkey_i *insert)
 275 {
 276         struct btree_trans trans;
 277         struct btree_iter *iter;
 278         struct bkey_s_c k;
 279         u8 max_stale;
 280         int ret = 0;
 281
 282         ret = bch2_gc_mark_key(c, bkey_i_to_s_c(insert), &max_stale, true);
 283         if (ret)
 284                 return ret;
 285
 286         bch2_trans_init(&trans, c, 0, 0);
 287
 288         for_each_btree_key(&trans, iter, id, bkey_start_pos(&insert->k),
 289                            BTREE_ITER_SLOTS, k, ret) {
 290                 percpu_down_read_preempt_disable(&c->mark_lock);
 291                 ret = bch2_mark_overwrite(&trans, iter, k, insert, NULL,
 292                                          BCH_BUCKET_MARK_GC|
 293                                          BCH_BUCKET_MARK_NOATOMIC);
 294                 percpu_up_read_preempt_enable(&c->mark_lock);
 295
 296                 if (!ret)
 297                         break;
 298         }
 299
 300         return bch2_trans_exit(&trans) ?: ret;
 301 }
 302
 303 static int bch2_gc_btrees(struct bch_fs *c, struct journal_keys *journal_keys,
 304                           bool initial, bool metadata_only)
 305 {
 306         enum btree_id ids[BTREE_ID_NR];
 307         unsigned i;
 308
 309         for (i = 0; i < BTREE_ID_NR; i++)
 310                 ids[i] = i;
 311         bubble_sort(ids, BTREE_ID_NR, btree_id_gc_phase_cmp);
 312
 313         for (i = 0; i < BTREE_ID_NR; i++) {
 314                 enum btree_id id = ids[i];
 315                 enum btree_node_type type = __btree_node_type(0, id);
 316
 317                 int ret = bch2_gc_btree(c, id, initial, metadata_only);
 318                 if (ret)
 319                         return ret;
 320
 321                 if (journal_keys && !metadata_only &&
 322                     btree_node_type_needs_gc(type)) {
 323                         struct journal_key *j;
 324                         int ret;
 325
 326                         for_each_journal_key(*journal_keys, j)
 327                                 if (j->btree_id == id) {
 328                                         ret = mark_journal_key(c, id, j->k);
 329                                         if (ret)
 330                                                 return ret;
 331                                 }
 332                 }
 333         }
 334
 335         return 0;
 336 }
 337
 338 static void mark_metadata_sectors(struct bch_fs *c, struct bch_dev *ca,
 339                                   u64 start, u64 end,
 340                                   enum bch_data_type type,
 341                                   unsigned flags)
 342 {
 343         u64 b = sector_to_bucket(ca, start);
 344
 345         do {
 346                 unsigned sectors =
 347                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
 348
 349                 bch2_mark_metadata_bucket(c, ca, b, type, sectors,
 350                                           gc_phase(GC_PHASE_SB), flags);
 351                 b++;
 352                 start += sectors;
 353         } while (start < end);
 354 }
 355
 356 void bch2_mark_dev_superblock(struct bch_fs *c, struct bch_dev *ca,
 357                               unsigned flags)
 358 {
 359         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
 360         unsigned i;
 361         u64 b;
 362
 363         /*
 364          * This conditional is kind of gross, but we may be called from the
 365          * device add path, before the new device has actually been added to the
 366          * running filesystem:
 367          */
 368         if (c) {
 369                 lockdep_assert_held(&c->sb_lock);
 370                 percpu_down_read_preempt_disable(&c->mark_lock);
 371         } else {
 372                 preempt_disable();
 373         }
 374
 375         for (i = 0; i < layout->nr_superblocks; i++) {
 376                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
 377
 378                 if (offset == BCH_SB_SECTOR)
 379                         mark_metadata_sectors(c, ca, 0, BCH_SB_SECTOR,
 380                                               BCH_DATA_SB, flags);
 381
 382                 mark_metadata_sectors(c, ca, offset,
 383                                       offset + (1 << layout->sb_max_size_bits),
 384                                       BCH_DATA_SB, flags);
 385         }
 386
 387         for (i = 0; i < ca->journal.nr; i++) {
 388                 b = ca->journal.buckets[i];
 389                 bch2_mark_metadata_bucket(c, ca, b, BCH_DATA_JOURNAL,
 390                                           ca->mi.bucket_size,
 391                                           gc_phase(GC_PHASE_SB), flags);
 392         }
