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 nr_replicas = READ_ONCE(c->replicas.nr);
 141         unsigned seq, i;
 142 retry:
 143         ret = kmalloc(__fs_usage_online_u64s(nr_replicas) * sizeof(u64), GFP_NOFS);
 144         if (unlikely(!ret))
 145                 return NULL;
 146
 147         percpu_down_read(&c->mark_lock);
 148
 149         if (nr_replicas != c->replicas.nr) {
 150                 nr_replicas = c->replicas.nr;
 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,
 161                               __fs_usage_u64s(nr_replicas) * sizeof(u64),
 162                               "embedded variable length struct");
 163                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 164                         acc_u64s_percpu((u64 *) &ret->u, (u64 __percpu *) c->usage[i],
 165                                         __fs_usage_u64s(nr_replicas));
 166         } while (read_seqcount_retry(&c->usage_lock, seq));
 167
 168         return ret;
 169 }
 170
 171 void bch2_fs_usage_acc_to_base(struct bch_fs *c, unsigned idx)
 172 {
 173         struct bch_dev *ca;
 174         unsigned i, u64s = fs_usage_u64s(c);
 175
 176         BUG_ON(idx >= ARRAY_SIZE(c->usage));
 177
 178         preempt_disable();
 179         write_seqcount_begin(&c->usage_lock);
 180
 181         acc_u64s_percpu((u64 *) c->usage_base,
 182                         (u64 __percpu *) c->usage[idx], u64s);
 183         percpu_memset(c->usage[idx], 0, u64s * sizeof(u64));
 184
 185         rcu_read_lock();
 186         for_each_member_device_rcu(ca, c, i, NULL) {
 187                 u64s = dev_usage_u64s();
 188
 189                 acc_u64s_percpu((u64 *) ca->usage_base,
 190                                 (u64 __percpu *) ca->usage[idx], u64s);
 191                 percpu_memset(ca->usage[idx], 0, u64s * sizeof(u64));
 192         }
 193         rcu_read_unlock();
 194
 195         write_seqcount_end(&c->usage_lock);
 196         preempt_enable();
 197 }
 198
 199 void bch2_fs_usage_to_text(struct printbuf *out,
 200                            struct bch_fs *c,
 201                            struct bch_fs_usage_online *fs_usage)
 202 {
 203         unsigned i;
 204
 205         prt_printf(out, "capacity:\t\t\t%llu\n", c->capacity);
 206
 207         prt_printf(out, "hidden:\t\t\t\t%llu\n",
 208                fs_usage->u.hidden);
 209         prt_printf(out, "data:\t\t\t\t%llu\n",
 210                fs_usage->u.data);
 211         prt_printf(out, "cached:\t\t\t\t%llu\n",
 212                fs_usage->u.cached);
 213         prt_printf(out, "reserved:\t\t\t%llu\n",
 214                fs_usage->u.reserved);
 215         prt_printf(out, "nr_inodes:\t\t\t%llu\n",
 216                fs_usage->u.nr_inodes);
 217         prt_printf(out, "online reserved:\t\t%llu\n",
 218                fs_usage->online_reserved);
 219
 220         for (i = 0;
 221              i < ARRAY_SIZE(fs_usage->u.persistent_reserved);
 222              i++) {
 223                 prt_printf(out, "%u replicas:\n", i + 1);
 224                 prt_printf(out, "\treserved:\t\t%llu\n",
 225                        fs_usage->u.persistent_reserved[i]);
 226         }
 227
 228         for (i = 0; i < c->replicas.nr; i++) {
 229                 struct bch_replicas_entry *e =
 230                         cpu_replicas_entry(&c->replicas, i);
 231
 232                 prt_printf(out, "\t");
 233                 bch2_replicas_entry_to_text(out, e);
 234                 prt_printf(out, ":\t%llu\n", fs_usage->u.replicas[i]);
 235         }
 236 }
 237
 238 static u64 reserve_factor(u64 r)
 239 {
 240         return r + (round_up(r, (1 << RESERVE_FACTOR)) >> RESERVE_FACTOR);
 241 }
 242
 243 u64 bch2_fs_sectors_used(struct bch_fs *c, struct bch_fs_usage_online *fs_usage)
 244 {
 245         return min(fs_usage->u.hidden +
 246                    fs_usage->u.btree +
 247                    fs_usage->u.data +
 248                    reserve_factor(fs_usage->u.reserved +
 249                                   fs_usage->online_reserved),
 250                    c->capacity);
 251 }
 252
 253 static struct bch_fs_usage_short
 254 __bch2_fs_usage_read_short(struct bch_fs *c)
 255 {
 256         struct bch_fs_usage_short ret;
 257         u64 data, reserved;
 258
 259         ret.capacity = c->capacity -
 260                 bch2_fs_usage_read_one(c, &c->usage_base->hidden);
 261
 262         data            = bch2_fs_usage_read_one(c, &c->usage_base->data) +
 263                 bch2_fs_usage_read_one(c, &c->usage_base->btree);
 264         reserved        = bch2_fs_usage_read_one(c, &c->usage_base->reserved) +
 265                 percpu_u64_get(c->online_reserved);
 266
