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

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Code for manipulating bucket marks for garbage collection.
   4  *
   5  * Copyright 2014 Datera, Inc.
   6  */
   7
   8 #include "bcachefs.h"
   9 #include "alloc_background.h"
  10 #include "backpointers.h"
  11 #include "bset.h"
  12 #include "btree_gc.h"
  13 #include "btree_update.h"
  14 #include "buckets.h"
  15 #include "buckets_waiting_for_journal.h"
  16 #include "ec.h"
  17 #include "error.h"
  18 #include "inode.h"
  19 #include "movinggc.h"
  20 #include "recovery.h"
  21 #include "reflink.h"
  22 #include "replicas.h"
  23 #include "subvolume.h"
  24
  25 #include <linux/preempt.h>
  26 #include <trace/events/bcachefs.h>
  27
  28 static inline void fs_usage_data_type_to_base(struct bch_fs_usage *fs_usage,
  29                                               enum bch_data_type data_type,
  30                                               s64 sectors)
  31 {
  32         switch (data_type) {
  33         case BCH_DATA_btree:
  34                 fs_usage->btree         += sectors;
  35                 break;
  36         case BCH_DATA_user:
  37         case BCH_DATA_parity:
  38                 fs_usage->data          += sectors;
  39                 break;
  40         case BCH_DATA_cached:
  41                 fs_usage->cached        += sectors;
  42                 break;
  43         default:
  44                 break;
  45         }
  46 }
  47
  48 void bch2_fs_usage_initialize(struct bch_fs *c)
  49 {
  50         struct bch_fs_usage *usage;
  51         struct bch_dev *ca;
  52         unsigned i;
  53
  54         percpu_down_write(&c->mark_lock);
  55         usage = c->usage_base;
  56
  57         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
  58                 bch2_fs_usage_acc_to_base(c, i);
  59
  60         for (i = 0; i < BCH_REPLICAS_MAX; i++)
  61                 usage->reserved += usage->persistent_reserved[i];
  62
  63         for (i = 0; i < c->replicas.nr; i++) {
  64                 struct bch_replicas_entry *e =
  65                         cpu_replicas_entry(&c->replicas, i);
  66
  67                 fs_usage_data_type_to_base(usage, e->data_type, usage->replicas[i]);
  68         }
  69
  70         for_each_member_device(ca, c, i) {
  71                 struct bch_dev_usage dev = bch2_dev_usage_read(ca);
  72
  73                 usage->hidden += (dev.d[BCH_DATA_sb].buckets +
  74                                   dev.d[BCH_DATA_journal].buckets) *
  75                         ca->mi.bucket_size;
  76         }
  77
  78         percpu_up_write(&c->mark_lock);
  79 }
  80
  81 static inline struct bch_dev_usage *dev_usage_ptr(struct bch_dev *ca,
  82                                                   unsigned journal_seq,
  83                                                   bool gc)
  84 {
  85         BUG_ON(!gc && !journal_seq);
  86
  87         return this_cpu_ptr(gc
  88                             ? ca->usage_gc
  89                             : ca->usage[journal_seq & JOURNAL_BUF_MASK]);
  90 }
  91
  92 void bch2_dev_usage_read_fast(struct bch_dev *ca, struct bch_dev_usage *usage)
  93 {
  94         struct bch_fs *c = ca->fs;
  95         unsigned seq, i, u64s = dev_usage_u64s();
  96
  97         do {
  98                 seq = read_seqcount_begin(&c->usage_lock);
  99                 memcpy(usage, ca->usage_base, u64s * sizeof(u64));
 100                 for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
 101                         acc_u64s_percpu((u64 *) usage, (u64 __percpu *) ca->usage[i], u64s);
 102         } while (read_seqcount_retry(&c->usage_lock, seq));
 103 }
 104
 105 static inline struct bch_fs_usage *fs_usage_ptr(struct bch_fs *c,
 106                                                 unsigned journal_seq,
 107                                                 bool gc)
 108 {
 109         percpu_rwsem_assert_held(&c->mark_lock);
 110         BUG_ON(!gc && !journal_seq);
 111
 112         return this_cpu_ptr(gc
 113                             ? c->usage_gc
 114                             : c->usage[journal_seq & JOURNAL_BUF_MASK]);
 115 }
 116
 117 u64 bch2_fs_usage_read_one(struct bch_fs *c, u64 *v)
 118 {
 119         ssize_t offset = v - (u64 *) c->usage_base;
 120         unsigned i, seq;
 121         u64 ret;
 122
 123         BUG_ON(offset < 0 || offset >= fs_usage_u64s(c));
 124         percpu_rwsem_assert_held(&c->mark_lock);
 125
 126         do {
 127                 seq = read_seqcount_begin(&c->usage_lock);
 128                 ret = *v;
 129
 130                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 131                         ret += percpu_u64_get((u64 __percpu *) c->usage[i] + offset);
 132         } while (read_seqcount_retry(&c->usage_lock, seq));
 133
 134         return ret;
 135 }
 136
 137 struct bch_fs_usage_online *bch2_fs_usage_read(struct bch_fs *c)
 138 {
 139         struct bch_fs_usage_online *ret;
 140         unsigned seq, i, v, u64s = fs_usage_u64s(c) + 1;
 141 retry:
 142         ret = kmalloc(u64s * sizeof(u64), GFP_NOFS);
 143         if (unlikely(!ret))
 144                 return NULL;
 145
 146         percpu_down_read(&c->mark_lock);
 147
 148         v = fs_usage_u64s(c) + 1;
 149         if (unlikely(u64s != v)) {
 150                 u64s = v;
 151                 percpu_up_read(&c->mark_lock);
 152                 kfree(ret);
 153                 goto retry;
 154         }
 155
 156         ret->online_reserved = percpu_u64_get(c->online_reserved);
 157
 158         do {
 159                 seq = read_seqcount_begin(&c->usage_lock);
 160                 unsafe_memcpy(&ret->u, c->usage_base, u64s * sizeof(u64),
 161                               "embedded variable length struct");
 162                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 163                         acc_u64s_percpu((u64 *) &ret->u, (u64 __percpu *) c->usage[i], u64s);
 164         } while (read_seqcount_retry(&c->usage_lock, seq));
 165
 166         return ret;
 167 }
 168
 169 void bch2_fs_usage_acc_to_base(struct bch_fs *c, unsigned idx)
 170 {
 171         struct bch_dev *ca;
 172         unsigned i, u64s = fs_usage_u64s(c);
 173
 174         BUG_ON(idx >= ARRAY_SIZE(c->usage));
 175
 176         preempt_disable();
 177         write_seqcount_begin(&c->usage_lock);
 178
 179         acc_u64s_percpu((u64 *) c->usage_base,
 180                         (u64 __percpu *) c->usage[idx], u64s);
 181         percpu_memset(c->usage[idx], 0, u64s * sizeof(u64));
 182
 183         rcu_read_lock();
 184         for_each_member_device_rcu(ca, c, i, NULL) {
 185                 u64s = dev_usage_u64s();
 186
 187                 acc_u64s_percpu((u64 *) ca->usage_base,
 188                                 (u64 __percpu *) ca->usage[idx], u64s);
 189                 percpu_memset(ca->usage[idx], 0, u64s * sizeof(u64));
 190         }
 191         rcu_read_unlock();
 192
 193         write_seqcount_end(&c->usage_lock);
 194         preempt_enable();
 195 }
 196
 197 void bch2_fs_usage_to_text(struct printbuf *out,
 198                            struct bch_fs *c,
 199                            struct bch_fs_usage_online *fs_usage)
 200 {
 201         unsigned i;
 202
 203         prt_printf(out, "capacity:\t\t\t%llu\n", c->capacity);
 204
 205         prt_printf(out, "hidden:\t\t\t\t%llu\n",
 206                fs_usage->u.hidden);
 207         prt_printf(out, "data:\t\t\t\t%llu\n",
 208                fs_usage->u.data);
 209         prt_printf(out, "cached:\t\t\t\t%llu\n",
 210                fs_usage->u.cached);
 211         prt_printf(out, "reserved:\t\t\t%llu\n",
 212                fs_usage->u.reserved);
 213         prt_printf(out, "nr_inodes:\t\t\t%llu\n",
 214                fs_usage->u.nr_inodes);
 215         prt_printf(out, "online reserved:\t\t%llu\n",
 216                fs_usage->online_reserved);
 217
 218         for (i = 0;
 219              i < ARRAY_SIZE(fs_usage->u.persistent_reserved);
 220              i++) {
 221                 prt_printf(out, "%u replicas:\n", i + 1);
 222                 prt_printf(out, "\treserved:\t\t%llu\n",
 223                        fs_usage->u.persistent_reserved[i]);
 224         }
 225
 226         for (i = 0; i < c->replicas.nr; i++) {
 227                 struct bch_replicas_entry *e =
 228                         cpu_replicas_entry(&c->replicas, i);
 229
 230                 prt_printf(out, "\t");
 231                 bch2_replicas_entry_to_text(out, e);
 232                 prt_printf(out, ":\t%llu\n", fs_usage->u.replicas[i]);
 233         }
 234 }
 235
 236 static u64 reserve_factor(u64 r)
 237 {
 238         return r + (round_up(r, (1 << RESERVE_FACTOR)) >> RESERVE_FACTOR);
 239 }
 240
 241 u64 bch2_fs_sectors_used(struct bch_fs *c, struct bch_fs_usage_online *fs_usage)
 242 {
 243         return min(fs_usage->u.hidden +
 244                    fs_usage->u.btree +
 245                    fs_usage->u.data +
 246                    reserve_factor(fs_usage->u.reserved +
 247                                   fs_usage->online_reserved),
 248                    c->capacity);
 249 }
 250
 251 static struct bch_fs_usage_short
 252 __bch2_fs_usage_read_short(struct bch_fs *c)
 253 {
 254         struct bch_fs_usage_short ret;
 255         u64 data, reserved;
 256
 257         ret.capacity = c->capacity -
 258                 bch2_fs_usage_read_one(c, &c->usage_base->hidden);
 259
 260         data            = bch2_fs_usage_read_one(c, &c->usage_base->data) +
 261                 bch2_fs_usage_read_one(c, &c->usage_base->btree);
 262         reserved        = bch2_fs_usage_read_one(c, &c->usage_base->reserved) +
 263                 percpu_u64_get(c->online_reserved);
 264
 265         ret.used        = min(ret.capacity, data + reserve_factor(reserved));
 266         ret.free        = ret.capacity - ret.used;
 267
 268         ret.nr_inodes   = bch2_fs_usage_read_one(c, &c->usage_base->nr_inodes);
 269
 270         return ret;
 271 }
 272
 273 struct bch_fs_usage_short
 274 bch2_fs_usage_read_short(struct bch_fs *c)
 275 {
 276         struct bch_fs_usage_short ret;
 277
 278         percpu_down_read(&c->mark_lock);
 279         ret = __bch2_fs_usage_read_short(c);
 280         percpu_up_read(&c->mark_lock);
 281
 282         return ret;
 283 }
 284
 285 void bch2_dev_usage_init(struct bch_dev *ca)
 286 {
 287         ca->usage_base->d[BCH_DATA_free].buckets = ca->mi.nbuckets - ca->mi.first_bucket;
 288 }
 289
 290 static inline int bucket_sectors_fragmented(struct bch_dev *ca,