 393
 394         if (c) {
 395                 percpu_up_read_preempt_enable(&c->mark_lock);
 396         } else {
 397                 preempt_enable();
 398         }
 399 }
 400
 401 static void bch2_mark_superblocks(struct bch_fs *c)
 402 {
 403         struct bch_dev *ca;
 404         unsigned i;
 405
 406         mutex_lock(&c->sb_lock);
 407         gc_pos_set(c, gc_phase(GC_PHASE_SB));
 408
 409         for_each_online_member(ca, c, i)
 410                 bch2_mark_dev_superblock(c, ca, BCH_BUCKET_MARK_GC);
 411         mutex_unlock(&c->sb_lock);
 412 }
 413
 414 /* Also see bch2_pending_btree_node_free_insert_done() */
 415 static void bch2_mark_pending_btree_node_frees(struct bch_fs *c)
 416 {
 417         struct btree_update *as;
 418         struct pending_btree_node_free *d;
 419
 420         mutex_lock(&c->btree_interior_update_lock);
 421         gc_pos_set(c, gc_phase(GC_PHASE_PENDING_DELETE));
 422
 423         for_each_pending_btree_node_free(c, as, d)
 424                 if (d->index_update_done)
 425                         bch2_mark_key(c, bkey_i_to_s_c(&d->key), 0, NULL, 0,
 426                                       BCH_BUCKET_MARK_GC);
 427
 428         mutex_unlock(&c->btree_interior_update_lock);
 429 }
 430
 431 static void bch2_mark_allocator_buckets(struct bch_fs *c)
 432 {
 433         struct bch_dev *ca;
 434         struct open_bucket *ob;
 435         size_t i, j, iter;
 436         unsigned ci;
 437
 438         percpu_down_read_preempt_disable(&c->mark_lock);
 439
 440         spin_lock(&c->freelist_lock);
 441         gc_pos_set(c, gc_pos_alloc(c, NULL));
 442
 443         for_each_member_device(ca, c, ci) {
 444                 fifo_for_each_entry(i, &ca->free_inc, iter)
 445                         bch2_mark_alloc_bucket(c, ca, i, true,
 446                                                gc_pos_alloc(c, NULL),
 447                                                BCH_BUCKET_MARK_GC);
 448
 449
 450
 451                 for (j = 0; j < RESERVE_NR; j++)
 452                         fifo_for_each_entry(i, &ca->free[j], iter)
 453                                 bch2_mark_alloc_bucket(c, ca, i, true,
 454                                                        gc_pos_alloc(c, NULL),
 455                                                        BCH_BUCKET_MARK_GC);
 456         }
 457
 458         spin_unlock(&c->freelist_lock);
 459
 460         for (ob = c->open_buckets;
 461              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
 462              ob++) {
 463                 spin_lock(&ob->lock);
 464                 if (ob->valid) {
 465                         gc_pos_set(c, gc_pos_alloc(c, ob));
 466                         ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 467                         bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr), true,
 468                                                gc_pos_alloc(c, ob),
 469                                                BCH_BUCKET_MARK_GC);
 470                 }
 471                 spin_unlock(&ob->lock);
 472         }
 473
 474         percpu_up_read_preempt_enable(&c->mark_lock);
 475 }
 476
 477 static void bch2_gc_free(struct bch_fs *c)
 478 {
 479         struct bch_dev *ca;
 480         unsigned i;
 481
 482         genradix_free(&c->stripes[1]);
 483
 484         for_each_member_device(ca, c, i) {
 485                 kvpfree(rcu_dereference_protected(ca->buckets[1], 1),
 486                         sizeof(struct bucket_array) +
 487                         ca->mi.nbuckets * sizeof(struct bucket));
 488                 ca->buckets[1] = NULL;
 489
 490                 free_percpu(ca->usage[1]);
 491                 ca->usage[1] = NULL;
 492         }
 493
 494         free_percpu(c->usage_gc);
 495         c->usage_gc = NULL;
 496 }
 497
 498 static int bch2_gc_done(struct bch_fs *c,
 499                         bool initial, bool metadata_only)
 500 {
 501         struct bch_dev *ca;
 502         bool verify = !metadata_only &&
 503                 (!initial ||
 504                  (c->sb.compat & (1ULL << BCH_COMPAT_FEAT_ALLOC_INFO)));
 505         unsigned i;
 506         int ret = 0;
 507
 508 #define copy_field(_f, _msg, ...)                                       \
 509         if (dst->_f != src->_f) {                                       \
 510                 if (verify)                                             \
 511                         fsck_err(c, _msg ": got %llu, should be %llu"   \
 512                                 , ##__VA_ARGS__, dst->_f, src->_f);     \
 513                 dst->_f = src->_f;                                      \
 514         }
 515 #define copy_stripe_field(_f, _msg, ...)                                \
