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 #include "trace.h"
  25
  26 #include <linux/preempt.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                     enum btree_id btree, unsigned level,
 490                     struct bkey_s_c old, struct bkey_s_c new,
 491                     unsigned flags)
 492 {
 493         bool gc = flags & BTREE_TRIGGER_GC;
 494         u64 journal_seq = trans->journal_res.seq;
 495         u64 bucket_journal_seq;
 496         struct bch_fs *c = trans->c;
 497         struct bch_alloc_v4 old_a_convert, new_a_convert;
 498         const struct bch_alloc_v4 *old_a, *new_a;
 499         struct bch_dev *ca;
 500         int ret = 0;
 501
 502         /*
 503          * alloc btree is read in by bch2_alloc_read, not gc:
 504          */
 505         if ((flags & BTREE_TRIGGER_GC) &&
 506             !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
 507                 return 0;
 508
 509         if (bch2_trans_inconsistent_on(!bch2_dev_bucket_exists(c, new.k->p), trans,
 510                                        "alloc key for invalid device or bucket"))
 511                 return -EIO;
 512
 513         ca = bch_dev_bkey_exists(c, new.k->p.inode);
 514
 515         old_a = bch2_alloc_to_v4(old, &old_a_convert);
 516         new_a = bch2_alloc_to_v4(new, &new_a_convert);
 517
 518         bucket_journal_seq = new_a->journal_seq;
 519
 520         if ((flags & BTREE_TRIGGER_INSERT) &&
 521             data_type_is_empty(old_a->data_type) !=
 522             data_type_is_empty(new_a->data_type) &&
 523             new.k->type == KEY_TYPE_alloc_v4) {
 524                 struct bch_alloc_v4 *v = (struct bch_alloc_v4 *) new.v;
 525
 526                 EBUG_ON(!journal_seq);
 527
 528                 /*
 529                  * If the btree updates referring to a bucket weren't flushed
 530                  * before the bucket became empty again, then the we don't have
 531                  * to wait on a journal flush before we can reuse the bucket:
 532                  */
 533                 v->journal_seq = bucket_journal_seq =
 534                         data_type_is_empty(new_a->data_type) &&
 535                         (journal_seq == v->journal_seq ||
 536                          bch2_journal_noflush_seq(&c->journal, v->journal_seq))
 537                         ? 0 : journal_seq;
 538         }
 539
 540         if (!data_type_is_empty(old_a->data_type) &&
 541             data_type_is_empty(new_a->data_type) &&
 542             bucket_journal_seq) {
 543                 ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 544                                 c->journal.flushed_seq_ondisk,
 545                                 new.k->p.inode, new.k->p.offset,
 546                                 bucket_journal_seq);
 547                 if (ret) {
 548                         bch2_fs_fatal_error(c,
 549                                 "error setting bucket_needs_journal_commit: %i", ret);
 550                         return ret;
 551                 }
 552         }
 553
 554         percpu_down_read(&c->mark_lock);
 555         if (!gc && new_a->gen != old_a->gen)
 556                 *bucket_gen(ca, new.k->p.offset) = new_a->gen;
 557
 558         bch2_dev_usage_update(c, ca, *old_a, *new_a, journal_seq, gc);
 559
 560         if (gc) {
 561                 struct bucket *g = gc_bucket(ca, new.k->p.offset);
 562
 563                 bucket_lock(g);
 564
 565                 g->gen_valid            = 1;
 566                 g->gen                  = new_a->gen;
 567                 g->data_type            = new_a->data_type;
 568                 g->stripe               = new_a->stripe;
 569                 g->stripe_redundancy    = new_a->stripe_redundancy;
 570                 g->dirty_sectors        = new_a->dirty_sectors;
 571                 g->cached_sectors       = new_a->cached_sectors;
 572
 573                 bucket_unlock(g);
 574         }
 575         percpu_up_read(&c->mark_lock);
 576
 577         /*
 578          * need to know if we're getting called from the invalidate path or
 579          * not:
 580          */
 581
 582         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 583             old_a->cached_sectors) {
 584                 ret = update_cached_sectors(c, new, ca->dev_idx,
 585                                             -((s64) old_a->cached_sectors),
 586                                             journal_seq, gc);
 587                 if (ret) {
 588                         bch2_fs_fatal_error(c, "%s(): no replicas entry while updating cached sectors",
 589                                             __func__);
 590                         return ret;
 591                 }
 592         }
 593
 594         if (new_a->data_type == BCH_DATA_free &&
 595             (!new_a->journal_seq || new_a->journal_seq < c->journal.flushed_seq_ondisk))
 596                 closure_wake_up(&c->freelist_wait);
 597
 598         if (new_a->data_type == BCH_DATA_need_discard &&
 599             (!bucket_journal_seq || bucket_journal_seq < c->journal.flushed_seq_ondisk))
 600                 bch2_do_discards(c);
 601
 602         if (old_a->data_type != BCH_DATA_cached &&
 603             new_a->data_type == BCH_DATA_cached &&
 604             should_invalidate_buckets(ca, bch2_dev_usage_read(ca)))
 605                 bch2_do_invalidates(c);
 606
 607         if (new_a->data_type == BCH_DATA_need_gc_gens)
 608                 bch2_do_gc_gens(c);
 609
 610         return 0;
 611 }
 612
 613 int bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 614                               size_t b, enum bch_data_type data_type,
 615                               unsigned sectors, struct gc_pos pos,
 616                               unsigned flags)
 617 {
 618         struct bucket old, new, *g;
 619         int ret = 0;
 620
 621         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 622         BUG_ON(data_type != BCH_DATA_sb &&
 623                data_type != BCH_DATA_journal);
 624
 625         /*
 626          * Backup superblock might be past the end of our normal usable space:
 627          */
 628         if (b >= ca->mi.nbuckets)
 629                 return 0;
 630
 631         percpu_down_read(&c->mark_lock);
 632         g = gc_bucket(ca, b);
 633
 634         bucket_lock(g);
 635         old = *g;
 636
 637         if (bch2_fs_inconsistent_on(g->data_type &&
 638                         g->data_type != data_type, c,
 639                         "different types of data in same bucket: %s, %s",
 640                         bch2_data_types[g->data_type],