 291                                             struct bch_alloc_v4 a)
 292 {
 293         return a.dirty_sectors
 294                 ? max(0, (int) ca->mi.bucket_size - (int) a.dirty_sectors)
 295                 : 0;
 296 }
 297
 298 static void bch2_dev_usage_update(struct bch_fs *c, struct bch_dev *ca,
 299                                   struct bch_alloc_v4 old,
 300                                   struct bch_alloc_v4 new,
 301                                   u64 journal_seq, bool gc)
 302 {
 303         struct bch_fs_usage *fs_usage;
 304         struct bch_dev_usage *u;
 305
 306         preempt_disable();
 307         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 308
 309         if (data_type_is_hidden(old.data_type))
 310                 fs_usage->hidden -= ca->mi.bucket_size;
 311         if (data_type_is_hidden(new.data_type))
 312                 fs_usage->hidden += ca->mi.bucket_size;
 313
 314         u = dev_usage_ptr(ca, journal_seq, gc);
 315
 316         u->d[old.data_type].buckets--;
 317         u->d[new.data_type].buckets++;
 318
 319         u->buckets_ec -= (int) !!old.stripe;
 320         u->buckets_ec += (int) !!new.stripe;
 321
 322         u->d[old.data_type].sectors -= old.dirty_sectors;
 323         u->d[new.data_type].sectors += new.dirty_sectors;
 324
 325         u->d[BCH_DATA_cached].sectors += new.cached_sectors;
 326         u->d[BCH_DATA_cached].sectors -= old.cached_sectors;
 327
 328         u->d[old.data_type].fragmented -= bucket_sectors_fragmented(ca, old);
 329         u->d[new.data_type].fragmented += bucket_sectors_fragmented(ca, new);
 330
 331         preempt_enable();
 332 }
 333
 334 static void bch2_dev_usage_update_m(struct bch_fs *c, struct bch_dev *ca,
 335                                     struct bucket old, struct bucket new,
 336                                     u64 journal_seq, bool gc)
 337 {
 338         struct bch_alloc_v4 old_a = {
 339                 .gen            = old.gen,
 340                 .data_type      = old.data_type,
 341                 .dirty_sectors  = old.dirty_sectors,
 342                 .cached_sectors = old.cached_sectors,
 343                 .stripe         = old.stripe,
 344         };
 345         struct bch_alloc_v4 new_a = {
 346                 .gen            = new.gen,
 347                 .data_type      = new.data_type,
 348                 .dirty_sectors  = new.dirty_sectors,
 349                 .cached_sectors = new.cached_sectors,
 350                 .stripe         = new.stripe,
 351         };
 352
 353         bch2_dev_usage_update(c, ca, old_a, new_a, journal_seq, gc);
 354 }
 355
 356 static inline int __update_replicas(struct bch_fs *c,
 357                                     struct bch_fs_usage *fs_usage,
 358                                     struct bch_replicas_entry *r,
 359                                     s64 sectors)
 360 {
 361         int idx = bch2_replicas_entry_idx(c, r);
 362
 363         if (idx < 0)
 364                 return -1;
 365
 366         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 367         fs_usage->replicas[idx]         += sectors;
 368         return 0;
 369 }
 370
 371 static inline int update_replicas(struct bch_fs *c, struct bkey_s_c k,
 372                         struct bch_replicas_entry *r, s64 sectors,
 373                         unsigned journal_seq, bool gc)
 374 {
 375         struct bch_fs_usage __percpu *fs_usage;
 376         int idx, ret = 0;
 377         struct printbuf buf = PRINTBUF;
 378
 379         percpu_down_read(&c->mark_lock);
 380         buf.atomic++;
 381
 382         idx = bch2_replicas_entry_idx(c, r);
 383         if (idx < 0 &&
 384             fsck_err(c, "no replicas entry\n"
 385                      "  while marking %s",
 386                      (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 387                 percpu_up_read(&c->mark_lock);
 388                 ret = bch2_mark_replicas(c, r);
 389                 percpu_down_read(&c->mark_lock);
 390
 391                 if (ret)
 392                         goto err;
 393                 idx = bch2_replicas_entry_idx(c, r);
 394         }
 395         if (idx < 0) {
 396                 ret = -1;
 397                 goto err;
 398         }
 399
 400         preempt_disable();
 401         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 402         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 403         fs_usage->replicas[idx]         += sectors;
 404         preempt_enable();
 405 err:
 406 fsck_err:
 407         percpu_up_read(&c->mark_lock);
 408         printbuf_exit(&buf);
 409         return ret;
 410 }
 411
 412 static inline int update_cached_sectors(struct bch_fs *c,
 413                         struct bkey_s_c k,
 414                         unsigned dev, s64 sectors,
 415                         unsigned journal_seq, bool gc)
 416 {
 417         struct bch_replicas_padded r;
 418
 419         bch2_replicas_entry_cached(&r.e, dev);
 420
 421         return update_replicas(c, k, &r.e, sectors, journal_seq, gc);
 422 }
 423
 424 static struct replicas_delta_list *
 425 replicas_deltas_realloc(struct btree_trans *trans, unsigned more)
 426 {
 427         struct replicas_delta_list *d = trans->fs_usage_deltas;
 428         unsigned new_size = d ? (d->size + more) * 2 : 128;
 429         unsigned alloc_size = sizeof(*d) + new_size;
 430
 431         WARN_ON_ONCE(alloc_size > REPLICAS_DELTA_LIST_MAX);
 432
 433         if (!d || d->used + more > d->size) {
 434                 d = krealloc(d, alloc_size, GFP_NOIO|__GFP_ZERO);
 435
 436                 BUG_ON(!d && alloc_size > REPLICAS_DELTA_LIST_MAX);
 437
 438                 if (!d) {
 439                         d = mempool_alloc(&trans->c->replicas_delta_pool, GFP_NOIO);
 440                         memset(d, 0, REPLICAS_DELTA_LIST_MAX);
 441
 442                         if (trans->fs_usage_deltas)
 443                                 memcpy(d, trans->fs_usage_deltas,
 444                                        trans->fs_usage_deltas->size + sizeof(*d));
 445
 446                         new_size = REPLICAS_DELTA_LIST_MAX - sizeof(*d);
 447                         kfree(trans->fs_usage_deltas);
 448                 }
 449
 450                 d->size = new_size;
 451                 trans->fs_usage_deltas = d;
 452         }
 453         return d;
 454 }
 455
 456 static inline void update_replicas_list(struct btree_trans *trans,
 457                                         struct bch_replicas_entry *r,
 458                                         s64 sectors)
 459 {
 460         struct replicas_delta_list *d;
 461         struct replicas_delta *n;
 462         unsigned b;
 463
 464         if (!sectors)
 465                 return;
 466
 467         b = replicas_entry_bytes(r) + 8;
 468         d = replicas_deltas_realloc(trans, b);
 469
 470         n = (void *) d->d + d->used;
 471         n->delta = sectors;
 472         memcpy((void *) n + offsetof(struct replicas_delta, r),
 473                r, replicas_entry_bytes(r));
 474         bch2_replicas_entry_sort(&n->r);
 475         d->used += b;
 476 }
 477
 478 static inline void update_cached_sectors_list(struct btree_trans *trans,
 479                                               unsigned dev, s64 sectors)
 480 {
 481         struct bch_replicas_padded r;
 482
 483         bch2_replicas_entry_cached(&r.e, dev);
 484
 485         update_replicas_list(trans, &r.e, sectors);
 486 }
 487
 488 int bch2_mark_alloc(struct btree_trans *trans,
 489                     struct bkey_s_c old, struct bkey_s_c new,
 490                     unsigned flags)
 491 {
 492         bool gc = flags & BTREE_TRIGGER_GC;
 493         u64 journal_seq = trans->journal_res.seq;
 494         u64 bucket_journal_seq;
 495         struct bch_fs *c = trans->c;
 496         struct bch_alloc_v4 old_a_convert, new_a_convert;
 497         const struct bch_alloc_v4 *old_a, *new_a;
 498         struct bch_dev *ca;
 499         int ret = 0;
 500
 501         /*
 502          * alloc btree is read in by bch2_alloc_read, not gc:
 503          */
 504         if ((flags & BTREE_TRIGGER_GC) &&
 505             !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
 506                 return 0;
 507
 508         if (bch2_trans_inconsistent_on(!bch2_dev_bucket_exists(c, new.k->p), trans,
 509                                        "alloc key for invalid device or bucket"))
 510                 return -EIO;
 511
 512         ca = bch_dev_bkey_exists(c, new.k->p.inode);
 513
 514         old_a = bch2_alloc_to_v4(old, &old_a_convert);
 515         new_a = bch2_alloc_to_v4(new, &new_a_convert);
 516
 517         bucket_journal_seq = new_a->journal_seq;
 518
 519         if ((flags & BTREE_TRIGGER_INSERT) &&
 520             data_type_is_empty(old_a->data_type) !=
 521             data_type_is_empty(new_a->data_type) &&
 522             new.k->type == KEY_TYPE_alloc_v4) {
 523                 struct bch_alloc_v4 *v = (struct bch_alloc_v4 *) new.v;
 524
 525                 EBUG_ON(!journal_seq);
 526
 527                 /*
 528                  * If the btree updates referring to a bucket weren't flushed
 529                  * before the bucket became empty again, then the we don't have
 530                  * to wait on a journal flush before we can reuse the bucket:
 531                  */
 532                 v->journal_seq = bucket_journal_seq =
 533                         data_type_is_empty(new_a->data_type) &&
 534                         (journal_seq == v->journal_seq ||
 535                          bch2_journal_noflush_seq(&c->journal, v->journal_seq))
 536                         ? 0 : journal_seq;
 537         }
 538
 539         if (!data_type_is_empty(old_a->data_type) &&
 540             data_type_is_empty(new_a->data_type) &&
 541             bucket_journal_seq) {
 542                 ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 543                                 c->journal.flushed_seq_ondisk,
 544                                 new.k->p.inode, new.k->p.offset,
 545                                 bucket_journal_seq);
 546                 if (ret) {
 547                         bch2_fs_fatal_error(c,
 548                                 "error setting bucket_needs_journal_commit: %i", ret);
 549                         return ret;
 550                 }
 551         }
 552
 553         percpu_down_read(&c->mark_lock);
 554         if (!gc && new_a->gen != old_a->gen)
 555                 *bucket_gen(ca, new.k->p.offset) = new_a->gen;