 267         ret.used        = min(ret.capacity, data + reserve_factor(reserved));
 268         ret.free        = ret.capacity - ret.used;
 269
 270         ret.nr_inodes   = bch2_fs_usage_read_one(c, &c->usage_base->nr_inodes);
 271
 272         return ret;
 273 }
 274
 275 struct bch_fs_usage_short
 276 bch2_fs_usage_read_short(struct bch_fs *c)
 277 {
 278         struct bch_fs_usage_short ret;
 279
 280         percpu_down_read(&c->mark_lock);
 281         ret = __bch2_fs_usage_read_short(c);
 282         percpu_up_read(&c->mark_lock);
 283
 284         return ret;
 285 }
 286
 287 void bch2_dev_usage_init(struct bch_dev *ca)
 288 {
 289         ca->usage_base->d[BCH_DATA_free].buckets = ca->mi.nbuckets - ca->mi.first_bucket;
 290 }
 291
 292 static inline int bucket_sectors_fragmented(struct bch_dev *ca,
 293                                             struct bch_alloc_v4 a)
 294 {
 295         return a.dirty_sectors
 296                 ? max(0, (int) ca->mi.bucket_size - (int) a.dirty_sectors)
 297                 : 0;
 298 }
 299
 300 static void bch2_dev_usage_update(struct bch_fs *c, struct bch_dev *ca,
 301                                   struct bch_alloc_v4 old,
 302                                   struct bch_alloc_v4 new,
 303                                   u64 journal_seq, bool gc)
 304 {
 305         struct bch_fs_usage *fs_usage;
 306         struct bch_dev_usage *u;
 307
 308         preempt_disable();
 309         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 310
 311         if (data_type_is_hidden(old.data_type))
 312                 fs_usage->hidden -= ca->mi.bucket_size;
 313         if (data_type_is_hidden(new.data_type))
 314                 fs_usage->hidden += ca->mi.bucket_size;
 315
 316         u = dev_usage_ptr(ca, journal_seq, gc);
 317
 318         u->d[old.data_type].buckets--;
 319         u->d[new.data_type].buckets++;
 320
 321         u->buckets_ec -= (int) !!old.stripe;
 322         u->buckets_ec += (int) !!new.stripe;
 323
 324         u->d[old.data_type].sectors -= old.dirty_sectors;
 325         u->d[new.data_type].sectors += new.dirty_sectors;
 326
 327         u->d[BCH_DATA_cached].sectors += new.cached_sectors;
 328         u->d[BCH_DATA_cached].sectors -= old.cached_sectors;
 329
 330         u->d[old.data_type].fragmented -= bucket_sectors_fragmented(ca, old);
 331         u->d[new.data_type].fragmented += bucket_sectors_fragmented(ca, new);
 332
 333         preempt_enable();
 334 }
 335
 336 static void bch2_dev_usage_update_m(struct bch_fs *c, struct bch_dev *ca,
 337                                     struct bucket old, struct bucket new,
 338                                     u64 journal_seq, bool gc)
 339 {
 340         struct bch_alloc_v4 old_a = {
 341                 .gen            = old.gen,
 342                 .data_type      = old.data_type,
 343                 .dirty_sectors  = old.dirty_sectors,
 344                 .cached_sectors = old.cached_sectors,
 345                 .stripe         = old.stripe,
 346         };
 347         struct bch_alloc_v4 new_a = {
 348                 .gen            = new.gen,
 349                 .data_type      = new.data_type,
 350                 .dirty_sectors  = new.dirty_sectors,
 351                 .cached_sectors = new.cached_sectors,
 352                 .stripe         = new.stripe,
 353         };
 354
 355         bch2_dev_usage_update(c, ca, old_a, new_a, journal_seq, gc);
 356 }
 357
 358 static inline int __update_replicas(struct bch_fs *c,
 359                                     struct bch_fs_usage *fs_usage,
 360                                     struct bch_replicas_entry *r,
 361                                     s64 sectors)
 362 {
 363         int idx = bch2_replicas_entry_idx(c, r);
 364
 365         if (idx < 0)
 366                 return -1;
 367
 368         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 369         fs_usage->replicas[idx]         += sectors;
 370         return 0;
 371 }
 372
 373 static inline int update_replicas(struct bch_fs *c, struct bkey_s_c k,
 374                         struct bch_replicas_entry *r, s64 sectors,
 375                         unsigned journal_seq, bool gc)
 376 {
 377         struct bch_fs_usage __percpu *fs_usage;
 378         int idx, ret = 0;
 379         struct printbuf buf = PRINTBUF;
 380
 381         percpu_down_read(&c->mark_lock);
 382         buf.atomic++;
 383
 384         idx = bch2_replicas_entry_idx(c, r);
 385         if (idx < 0 &&
 386             fsck_err(c, "no replicas entry\n"
 387                      "  while marking %s",
 388                      (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 389                 percpu_up_read(&c->mark_lock);