 516         if (dst->_f != src->_f) {                                       \
 517                 if (verify)                                             \
 518                         fsck_err(c, "stripe %zu has wrong "_msg         \
 519                                 ": got %u, should be %u",               \
 520                                 dst_iter.pos, ##__VA_ARGS__,            \
 521                                 dst->_f, src->_f);                      \
 522                 dst->_f = src->_f;                                      \
 523                 dst->dirty = true;                                      \
 524         }
 525 #define copy_bucket_field(_f)                                           \
 526         if (dst->b[b].mark._f != src->b[b].mark._f) {                   \
 527                 if (verify)                                             \
 528                         fsck_err(c, "dev %u bucket %zu has wrong " #_f  \
 529                                 ": got %u, should be %u", i, b,         \
 530                                 dst->b[b].mark._f, src->b[b].mark._f);  \
 531                 dst->b[b]._mark._f = src->b[b].mark._f;                 \
 532                 dst->b[b]._mark.dirty = true;                           \
 533         }
 534 #define copy_dev_field(_f, _msg, ...)                                   \
 535         copy_field(_f, "dev %u has wrong " _msg, i, ##__VA_ARGS__)
 536 #define copy_fs_field(_f, _msg, ...)                                    \
 537         copy_field(_f, "fs has wrong " _msg, ##__VA_ARGS__)
 538
 539         if (!metadata_only) {
 540                 struct genradix_iter dst_iter = genradix_iter_init(&c->stripes[0], 0);
 541                 struct genradix_iter src_iter = genradix_iter_init(&c->stripes[1], 0);
 542                 struct stripe *dst, *src;
 543                 unsigned i;
 544
 545                 c->ec_stripes_heap.used = 0;
 546
 547                 while ((dst = genradix_iter_peek(&dst_iter, &c->stripes[0])) &&
 548                        (src = genradix_iter_peek(&src_iter, &c->stripes[1]))) {
 549                         BUG_ON(src_iter.pos != dst_iter.pos);
 550
 551                         copy_stripe_field(alive,        "alive");
 552                         copy_stripe_field(sectors,      "sectors");
 553                         copy_stripe_field(algorithm,    "algorithm");
 554                         copy_stripe_field(nr_blocks,    "nr_blocks");
 555                         copy_stripe_field(nr_redundant, "nr_redundant");
 556                         copy_stripe_field(blocks_nonempty,
 557                                           "blocks_nonempty");
 558
 559                         for (i = 0; i < ARRAY_SIZE(dst->block_sectors); i++)
 560                                 copy_stripe_field(block_sectors[i],
 561                                                   "block_sectors[%u]", i);
 562
 563                         if (dst->alive)
 564                                 bch2_stripes_heap_insert(c, dst, dst_iter.pos);
 565
 566                         genradix_iter_advance(&dst_iter, &c->stripes[0]);
 567                         genradix_iter_advance(&src_iter, &c->stripes[1]);
 568                 }
 569         }
 570
 571         for_each_member_device(ca, c, i) {
 572                 struct bucket_array *dst = __bucket_array(ca, 0);
 573                 struct bucket_array *src = __bucket_array(ca, 1);
 574                 size_t b;
 575
 576                 for (b = 0; b < src->nbuckets; b++) {
 577                         copy_bucket_field(gen);
 578                         copy_bucket_field(data_type);
 579                         copy_bucket_field(owned_by_allocator);
 580                         copy_bucket_field(stripe);
 581                         copy_bucket_field(dirty_sectors);
 582                         copy_bucket_field(cached_sectors);
 583
 584                         if (dst->b[b].oldest_gen != src->b[b].oldest_gen) {
 585                                 dst->b[b].oldest_gen = src->b[b].oldest_gen;
 586                                 dst->b[b]._mark.dirty = true;
 587                         }
 588                 }
 589         };
 590
 591         bch2_fs_usage_acc_to_base(c, 0);
 592         bch2_fs_usage_acc_to_base(c, 1);
 593
 594         bch2_dev_usage_from_buckets(c);
 595
 596         {
 597                 unsigned nr = fs_usage_u64s(c);
 598                 struct bch_fs_usage *dst = c->usage_base;
 599                 struct bch_fs_usage *src = (void *)
 600                         bch2_acc_percpu_u64s((void *) c->usage_gc, nr);
 601
 602                 copy_fs_field(hidden,           "hidden");