 556
 557         bch2_dev_usage_update(c, ca, *old_a, *new_a, journal_seq, gc);
 558
 559         if (gc) {
 560                 struct bucket *g = gc_bucket(ca, new.k->p.offset);
 561
 562                 bucket_lock(g);
 563
 564                 g->gen_valid            = 1;
 565                 g->gen                  = new_a->gen;
 566                 g->data_type            = new_a->data_type;
 567                 g->stripe               = new_a->stripe;
 568                 g->stripe_redundancy    = new_a->stripe_redundancy;
 569                 g->dirty_sectors        = new_a->dirty_sectors;
 570                 g->cached_sectors       = new_a->cached_sectors;
 571
 572                 bucket_unlock(g);
 573         }
 574         percpu_up_read(&c->mark_lock);
 575
 576         /*
 577          * need to know if we're getting called from the invalidate path or
 578          * not:
 579          */
 580
 581         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 582             old_a->cached_sectors) {
 583                 ret = update_cached_sectors(c, new, ca->dev_idx,
 584                                             -((s64) old_a->cached_sectors),
 585                                             journal_seq, gc);
 586                 if (ret) {
 587                         bch2_fs_fatal_error(c, "%s(): no replicas entry while updating cached sectors",
 588                                             __func__);
 589                         return ret;
 590                 }
 591         }
 592
 593         if (new_a->data_type == BCH_DATA_free &&
 594             (!new_a->journal_seq || new_a->journal_seq < c->journal.flushed_seq_ondisk))
 595                 closure_wake_up(&c->freelist_wait);
 596
 597         if (new_a->data_type == BCH_DATA_need_discard &&
 598             (!bucket_journal_seq || bucket_journal_seq < c->journal.flushed_seq_ondisk))
 599                 bch2_do_discards(c);
 600
 601         if (old_a->data_type != BCH_DATA_cached &&
 602             new_a->data_type == BCH_DATA_cached &&
 603             should_invalidate_buckets(ca, bch2_dev_usage_read(ca)))
 604                 bch2_do_invalidates(c);
 605
 606         if (new_a->data_type == BCH_DATA_need_gc_gens)
 607                 bch2_do_gc_gens(c);
 608
 609         return 0;
 610 }
 611
 612 int bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 613                               size_t b, enum bch_data_type data_type,
 614                               unsigned sectors, struct gc_pos pos,
 615                               unsigned flags)
 616 {
 617         struct bucket old, new, *g;
 618         int ret = 0;
 619
 620         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 621         BUG_ON(data_type != BCH_DATA_sb &&
 622                data_type != BCH_DATA_journal);
 623
 624         /*
 625          * Backup superblock might be past the end of our normal usable space:
 626          */
 627         if (b >= ca->mi.nbuckets)
 628                 return 0;
 629
 630         percpu_down_read(&c->mark_lock);
 631         g = gc_bucket(ca, b);
 632
 633         bucket_lock(g);
 634         old = *g;
 635
 636         if (bch2_fs_inconsistent_on(g->data_type &&
 637                         g->data_type != data_type, c,
 638                         "different types of data in same bucket: %s, %s",
 639                         bch2_data_types[g->data_type],
 640                         bch2_data_types[data_type])) {
 641                 ret = -EIO;
 642                 goto err;
 643         }
 644
 645         if (bch2_fs_inconsistent_on((u64) g->dirty_sectors + sectors > ca->mi.bucket_size, c,
 646                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > bucket size",
 647                         ca->dev_idx, b, g->gen,
 648                         bch2_data_types[g->data_type ?: data_type],
 649                         g->dirty_sectors, sectors)) {
 650                 ret = -EIO;
 651                 goto err;
 652         }
 653
 654
 655         g->data_type = data_type;
 656         g->dirty_sectors += sectors;
 657         new = *g;
 658 err:
 659         bucket_unlock(g);
 660         if (!ret)
 661                 bch2_dev_usage_update_m(c, ca, old, new, 0, true);
 662         percpu_up_read(&c->mark_lock);
 663         return ret;
 664 }
 665
 666 static int check_bucket_ref(struct btree_trans *trans,
 667                             struct bkey_s_c k,
 668                             const struct bch_extent_ptr *ptr,
 669                             s64 sectors, enum bch_data_type ptr_data_type,
 670                             u8 b_gen, u8 bucket_data_type,
 671                             u32 dirty_sectors, u32 cached_sectors)
 672 {
 673         struct bch_fs *c = trans->c;
 674         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 675         size_t bucket_nr = PTR_BUCKET_NR(ca, ptr);
 676         u16 bucket_sectors = !ptr->cached
 677                 ? dirty_sectors
 678                 : cached_sectors;
 679         struct printbuf buf = PRINTBUF;
 680         int ret = 0;
 681
 682         if (bucket_data_type == BCH_DATA_cached)
 683                 bucket_data_type = BCH_DATA_user;
 684
 685         if ((bucket_data_type == BCH_DATA_stripe && ptr_data_type == BCH_DATA_user) ||
 686             (bucket_data_type == BCH_DATA_user   && ptr_data_type == BCH_DATA_stripe))
 687                 bucket_data_type = ptr_data_type = BCH_DATA_stripe;
 688
 689         if (gen_after(ptr->gen, b_gen)) {
 690                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 691                         "bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\n"
 692                         "while marking %s",
 693                         ptr->dev, bucket_nr, b_gen,
 694                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 695                         ptr->gen,
 696                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 697                 ret = -EIO;
 698                 goto err;
 699         }
 700
 701         if (gen_cmp(b_gen, ptr->gen) > BUCKET_GC_GEN_MAX) {
 702                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 703                         "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 704                         "while marking %s",
 705                         ptr->dev, bucket_nr, b_gen,
 706                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 707                         ptr->gen,
 708                         (printbuf_reset(&buf),
 709                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 710                 ret = -EIO;
 711                 goto err;
 712         }
 713
 714         if (b_gen != ptr->gen && !ptr->cached) {
 715                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 716                         "bucket %u:%zu gen %u (mem gen %u) data type %s: stale dirty ptr (gen %u)\n"
 717                         "while marking %s",
 718                         ptr->dev, bucket_nr, b_gen,
 719                         *bucket_gen(ca, bucket_nr),
 720                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 721                         ptr->gen,
 722                         (printbuf_reset(&buf),
 723                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 724                 ret = -EIO;
 725                 goto err;
 726         }
 727
 728         if (b_gen != ptr->gen) {
 729                 ret = 1;
 730                 goto out;
 731         }
 732
 733         if (!data_type_is_empty(bucket_data_type) &&
 734             ptr_data_type &&
 735             bucket_data_type != ptr_data_type) {
 736                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 737                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 738                         "while marking %s",
 739                         ptr->dev, bucket_nr, b_gen,
 740                         bch2_data_types[bucket_data_type],
 741                         bch2_data_types[ptr_data_type],
 742                         (printbuf_reset(&buf),
 743                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 744                 ret = -EIO;
 745                 goto err;
 746         }
 747
 748         if ((unsigned) (bucket_sectors + sectors) > U32_MAX) {
 749                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 750                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U16_MAX\n"
 751                         "while marking %s",
 752                         ptr->dev, bucket_nr, b_gen,
 753                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 754                         bucket_sectors, sectors,
 755                         (printbuf_reset(&buf),
 756                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 757                 ret = -EIO;
 758                 goto err;
 759         }
 760 out:
 761         printbuf_exit(&buf);
 762         return ret;
 763 err:
 764         bch2_dump_trans_updates(trans);
 765         goto out;
 766 }
 767
 768 static int mark_stripe_bucket(struct btree_trans *trans,
 769                               struct bkey_s_c k,
 770                               unsigned ptr_idx,
 771                               unsigned flags)
 772 {
 773         struct bch_fs *c = trans->c;
 774         u64 journal_seq = trans->journal_res.seq;
 775         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 776         unsigned nr_data = s->nr_blocks - s->nr_redundant;
 777         bool parity = ptr_idx >= nr_data;
 778         enum bch_data_type data_type = parity ? BCH_DATA_parity : 0;
 779         s64 sectors = parity ? le16_to_cpu(s->sectors) : 0;
 780         const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
 781         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 782         struct bucket old, new, *g;
 783         struct printbuf buf = PRINTBUF;
 784         int ret = 0;
 785
 786         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 787
 788         /* * XXX doesn't handle deletion */
 789
 790         percpu_down_read(&c->mark_lock);
 791         buf.atomic++;
 792         g = PTR_GC_BUCKET(ca, ptr);
 793
 794         if (g->dirty_sectors ||
 795             (g->stripe && g->stripe != k.k->p.offset)) {
 796                 bch2_fs_inconsistent(c,
 797                               "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
 798                               ptr->dev, PTR_BUCKET_NR(ca, ptr), g->gen,