 390                 ret = bch2_mark_replicas(c, r);
 391                 percpu_down_read(&c->mark_lock);
 392
 393                 if (ret)
 394                         goto err;
 395                 idx = bch2_replicas_entry_idx(c, r);
 396         }
 397         if (idx < 0) {
 398                 ret = -1;
 399                 goto err;
 400         }
 401
 402         preempt_disable();
 403         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 404         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 405         fs_usage->replicas[idx]         += sectors;
 406         preempt_enable();
 407 err:
 408 fsck_err:
 409         percpu_up_read(&c->mark_lock);
 410         printbuf_exit(&buf);
 411         return ret;
 412 }
 413
 414 static inline int update_cached_sectors(struct bch_fs *c,
 415                         struct bkey_s_c k,
 416                         unsigned dev, s64 sectors,
 417                         unsigned journal_seq, bool gc)
 418 {
 419         struct bch_replicas_padded r;
 420
 421         bch2_replicas_entry_cached(&r.e, dev);
 422
 423         return update_replicas(c, k, &r.e, sectors, journal_seq, gc);
 424 }
 425
 426 static struct replicas_delta_list *
 427 replicas_deltas_realloc(struct btree_trans *trans, unsigned more)
 428 {
 429         struct replicas_delta_list *d = trans->fs_usage_deltas;
 430         unsigned new_size = d ? (d->size + more) * 2 : 128;
 431         unsigned alloc_size = sizeof(*d) + new_size;
 432
 433         WARN_ON_ONCE(alloc_size > REPLICAS_DELTA_LIST_MAX);
 434
 435         if (!d || d->used + more > d->size) {
 436                 d = krealloc(d, alloc_size, GFP_NOIO|__GFP_ZERO);
 437
 438                 BUG_ON(!d && alloc_size > REPLICAS_DELTA_LIST_MAX);
 439
 440                 if (!d) {
 441                         d = mempool_alloc(&trans->c->replicas_delta_pool, GFP_NOIO);
 442                         memset(d, 0, REPLICAS_DELTA_LIST_MAX);
 443
 444                         if (trans->fs_usage_deltas)
 445                                 memcpy(d, trans->fs_usage_deltas,
 446                                        trans->fs_usage_deltas->size + sizeof(*d));
 447
 448                         new_size = REPLICAS_DELTA_LIST_MAX - sizeof(*d);
 449                         kfree(trans->fs_usage_deltas);
 450                 }
 451
 452                 d->size = new_size;
 453                 trans->fs_usage_deltas = d;
 454         }
 455         return d;
 456 }
 457
 458 static inline void update_replicas_list(struct btree_trans *trans,
 459                                         struct bch_replicas_entry *r,
 460                                         s64 sectors)
 461 {
 462         struct replicas_delta_list *d;
 463         struct replicas_delta *n;
 464         unsigned b;
 465
 466         if (!sectors)
 467                 return;
 468
 469         b = replicas_entry_bytes(r) + 8;
 470         d = replicas_deltas_realloc(trans, b);
 471
 472         n = (void *) d->d + d->used;
 473         n->delta = sectors;
 474         memcpy((void *) n + offsetof(struct replicas_delta, r),
 475                r, replicas_entry_bytes(r));
 476         bch2_replicas_entry_sort(&n->r);
 477         d->used += b;
 478 }
 479
 480 static inline void update_cached_sectors_list(struct btree_trans *trans,
 481                                               unsigned dev, s64 sectors)
 482 {
 483         struct bch_replicas_padded r;
 484
 485         bch2_replicas_entry_cached(&r.e, dev);
 486
 487         update_replicas_list(trans, &r.e, sectors);
 488 }
 489
 490 int bch2_mark_alloc(struct btree_trans *trans,
 491                     enum btree_id btree, unsigned level,
 492                     struct bkey_s_c old, struct bkey_s_c new,
 493                     unsigned flags)
 494 {
 495         bool gc = flags & BTREE_TRIGGER_GC;
 496         u64 journal_seq = trans->journal_res.seq;
 497         u64 bucket_journal_seq;
 498         struct bch_fs *c = trans->c;
 499         struct bch_alloc_v4 old_a_convert, new_a_convert;
 500         const struct bch_alloc_v4 *old_a, *new_a;
 501         struct bch_dev *ca;
 502         int ret = 0;
 503
 504         /*
 505          * alloc btree is read in by bch2_alloc_read, not gc:
 506          */
 507         if ((flags & BTREE_TRIGGER_GC) &&
 508             !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
 509                 return 0;
 510
 511         if (bch2_trans_inconsistent_on(!bch2_dev_bucket_exists(c, new.k->p), trans,
 512                                        "alloc key for invalid device or bucket"))
 513                 return -EIO;
 514
 515         ca = bch_dev_bkey_exists(c, new.k->p.inode);
 516
 517         old_a = bch2_alloc_to_v4(old, &old_a_convert);