 641                         bch2_data_types[data_type])) {
 642                 ret = -EIO;
 643                 goto err;
 644         }
 645
 646         if (bch2_fs_inconsistent_on((u64) g->dirty_sectors + sectors > ca->mi.bucket_size, c,
 647                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > bucket size",
 648                         ca->dev_idx, b, g->gen,
 649                         bch2_data_types[g->data_type ?: data_type],
 650                         g->dirty_sectors, sectors)) {
 651                 ret = -EIO;
 652                 goto err;
 653         }
 654
 655
 656         g->data_type = data_type;
 657         g->dirty_sectors += sectors;
 658         new = *g;
 659 err:
 660         bucket_unlock(g);
 661         if (!ret)
 662                 bch2_dev_usage_update_m(c, ca, old, new, 0, true);
 663         percpu_up_read(&c->mark_lock);
 664         return ret;
 665 }
 666
 667 static int check_bucket_ref(struct btree_trans *trans,
 668                             struct bkey_s_c k,
 669                             const struct bch_extent_ptr *ptr,
 670                             s64 sectors, enum bch_data_type ptr_data_type,
 671                             u8 b_gen, u8 bucket_data_type,
 672                             u32 dirty_sectors, u32 cached_sectors)
 673 {
 674         struct bch_fs *c = trans->c;
 675         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 676         size_t bucket_nr = PTR_BUCKET_NR(ca, ptr);
 677         u16 bucket_sectors = !ptr->cached
 678                 ? dirty_sectors
 679                 : cached_sectors;
 680         struct printbuf buf = PRINTBUF;
 681         int ret = 0;
 682
 683         if (bucket_data_type == BCH_DATA_cached)
 684                 bucket_data_type = BCH_DATA_user;
 685
 686         if ((bucket_data_type == BCH_DATA_stripe && ptr_data_type == BCH_DATA_user) ||
 687             (bucket_data_type == BCH_DATA_user   && ptr_data_type == BCH_DATA_stripe))
 688                 bucket_data_type = ptr_data_type = BCH_DATA_stripe;
 689
 690         if (gen_after(ptr->gen, b_gen)) {
 691                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 692                         "bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\n"
 693                         "while marking %s",
 694                         ptr->dev, bucket_nr, b_gen,
 695                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 696                         ptr->gen,
 697                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 698                 ret = -EIO;
 699                 goto err;
 700         }
 701
 702         if (gen_cmp(b_gen, ptr->gen) > BUCKET_GC_GEN_MAX) {
 703                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 704                         "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 705                         "while marking %s",
 706                         ptr->dev, bucket_nr, b_gen,
 707                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 708                         ptr->gen,
 709                         (printbuf_reset(&buf),
 710                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 711                 ret = -EIO;
 712                 goto err;
 713         }
 714
 715         if (b_gen != ptr->gen && !ptr->cached) {
 716                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 717                         "bucket %u:%zu gen %u (mem gen %u) data type %s: stale dirty ptr (gen %u)\n"
 718                         "while marking %s",
 719                         ptr->dev, bucket_nr, b_gen,
 720                         *bucket_gen(ca, bucket_nr),
 721                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 722                         ptr->gen,
 723                         (printbuf_reset(&buf),
 724                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 725                 ret = -EIO;
 726                 goto err;
 727         }
 728
 729         if (b_gen != ptr->gen) {
 730                 ret = 1;
 731                 goto out;
 732         }
 733
 734         if (!data_type_is_empty(bucket_data_type) &&
 735             ptr_data_type &&
 736             bucket_data_type != ptr_data_type) {
 737                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 738                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 739                         "while marking %s",
 740                         ptr->dev, bucket_nr, b_gen,
 741                         bch2_data_types[bucket_data_type],
 742                         bch2_data_types[ptr_data_type],
 743                         (printbuf_reset(&buf),
 744                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 745                 ret = -EIO;
 746                 goto err;
 747         }
 748
 749         if ((unsigned) (bucket_sectors + sectors) > U32_MAX) {
 750                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 751                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U16_MAX\n"
 752                         "while marking %s",
 753                         ptr->dev, bucket_nr, b_gen,
 754                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 755                         bucket_sectors, sectors,
 756                         (printbuf_reset(&buf),
 757                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 758                 ret = -EIO;
 759                 goto err;
 760         }
 761 out:
 762         printbuf_exit(&buf);
 763         return ret;
 764 err:
 765         bch2_dump_trans_updates(trans);
 766         goto out;
 767 }
 768
 769 static int mark_stripe_bucket(struct btree_trans *trans,
 770                               struct bkey_s_c k,
 771                               unsigned ptr_idx,
 772                               unsigned flags)
 773 {
 774         struct bch_fs *c = trans->c;
 775         u64 journal_seq = trans->journal_res.seq;
 776         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 777         unsigned nr_data = s->nr_blocks - s->nr_redundant;
 778         bool parity = ptr_idx >= nr_data;
 779         enum bch_data_type data_type = parity ? BCH_DATA_parity : BCH_DATA_stripe;
 780         s64 sectors = parity ? le16_to_cpu(s->sectors) : 0;
 781         const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
 782         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 783         struct bucket old, new, *g;
 784         struct printbuf buf = PRINTBUF;
 785         int ret = 0;
 786
 787         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 788
 789         /* * XXX doesn't handle deletion */