 603                 copy_fs_field(btree,            "btree");
 604
 605                 if (!metadata_only) {
 606                         copy_fs_field(data,     "data");
 607                         copy_fs_field(cached,   "cached");
 608                         copy_fs_field(reserved, "reserved");
 609                         copy_fs_field(nr_inodes,"nr_inodes");
 610
 611                         for (i = 0; i < BCH_REPLICAS_MAX; i++)
 612                                 copy_fs_field(persistent_reserved[i],
 613                                               "persistent_reserved[%i]", i);
 614                 }
 615
 616                 for (i = 0; i < c->replicas.nr; i++) {
 617                         struct bch_replicas_entry *e =
 618                                 cpu_replicas_entry(&c->replicas, i);
 619                         char buf[80];
 620
 621                         if (metadata_only &&
 622                             (e->data_type == BCH_DATA_USER ||
 623                              e->data_type == BCH_DATA_CACHED))
 624                                 continue;
 625
 626                         bch2_replicas_entry_to_text(&PBUF(buf), e);
 627
 628                         copy_fs_field(replicas[i], "%s", buf);
 629                 }
 630         }
 631
 632 #undef copy_fs_field
 633 #undef copy_dev_field
 634 #undef copy_bucket_field
 635 #undef copy_stripe_field
 636 #undef copy_field
 637 fsck_err:
 638         return ret;
 639 }
 640
 641 static int bch2_gc_start(struct bch_fs *c,
 642                          bool metadata_only)
 643 {
 644         struct bch_dev *ca;
 645         unsigned i;
 646
 647         /*
 648          * indicate to stripe code that we need to allocate for the gc stripes
 649          * radix tree, too
 650          */
 651         gc_pos_set(c, gc_phase(GC_PHASE_START));
 652
 653         BUG_ON(c->usage_gc);
 654
 655         c->usage_gc = __alloc_percpu_gfp(fs_usage_u64s(c) * sizeof(u64),
 656                                          sizeof(u64), GFP_KERNEL);
 657         if (!c->usage_gc)
 658                 return -ENOMEM;
 659
 660         for_each_member_device(ca, c, i) {
 661                 BUG_ON(ca->buckets[1]);
 662                 BUG_ON(ca->usage[1]);
 663
 664                 ca->buckets[1] = kvpmalloc(sizeof(struct bucket_array) +
 665                                 ca->mi.nbuckets * sizeof(struct bucket),
 666                                 GFP_KERNEL|__GFP_ZERO);
 667                 if (!ca->buckets[1]) {
 668                         percpu_ref_put(&ca->ref);
 669                         return -ENOMEM;
 670                 }
 671
 672                 ca->usage[1] = alloc_percpu(struct bch_dev_usage);
 673                 if (!ca->usage[1]) {
 674                         percpu_ref_put(&ca->ref);
 675                         return -ENOMEM;
 676                 }
 677         }
 678
 679         for_each_member_device(ca, c, i) {
 680                 struct bucket_array *dst = __bucket_array(ca, 1);
 681                 struct bucket_array *src = __bucket_array(ca, 0);
 682                 size_t b;
 683
 684                 dst->first_bucket       = src->first_bucket;
 685                 dst->nbuckets           = src->nbuckets;
 686
 687                 for (b = 0; b < src->nbuckets; b++) {
 688                         struct bucket *d = &dst->b[b];
 689                         struct bucket *s = &src->b[b];
 690
 691                         d->_mark.gen = dst->b[b].oldest_gen = s->mark.gen;
 692                         d->gen_valid = s->gen_valid;
 693
 694                         if (metadata_only &&
 695                             (s->mark.data_type == BCH_DATA_USER ||
 696                              s->mark.data_type == BCH_DATA_CACHED)) {
 697                                 d->_mark = s->mark;
 698                                 d->_mark.owned_by_allocator = 0;
 699                         }
 700                 }
 701         };
 702
 703         return bch2_ec_mem_alloc(c, true);
 704 }
 705
 706 /**
 707  * bch2_gc - walk _all_ references to buckets, and recompute them:
 708  *
 709  * Order matters here:
 710  *  - Concurrent GC relies on the fact that we have a total ordering for
 711  *    everything that GC walks - see  gc_will_visit_node(),
 712  *    gc_will_visit_root()
 713  *
 714  *  - also, references move around in the course of index updates and
 715  *    various other crap: everything needs to agree on the ordering
 716  *    references are allowed to move around in - e.g., we're allowed to
 717  *    start with a reference owned by an open_bucket (the allocator) and
 718  *    move it to the btree, but not the reverse.