 799                               (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 800                 ret = -EINVAL;
 801                 goto err;
 802         }
 803
 804         bucket_lock(g);
 805         old = *g;
 806
 807         ret = check_bucket_ref(trans, k, ptr, sectors, data_type,
 808                                g->gen, g->data_type,
 809                                g->dirty_sectors, g->cached_sectors);
 810         if (ret)
 811                 goto err;
 812
 813         if (data_type)
 814                 g->data_type = data_type;
 815         g->dirty_sectors += sectors;
 816
 817         g->stripe               = k.k->p.offset;
 818         g->stripe_redundancy    = s->nr_redundant;
 819         new = *g;
 820 err:
 821         bucket_unlock(g);
 822         if (!ret)
 823                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 824         percpu_up_read(&c->mark_lock);
 825         printbuf_exit(&buf);
 826         return ret;
 827 }
 828
 829 static int __mark_pointer(struct btree_trans *trans,
 830                           struct bkey_s_c k,
 831                           const struct bch_extent_ptr *ptr,
 832                           s64 sectors, enum bch_data_type ptr_data_type,
 833                           u8 bucket_gen, u8 *bucket_data_type,
 834                           u32 *dirty_sectors, u32 *cached_sectors)
 835 {
 836         u32 *dst_sectors = !ptr->cached
 837                 ? dirty_sectors
 838                 : cached_sectors;
 839         int ret = check_bucket_ref(trans, k, ptr, sectors, ptr_data_type,
 840                                    bucket_gen, *bucket_data_type,
 841                                    *dirty_sectors, *cached_sectors);
 842
 843         if (ret)
 844                 return ret;
 845
 846         *dst_sectors += sectors;
 847         *bucket_data_type = *dirty_sectors || *cached_sectors
 848                 ? ptr_data_type : 0;
 849         return 0;
 850 }
 851
 852 static int bch2_mark_pointer(struct btree_trans *trans,
 853                              struct bkey_s_c k,
 854                              struct extent_ptr_decoded p,
 855                              s64 sectors, enum bch_data_type data_type,
 856                              unsigned flags)
 857 {
 858         u64 journal_seq = trans->journal_res.seq;
 859         struct bch_fs *c = trans->c;
 860         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 861         struct bucket old, new, *g;
 862         u8 bucket_data_type;
 863         int ret = 0;
 864
 865         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 866
 867         percpu_down_read(&c->mark_lock);
 868         g = PTR_GC_BUCKET(ca, &p.ptr);
 869         bucket_lock(g);
 870         old = *g;
 871
 872         bucket_data_type = g->data_type;
 873         ret = __mark_pointer(trans, k, &p.ptr, sectors,
 874                              data_type, g->gen,
 875                              &bucket_data_type,
 876                              &g->dirty_sectors,
 877                              &g->cached_sectors);
 878         if (!ret)
 879                 g->data_type = bucket_data_type;
 880
 881         new = *g;
 882         bucket_unlock(g);
 883         if (!ret)
 884                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 885         percpu_up_read(&c->mark_lock);
 886
 887         return ret;
 888 }
 889
 890 static int bch2_mark_stripe_ptr(struct btree_trans *trans,
 891                                 struct bkey_s_c k,
 892                                 struct bch_extent_stripe_ptr p,
 893                                 enum bch_data_type data_type,
 894                                 s64 sectors,
 895                                 unsigned flags)
 896 {
 897         struct bch_fs *c = trans->c;
 898         struct bch_replicas_padded r;
 899         struct gc_stripe *m;
 900
 901         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 902
 903         m = genradix_ptr_alloc(&c->gc_stripes, p.idx, GFP_KERNEL);
 904         if (!m) {
 905                 bch_err(c, "error allocating memory for gc_stripes, idx %llu",
 906                         (u64) p.idx);
 907                 return -ENOMEM;
 908         }
 909
 910         spin_lock(&c->ec_stripes_heap_lock);
 911
 912         if (!m || !m->alive) {
 913                 spin_unlock(&c->ec_stripes_heap_lock);
 914                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
 915                                     (u64) p.idx);
 916                 bch2_inconsistent_error(c);
 917                 return -EIO;
 918         }
 919
 920         m->block_sectors[p.block] += sectors;
 921
 922         r = m->r;
 923         spin_unlock(&c->ec_stripes_heap_lock);
 924
 925         r.e.data_type = data_type;
 926         update_replicas(c, k, &r.e, sectors, trans->journal_res.seq, true);
 927
 928         return 0;
 929 }
 930
 931 int bch2_mark_extent(struct btree_trans *trans,
 932                      struct bkey_s_c old, struct bkey_s_c new,
 933                      unsigned flags)
 934 {
 935         u64 journal_seq = trans->journal_res.seq;
 936         struct bch_fs *c = trans->c;
 937         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old : new;
 938         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 939         const union bch_extent_entry *entry;
 940         struct extent_ptr_decoded p;
 941         struct bch_replicas_padded r;
 942         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
 943                 ? BCH_DATA_btree
 944                 : BCH_DATA_user;
 945         s64 sectors = bkey_is_btree_ptr(k.k)
 946                 ? btree_sectors(c)
 947                 : k.k->size;
 948         s64 dirty_sectors = 0;
 949         bool stale;
 950         int ret;
 951
 952         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 953
 954         r.e.data_type   = data_type;
 955         r.e.nr_devs     = 0;
 956         r.e.nr_required = 1;
 957
 958         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 959                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
 960
 961                 if (flags & BTREE_TRIGGER_OVERWRITE)
 962                         disk_sectors = -disk_sectors;
 963
 964                 ret = bch2_mark_pointer(trans, k, p, disk_sectors,
 965                                         data_type, flags);
 966                 if (ret < 0)
 967                         return ret;
 968
 969                 stale = ret > 0;
 970
 971                 if (p.ptr.cached) {
 972                         if (!stale) {
 973                                 ret = update_cached_sectors(c, k, p.ptr.dev,
 974                                                 disk_sectors, journal_seq, true);
 975                                 if (ret) {
 976                                         bch2_fs_fatal_error(c, "%s(): no replicas entry while updating cached sectors",
 977                                                             __func__);
 978                                         return ret;
 979                                 }
 980                         }
 981                 } else if (!p.has_ec) {
 982                         dirty_sectors          += disk_sectors;
 983                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
 984                 } else {
 985                         ret = bch2_mark_stripe_ptr(trans, k, p.ec, data_type,
 986                                         disk_sectors, flags);
 987                         if (ret)
 988                                 return ret;
 989
 990                         /*
 991                          * There may be other dirty pointers in this extent, but
 992                          * if so they're not required for mounting if we have an
 993                          * erasure coded pointer in this extent:
 994                          */
 995                         r.e.nr_required = 0;
 996                 }
 997         }
 998
 999         if (r.e.nr_devs) {
1000                 ret = update_replicas(c, k, &r.e, dirty_sectors, journal_seq, true);
1001                 if (ret) {
1002                         struct printbuf buf = PRINTBUF;
1003
1004                         bch2_bkey_val_to_text(&buf, c, k);
1005                         bch2_fs_fatal_error(c, "%s(): no replicas entry for %s", __func__, buf.buf);
1006                         printbuf_exit(&buf);
1007                         return ret;
1008                 }
1009         }
1010
1011         return 0;
1012 }
1013
1014 int bch2_mark_stripe(struct btree_trans *trans,
1015                      struct bkey_s_c old, struct bkey_s_c new,
1016                      unsigned flags)
1017 {
1018         bool gc = flags & BTREE_TRIGGER_GC;
1019         u64 journal_seq = trans->journal_res.seq;
1020         struct bch_fs *c = trans->c;
1021         u64 idx = new.k->p.offset;
1022         const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
1023                 ? bkey_s_c_to_stripe(old).v : NULL;
1024         const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
1025                 ? bkey_s_c_to_stripe(new).v : NULL;
1026         unsigned i;
1027         int ret;
1028
1029         BUG_ON(gc && old_s);
1030
1031         if (!gc) {
1032                 struct stripe *m = genradix_ptr(&c->stripes, idx);
1033
1034                 if (!m || (old_s && !m->alive)) {
1035                         struct printbuf buf1 = PRINTBUF;
1036                         struct printbuf buf2 = PRINTBUF;
1037
1038                         bch2_bkey_val_to_text(&buf1, c, old);
1039                         bch2_bkey_val_to_text(&buf2, c, new);
1040                         bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
1041                                             "old %s\n"
1042                                             "new %s", idx, buf1.buf, buf2.buf);
1043                         printbuf_exit(&buf2);
1044                         printbuf_exit(&buf1);
1045                         bch2_inconsistent_error(c);
1046                         return -1;
1047                 }
1048
1049                 if (!new_s) {
1050                         spin_lock(&c->ec_stripes_heap_lock);
1051                         bch2_stripes_heap_del(c, m, idx);
1052                         spin_unlock(&c->ec_stripes_heap_lock);
1053
1054                         memset(m, 0, sizeof(*m));
1055                 } else {
1056                         m->alive        = true;
1057                         m->sectors      = le16_to_cpu(new_s->sectors);