 790
 791         percpu_down_read(&c->mark_lock);
 792         buf.atomic++;
 793         g = PTR_GC_BUCKET(ca, ptr);
 794
 795         if (g->dirty_sectors ||
 796             (g->stripe && g->stripe != k.k->p.offset)) {
 797                 bch2_fs_inconsistent(c,
 798                               "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
 799                               ptr->dev, PTR_BUCKET_NR(ca, ptr), g->gen,
 800                               (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 801                 ret = -EINVAL;
 802                 goto err;
 803         }
 804
 805         bucket_lock(g);
 806         old = *g;
 807
 808         ret = check_bucket_ref(trans, k, ptr, sectors, data_type,
 809                                g->gen, g->data_type,
 810                                g->dirty_sectors, g->cached_sectors);
 811         if (ret)
 812                 goto err;
 813
 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                              enum btree_id btree_id, unsigned level,
 854                              struct bkey_s_c k,
 855                              struct extent_ptr_decoded p,
 856                              s64 sectors,
 857                              unsigned flags)
 858 {
 859         u64 journal_seq = trans->journal_res.seq;
 860         struct bch_fs *c = trans->c;
 861         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 862         struct bucket old, new, *g;
 863         enum bch_data_type data_type = bkey_ptr_data_type(btree_id, level, k, p);
 864         u8 bucket_data_type;
 865         int ret = 0;
 866
 867         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 868
 869         percpu_down_read(&c->mark_lock);
 870         g = PTR_GC_BUCKET(ca, &p.ptr);
 871         bucket_lock(g);
 872         old = *g;
 873
 874         bucket_data_type = g->data_type;
 875         ret = __mark_pointer(trans, k, &p.ptr, sectors,
 876                              data_type, g->gen,
 877                              &bucket_data_type,
 878                              &g->dirty_sectors,
 879                              &g->cached_sectors);
 880         if (!ret)
 881                 g->data_type = bucket_data_type;
 882
 883         new = *g;
 884         bucket_unlock(g);
 885         if (!ret)
 886                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 887         percpu_up_read(&c->mark_lock);
 888
 889         return ret;
 890 }
 891
 892 static int bch2_mark_stripe_ptr(struct btree_trans *trans,
 893                                 struct bkey_s_c k,
 894                                 struct bch_extent_stripe_ptr p,
 895                                 enum bch_data_type data_type,
 896                                 s64 sectors,
 897                                 unsigned flags)
 898 {
 899         struct bch_fs *c = trans->c;
 900         struct bch_replicas_padded r;
 901         struct gc_stripe *m;
 902
 903         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 904
 905         m = genradix_ptr_alloc(&c->gc_stripes, p.idx, GFP_KERNEL);
 906         if (!m) {
 907                 bch_err(c, "error allocating memory for gc_stripes, idx %llu",
 908                         (u64) p.idx);
 909                 return -BCH_ERR_ENOMEM_mark_stripe_ptr;
 910         }
 911
 912         mutex_lock(&c->ec_stripes_heap_lock);
 913
 914         if (!m || !m->alive) {
 915                 mutex_unlock(&c->ec_stripes_heap_lock);
 916                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
 917                                     (u64) p.idx);
 918                 bch2_inconsistent_error(c);
 919                 return -EIO;
 920         }
 921
 922         m->block_sectors[p.block] += sectors;
 923
 924         r = m->r;
 925         mutex_unlock(&c->ec_stripes_heap_lock);
 926
 927         r.e.data_type = data_type;
 928         update_replicas(c, k, &r.e, sectors, trans->journal_res.seq, true);
 929
 930         return 0;
 931 }
 932
 933 int bch2_mark_extent(struct btree_trans *trans,
 934                      enum btree_id btree_id, unsigned level,
 935                      struct bkey_s_c old, struct bkey_s_c new,
 936                      unsigned flags)
 937 {
 938         u64 journal_seq = trans->journal_res.seq;
 939         struct bch_fs *c = trans->c;
 940         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old : new;
 941         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 942         const union bch_extent_entry *entry;
 943         struct extent_ptr_decoded p;
 944         struct bch_replicas_padded r;
 945         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
 946                 ? BCH_DATA_btree
 947                 : BCH_DATA_user;
 948         s64 sectors = bkey_is_btree_ptr(k.k)
 949                 ? btree_sectors(c)
 950                 : k.k->size;
 951         s64 dirty_sectors = 0;
 952         bool stale;
 953         int ret;
 954
 955         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 956
 957         r.e.data_type   = data_type;
 958         r.e.nr_devs     = 0;
 959         r.e.nr_required = 1;
 960
 961         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 962                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
 963
 964                 if (flags & BTREE_TRIGGER_OVERWRITE)
 965                         disk_sectors = -disk_sectors;
 966
 967                 ret = bch2_mark_pointer(trans, btree_id, level, k, p, disk_sectors, flags);
 968                 if (ret < 0)
 969                         return ret;
 970
 971                 stale = ret > 0;
 972
 973                 if (p.ptr.cached) {
 974                         if (!stale) {
 975                                 ret = update_cached_sectors(c, k, p.ptr.dev,
 976                                                 disk_sectors, journal_seq, true);
 977                                 if (ret) {
 978                                         bch2_fs_fatal_error(c, "%s(): no replicas entry while updating cached sectors",
 979                                                             __func__);
 980                                         return ret;
 981                                 }
 982                         }
 983                 } else if (!p.has_ec) {
 984                         dirty_sectors          += disk_sectors;
 985                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
 986                 } else {
 987                         ret = bch2_mark_stripe_ptr(trans, k, p.ec, data_type,
 988                                         disk_sectors, flags);
 989                         if (ret)