 719  *
 720  *    This is necessary to ensure that gc doesn't miss references that
 721  *    move around - if references move backwards in the ordering GC
 722  *    uses, GC could skip past them
 723  */
 724 int bch2_gc(struct bch_fs *c, struct journal_keys *journal_keys,
 725             bool initial, bool metadata_only)
 726 {
 727         struct bch_dev *ca;
 728         u64 start_time = local_clock();
 729         unsigned i, iter = 0;
 730         int ret;
 731
 732         trace_gc_start(c);
 733
 734         down_write(&c->gc_lock);
 735 again:
 736         percpu_down_write(&c->mark_lock);
 737         ret = bch2_gc_start(c, metadata_only);
 738         percpu_up_write(&c->mark_lock);
 739
 740         if (ret)
 741                 goto out;
 742
 743         bch2_mark_superblocks(c);
 744
 745         ret = bch2_gc_btrees(c, journal_keys, initial, metadata_only);
 746         if (ret)
 747                 goto out;
 748
 749         bch2_mark_pending_btree_node_frees(c);
 750         bch2_mark_allocator_buckets(c);
 751
 752         c->gc_count++;
 753 out:
 754         if (!ret &&
 755             (test_bit(BCH_FS_FIXED_GENS, &c->flags) ||
 756              (!iter && test_restart_gc(c)))) {
 757                 /*
 758                  * XXX: make sure gens we fixed got saved
 759                  */
 760                 if (iter++ <= 2) {
 761                         bch_info(c, "Fixed gens, restarting mark and sweep:");
 762                         clear_bit(BCH_FS_FIXED_GENS, &c->flags);
 763                         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
 764
 765                         percpu_down_write(&c->mark_lock);
 766                         bch2_gc_free(c);
 767                         percpu_up_write(&c->mark_lock);
 768
 769                         goto again;
 770                 }
 771
 772                 bch_info(c, "Unable to fix bucket gens, looping");
 773                 ret = -EINVAL;
 774         }
 775
 776         if (!ret) {
 777                 bch2_journal_block(&c->journal);
 778
 779                 percpu_down_write(&c->mark_lock);
 780                 ret = bch2_gc_done(c, initial, metadata_only);
 781
 782                 bch2_journal_unblock(&c->journal);
 783         } else {
 784                 percpu_down_write(&c->mark_lock);
 785         }
 786
 787         /* Indicates that gc is no longer in progress: */
 788         __gc_pos_set(c, gc_phase(GC_PHASE_NOT_RUNNING));
 789
 790         bch2_gc_free(c);
 791         percpu_up_write(&c->mark_lock);
 792
 793         up_write(&c->gc_lock);
 794
 795         trace_gc_end(c);
 796         bch2_time_stats_update(&c->times[BCH_TIME_btree_gc], start_time);
 797
 798         /*
 799          * Wake up allocator in case it was waiting for buckets
 800          * because of not being able to inc gens
 801          */
 802         for_each_member_device(ca, c, i)
 803                 bch2_wake_allocator(ca);
 804
 805         /*
 806          * At startup, allocations can happen directly instead of via the
 807          * allocator thread - issue wakeup in case they blocked on gc_lock:
 808          */
 809         closure_wake_up(&c->freelist_wait);
 810         return ret;
 811 }
 812
 813 /* Btree coalescing */
 814
 815 static void recalc_packed_keys(struct btree *b)
 816 {
 817         struct bset *i = btree_bset_first(b);
 818         struct bkey_packed *k;
 819
 820         memset(&b->nr, 0, sizeof(b->nr));
 821
 822         BUG_ON(b->nsets != 1);
 823
 824         vstruct_for_each(i, k)
 825                 btree_keys_account_key_add(&b->nr, 0, k);
 826 }
 827
 828 static void bch2_coalesce_nodes(struct bch_fs *c, struct btree_iter *iter,
 829                                 struct btree *old_nodes[GC_MERGE_NODES])
 830 {
 831         struct btree *parent = btree_node_parent(iter, old_nodes[0]);
 832         unsigned i, nr_old_nodes, nr_new_nodes, u64s = 0;
 833         unsigned blocks = btree_blocks(c) * 2 / 3;
 834         struct btree *new_nodes[GC_MERGE_NODES];
 835         struct btree_update *as;
 836         struct keylist keylist;
 837         struct bkey_format_state format_state;
 838         struct bkey_format new_format;
 839
 840         memset(new_nodes, 0, sizeof(new_nodes));
 841         bch2_keylist_init(&keylist, NULL);
 842
 843         /* Count keys that are not deleted */
 844         for (i = 0; i < GC_MERGE_NODES && old_nodes[i]; i++)
 845                 u64s += old_nodes[i]->nr.live_u64s;
 846
 847         nr_old_nodes = nr_new_nodes = i;
 848
 849         /* Check if all keys in @old_nodes could fit in one fewer node */
 850         if (nr_old_nodes <= 1 ||
 851             __vstruct_blocks(struct btree_node, c->block_bits,
 852                              DIV_ROUND_UP(u64s, nr_old_nodes - 1)) > blocks)
 853                 return;
 854
 855         /* Find a format that all keys in @old_nodes can pack into */
 856         bch2_bkey_format_init(&format_state);
 857
 858         for (i = 0; i < nr_old_nodes; i++)
 859                 __bch2_btree_calc_format(&format_state, old_nodes[i]);
 860
 861         new_format = bch2_bkey_format_done(&format_state);
 862
 863         /* Check if repacking would make any nodes too big to fit */
 864         for (i = 0; i < nr_old_nodes; i++)
 865                 if (!bch2_btree_node_format_fits(c, old_nodes[i], &new_format)) {
 866                         trace_btree_gc_coalesce_fail(c,