 518         new_a = bch2_alloc_to_v4(new, &new_a_convert);
 519
 520         bucket_journal_seq = new_a->journal_seq;
 521
 522         if ((flags & BTREE_TRIGGER_INSERT) &&
 523             data_type_is_empty(old_a->data_type) !=
 524             data_type_is_empty(new_a->data_type) &&
 525             new.k->type == KEY_TYPE_alloc_v4) {
 526                 struct bch_alloc_v4 *v = (struct bch_alloc_v4 *) new.v;
 527
 528                 EBUG_ON(!journal_seq);
 529
 530                 /*
 531                  * If the btree updates referring to a bucket weren't flushed
 532                  * before the bucket became empty again, then the we don't have
 533                  * to wait on a journal flush before we can reuse the bucket:
 534                  */
 535                 v->journal_seq = bucket_journal_seq =
 536                         data_type_is_empty(new_a->data_type) &&
 537                         (journal_seq == v->journal_seq ||
 538                          bch2_journal_noflush_seq(&c->journal, v->journal_seq))
 539                         ? 0 : journal_seq;
 540         }
 541
 542         if (!data_type_is_empty(old_a->data_type) &&
 543             data_type_is_empty(new_a->data_type) &&
 544             bucket_journal_seq) {
 545                 ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 546                                 c->journal.flushed_seq_ondisk,
 547                                 new.k->p.inode, new.k->p.offset,
 548                                 bucket_journal_seq);
 549                 if (ret) {
 550                         bch2_fs_fatal_error(c,
 551                                 "error setting bucket_needs_journal_commit: %i", ret);
 552                         return ret;
 553                 }
 554         }
 555
 556         percpu_down_read(&c->mark_lock);
 557         if (!gc && new_a->gen != old_a->gen)
 558                 *bucket_gen(ca, new.k->p.offset) = new_a->gen;
 559
 560         bch2_dev_usage_update(c, ca, *old_a, *new_a, journal_seq, gc);
 561
 562         if (gc) {
 563                 struct bucket *g = gc_bucket(ca, new.k->p.offset);
 564
 565                 bucket_lock(g);
 566
 567                 g->gen_valid            = 1;
 568                 g->gen                  = new_a->gen;
 569                 g->data_type            = new_a->data_type;
 570                 g->stripe               = new_a->stripe;
 571                 g->stripe_redundancy    = new_a->stripe_redundancy;
 572                 g->dirty_sectors        = new_a->dirty_sectors;
 573                 g->cached_sectors       = new_a->cached_sectors;
 574
 575                 bucket_unlock(g);
 576         }
 577         percpu_up_read(&c->mark_lock);
 578
 579         /*
 580          * need to know if we're getting called from the invalidate path or
 581          * not:
 582          */
 583
 584         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 585             old_a->cached_sectors) {
 586                 ret = update_cached_sectors(c, new, ca->dev_idx,
 587                                             -((s64) old_a->cached_sectors),
 588                                             journal_seq, gc);
 589                 if (ret) {
 590                         bch2_fs_fatal_error(c, "%s(): no replicas entry while updating cached sectors",
 591                                             __func__);
 592                         return ret;
 593                 }
 594         }
 595
 596         if (new_a->data_type == BCH_DATA_free &&
 597             (!new_a->journal_seq || new_a->journal_seq < c->journal.flushed_seq_ondisk))
 598                 closure_wake_up(&c->freelist_wait);
 599
 600         if (new_a->data_type == BCH_DATA_need_discard &&
 601             (!bucket_journal_seq || bucket_journal_seq < c->journal.flushed_seq_ondisk))
 602                 bch2_do_discards(c);
 603
 604         if (old_a->data_type != BCH_DATA_cached &&
 605             new_a->data_type == BCH_DATA_cached &&
 606             should_invalidate_buckets(ca, bch2_dev_usage_read(ca)))
 607                 bch2_do_invalidates(c);
 608
 609         if (new_a->data_type == BCH_DATA_need_gc_gens)
 610                 bch2_do_gc_gens(c);
 611
 612         return 0;
 613 }
 614
 615 int bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 616                               size_t b, enum bch_data_type data_type,
 617                               unsigned sectors, struct gc_pos pos,
 618                               unsigned flags)
 619 {
 620         struct bucket old, new, *g;
 621         int ret = 0;
 622
 623         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 624         BUG_ON(data_type != BCH_DATA_sb &&