1058                         m->algorithm    = new_s->algorithm;
1059                         m->nr_blocks    = new_s->nr_blocks;
1060                         m->nr_redundant = new_s->nr_redundant;
1061                         m->blocks_nonempty = 0;
1062
1063                         for (i = 0; i < new_s->nr_blocks; i++)
1064                                 m->blocks_nonempty += !!stripe_blockcount_get(new_s, i);
1065
1066                         spin_lock(&c->ec_stripes_heap_lock);
1067                         bch2_stripes_heap_update(c, m, idx);
1068                         spin_unlock(&c->ec_stripes_heap_lock);
1069                 }
1070         } else {
1071                 struct gc_stripe *m =
1072                         genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
1073
1074                 if (!m) {
1075                         bch_err(c, "error allocating memory for gc_stripes, idx %llu",
1076                                 idx);
1077                         return -ENOMEM;
1078                 }
1079                 /*
1080                  * This will be wrong when we bring back runtime gc: we should
1081                  * be unmarking the old key and then marking the new key
1082                  */
1083                 m->alive        = true;
1084                 m->sectors      = le16_to_cpu(new_s->sectors);
1085                 m->nr_blocks    = new_s->nr_blocks;
1086                 m->nr_redundant = new_s->nr_redundant;
1087
1088                 for (i = 0; i < new_s->nr_blocks; i++)
1089                         m->ptrs[i] = new_s->ptrs[i];
1090
1091                 bch2_bkey_to_replicas(&m->r.e, new);
1092
1093                 /*
1094                  * gc recalculates this field from stripe ptr
1095                  * references:
1096                  */
1097                 memset(m->block_sectors, 0, sizeof(m->block_sectors));
1098
1099                 for (i = 0; i < new_s->nr_blocks; i++) {
1100                         ret = mark_stripe_bucket(trans, new, i, flags);
1101                         if (ret)
1102                                 return ret;
1103                 }
1104
1105                 ret = update_replicas(c, new, &m->r.e,
1106                                       ((s64) m->sectors * m->nr_redundant),
1107                                       journal_seq, gc);
1108                 if (ret) {
1109                         struct printbuf buf = PRINTBUF;
1110
1111                         bch2_bkey_val_to_text(&buf, c, new);
1112                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
1113                         printbuf_exit(&buf);
1114                         return ret;
1115                 }
1116         }
1117
1118         return 0;
1119 }
1120
1121 int bch2_mark_inode(struct btree_trans *trans,
1122                     struct bkey_s_c old, struct bkey_s_c new,
1123                     unsigned flags)
1124 {
1125         struct bch_fs *c = trans->c;
1126         struct bch_fs_usage __percpu *fs_usage;
1127         u64 journal_seq = trans->journal_res.seq;
1128
1129         if (flags & BTREE_TRIGGER_INSERT) {
1130                 struct bch_inode_v3 *v = (struct bch_inode_v3 *) new.v;
1131
1132                 BUG_ON(!journal_seq);
1133                 BUG_ON(new.k->type != KEY_TYPE_inode_v3);
1134
1135                 v->bi_journal_seq = cpu_to_le64(journal_seq);
1136         }
1137
1138         if (flags & BTREE_TRIGGER_GC) {
1139                 percpu_down_read(&c->mark_lock);
1140                 preempt_disable();
1141
1142                 fs_usage = fs_usage_ptr(c, journal_seq, flags & BTREE_TRIGGER_GC);
1143                 fs_usage->nr_inodes += bkey_is_inode(new.k);
1144                 fs_usage->nr_inodes -= bkey_is_inode(old.k);
1145
1146                 preempt_enable();
1147                 percpu_up_read(&c->mark_lock);
1148         }
1149         return 0;
1150 }
1151
1152 int bch2_mark_reservation(struct btree_trans *trans,
1153                           struct bkey_s_c old, struct bkey_s_c new,
1154                           unsigned flags)
1155 {
1156         struct bch_fs *c = trans->c;
1157         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old : new;
1158         struct bch_fs_usage __percpu *fs_usage;
1159         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1160         s64 sectors = (s64) k.k->size;
1161
1162         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1163
1164         if (flags & BTREE_TRIGGER_OVERWRITE)
1165                 sectors = -sectors;
1166         sectors *= replicas;
1167
1168         percpu_down_read(&c->mark_lock);
1169         preempt_disable();
1170
1171         fs_usage = fs_usage_ptr(c, trans->journal_res.seq, flags & BTREE_TRIGGER_GC);
1172         replicas = clamp_t(unsigned, replicas, 1,
1173                            ARRAY_SIZE(fs_usage->persistent_reserved));
1174
1175         fs_usage->reserved                              += sectors;
1176         fs_usage->persistent_reserved[replicas - 1]     += sectors;
1177
1178         preempt_enable();
1179         percpu_up_read(&c->mark_lock);
1180
1181         return 0;
1182 }
1183
1184 static s64 __bch2_mark_reflink_p(struct btree_trans *trans,
1185                                  struct bkey_s_c_reflink_p p,
1186                                  u64 start, u64 end,
1187                                  u64 *idx, unsigned flags, size_t r_idx)
1188 {
1189         struct bch_fs *c = trans->c;
1190         struct reflink_gc *r;
1191         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1192         u64 next_idx = end;
1193         s64 ret = 0;
1194         struct printbuf buf = PRINTBUF;
1195
1196         if (r_idx >= c->reflink_gc_nr)
1197                 goto not_found;
1198
1199         r = genradix_ptr(&c->reflink_gc_table, r_idx);
1200         next_idx = min(next_idx, r->offset - r->size);
1201         if (*idx < next_idx)
1202                 goto not_found;
1203
1204         BUG_ON((s64) r->refcount + add < 0);
1205
1206         r->refcount += add;
1207         *idx = r->offset;
1208         return 0;
1209 not_found:
1210         if (fsck_err(c, "pointer to missing indirect extent\n"
1211                      "  %s\n"
1212                      "  missing range %llu-%llu",
1213                      (bch2_bkey_val_to_text(&buf, c, p.s_c), buf.buf),
1214                      *idx, next_idx)) {
1215                 struct bkey_i_error *new;
1216
1217                 new = bch2_trans_kmalloc(trans, sizeof(*new));
1218                 ret = PTR_ERR_OR_ZERO(new);
1219                 if (ret)
1220                         goto err;
1221
1222                 bkey_init(&new->k);
1223                 new->k.type     = KEY_TYPE_error;
1224                 new->k.p                = bkey_start_pos(p.k);
1225                 new->k.p.offset += *idx - start;
1226                 bch2_key_resize(&new->k, next_idx - *idx);
1227                 ret = __bch2_btree_insert(trans, BTREE_ID_extents, &new->k_i,
1228                                           BTREE_TRIGGER_NORUN);
1229         }
1230
1231         *idx = next_idx;
1232 err:
1233 fsck_err:
1234         printbuf_exit(&buf);
1235         return ret;
1236 }
1237
1238 int bch2_mark_reflink_p(struct btree_trans *trans,
1239                         struct bkey_s_c old, struct bkey_s_c new,
1240                         unsigned flags)
1241 {
1242         struct bch_fs *c = trans->c;
1243         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old : new;
1244         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1245         struct reflink_gc *ref;
1246         size_t l, r, m;
1247         u64 idx = le64_to_cpu(p.v->idx), start = idx;
1248         u64 end = le64_to_cpu(p.v->idx) + p.k->size;
1249         int ret = 0;
1250
1251         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1252
1253         if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix) {
1254                 idx -= le32_to_cpu(p.v->front_pad);
1255                 end += le32_to_cpu(p.v->back_pad);
1256         }
1257
1258         l = 0;
1259         r = c->reflink_gc_nr;
1260         while (l < r) {
1261                 m = l + (r - l) / 2;
1262
1263                 ref = genradix_ptr(&c->reflink_gc_table, m);
1264                 if (ref->offset <= idx)
1265                         l = m + 1;
1266                 else
1267                         r = m;
1268         }
1269
1270         while (idx < end && !ret)
1271                 ret = __bch2_mark_reflink_p(trans, p, start, end,
1272                                             &idx, flags, l++);
1273
1274         return ret;
1275 }
1276
1277 void bch2_trans_fs_usage_revert(struct btree_trans *trans,
1278                                 struct replicas_delta_list *deltas)
1279 {
1280         struct bch_fs *c = trans->c;
1281         struct bch_fs_usage *dst;
1282         struct replicas_delta *d, *top = (void *) deltas->d + deltas->used;
1283         s64 added = 0;
1284         unsigned i;
1285
1286         percpu_down_read(&c->mark_lock);
1287         preempt_disable();
1288         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1289
1290         /* revert changes: */
1291         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1292                 switch (d->r.data_type) {
1293                 case BCH_DATA_btree:
1294                 case BCH_DATA_user:
1295                 case BCH_DATA_parity:
1296                         added += d->delta;
1297                 }
1298                 BUG_ON(__update_replicas(c, dst, &d->r, -d->delta));
1299         }
1300
1301         dst->nr_inodes -= deltas->nr_inodes;
1302
1303         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1304                 added                           -= deltas->persistent_reserved[i];
1305                 dst->reserved                   -= deltas->persistent_reserved[i];
1306                 dst->persistent_reserved[i]     -= deltas->persistent_reserved[i];
1307         }
1308
1309         if (added > 0) {
1310                 trans->disk_res->sectors += added;
1311                 this_cpu_add(*c->online_reserved, added);
1312         }
1313
1314         preempt_enable();
1315         percpu_up_read(&c->mark_lock);
1316 }
1317
1318 int bch2_trans_fs_usage_apply(struct btree_trans *trans,
1319                               struct replicas_delta_list *deltas)
1320 {
1321         struct bch_fs *c = trans->c;
1322         static int warned_disk_usage = 0;
1323         bool warn = false;
1324         unsigned disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1325         struct replicas_delta *d = deltas->d, *d2;