 990                                 return ret;
 991
 992                         /*
 993                          * There may be other dirty pointers in this extent, but
 994                          * if so they're not required for mounting if we have an
 995                          * erasure coded pointer in this extent:
 996                          */
 997                         r.e.nr_required = 0;
 998                 }
 999         }
1000
1001         if (r.e.nr_devs) {
1002                 ret = update_replicas(c, k, &r.e, dirty_sectors, journal_seq, true);
1003                 if (ret) {
1004                         struct printbuf buf = PRINTBUF;
1005
1006                         bch2_bkey_val_to_text(&buf, c, k);
1007                         bch2_fs_fatal_error(c, "%s(): no replicas entry for %s", __func__, buf.buf);
1008                         printbuf_exit(&buf);
1009                         return ret;
1010                 }
1011         }
1012
1013         return 0;
1014 }
1015
1016 int bch2_mark_stripe(struct btree_trans *trans,
1017                      enum btree_id btree_id, unsigned level,
1018                      struct bkey_s_c old, struct bkey_s_c new,
1019                      unsigned flags)
1020 {
1021         bool gc = flags & BTREE_TRIGGER_GC;
1022         u64 journal_seq = trans->journal_res.seq;
1023         struct bch_fs *c = trans->c;
1024         u64 idx = new.k->p.offset;
1025         const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
1026                 ? bkey_s_c_to_stripe(old).v : NULL;
1027         const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
1028                 ? bkey_s_c_to_stripe(new).v : NULL;
1029         unsigned i;
1030         int ret;
1031
1032         BUG_ON(gc && old_s);
1033
1034         if (!gc) {
1035                 struct stripe *m = genradix_ptr(&c->stripes, idx);
1036
1037                 if (!m) {
1038                         struct printbuf buf1 = PRINTBUF;
1039                         struct printbuf buf2 = PRINTBUF;
1040
1041                         bch2_bkey_val_to_text(&buf1, c, old);
1042                         bch2_bkey_val_to_text(&buf2, c, new);
1043                         bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
1044                                             "old %s\n"
1045                                             "new %s", idx, buf1.buf, buf2.buf);
1046                         printbuf_exit(&buf2);
1047                         printbuf_exit(&buf1);
1048                         bch2_inconsistent_error(c);
1049                         return -1;
1050                 }
1051
1052                 if (!new_s) {
1053                         bch2_stripes_heap_del(c, m, idx);
1054
1055                         memset(m, 0, sizeof(*m));
1056                 } else {
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                         if (!old_s)
1067                                 bch2_stripes_heap_insert(c, m, idx);
1068                         else
1069                                 bch2_stripes_heap_update(c, m, idx);
1070                 }
1071         } else {
1072                 struct gc_stripe *m =
1073                         genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
1074
1075                 if (!m) {
1076                         bch_err(c, "error allocating memory for gc_stripes, idx %llu",
1077                                 idx);
1078                         return -BCH_ERR_ENOMEM_mark_stripe;
1079                 }
1080                 /*
1081                  * This will be wrong when we bring back runtime gc: we should
1082                  * be unmarking the old key and then marking the new key
1083                  */
1084                 m->alive        = true;
1085                 m->sectors      = le16_to_cpu(new_s->sectors);
1086                 m->nr_blocks    = new_s->nr_blocks;
1087                 m->nr_redundant = new_s->nr_redundant;
1088
1089                 for (i = 0; i < new_s->nr_blocks; i++)
1090                         m->ptrs[i] = new_s->ptrs[i];
1091
1092                 bch2_bkey_to_replicas(&m->r.e, new);
1093
1094                 /*
1095                  * gc recalculates this field from stripe ptr
1096                  * references:
1097                  */
1098                 memset(m->block_sectors, 0, sizeof(m->block_sectors));
1099
1100                 for (i = 0; i < new_s->nr_blocks; i++) {
1101                         ret = mark_stripe_bucket(trans, new, i, flags);
1102                         if (ret)
1103                                 return ret;
1104                 }
1105
1106                 ret = update_replicas(c, new, &m->r.e,
1107                                       ((s64) m->sectors * m->nr_redundant),
1108                                       journal_seq, gc);
1109                 if (ret) {
1110                         struct printbuf buf = PRINTBUF;
1111
1112                         bch2_bkey_val_to_text(&buf, c, new);
1113                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
1114                         printbuf_exit(&buf);
1115                         return ret;
1116                 }
1117         }
1118
1119         return 0;
1120 }
1121
1122 int bch2_mark_inode(struct btree_trans *trans,
1123                     enum btree_id btree_id, unsigned level,
1124                     struct bkey_s_c old, struct bkey_s_c new,
1125                     unsigned flags)
1126 {
1127         struct bch_fs *c = trans->c;
1128         struct bch_fs_usage __percpu *fs_usage;
1129         u64 journal_seq = trans->journal_res.seq;
1130
1131         if (flags & BTREE_TRIGGER_INSERT) {
1132                 struct bch_inode_v3 *v = (struct bch_inode_v3 *) new.v;
1133
1134                 BUG_ON(!journal_seq);
1135                 BUG_ON(new.k->type != KEY_TYPE_inode_v3);
1136
1137                 v->bi_journal_seq = cpu_to_le64(journal_seq);
1138         }
1139
1140         if (flags & BTREE_TRIGGER_GC) {
1141                 percpu_down_read(&c->mark_lock);
1142                 preempt_disable();
1143
1144                 fs_usage = fs_usage_ptr(c, journal_seq, flags & BTREE_TRIGGER_GC);
1145                 fs_usage->nr_inodes += bkey_is_inode(new.k);
1146                 fs_usage->nr_inodes -= bkey_is_inode(old.k);
1147
1148                 preempt_enable();
1149                 percpu_up_read(&c->mark_lock);
1150         }
1151         return 0;
1152 }
1153
1154 int bch2_mark_reservation(struct btree_trans *trans,
1155                           enum btree_id btree_id, unsigned level,
1156                           struct bkey_s_c old, struct bkey_s_c new,