 625                data_type != BCH_DATA_journal);
 626
 627         /*
 628          * Backup superblock might be past the end of our normal usable space:
 629          */
 630         if (b >= ca->mi.nbuckets)
 631                 return 0;
 632
 633         percpu_down_read(&c->mark_lock);
 634         g = gc_bucket(ca, b);
 635
 636         bucket_lock(g);
 637         old = *g;
 638
 639         if (bch2_fs_inconsistent_on(g->data_type &&
 640                         g->data_type != data_type, c,
 641                         "different types of data in same bucket: %s, %s",
 642                         bch2_data_types[g->data_type],
 643                         bch2_data_types[data_type])) {
 644                 ret = -EIO;
 645                 goto err;
 646         }
 647
 648         if (bch2_fs_inconsistent_on((u64) g->dirty_sectors + sectors > ca->mi.bucket_size, c,
 649                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > bucket size",
 650                         ca->dev_idx, b, g->gen,
 651                         bch2_data_types[g->data_type ?: data_type],
 652                         g->dirty_sectors, sectors)) {
 653                 ret = -EIO;
 654                 goto err;
 655         }
 656
 657
 658         g->data_type = data_type;
 659         g->dirty_sectors += sectors;
 660         new = *g;
 661 err:
 662         bucket_unlock(g);
 663         if (!ret)
 664                 bch2_dev_usage_update_m(c, ca, old, new, 0, true);
 665         percpu_up_read(&c->mark_lock);
 666         return ret;
 667 }
 668
 669 static int check_bucket_ref(struct btree_trans *trans,
 670                             struct bkey_s_c k,
 671                             const struct bch_extent_ptr *ptr,
 672                             s64 sectors, enum bch_data_type ptr_data_type,
 673                             u8 b_gen, u8 bucket_data_type,
 674                             u32 dirty_sectors, u32 cached_sectors)
 675 {
 676         struct bch_fs *c = trans->c;
 677         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 678         size_t bucket_nr = PTR_BUCKET_NR(ca, ptr);
 679         u16 bucket_sectors = !ptr->cached
 680                 ? dirty_sectors
 681                 : cached_sectors;
 682         struct printbuf buf = PRINTBUF;
 683         int ret = 0;
 684
 685         if (bucket_data_type == BCH_DATA_cached)
 686                 bucket_data_type = BCH_DATA_user;
 687
 688         if ((bucket_data_type == BCH_DATA_stripe && ptr_data_type == BCH_DATA_user) ||
 689             (bucket_data_type == BCH_DATA_user   && ptr_data_type == BCH_DATA_stripe))
 690                 bucket_data_type = ptr_data_type = BCH_DATA_stripe;
 691
 692         if (gen_after(ptr->gen, b_gen)) {
 693                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 694                         "bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\n"
 695                         "while marking %s",
 696                         ptr->dev, bucket_nr, b_gen,
 697                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 698                         ptr->gen,
 699                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 700                 ret = -EIO;
 701                 goto err;
 702         }
 703
 704         if (gen_cmp(b_gen, ptr->gen) > BUCKET_GC_GEN_MAX) {
 705                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 706                         "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 707                         "while marking %s",
 708                         ptr->dev, bucket_nr, b_gen,
 709                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 710                         ptr->gen,
 711                         (printbuf_reset(&buf),
 712                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 713                 ret = -EIO;
 714                 goto err;
 715         }
 716
 717         if (b_gen != ptr->gen && !ptr->cached) {
 718                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 719                         "bucket %u:%zu gen %u (mem gen %u) data type %s: stale dirty ptr (gen %u)\n"
 720                         "while marking %s",
 721                         ptr->dev, bucket_nr, b_gen,
 722                         *bucket_gen(ca, bucket_nr),
 723                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 724                         ptr->gen,
 725                         (printbuf_reset(&buf),
 726                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 727                 ret = -EIO;
 728                 goto err;
 729         }
 730
 731         if (b_gen != ptr->gen) {
 732                 ret = 1;