 867                                         BTREE_GC_COALESCE_FAIL_FORMAT_FITS);
 868                         return;
 869                 }
 870
 871         if (bch2_keylist_realloc(&keylist, NULL, 0,
 872                         (BKEY_U64s + BKEY_EXTENT_U64s_MAX) * nr_old_nodes)) {
 873                 trace_btree_gc_coalesce_fail(c,
 874                                 BTREE_GC_COALESCE_FAIL_KEYLIST_REALLOC);
 875                 return;
 876         }
 877
 878         as = bch2_btree_update_start(c, iter->btree_id,
 879                         btree_update_reserve_required(c, parent) + nr_old_nodes,
 880                         BTREE_INSERT_NOFAIL|
 881                         BTREE_INSERT_USE_RESERVE,
 882                         NULL);
 883         if (IS_ERR(as)) {
 884                 trace_btree_gc_coalesce_fail(c,
 885                                 BTREE_GC_COALESCE_FAIL_RESERVE_GET);
 886                 bch2_keylist_free(&keylist, NULL);
 887                 return;
 888         }
 889
 890         trace_btree_gc_coalesce(c, old_nodes[0]);
 891
 892         for (i = 0; i < nr_old_nodes; i++)
 893                 bch2_btree_interior_update_will_free_node(as, old_nodes[i]);
 894
 895         /* Repack everything with @new_format and sort down to one bset */
 896         for (i = 0; i < nr_old_nodes; i++)
 897                 new_nodes[i] =
 898                         __bch2_btree_node_alloc_replacement(as, old_nodes[i],
 899                                                             new_format);
 900
 901         /*
 902          * Conceptually we concatenate the nodes together and slice them
 903          * up at different boundaries.
 904          */
 905         for (i = nr_new_nodes - 1; i > 0; --i) {
 906                 struct btree *n1 = new_nodes[i];
 907                 struct btree *n2 = new_nodes[i - 1];
 908
 909                 struct bset *s1 = btree_bset_first(n1);
 910                 struct bset *s2 = btree_bset_first(n2);
 911                 struct bkey_packed *k, *last = NULL;
 912
 913                 /* Calculate how many keys from @n2 we could fit inside @n1 */
 914                 u64s = 0;
 915
 916                 for (k = s2->start;
 917                      k < vstruct_last(s2) &&
 918                      vstruct_blocks_plus(n1->data, c->block_bits,
 919                                          u64s + k->u64s) <= blocks;
 920                      k = bkey_next(k)) {
 921                         last = k;
 922                         u64s += k->u64s;
 923                 }
 924
 925                 if (u64s == le16_to_cpu(s2->u64s)) {
 926                         /* n2 fits entirely in n1 */
 927                         n1->key.k.p = n1->data->max_key = n2->data->max_key;
 928
 929                         memcpy_u64s(vstruct_last(s1),
 930                                     s2->start,
 931                                     le16_to_cpu(s2->u64s));
 932                         le16_add_cpu(&s1->u64s, le16_to_cpu(s2->u64s));
 933
 934                         set_btree_bset_end(n1, n1->set);
 935
 936                         six_unlock_write(&n2->lock);
 937                         bch2_btree_node_free_never_inserted(c, n2);
 938                         six_unlock_intent(&n2->lock);
 939
 940                         memmove(new_nodes + i - 1,
 941                                 new_nodes + i,
 942                                 sizeof(new_nodes[0]) * (nr_new_nodes - i));
 943                         new_nodes[--nr_new_nodes] = NULL;
 944                 } else if (u64s) {
 945                         /* move part of n2 into n1 */
 946                         n1->key.k.p = n1->data->max_key =
 947                                 bkey_unpack_pos(n1, last);
 948
 949                         n2->data->min_key =
 950                                 btree_type_successor(iter->btree_id,
 951                                                      n1->data->max_key);
 952
 953                         memcpy_u64s(vstruct_last(s1),
 954                                     s2->start, u64s);
 955                         le16_add_cpu(&s1->u64s, u64s);
 956
 957                         memmove(s2->start,
 958                                 vstruct_idx(s2, u64s),
 959                                 (le16_to_cpu(s2->u64s) - u64s) * sizeof(u64));
 960                         s2->u64s = cpu_to_le16(le16_to_cpu(s2->u64s) - u64s);
 961
 962                         set_btree_bset_end(n1, n1->set);
 963                         set_btree_bset_end(n2, n2->set);
 964                 }
 965         }
 966
 967         for (i = 0; i < nr_new_nodes; i++) {
 968                 struct btree *n = new_nodes[i];
 969
 970                 recalc_packed_keys(n);
 971                 btree_node_reset_sib_u64s(n);
 972
 973                 bch2_btree_build_aux_trees(n);
 974                 six_unlock_write(&n->lock);
 975
 976                 bch2_btree_node_write(c, n, SIX_LOCK_intent);
 977         }
 978
 979         /*
 980          * The keys for the old nodes get deleted. We don't want to insert keys
 981          * that compare equal to the keys for the new nodes we'll also be
 982          * inserting - we can't because keys on a keylist must be strictly
 983          * greater than the previous keys, and we also don't need to since the
 984          * key for the new node will serve the same purpose (overwriting the key
 985          * for the old node).