1157                           unsigned flags)
1158 {
1159         struct bch_fs *c = trans->c;
1160         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old : new;
1161         struct bch_fs_usage __percpu *fs_usage;
1162         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1163         s64 sectors = (s64) k.k->size;
1164
1165         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1166
1167         if (flags & BTREE_TRIGGER_OVERWRITE)
1168                 sectors = -sectors;
1169         sectors *= replicas;
1170
1171         percpu_down_read(&c->mark_lock);
1172         preempt_disable();
1173
1174         fs_usage = fs_usage_ptr(c, trans->journal_res.seq, flags & BTREE_TRIGGER_GC);
1175         replicas = clamp_t(unsigned, replicas, 1,
1176                            ARRAY_SIZE(fs_usage->persistent_reserved));
1177
1178         fs_usage->reserved                              += sectors;
1179         fs_usage->persistent_reserved[replicas - 1]     += sectors;
1180
1181         preempt_enable();
1182         percpu_up_read(&c->mark_lock);
1183
1184         return 0;
1185 }
1186
1187 static s64 __bch2_mark_reflink_p(struct btree_trans *trans,
1188                                  struct bkey_s_c_reflink_p p,
1189                                  u64 start, u64 end,
1190                                  u64 *idx, unsigned flags, size_t r_idx)
1191 {
1192         struct bch_fs *c = trans->c;
1193         struct reflink_gc *r;
1194         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1195         u64 next_idx = end;
1196         s64 ret = 0;
1197         struct printbuf buf = PRINTBUF;
1198
1199         if (r_idx >= c->reflink_gc_nr)
1200                 goto not_found;
1201
1202         r = genradix_ptr(&c->reflink_gc_table, r_idx);
1203         next_idx = min(next_idx, r->offset - r->size);
1204         if (*idx < next_idx)
1205                 goto not_found;
1206
1207         BUG_ON((s64) r->refcount + add < 0);
1208
1209         r->refcount += add;
1210         *idx = r->offset;
1211         return 0;
1212 not_found:
1213         if (fsck_err(c, "pointer to missing indirect extent\n"
1214                      "  %s\n"
1215                      "  missing range %llu-%llu",
1216                      (bch2_bkey_val_to_text(&buf, c, p.s_c), buf.buf),
1217                      *idx, next_idx)) {
1218                 struct bkey_i_error *new;
1219
1220                 new = bch2_trans_kmalloc(trans, sizeof(*new));
1221                 ret = PTR_ERR_OR_ZERO(new);
1222                 if (ret)
1223                         goto err;
1224
1225                 bkey_init(&new->k);
1226                 new->k.type     = KEY_TYPE_error;
1227                 new->k.p                = bkey_start_pos(p.k);
1228                 new->k.p.offset += *idx - start;
1229                 bch2_key_resize(&new->k, next_idx - *idx);
1230                 ret = __bch2_btree_insert(trans, BTREE_ID_extents, &new->k_i,
1231                                           BTREE_TRIGGER_NORUN);
1232         }
1233
1234         *idx = next_idx;
1235 err:
1236 fsck_err:
1237         printbuf_exit(&buf);
1238         return ret;
1239 }
1240
1241 int bch2_mark_reflink_p(struct btree_trans *trans,
1242                         enum btree_id btree_id, unsigned level,
1243                         struct bkey_s_c old, struct bkey_s_c new,
1244                         unsigned flags)
1245 {
1246         struct bch_fs *c = trans->c;
1247         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old : new;
1248         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1249         struct reflink_gc *ref;
1250         size_t l, r, m;
1251         u64 idx = le64_to_cpu(p.v->idx), start = idx;
1252         u64 end = le64_to_cpu(p.v->idx) + p.k->size;
1253         int ret = 0;
1254
1255         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1256
1257         if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix) {
1258                 idx -= le32_to_cpu(p.v->front_pad);
1259                 end += le32_to_cpu(p.v->back_pad);
1260         }
1261
1262         l = 0;
1263         r = c->reflink_gc_nr;
1264         while (l < r) {
1265                 m = l + (r - l) / 2;
1266
1267                 ref = genradix_ptr(&c->reflink_gc_table, m);
1268                 if (ref->offset <= idx)
1269                         l = m + 1;
1270                 else
1271                         r = m;
1272         }
1273
1274         while (idx < end && !ret)
1275                 ret = __bch2_mark_reflink_p(trans, p, start, end,
1276                                             &idx, flags, l++);
1277
1278         return ret;
1279 }
1280
1281 void bch2_trans_fs_usage_revert(struct btree_trans *trans,
1282                                 struct replicas_delta_list *deltas)
1283 {
1284         struct bch_fs *c = trans->c;
1285         struct bch_fs_usage *dst;
1286         struct replicas_delta *d, *top = (void *) deltas->d + deltas->used;
1287         s64 added = 0;
1288         unsigned i;
1289
1290         percpu_down_read(&c->mark_lock);
1291         preempt_disable();
1292         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1293
1294         /* revert changes: */
1295         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1296                 switch (d->r.data_type) {
1297                 case BCH_DATA_btree:
1298                 case BCH_DATA_user:
1299                 case BCH_DATA_parity:
1300                         added += d->delta;
1301                 }
1302                 BUG_ON(__update_replicas(c, dst, &d->r, -d->delta));
1303         }
1304
1305         dst->nr_inodes -= deltas->nr_inodes;
1306
1307         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1308                 added                           -= deltas->persistent_reserved[i];
1309                 dst->reserved                   -= deltas->persistent_reserved[i];
1310                 dst->persistent_reserved[i]     -= deltas->persistent_reserved[i];
1311         }
1312
1313         if (added > 0) {
1314                 trans->disk_res->sectors += added;
1315                 this_cpu_add(*c->online_reserved, added);
1316         }
1317
1318         preempt_enable();
1319         percpu_up_read(&c->mark_lock);
1320 }
1321
1322 int bch2_trans_fs_usage_apply(struct btree_trans *trans,
1323                               struct replicas_delta_list *deltas)
1324 {
1325         struct bch_fs *c = trans->c;
1326         static int warned_disk_usage = 0;
1327         bool warn = false;
1328         unsigned disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1329         struct replicas_delta *d = deltas->d, *d2;