1326         struct replicas_delta *top = (void *) deltas->d + deltas->used;
1327         struct bch_fs_usage *dst;
1328         s64 added = 0, should_not_have_added;
1329         unsigned i;
1330
1331         percpu_down_read(&c->mark_lock);
1332         preempt_disable();
1333         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1334
1335         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1336                 switch (d->r.data_type) {
1337                 case BCH_DATA_btree:
1338                 case BCH_DATA_user:
1339                 case BCH_DATA_parity:
1340                         added += d->delta;
1341                 }
1342
1343                 if (__update_replicas(c, dst, &d->r, d->delta))
1344                         goto need_mark;
1345         }
1346
1347         dst->nr_inodes += deltas->nr_inodes;
1348
1349         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1350                 added                           += deltas->persistent_reserved[i];
1351                 dst->reserved                   += deltas->persistent_reserved[i];
1352                 dst->persistent_reserved[i]     += deltas->persistent_reserved[i];
1353         }
1354
1355         /*
1356          * Not allowed to reduce sectors_available except by getting a
1357          * reservation:
1358          */
1359         should_not_have_added = added - (s64) disk_res_sectors;
1360         if (unlikely(should_not_have_added > 0)) {
1361                 u64 old, new, v = atomic64_read(&c->sectors_available);
1362
1363                 do {
1364                         old = v;
1365                         new = max_t(s64, 0, old - should_not_have_added);
1366                 } while ((v = atomic64_cmpxchg(&c->sectors_available,
1367                                                old, new)) != old);
1368
1369                 added -= should_not_have_added;
1370                 warn = true;
1371         }
1372
1373         if (added > 0) {
1374                 trans->disk_res->sectors -= added;
1375                 this_cpu_sub(*c->online_reserved, added);
1376         }
1377
1378         preempt_enable();
1379         percpu_up_read(&c->mark_lock);
1380
1381         if (unlikely(warn) && !xchg(&warned_disk_usage, 1))
1382                 bch2_trans_inconsistent(trans,
1383                                         "disk usage increased %lli more than %u sectors reserved)",
1384                                         should_not_have_added, disk_res_sectors);
1385         return 0;
1386 need_mark:
1387         /* revert changes: */
1388         for (d2 = deltas->d; d2 != d; d2 = replicas_delta_next(d2))
1389                 BUG_ON(__update_replicas(c, dst, &d2->r, -d2->delta));
1390
1391         preempt_enable();
1392         percpu_up_read(&c->mark_lock);
1393         return -1;
1394 }
1395
1396 /* trans_mark: */
1397
1398 static inline int bch2_trans_mark_pointer(struct btree_trans *trans,
1399                                    enum btree_id btree_id, unsigned level,
1400                                    struct bkey_s_c k, struct extent_ptr_decoded p,
1401                                    unsigned flags)
1402 {
1403         bool insert = !(flags & BTREE_TRIGGER_OVERWRITE);
1404         struct btree_iter iter;
1405         struct bkey_i_alloc_v4 *a;
1406         struct bpos bucket_pos;
1407         struct bch_backpointer bp;
1408         s64 sectors;
1409         int ret;
1410
1411         bch2_extent_ptr_to_bp(trans->c, btree_id, level, k, p, &bucket_pos, &bp);
1412         sectors = bp.bucket_len;
1413         if (!insert)
1414                 sectors = -sectors;
1415
1416         a = bch2_trans_start_alloc_update(trans, &iter, bucket_pos);
1417         if (IS_ERR(a))
1418                 return PTR_ERR(a);
1419
1420         ret = __mark_pointer(trans, k, &p.ptr, sectors, bp.data_type,
1421                              a->v.gen, &a->v.data_type,
1422                              &a->v.dirty_sectors, &a->v.cached_sectors);
1423         if (ret)
1424                 goto err;
1425
1426         if (!p.ptr.cached) {
1427                 ret = bch2_bucket_backpointer_mod(trans, a, bp, k, insert);
1428                 if (ret)
1429                         goto err;
1430         }
1431
1432         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1433 err:
1434         bch2_trans_iter_exit(trans, &iter);
1435         return ret;
1436 }
1437
1438 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1439                         struct extent_ptr_decoded p,
1440                         s64 sectors, enum bch_data_type data_type)
1441 {
1442         struct btree_iter iter;
1443         struct bkey_i_stripe *s;
1444         struct bch_replicas_padded r;
1445         int ret = 0;
1446
1447         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes, POS(0, p.ec.idx),
1448                              BTREE_ITER_INTENT|
1449                              BTREE_ITER_WITH_UPDATES);
1450         s = bch2_bkey_get_mut_typed(trans, &iter, stripe);
1451         ret = PTR_ERR_OR_ZERO(s);
1452         if (unlikely(ret)) {
1453                 bch2_trans_inconsistent_on(ret == -ENOENT, trans,
1454                         "pointer to nonexistent stripe %llu",
1455                         (u64) p.ec.idx);
1456                 goto err;
1457         }
1458
1459         if (!bch2_ptr_matches_stripe(&s->v, p)) {
1460                 bch2_trans_inconsistent(trans,
1461                         "stripe pointer doesn't match stripe %llu",
1462                         (u64) p.ec.idx);
1463                 ret = -EIO;
1464                 goto err;
1465         }
1466
1467         stripe_blockcount_set(&s->v, p.ec.block,
1468                 stripe_blockcount_get(&s->v, p.ec.block) +
1469                 sectors);
1470
1471         ret = bch2_trans_update(trans, &iter, &s->k_i, 0);
1472         if (ret)
1473                 goto err;
1474
1475         bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1476         r.e.data_type = data_type;
1477         update_replicas_list(trans, &r.e, sectors);
1478 err:
1479         bch2_trans_iter_exit(trans, &iter);
1480         return ret;
1481 }
1482
1483 int bch2_trans_mark_extent(struct btree_trans *trans,
1484                            enum btree_id btree_id, unsigned level,
1485                            struct bkey_s_c old, struct bkey_i *new,
1486                            unsigned flags)
1487 {
1488         struct bch_fs *c = trans->c;
1489         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1490                 ? old
1491                 : bkey_i_to_s_c(new);
1492         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1493         const union bch_extent_entry *entry;
1494         struct extent_ptr_decoded p;
1495         struct bch_replicas_padded r;
1496         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
1497                 ? BCH_DATA_btree
1498                 : BCH_DATA_user;
1499         s64 sectors = bkey_is_btree_ptr(k.k)
1500                 ? btree_sectors(c)
1501                 : k.k->size;
1502         s64 dirty_sectors = 0;
1503         bool stale;
1504         int ret;
1505
1506         r.e.data_type   = data_type;
1507         r.e.nr_devs     = 0;
1508         r.e.nr_required = 1;
1509
1510         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1511                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
1512
1513                 if (flags & BTREE_TRIGGER_OVERWRITE)
1514                         disk_sectors = -disk_sectors;
1515
1516                 ret = bch2_trans_mark_pointer(trans, btree_id, level, k, p, flags);
1517                 if (ret < 0)
1518                         return ret;
1519
1520                 stale = ret > 0;
1521
1522                 if (p.ptr.cached) {
1523                         if (!stale)
1524                                 update_cached_sectors_list(trans, p.ptr.dev,
1525                                                            disk_sectors);
1526                 } else if (!p.has_ec) {
1527                         dirty_sectors          += disk_sectors;
1528                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1529                 } else {
1530                         ret = bch2_trans_mark_stripe_ptr(trans, p,
1531                                         disk_sectors, data_type);
1532                         if (ret)
1533                                 return ret;
1534
1535                         r.e.nr_required = 0;
1536                 }
1537         }
1538
1539         if (r.e.nr_devs)
1540                 update_replicas_list(trans, &r.e, dirty_sectors);
1541
1542         return 0;
1543 }
1544
1545 static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
1546                                          struct bkey_s_c_stripe s,
1547                                          unsigned idx, bool deleting)
1548 {
1549         struct bch_fs *c = trans->c;
1550         const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
1551         struct btree_iter iter;
1552         struct bkey_i_alloc_v4 *a;
1553         enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
1554                 ? BCH_DATA_parity : 0;
1555         s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
1556         int ret = 0;
1557
1558         if (deleting)
1559                 sectors = -sectors;
1560
1561         a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(c, ptr));
1562         if (IS_ERR(a))
1563                 return PTR_ERR(a);
1564
1565         ret = check_bucket_ref(trans, s.s_c, ptr, sectors, data_type,
1566                                a->v.gen, a->v.data_type,
1567                                a->v.dirty_sectors, a->v.cached_sectors);
1568         if (ret)
1569                 goto err;
1570
1571         if (!deleting) {
1572                 if (bch2_trans_inconsistent_on(a->v.stripe ||
1573                                                a->v.stripe_redundancy, trans,
1574                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
1575                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1576                                 bch2_data_types[a->v.data_type],
1577                                 a->v.dirty_sectors,
1578                                 a->v.stripe, s.k->p.offset)) {
1579                         ret = -EIO;
1580                         goto err;
1581                 }
1582
1583                 if (bch2_trans_inconsistent_on(data_type && a->v.dirty_sectors, trans,