 733                 goto out;
 734         }
 735
 736         if (!data_type_is_empty(bucket_data_type) &&
 737             ptr_data_type &&
 738             bucket_data_type != ptr_data_type) {
 739                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 740                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 741                         "while marking %s",
 742                         ptr->dev, bucket_nr, b_gen,
 743                         bch2_data_types[bucket_data_type],
 744                         bch2_data_types[ptr_data_type],
 745                         (printbuf_reset(&buf),
 746                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 747                 ret = -EIO;
 748                 goto err;
 749         }
 750
 751         if ((unsigned) (bucket_sectors + sectors) > U32_MAX) {
 752                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 753                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U16_MAX\n"
 754                         "while marking %s",
 755                         ptr->dev, bucket_nr, b_gen,
 756                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 757                         bucket_sectors, sectors,
 758                         (printbuf_reset(&buf),
 759                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 760                 ret = -EIO;
 761                 goto err;
 762         }
 763 out:
 764         printbuf_exit(&buf);
 765         return ret;
 766 err:
 767         bch2_dump_trans_updates(trans);
 768         goto out;
 769 }
 770
 771 static int mark_stripe_bucket(struct btree_trans *trans,
 772                               struct bkey_s_c k,
 773                               unsigned ptr_idx,
 774                               unsigned flags)
 775 {
 776         struct bch_fs *c = trans->c;
 777         u64 journal_seq = trans->journal_res.seq;
 778         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 779         unsigned nr_data = s->nr_blocks - s->nr_redundant;
 780         bool parity = ptr_idx >= nr_data;
 781         enum bch_data_type data_type = parity ? BCH_DATA_parity : BCH_DATA_stripe;
 782         s64 sectors = parity ? le16_to_cpu(s->sectors) : 0;
 783         const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
 784         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 785         struct bucket old, new, *g;
 786         struct printbuf buf = PRINTBUF;
 787         int ret = 0;
 788
 789         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 790
 791         /* * XXX doesn't handle deletion */
 792
 793         percpu_down_read(&c->mark_lock);
 794         buf.atomic++;
 795         g = PTR_GC_BUCKET(ca, ptr);
 796
 797         if (g->dirty_sectors ||
 798             (g->stripe && g->stripe != k.k->p.offset)) {
 799                 bch2_fs_inconsistent(c,
 800                               "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
 801                               ptr->dev, PTR_BUCKET_NR(ca, ptr), g->gen,
 802                               (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 803                 ret = -EINVAL;
 804                 goto err;
 805         }
 806
 807         bucket_lock(g);
 808         old = *g;
 809
 810         ret = check_bucket_ref(trans, k, ptr, sectors, data_type,
 811                                g->gen, g->data_type,
 812                                g->dirty_sectors, g->cached_sectors);
 813         if (ret)
 814                 goto err;
 815
 816         g->data_type = data_type;
 817         g->dirty_sectors += sectors;
 818
 819         g->stripe               = k.k->p.offset;
 820         g->stripe_redundancy    = s->nr_redundant;
 821         new = *g;
 822 err:
 823         bucket_unlock(g);
 824         if (!ret)
 825                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 826         percpu_up_read(&c->mark_lock);
 827         printbuf_exit(&buf);
 828         return ret;
 829 }
 830
 831 static int __mark_pointer(struct btree_trans *trans,
 832                           struct bkey_s_c k,
 833                           const struct bch_extent_ptr *ptr,
 834                           s64 sectors, enum bch_data_type ptr_data_type,
 835                           u8 bucket_gen, u8 *bucket_data_type,
 836                           u32 *dirty_sectors, u32 *cached_sectors)
 837 {
 838         u32 *dst_sectors = !ptr->cached
 839                 ? dirty_sectors
 840                 : cached_sectors;
 841         int ret = check_bucket_ref(trans, k, ptr, sectors, ptr_data_type,
 842                                    bucket_gen, *bucket_data_type,
 843                                    *dirty_sectors, *cached_sectors);