1330         struct replicas_delta *top = (void *) deltas->d + deltas->used;
1331         struct bch_fs_usage *dst;
1332         s64 added = 0, should_not_have_added;
1333         unsigned i;
1334
1335         percpu_down_read(&c->mark_lock);
1336         preempt_disable();
1337         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1338
1339         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1340                 switch (d->r.data_type) {
1341                 case BCH_DATA_btree:
1342                 case BCH_DATA_user:
1343                 case BCH_DATA_parity:
1344                         added += d->delta;
1345                 }
1346
1347                 if (__update_replicas(c, dst, &d->r, d->delta))
1348                         goto need_mark;
1349         }
1350
1351         dst->nr_inodes += deltas->nr_inodes;
1352
1353         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1354                 added                           += deltas->persistent_reserved[i];
1355                 dst->reserved                   += deltas->persistent_reserved[i];
1356                 dst->persistent_reserved[i]     += deltas->persistent_reserved[i];
1357         }
1358
1359         /*
1360          * Not allowed to reduce sectors_available except by getting a
1361          * reservation:
1362          */
1363         should_not_have_added = added - (s64) disk_res_sectors;
1364         if (unlikely(should_not_have_added > 0)) {
1365                 u64 old, new, v = atomic64_read(&c->sectors_available);
1366
1367                 do {
1368                         old = v;
1369                         new = max_t(s64, 0, old - should_not_have_added);
1370                 } while ((v = atomic64_cmpxchg(&c->sectors_available,
1371                                                old, new)) != old);
1372
1373                 added -= should_not_have_added;
1374                 warn = true;
1375         }
1376
1377         if (added > 0) {
1378                 trans->disk_res->sectors -= added;
1379                 this_cpu_sub(*c->online_reserved, added);
1380         }
1381
1382         preempt_enable();
1383         percpu_up_read(&c->mark_lock);
1384
1385         if (unlikely(warn) && !xchg(&warned_disk_usage, 1))
1386                 bch2_trans_inconsistent(trans,
1387                                         "disk usage increased %lli more than %u sectors reserved)",
1388                                         should_not_have_added, disk_res_sectors);
1389         return 0;
1390 need_mark:
1391         /* revert changes: */
1392         for (d2 = deltas->d; d2 != d; d2 = replicas_delta_next(d2))
1393                 BUG_ON(__update_replicas(c, dst, &d2->r, -d2->delta));
1394
1395         preempt_enable();
1396         percpu_up_read(&c->mark_lock);
1397         return -1;
1398 }
1399
1400 /* trans_mark: */
1401
1402 static inline int bch2_trans_mark_pointer(struct btree_trans *trans,
1403                                    enum btree_id btree_id, unsigned level,
1404                                    struct bkey_s_c k, struct extent_ptr_decoded p,
1405                                    unsigned flags)
1406 {
1407         bool insert = !(flags & BTREE_TRIGGER_OVERWRITE);
1408         struct btree_iter iter;
1409         struct bkey_i_alloc_v4 *a;
1410         struct bpos bucket;
1411         struct bch_backpointer bp;
1412         s64 sectors;
1413         int ret;
1414
1415         bch2_extent_ptr_to_bp(trans->c, btree_id, level, k, p, &bucket, &bp);
1416         sectors = bp.bucket_len;
1417         if (!insert)
1418                 sectors = -sectors;
1419
1420         a = bch2_trans_start_alloc_update(trans, &iter, bucket);
1421         if (IS_ERR(a))
1422                 return PTR_ERR(a);
1423
1424         ret = __mark_pointer(trans, k, &p.ptr, sectors, bp.data_type,
1425                              a->v.gen, &a->v.data_type,
1426                              &a->v.dirty_sectors, &a->v.cached_sectors);
1427         if (ret)
1428                 goto err;
1429
1430         if (!p.ptr.cached) {
1431                 ret = bch2_bucket_backpointer_mod(trans, bucket, bp, k, insert);
1432                 if (ret)
1433                         goto err;
1434         }
1435
1436         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1437 err:
1438         bch2_trans_iter_exit(trans, &iter);
1439         return ret;
1440 }
1441
1442 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1443                         struct extent_ptr_decoded p,
1444                         s64 sectors, enum bch_data_type data_type)
1445 {
1446         struct btree_iter iter;
1447         struct bkey_i_stripe *s;
1448         struct bch_replicas_padded r;
1449         int ret = 0;
1450
1451         s = bch2_bkey_get_mut_typed(trans, &iter,
1452                         BTREE_ID_stripes, POS(0, p.ec.idx),
1453                         BTREE_ITER_WITH_UPDATES, stripe);
1454         ret = PTR_ERR_OR_ZERO(s);
1455         if (unlikely(ret)) {
1456                 bch2_trans_inconsistent_on(ret == -ENOENT, trans,
1457                         "pointer to nonexistent stripe %llu",
1458                         (u64) p.ec.idx);
1459                 goto err;
1460         }
1461
1462         if (!bch2_ptr_matches_stripe(&s->v, p)) {
1463                 bch2_trans_inconsistent(trans,
1464                         "stripe pointer doesn't match stripe %llu",
1465                         (u64) p.ec.idx);
1466                 ret = -EIO;
1467                 goto err;
1468         }
1469
1470         stripe_blockcount_set(&s->v, p.ec.block,
1471                 stripe_blockcount_get(&s->v, p.ec.block) +
1472                 sectors);
1473
1474         bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1475         r.e.data_type = data_type;
1476         update_replicas_list(trans, &r.e, sectors);
1477 err:
1478         bch2_trans_iter_exit(trans, &iter);
1479         return ret;
1480 }
1481
1482 int bch2_trans_mark_extent(struct btree_trans *trans,
1483                            enum btree_id btree_id, unsigned level,
1484                            struct bkey_s_c old, struct bkey_i *new,
1485                            unsigned flags)
1486 {
1487         struct bch_fs *c = trans->c;
1488         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1489                 ? old
1490                 : bkey_i_to_s_c(new);
1491         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1492         const union bch_extent_entry *entry;
1493         struct extent_ptr_decoded p;
1494         struct bch_replicas_padded r;
1495         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