 986          */
 987         for (i = 0; i < nr_old_nodes; i++) {
 988                 struct bkey_i delete;
 989                 unsigned j;
 990
 991                 for (j = 0; j < nr_new_nodes; j++)
 992                         if (!bkey_cmp(old_nodes[i]->key.k.p,
 993                                       new_nodes[j]->key.k.p))
 994                                 goto next;
 995
 996                 bkey_init(&delete.k);
 997                 delete.k.p = old_nodes[i]->key.k.p;
 998                 bch2_keylist_add_in_order(&keylist, &delete);
 999 next:
1000                 i = i;
1001         }
1002
1003         /*
1004          * Keys for the new nodes get inserted: bch2_btree_insert_keys() only
1005          * does the lookup once and thus expects the keys to be in sorted order
1006          * so we have to make sure the new keys are correctly ordered with
1007          * respect to the deleted keys added in the previous loop
1008          */
1009         for (i = 0; i < nr_new_nodes; i++)
1010                 bch2_keylist_add_in_order(&keylist, &new_nodes[i]->key);
1011
1012         /* Insert the newly coalesced nodes */
1013         bch2_btree_insert_node(as, parent, iter, &keylist, 0);
1014
1015         BUG_ON(!bch2_keylist_empty(&keylist));
1016
1017         BUG_ON(iter->l[old_nodes[0]->level].b != old_nodes[0]);
1018
1019         bch2_btree_iter_node_replace(iter, new_nodes[0]);
1020
1021         for (i = 0; i < nr_new_nodes; i++)
1022                 bch2_open_buckets_put(c, &new_nodes[i]->ob);
1023
1024         /* Free the old nodes and update our sliding window */
1025         for (i = 0; i < nr_old_nodes; i++) {
1026                 bch2_btree_node_free_inmem(c, old_nodes[i], iter);
1027
1028                 /*
1029                  * the index update might have triggered a split, in which case
1030                  * the nodes we coalesced - the new nodes we just created -
1031                  * might not be sibling nodes anymore - don't add them to the
1032                  * sliding window (except the first):
1033                  */
1034                 if (!i) {
1035                         old_nodes[i] = new_nodes[i];
1036                 } else {
1037                         old_nodes[i] = NULL;
1038                         if (new_nodes[i])
1039                                 six_unlock_intent(&new_nodes[i]->lock);
1040                 }
1041         }
1042
1043         bch2_btree_update_done(as);
1044         bch2_keylist_free(&keylist, NULL);
1045 }
1046
1047 static int bch2_coalesce_btree(struct bch_fs *c, enum btree_id btree_id)
1048 {
1049         struct btree_trans trans;
1050         struct btree_iter *iter;
1051         struct btree *b;
1052         bool kthread = (current->flags & PF_KTHREAD) != 0;
1053         unsigned i;
1054
1055         /* Sliding window of adjacent btree nodes */
1056         struct btree *merge[GC_MERGE_NODES];
1057         u32 lock_seq[GC_MERGE_NODES];
1058
1059         bch2_trans_init(&trans, c, 0, 0);
1060
1061         /*
1062          * XXX: We don't have a good way of positively matching on sibling nodes
1063          * that have the same parent - this code works by handling the cases
1064          * where they might not have the same parent, and is thus fragile. Ugh.
1065          *
1066          * Perhaps redo this to use multiple linked iterators?