1584                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
1585                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1586                                 bch2_data_types[a->v.data_type],
1587                                 a->v.dirty_sectors,
1588                                 s.k->p.offset)) {
1589                         ret = -EIO;
1590                         goto err;
1591                 }
1592
1593                 a->v.stripe             = s.k->p.offset;
1594                 a->v.stripe_redundancy  = s.v->nr_redundant;
1595         } else {
1596                 if (bch2_trans_inconsistent_on(a->v.stripe != s.k->p.offset ||
1597                                                a->v.stripe_redundancy != s.v->nr_redundant, trans,
1598                                 "bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
1599                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1600                                 s.k->p.offset, a->v.stripe)) {
1601                         ret = -EIO;
1602                         goto err;
1603                 }
1604
1605                 a->v.stripe             = 0;
1606                 a->v.stripe_redundancy  = 0;
1607         }
1608
1609         a->v.dirty_sectors += sectors;
1610         if (data_type)
1611                 a->v.data_type = !deleting ? data_type : 0;
1612
1613         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1614         if (ret)
1615                 goto err;
1616 err:
1617         bch2_trans_iter_exit(trans, &iter);
1618         return ret;
1619 }
1620
1621 int bch2_trans_mark_stripe(struct btree_trans *trans,
1622                            enum btree_id btree_id, unsigned level,
1623                            struct bkey_s_c old, struct bkey_i *new,
1624                            unsigned flags)
1625 {
1626         const struct bch_stripe *old_s = NULL;
1627         struct bch_stripe *new_s = NULL;
1628         struct bch_replicas_padded r;
1629         unsigned i, nr_blocks;
1630         int ret = 0;
1631
1632         if (old.k->type == KEY_TYPE_stripe)
1633                 old_s = bkey_s_c_to_stripe(old).v;
1634         if (new->k.type == KEY_TYPE_stripe)
1635                 new_s = &bkey_i_to_stripe(new)->v;
1636
1637         /*
1638          * If the pointers aren't changing, we don't need to do anything:
1639          */
1640         if (new_s && old_s &&
1641             new_s->nr_blocks    == old_s->nr_blocks &&
1642             new_s->nr_redundant == old_s->nr_redundant &&
1643             !memcmp(old_s->ptrs, new_s->ptrs,
1644                     new_s->nr_blocks * sizeof(struct bch_extent_ptr)))
1645                 return 0;
1646
1647         BUG_ON(new_s && old_s &&
1648                (new_s->nr_blocks        != old_s->nr_blocks ||
1649                 new_s->nr_redundant     != old_s->nr_redundant));
1650
1651         nr_blocks = new_s ? new_s->nr_blocks : old_s->nr_blocks;
1652
1653         if (new_s) {
1654                 s64 sectors = le16_to_cpu(new_s->sectors);
1655
1656                 bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(new));
1657                 update_replicas_list(trans, &r.e, sectors * new_s->nr_redundant);
1658         }
1659
1660         if (old_s) {
1661                 s64 sectors = -((s64) le16_to_cpu(old_s->sectors));
1662
1663                 bch2_bkey_to_replicas(&r.e, old);
1664                 update_replicas_list(trans, &r.e, sectors * old_s->nr_redundant);
1665         }
1666
1667         for (i = 0; i < nr_blocks; i++) {
1668                 if (new_s && old_s &&
1669                     !memcmp(&new_s->ptrs[i],
1670                             &old_s->ptrs[i],
1671                             sizeof(new_s->ptrs[i])))
1672                         continue;
1673
1674                 if (new_s) {
1675                         ret = bch2_trans_mark_stripe_bucket(trans,
1676                                         bkey_i_to_s_c_stripe(new), i, false);
1677                         if (ret)
1678                                 break;
1679                 }
1680
1681                 if (old_s) {
1682                         ret = bch2_trans_mark_stripe_bucket(trans,
1683                                         bkey_s_c_to_stripe(old), i, true);
1684                         if (ret)
1685                                 break;
1686                 }
1687         }
1688
1689         return ret;
1690 }
1691
1692 int bch2_trans_mark_inode(struct btree_trans *trans,
1693                           enum btree_id btree_id, unsigned level,
1694                           struct bkey_s_c old,
1695                           struct bkey_i *new,
1696                           unsigned flags)
1697 {
1698         int nr = bkey_is_inode(&new->k) - bkey_is_inode(old.k);
1699
1700         if (nr) {
1701                 struct replicas_delta_list *d =
1702                         replicas_deltas_realloc(trans, 0);
1703                 d->nr_inodes += nr;
1704         }
1705
1706         return 0;
1707 }
1708
1709 int bch2_trans_mark_reservation(struct btree_trans *trans,
1710                                 enum btree_id btree_id, unsigned level,
1711                                 struct bkey_s_c old,
1712                                 struct bkey_i *new,
1713                                 unsigned flags)
1714 {
1715         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1716                 ? old
1717                 : bkey_i_to_s_c(new);
1718         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1719         s64 sectors = (s64) k.k->size;
1720         struct replicas_delta_list *d;
1721
1722         if (flags & BTREE_TRIGGER_OVERWRITE)
1723                 sectors = -sectors;
1724         sectors *= replicas;
1725
1726         d = replicas_deltas_realloc(trans, 0);
1727
1728         replicas = clamp_t(unsigned, replicas, 1,
1729                            ARRAY_SIZE(d->persistent_reserved));
1730
1731         d->persistent_reserved[replicas - 1] += sectors;
1732         return 0;
1733 }
1734
1735 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1736                         struct bkey_s_c_reflink_p p,
1737                         u64 *idx, unsigned flags)
1738 {
1739         struct bch_fs *c = trans->c;
1740         struct btree_iter iter;
1741         struct bkey_i *k;
1742         __le64 *refcount;
1743         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1744         struct printbuf buf = PRINTBUF;
1745         int ret;
1746
1747         bch2_trans_iter_init(trans, &iter, BTREE_ID_reflink, POS(0, *idx),
1748                              BTREE_ITER_INTENT|
1749                              BTREE_ITER_WITH_UPDATES);
1750         k = bch2_bkey_get_mut(trans, &iter);
1751         ret = PTR_ERR_OR_ZERO(k);
1752         if (ret)
1753                 goto err;
1754
1755         refcount = bkey_refcount(k);
1756         if (!refcount) {
1757                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1758                 bch2_trans_inconsistent(trans,
1759                         "nonexistent indirect extent at %llu while marking\n  %s",
1760                         *idx, buf.buf);
1761                 ret = -EIO;
1762                 goto err;
1763         }
1764
1765         if (!*refcount && (flags & BTREE_TRIGGER_OVERWRITE)) {
1766                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1767                 bch2_trans_inconsistent(trans,
1768                         "indirect extent refcount underflow at %llu while marking\n  %s",
1769                         *idx, buf.buf);
1770                 ret = -EIO;
1771                 goto err;
1772         }
1773
1774         if (flags & BTREE_TRIGGER_INSERT) {
1775                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1776                 u64 pad;
1777
1778                 pad = max_t(s64, le32_to_cpu(v->front_pad),
1779                             le64_to_cpu(v->idx) - bkey_start_offset(&k->k));
1780                 BUG_ON(pad > U32_MAX);
1781                 v->front_pad = cpu_to_le32(pad);
1782
1783                 pad = max_t(s64, le32_to_cpu(v->back_pad),
1784                             k->k.p.offset - p.k->size - le64_to_cpu(v->idx));
1785                 BUG_ON(pad > U32_MAX);
1786                 v->back_pad = cpu_to_le32(pad);
1787         }
1788
1789         le64_add_cpu(refcount, add);
1790
1791         bch2_btree_iter_set_pos_to_extent_start(&iter);
1792         ret = bch2_trans_update(trans, &iter, k, 0);
1793         if (ret)
1794                 goto err;
1795
1796         *idx = k->k.p.offset;
1797 err:
1798         bch2_trans_iter_exit(trans, &iter);
1799         printbuf_exit(&buf);
1800         return ret;
1801 }
1802
1803 int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1804                               enum btree_id btree_id, unsigned level,
1805                               struct bkey_s_c old,
1806                               struct bkey_i *new,
1807                               unsigned flags)
1808 {
1809         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1810                 ? old
1811                 : bkey_i_to_s_c(new);
1812         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1813         u64 idx, end_idx;
1814         int ret = 0;
1815
1816         if (flags & BTREE_TRIGGER_INSERT) {
1817                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1818
1819                 v->front_pad = v->back_pad = 0;
1820         }
1821
1822         idx     = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
1823         end_idx = le64_to_cpu(p.v->idx) + p.k->size +
1824                 le32_to_cpu(p.v->back_pad);
1825
1826         while (idx < end_idx && !ret)
1827                 ret = __bch2_trans_mark_reflink_p(trans, p, &idx, flags);
1828
1829         return ret;
1830 }
1831
1832 static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1833                                     struct bch_dev *ca, size_t b,
1834                                     enum bch_data_type type,
1835                                     unsigned sectors)
1836 {
1837         struct bch_fs *c = trans->c;
1838         struct btree_iter iter;
1839         struct bkey_i_alloc_v4 *a;
1840         int ret = 0;
1841
1842         /*
1843          * Backup superblock might be past the end of our normal usable space:
1844          */
1845         if (b >= ca->mi.nbuckets)
1846                 return 0;
1847
1848         a = bch2_trans_start_alloc_update(trans, &iter, POS(ca->dev_idx, b));
1849         if (IS_ERR(a))
1850                 return PTR_ERR(a);
1851