1496                 ? BCH_DATA_btree
1497                 : BCH_DATA_user;
1498         s64 sectors = bkey_is_btree_ptr(k.k)
1499                 ? btree_sectors(c)
1500                 : k.k->size;
1501         s64 dirty_sectors = 0;
1502         bool stale;
1503         int ret;
1504
1505         r.e.data_type   = data_type;
1506         r.e.nr_devs     = 0;
1507         r.e.nr_required = 1;
1508
1509         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1510                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
1511
1512                 if (flags & BTREE_TRIGGER_OVERWRITE)
1513                         disk_sectors = -disk_sectors;
1514
1515                 ret = bch2_trans_mark_pointer(trans, btree_id, level, k, p, flags);
1516                 if (ret < 0)
1517                         return ret;
1518
1519                 stale = ret > 0;
1520
1521                 if (p.ptr.cached) {
1522                         if (!stale)
1523                                 update_cached_sectors_list(trans, p.ptr.dev,
1524                                                            disk_sectors);
1525                 } else if (!p.has_ec) {
1526                         dirty_sectors          += disk_sectors;
1527                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1528                 } else {
1529                         ret = bch2_trans_mark_stripe_ptr(trans, p,
1530                                         disk_sectors, data_type);
1531                         if (ret)
1532                                 return ret;
1533
1534                         r.e.nr_required = 0;
1535                 }
1536         }
1537
1538         if (r.e.nr_devs)
1539                 update_replicas_list(trans, &r.e, dirty_sectors);
1540
1541         return 0;
1542 }
1543
1544 static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
1545                                          struct bkey_s_c_stripe s,
1546                                          unsigned idx, bool deleting)
1547 {
1548         struct bch_fs *c = trans->c;
1549         const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
1550         struct btree_iter iter;
1551         struct bkey_i_alloc_v4 *a;
1552         enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
1553                 ? BCH_DATA_parity : 0;
1554         s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
1555         int ret = 0;
1556
1557         if (deleting)
1558                 sectors = -sectors;
1559
1560         a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(c, ptr));
1561         if (IS_ERR(a))
1562                 return PTR_ERR(a);
1563
1564         ret = check_bucket_ref(trans, s.s_c, ptr, sectors, data_type,
1565                                a->v.gen, a->v.data_type,
1566                                a->v.dirty_sectors, a->v.cached_sectors);
1567         if (ret)
1568                 goto err;
1569
1570         if (!deleting) {
1571                 if (bch2_trans_inconsistent_on(a->v.stripe ||
1572                                                a->v.stripe_redundancy, trans,
1573                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
1574                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1575                                 bch2_data_types[a->v.data_type],
1576                                 a->v.dirty_sectors,
1577                                 a->v.stripe, s.k->p.offset)) {
1578                         ret = -EIO;
1579                         goto err;
1580                 }
1581
1582                 if (bch2_trans_inconsistent_on(data_type && a->v.dirty_sectors, trans,
1583                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
1584                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1585                                 bch2_data_types[a->v.data_type],
1586                                 a->v.dirty_sectors,
1587                                 s.k->p.offset)) {
1588                         ret = -EIO;
1589                         goto err;
1590                 }
1591
1592                 a->v.stripe             = s.k->p.offset;
1593                 a->v.stripe_redundancy  = s.v->nr_redundant;
1594                 a->v.data_type          = BCH_DATA_stripe;
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                 a->v.data_type          = alloc_data_type(a->v, BCH_DATA_user);
1608         }
1609
1610         a->v.dirty_sectors += sectors;
1611         if (data_type)
1612                 a->v.data_type = !deleting ? data_type : 0;
1613
1614         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1615         if (ret)
1616                 goto err;
1617 err:
1618         bch2_trans_iter_exit(trans, &iter);
1619         return ret;
1620 }
1621
1622 int bch2_trans_mark_stripe(struct btree_trans *trans,
1623                            enum btree_id btree_id, unsigned level,
1624                            struct bkey_s_c old, struct bkey_i *new,
1625                            unsigned flags)
1626 {
1627         const struct bch_stripe *old_s = NULL;
1628         struct bch_stripe *new_s = NULL;
1629         struct bch_replicas_padded r;
1630         unsigned i, nr_blocks;
1631         int ret = 0;
1632
1633         if (old.k->type == KEY_TYPE_stripe)
1634                 old_s = bkey_s_c_to_stripe(old).v;
1635         if (new->k.type == KEY_TYPE_stripe)
1636                 new_s = &bkey_i_to_stripe(new)->v;
1637
1638         /*
1639          * If the pointers aren't changing, we don't need to do anything:
1640          */
1641         if (new_s && old_s &&
1642             new_s->nr_blocks    == old_s->nr_blocks &&
1643             new_s->nr_redundant == old_s->nr_redundant &&
1644             !memcmp(old_s->ptrs, new_s->ptrs,
1645                     new_s->nr_blocks * sizeof(struct bch_extent_ptr)))
1646                 return 0;
1647
1648         BUG_ON(new_s && old_s &&
1649                (new_s->nr_blocks        != old_s->nr_blocks ||
1650                 new_s->nr_redundant     != old_s->nr_redundant));
1651
1652         nr_blocks = new_s ? new_s->nr_blocks : old_s->nr_blocks;
1653
1654         if (new_s) {
1655                 s64 sectors = le16_to_cpu(new_s->sectors);
1656
1657                 bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(new));
1658                 update_replicas_list(trans, &r.e, sectors * new_s->nr_redundant);
1659         }
1660
1661         if (old_s) {
1662                 s64 sectors = -((s64) le16_to_cpu(old_s->sectors));
1663
1664                 bch2_bkey_to_replicas(&r.e, old);
1665                 update_replicas_list(trans, &r.e, sectors * old_s->nr_redundant);
1666         }