1067          */
1068         memset(merge, 0, sizeof(merge));
1069
1070         __for_each_btree_node(&trans, iter, btree_id, POS_MIN,
1071                               BTREE_MAX_DEPTH, 0,
1072                               BTREE_ITER_PREFETCH, b) {
1073                 memmove(merge + 1, merge,
1074                         sizeof(merge) - sizeof(merge[0]));
1075                 memmove(lock_seq + 1, lock_seq,
1076                         sizeof(lock_seq) - sizeof(lock_seq[0]));
1077
1078                 merge[0] = b;
1079
1080                 for (i = 1; i < GC_MERGE_NODES; i++) {
1081                         if (!merge[i] ||
1082                             !six_relock_intent(&merge[i]->lock, lock_seq[i]))
1083                                 break;
1084
1085                         if (merge[i]->level != merge[0]->level) {
1086                                 six_unlock_intent(&merge[i]->lock);
1087                                 break;
1088                         }
1089                 }
1090                 memset(merge + i, 0, (GC_MERGE_NODES - i) * sizeof(merge[0]));
1091
1092                 bch2_coalesce_nodes(c, iter, merge);
1093
1094                 for (i = 1; i < GC_MERGE_NODES && merge[i]; i++) {
1095                         lock_seq[i] = merge[i]->lock.state.seq;
1096                         six_unlock_intent(&merge[i]->lock);
1097                 }
1098
1099                 lock_seq[0] = merge[0]->lock.state.seq;
1100
1101                 if (kthread && kthread_should_stop()) {
1102                         bch2_trans_exit(&trans);
1103                         return -ESHUTDOWN;
1104                 }
1105
1106                 bch2_trans_cond_resched(&trans);
1107
1108                 /*
1109                  * If the parent node wasn't relocked, it might have been split
1110                  * and the nodes in our sliding window might not have the same
1111                  * parent anymore - blow away the sliding window:
1112                  */
1113                 if (btree_iter_node(iter, iter->level + 1) &&
1114                     !btree_node_intent_locked(iter, iter->level + 1))
1115                         memset(merge + 1, 0,
1116                                (GC_MERGE_NODES - 1) * sizeof(merge[0]));
1117         }
1118         return bch2_trans_exit(&trans);
1119 }
1120
1121 /**
1122  * bch_coalesce - coalesce adjacent nodes with low occupancy
1123  */
1124 void bch2_coalesce(struct bch_fs *c)
1125 {
1126         enum btree_id id;
1127
1128         down_read(&c->gc_lock);
1129         trace_gc_coalesce_start(c);
1130
1131         for (id = 0; id < BTREE_ID_NR; id++) {
1132                 int ret = c->btree_roots[id].b
1133                         ? bch2_coalesce_btree(c, id)
1134                         : 0;
1135
1136                 if (ret) {
1137                         if (ret != -ESHUTDOWN)
1138                                 bch_err(c, "btree coalescing failed: %d", ret);
1139                         return;
1140                 }
1141         }
1142
1143         trace_gc_coalesce_end(c);
1144         up_read(&c->gc_lock);
1145 }
1146
1147 static int bch2_gc_thread(void *arg)
1148 {
1149         struct bch_fs *c = arg;
1150         struct io_clock *clock = &c->io_clock[WRITE];
1151         unsigned long last = atomic_long_read(&clock->now);
1152         unsigned last_kick = atomic_read(&c->kick_gc);
1153         int ret;
1154
1155         set_freezable();
1156
1157         while (1) {
1158                 while (1) {
1159                         set_current_state(TASK_INTERRUPTIBLE);
1160
1161                         if (kthread_should_stop()) {
1162                                 __set_current_state(TASK_RUNNING);
1163                                 return 0;
1164                         }
1165
1166                         if (atomic_read(&c->kick_gc) != last_kick)
1167                                 break;
1168
1169                         if (c->btree_gc_periodic) {
1170                                 unsigned long next = last + c->capacity / 16;
1171
1172                                 if (atomic_long_read(&clock->now) >= next)
1173                                         break;
1174
1175                                 bch2_io_clock_schedule_timeout(clock, next);
1176                         } else {
1177                                 schedule();
1178                         }
1179
1180                         try_to_freeze();
1181                 }
1182                 __set_current_state(TASK_RUNNING);
1183
1184                 last = atomic_long_read(&clock->now);
1185                 last_kick = atomic_read(&c->kick_gc);
1186
1187                 ret = bch2_gc(c, NULL, false, false);
1188                 if (ret)
1189                         bch_err(c, "btree gc failed: %i", ret);
1190
1191                 debug_check_no_locks_held();
1192         }
1193
1194         return 0;
1195 }
1196
1197 void bch2_gc_thread_stop(struct bch_fs *c)
1198 {
1199         struct task_struct *p;
1200
1201         p = c->gc_thread;
1202         c->gc_thread = NULL;
1203
1204         if (p) {
1205                 kthread_stop(p);
1206                 put_task_struct(p);
1207         }
1208 }
1209
1210 int bch2_gc_thread_start(struct bch_fs *c)
1211 {
1212         struct task_struct *p;
1213
1214         BUG_ON(c->gc_thread);
1215
1216         p = kthread_create(bch2_gc_thread, c, "bch_gc");
1217         if (IS_ERR(p))
1218                 return PTR_ERR(p);
1219
1220         get_task_struct(p);
1221         c->gc_thread = p;
1222         wake_up_process(p);
1223         return 0;
1224 }