 844
 845         if (ret)
 846                 return ret;
 847
 848         *dst_sectors += sectors;
 849         *bucket_data_type = *dirty_sectors || *cached_sectors
 850                 ? ptr_data_type : 0;
 851         return 0;
 852 }
 853
 854 static int bch2_mark_pointer(struct btree_trans *trans,
 855                              enum btree_id btree_id, unsigned level,
 856                              struct bkey_s_c k,
 857                              struct extent_ptr_decoded p,
 858                              s64 sectors,
 859                              unsigned flags)
 860 {
 861         u64 journal_seq = trans->journal_res.seq;
 862         struct bch_fs *c = trans->c;
 863         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 864         struct bucket old, new, *g;
 865         enum bch_data_type data_type = bkey_ptr_data_type(btree_id, level, k, p);
 866         u8 bucket_data_type;
 867         int ret = 0;
 868
 869         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 870
 871         percpu_down_read(&c->mark_lock);
 872         g = PTR_GC_BUCKET(ca, &p.ptr);
 873         bucket_lock(g);
 874         old = *g;
 875
 876         bucket_data_type = g->data_type;
 877         ret = __mark_pointer(trans, k, &p.ptr, sectors,
 878                              data_type, g->gen,
 879                              &bucket_data_type,
 880                              &g->dirty_sectors,
 881                              &g->cached_sectors);
 882         if (!ret)
 883                 g->data_type = bucket_data_type;
 884
 885         new = *g;
 886         bucket_unlock(g);
 887         if (!ret)
 888                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 889         percpu_up_read(&c->mark_lock);
 890
 891         return ret;
 892 }
 893
 894 static int bch2_mark_stripe_ptr(struct btree_trans *trans,
 895                                 struct bkey_s_c k,
 896                                 struct bch_extent_stripe_ptr p,
 897                                 enum bch_data_type data_type,
 898                                 s64 sectors,
 899                                 unsigned flags)
 900 {
 901         struct bch_fs *c = trans->c;
 902         struct bch_replicas_padded r;
 903         struct gc_stripe *m;
 904
 905         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 906
 907         m = genradix_ptr_alloc(&c->gc_stripes, p.idx, GFP_KERNEL);
 908         if (!m) {
 909                 bch_err(c, "error allocating memory for gc_stripes, idx %llu",
 910                         (u64) p.idx);
 911                 return -BCH_ERR_ENOMEM_mark_stripe_ptr;
 912         }
 913
 914         mutex_lock(&c->ec_stripes_heap_lock);
 915
 916         if (!m || !m->alive) {
 917                 mutex_unlock(&c->ec_stripes_heap_lock);
 918                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
 919                                     (u64) p.idx);
 920                 bch2_inconsistent_error(c);
 921                 return -EIO;
 922         }
 923
 924         m->block_sectors[p.block] += sectors;
 925
 926         r = m->r;
 927         mutex_unlock(&c->ec_stripes_heap_lock);
 928
 929         r.e.data_type = data_type;
 930         update_replicas(c, k, &r.e, sectors, trans->journal_res.seq, true);
 931
 932         return 0;
 933 }
 934
 935 int bch2_mark_extent(struct btree_trans *trans,
 936                      enum btree_id btree_id, unsigned level,
 937                      struct bkey_s_c old, struct bkey_s_c new,
 938                      unsigned flags)
 939 {
 940         u64 journal_seq = trans->journal_res.seq;
 941         struct bch_fs *c = trans->c;
 942         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old : new;
 943         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 944         const union bch_extent_entry *entry;
 945         struct extent_ptr_decoded p;
 946         struct bch_replicas_padded r;
 947         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
 948                 ? BCH_DATA_btree
 949                 : BCH_DATA_user;
 950         s64 sectors = bkey_is_btree_ptr(k.k)
 951                 ? btree_sectors(c)
 952                 : k.k->size;
 953         s64 dirty_sectors = 0;
 954         bool stale;
 955         int ret;
 956
 957         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 958
 959         r.e.data_type   = data_type;
 960         r.e.nr_devs     = 0;
 961         r.e.nr_required = 1;
 962
 963         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 964                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
 965
 966                 if (flags & BTREE_TRIGGER_OVERWRITE)