1852         if (a->v.data_type && a->v.data_type != type) {
1853                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1854                         "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
1855                         "while marking %s",
1856                         iter.pos.inode, iter.pos.offset, a->v.gen,
1857                         bch2_data_types[a->v.data_type],
1858                         bch2_data_types[type],
1859                         bch2_data_types[type]);
1860                 ret = -EIO;
1861                 goto out;
1862         }
1863
1864         a->v.data_type          = type;
1865         a->v.dirty_sectors      = sectors;
1866
1867         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1868         if (ret)
1869                 goto out;
1870 out:
1871         bch2_trans_iter_exit(trans, &iter);
1872         return ret;
1873 }
1874
1875 int bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1876                                     struct bch_dev *ca, size_t b,
1877                                     enum bch_data_type type,
1878                                     unsigned sectors)
1879 {
1880         return commit_do(trans, NULL, NULL, 0,
1881                         __bch2_trans_mark_metadata_bucket(trans, ca, b, type, sectors));
1882 }
1883
1884 static int bch2_trans_mark_metadata_sectors(struct btree_trans *trans,
1885                                             struct bch_dev *ca,
1886                                             u64 start, u64 end,
1887                                             enum bch_data_type type,
1888                                             u64 *bucket, unsigned *bucket_sectors)
1889 {
1890         do {
1891                 u64 b = sector_to_bucket(ca, start);
1892                 unsigned sectors =
1893                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1894
1895                 if (b != *bucket && *bucket_sectors) {
1896                         int ret = bch2_trans_mark_metadata_bucket(trans, ca, *bucket,
1897                                                                   type, *bucket_sectors);
1898                         if (ret)
1899                                 return ret;
1900
1901                         *bucket_sectors = 0;
1902                 }
1903
1904                 *bucket         = b;
1905                 *bucket_sectors += sectors;
1906                 start += sectors;
1907         } while (start < end);
1908
1909         return 0;
1910 }
1911
1912 static int __bch2_trans_mark_dev_sb(struct btree_trans *trans,
1913                                     struct bch_dev *ca)
1914 {
1915         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1916         u64 bucket = 0;
1917         unsigned i, bucket_sectors = 0;
1918         int ret;
1919
1920         for (i = 0; i < layout->nr_superblocks; i++) {
1921                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1922
1923                 if (offset == BCH_SB_SECTOR) {
1924                         ret = bch2_trans_mark_metadata_sectors(trans, ca,
1925                                                 0, BCH_SB_SECTOR,
1926                                                 BCH_DATA_sb, &bucket, &bucket_sectors);
1927                         if (ret)
1928                                 return ret;
1929                 }
1930
1931                 ret = bch2_trans_mark_metadata_sectors(trans, ca, offset,
1932                                       offset + (1 << layout->sb_max_size_bits),
1933                                       BCH_DATA_sb, &bucket, &bucket_sectors);
1934                 if (ret)
1935                         return ret;
1936         }
1937
1938         if (bucket_sectors) {
1939                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1940                                 bucket, BCH_DATA_sb, bucket_sectors);
1941                 if (ret)
1942                         return ret;
1943         }
1944
1945         for (i = 0; i < ca->journal.nr; i++) {
1946                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1947                                 ca->journal.buckets[i],
1948                                 BCH_DATA_journal, ca->mi.bucket_size);
1949                 if (ret)
1950                         return ret;
1951         }
1952
1953         return 0;
1954 }
1955
1956 int bch2_trans_mark_dev_sb(struct bch_fs *c, struct bch_dev *ca)
1957 {
1958         return bch2_trans_run(c, __bch2_trans_mark_dev_sb(&trans, ca));
1959 }
1960
1961 /* Disk reservations: */
1962
1963 #define SECTORS_CACHE   1024
1964
1965 int __bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
1966                               u64 sectors, int flags)
1967 {
1968         struct bch_fs_pcpu *pcpu;
1969         u64 old, v, get;
1970         s64 sectors_available;
1971         int ret;
1972
1973         percpu_down_read(&c->mark_lock);
1974         preempt_disable();
1975         pcpu = this_cpu_ptr(c->pcpu);
1976
1977         if (sectors <= pcpu->sectors_available)
1978                 goto out;
1979
1980         v = atomic64_read(&c->sectors_available);
1981         do {
1982                 old = v;
1983                 get = min((u64) sectors + SECTORS_CACHE, old);
1984
1985                 if (get < sectors) {
1986                         preempt_enable();
1987                         goto recalculate;
1988                 }
1989         } while ((v = atomic64_cmpxchg(&c->sectors_available,
1990                                        old, old - get)) != old);
1991
1992         pcpu->sectors_available         += get;
1993
1994 out:
1995         pcpu->sectors_available         -= sectors;
1996         this_cpu_add(*c->online_reserved, sectors);
1997         res->sectors                    += sectors;
1998
1999         preempt_enable();
2000         percpu_up_read(&c->mark_lock);
2001         return 0;
2002
2003 recalculate:
2004         mutex_lock(&c->sectors_available_lock);
2005
2006         percpu_u64_set(&c->pcpu->sectors_available, 0);
2007         sectors_available = avail_factor(__bch2_fs_usage_read_short(c).free);
2008
2009         if (sectors <= sectors_available ||
2010             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
2011                 atomic64_set(&c->sectors_available,
2012                              max_t(s64, 0, sectors_available - sectors));
2013                 this_cpu_add(*c->online_reserved, sectors);
2014                 res->sectors                    += sectors;
2015                 ret = 0;
2016         } else {
2017                 atomic64_set(&c->sectors_available, sectors_available);
2018                 ret = -BCH_ERR_ENOSPC_disk_reservation;
2019         }
2020
2021         mutex_unlock(&c->sectors_available_lock);
2022         percpu_up_read(&c->mark_lock);
2023
2024         return ret;
2025 }
2026
2027 /* Startup/shutdown: */
2028
2029 static void bucket_gens_free_rcu(struct rcu_head *rcu)
2030 {
2031         struct bucket_gens *buckets =
2032                 container_of(rcu, struct bucket_gens, rcu);
2033
2034         kvpfree(buckets, sizeof(*buckets) + buckets->nbuckets);
2035 }
2036
2037 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
2038 {
2039         struct bucket_gens *bucket_gens = NULL, *old_bucket_gens = NULL;
2040         unsigned long *buckets_nouse = NULL;
2041         bool resize = ca->bucket_gens != NULL;
2042         int ret = -ENOMEM;
2043
2044         if (!(bucket_gens       = kvpmalloc(sizeof(struct bucket_gens) + nbuckets,
2045                                             GFP_KERNEL|__GFP_ZERO)) ||
2046             (c->opts.buckets_nouse &&
2047              !(buckets_nouse    = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2048                                             sizeof(unsigned long),
2049                                             GFP_KERNEL|__GFP_ZERO))))
2050                 goto err;
2051
2052         bucket_gens->first_bucket = ca->mi.first_bucket;
2053         bucket_gens->nbuckets   = nbuckets;
2054
2055         bch2_copygc_stop(c);
2056
2057         if (resize) {
2058                 down_write(&c->gc_lock);
2059                 down_write(&ca->bucket_lock);
2060                 percpu_down_write(&c->mark_lock);
2061         }
2062
2063         old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
2064
2065         if (resize) {
2066                 size_t n = min(bucket_gens->nbuckets, old_bucket_gens->nbuckets);
2067
2068                 memcpy(bucket_gens->b,
2069                        old_bucket_gens->b,
2070                        n);
2071                 if (buckets_nouse)
2072                         memcpy(buckets_nouse,
2073                                ca->buckets_nouse,
2074                                BITS_TO_LONGS(n) * sizeof(unsigned long));
2075         }
2076
2077         rcu_assign_pointer(ca->bucket_gens, bucket_gens);
2078         bucket_gens     = old_bucket_gens;
2079
2080         swap(ca->buckets_nouse, buckets_nouse);
2081
2082         nbuckets = ca->mi.nbuckets;
2083
2084         if (resize) {
2085                 percpu_up_write(&c->mark_lock);
2086                 up_write(&ca->bucket_lock);
2087                 up_write(&c->gc_lock);
2088         }
2089
2090         ret = 0;
2091 err:
2092         kvpfree(buckets_nouse,
2093                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2094         if (bucket_gens)
2095                 call_rcu(&bucket_gens->rcu, bucket_gens_free_rcu);
2096
2097         return ret;
2098 }
2099
2100 void bch2_dev_buckets_free(struct bch_dev *ca)
2101 {
2102         unsigned i;
2103
2104         kvpfree(ca->buckets_nouse,
2105                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2106         kvpfree(rcu_dereference_protected(ca->bucket_gens, 1),
2107                 sizeof(struct bucket_gens) + ca->mi.nbuckets);
2108
2109         for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
2110                 free_percpu(ca->usage[i]);
2111         kfree(ca->usage_base);
2112 }
2113
2114 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2115 {
2116         unsigned i;
2117
2118         ca->usage_base = kzalloc(sizeof(struct bch_dev_usage), GFP_KERNEL);
2119         if (!ca->usage_base)
2120                 return -ENOMEM;
2121
2122         for (i = 0; i < ARRAY_SIZE(ca->usage); i++) {
2123                 ca->usage[i] = alloc_percpu(struct bch_dev_usage);
2124                 if (!ca->usage[i])
2125                         return -ENOMEM;
2126         }
2127
2128         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);
2129 }