1667
1668         for (i = 0; i < nr_blocks; i++) {
1669                 if (new_s && old_s &&
1670                     !memcmp(&new_s->ptrs[i],
1671                             &old_s->ptrs[i],
1672                             sizeof(new_s->ptrs[i])))
1673                         continue;
1674
1675                 if (new_s) {
1676                         ret = bch2_trans_mark_stripe_bucket(trans,
1677                                         bkey_i_to_s_c_stripe(new), i, false);
1678                         if (ret)
1679                                 break;
1680                 }
1681
1682                 if (old_s) {
1683                         ret = bch2_trans_mark_stripe_bucket(trans,
1684                                         bkey_s_c_to_stripe(old), i, true);
1685                         if (ret)
1686                                 break;
1687                 }
1688         }
1689
1690         return ret;
1691 }
1692
1693 int bch2_trans_mark_inode(struct btree_trans *trans,
1694                           enum btree_id btree_id, unsigned level,
1695                           struct bkey_s_c old,
1696                           struct bkey_i *new,
1697                           unsigned flags)
1698 {
1699         int nr = bkey_is_inode(&new->k) - bkey_is_inode(old.k);
1700
1701         if (nr) {
1702                 struct replicas_delta_list *d =
1703                         replicas_deltas_realloc(trans, 0);
1704                 d->nr_inodes += nr;
1705         }
1706
1707         return 0;
1708 }
1709
1710 int bch2_trans_mark_reservation(struct btree_trans *trans,
1711                                 enum btree_id btree_id, unsigned level,
1712                                 struct bkey_s_c old,
1713                                 struct bkey_i *new,
1714                                 unsigned flags)
1715 {
1716         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1717                 ? old
1718                 : bkey_i_to_s_c(new);
1719         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1720         s64 sectors = (s64) k.k->size;
1721         struct replicas_delta_list *d;
1722
1723         if (flags & BTREE_TRIGGER_OVERWRITE)
1724                 sectors = -sectors;
1725         sectors *= replicas;
1726
1727         d = replicas_deltas_realloc(trans, 0);
1728
1729         replicas = clamp_t(unsigned, replicas, 1,
1730                            ARRAY_SIZE(d->persistent_reserved));
1731
1732         d->persistent_reserved[replicas - 1] += sectors;
1733         return 0;
1734 }
1735
1736 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1737                         struct bkey_s_c_reflink_p p,
1738                         u64 *idx, unsigned flags)
1739 {
1740         struct bch_fs *c = trans->c;
1741         struct btree_iter iter;
1742         struct bkey_i *k;
1743         __le64 *refcount;
1744         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1745         struct printbuf buf = PRINTBUF;
1746         int ret;
1747
1748         k = bch2_bkey_get_mut_noupdate(trans, &iter,
1749                         BTREE_ID_reflink, POS(0, *idx),
1750                         BTREE_ITER_WITH_UPDATES);
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 && 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;
2043
2044         if (!(bucket_gens       = kvpmalloc(sizeof(struct bucket_gens) + nbuckets,
2045                                             GFP_KERNEL|__GFP_ZERO))) {
2046                 ret = -BCH_ERR_ENOMEM_bucket_gens;
2047                 goto err;
2048         }
2049
2050         if ((c->opts.buckets_nouse &&
2051              !(buckets_nouse    = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2052                                             sizeof(unsigned long),
2053                                             GFP_KERNEL|__GFP_ZERO)))) {
2054                 ret = -BCH_ERR_ENOMEM_buckets_nouse;
2055                 goto err;
2056         }
2057
2058         bucket_gens->first_bucket = ca->mi.first_bucket;
2059         bucket_gens->nbuckets   = nbuckets;
2060
2061         bch2_copygc_stop(c);
2062
2063         if (resize) {
2064                 down_write(&c->gc_lock);
2065                 down_write(&ca->bucket_lock);
2066                 percpu_down_write(&c->mark_lock);
2067         }
2068
2069         old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
2070
2071         if (resize) {
2072                 size_t n = min(bucket_gens->nbuckets, old_bucket_gens->nbuckets);
2073
2074                 memcpy(bucket_gens->b,
2075                        old_bucket_gens->b,
2076                        n);
2077                 if (buckets_nouse)
2078                         memcpy(buckets_nouse,
2079                                ca->buckets_nouse,
2080                                BITS_TO_LONGS(n) * sizeof(unsigned long));
2081         }
2082
2083         rcu_assign_pointer(ca->bucket_gens, bucket_gens);
2084         bucket_gens     = old_bucket_gens;
2085
2086         swap(ca->buckets_nouse, buckets_nouse);
2087
2088         nbuckets = ca->mi.nbuckets;
2089
2090         if (resize) {
2091                 percpu_up_write(&c->mark_lock);
2092                 up_write(&ca->bucket_lock);
2093                 up_write(&c->gc_lock);
2094         }
2095
2096         ret = 0;
2097 err:
2098         kvpfree(buckets_nouse,
2099                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2100         if (bucket_gens)
2101                 call_rcu(&bucket_gens->rcu, bucket_gens_free_rcu);
2102
2103         return ret;
2104 }
2105
2106 void bch2_dev_buckets_free(struct bch_dev *ca)
2107 {
2108         unsigned i;
2109
2110         kvpfree(ca->buckets_nouse,
2111                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2112         kvpfree(rcu_dereference_protected(ca->bucket_gens, 1),
2113                 sizeof(struct bucket_gens) + ca->mi.nbuckets);
2114
2115         for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
2116                 free_percpu(ca->usage[i]);
2117         kfree(ca->usage_base);
2118 }
2119
2120 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2121 {
2122         unsigned i;
2123
2124         ca->usage_base = kzalloc(sizeof(struct bch_dev_usage), GFP_KERNEL);
2125         if (!ca->usage_base)
2126                 return -BCH_ERR_ENOMEM_usage_init;
2127
2128         for (i = 0; i < ARRAY_SIZE(ca->usage); i++) {
2129                 ca->usage[i] = alloc_percpu(struct bch_dev_usage);
2130                 if (!ca->usage[i])
2131                         return -BCH_ERR_ENOMEM_usage_init;
2132         }
2133
2134         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);
2135 }