 967                         disk_sectors = -disk_sectors;
 968
 969                 ret = bch2_mark_pointer(trans, btree_id, level, k, p, disk_sectors, flags);
 970                 if (ret < 0)
 971                         return ret;
 972
 973                 stale = ret > 0;
 974
 975                 if (p.ptr.cached) {
 976                         if (!stale) {
 977                                 ret = update_cached_sectors(c, k, p.ptr.dev,
 978                                                 disk_sectors, journal_seq, true);
 979                                 if (ret) {
 980                                         bch2_fs_fatal_error(c, "%s(): no replicas entry while updating cached sectors",
 981                                                             __func__);
 982                                         return ret;
 983                                 }
 984                         }
 985                 } else if (!p.has_ec) {
 986                         dirty_sectors          += disk_sectors;
 987                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
 988                 } else {
 989                         ret = bch2_mark_stripe_ptr(trans, k, p.ec, data_type,
 990                                         disk_sectors, flags);
 991                         if (ret)
 992                                 return ret;
 993
 994                         /*
 995                          * There may be other dirty pointers in this extent, but
 996                          * if so they're not required for mounting if we have an
 997                          * erasure coded pointer in this extent:
 998                          */
 999                         r.e.nr_required = 0;
1000                 }
1001         }
1002
1003         if (r.e.nr_devs) {
1004                 ret = update_replicas(c, k, &r.e, dirty_sectors, journal_seq, true);
1005                 if (ret) {
1006                         struct printbuf buf = PRINTBUF;
1007
1008                         bch2_bkey_val_to_text(&buf, c, k);
1009                         bch2_fs_fatal_error(c, "%s(): no replicas entry for %s", __func__, buf.buf);
1010                         printbuf_exit(&buf);
1011                         return ret;
1012                 }
1013         }
1014
1015         return 0;
1016 }
1017
1018 int bch2_mark_stripe(struct btree_trans *trans,
1019                      enum btree_id btree_id, unsigned level,
1020                      struct bkey_s_c old, struct bkey_s_c new,
1021                      unsigned flags)
1022 {
1023         bool gc = flags & BTREE_TRIGGER_GC;
1024         u64 journal_seq = trans->journal_res.seq;
1025         struct bch_fs *c = trans->c;
1026         u64 idx = new.k->p.offset;
1027         const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
1028                 ? bkey_s_c_to_stripe(old).v : NULL;
1029         const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
1030                 ? bkey_s_c_to_stripe(new).v : NULL;
1031         unsigned i;
1032         int ret;
1033
1034         BUG_ON(gc && old_s);
1035
1036         if (!gc) {
1037                 struct stripe *m = genradix_ptr(&c->stripes, idx);
1038
1039                 if (!m) {
1040                         struct printbuf buf1 = PRINTBUF;
1041                         struct printbuf buf2 = PRINTBUF;
1042
1043                         bch2_bkey_val_to_text(&buf1, c, old);
1044                         bch2_bkey_val_to_text(&buf2, c, new);
1045                         bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
1046                                             "old %s\n"
1047                                             "new %s", idx, buf1.buf, buf2.buf);
1048                         printbuf_exit(&buf2);
1049                         printbuf_exit(&buf1);
1050                         bch2_inconsistent_error(c);
1051                         return -1;
1052                 }
1053
1054                 if (!new_s) {
1055                         bch2_stripes_heap_del(c, m, idx);
1056
1057                         memset(m, 0, sizeof(*m));
1058                 } else {
1059                         m->sectors      = le16_to_cpu(new_s->sectors);
1060                         m->algorithm    = new_s->algorithm;
1061                         m->nr_blocks    = new_s->nr_blocks;
1062                         m->nr_redundant = new_s->nr_redundant;
1063                         m->blocks_nonempty = 0;
1064
1065                         for (i = 0; i < new_s->nr_blocks; i++)
1066                                 m->blocks_nonempty += !!stripe_blockcount_get(new_s, i);
1067
1068                         if (!old_s)
1069                                 bch2_stripes_heap_insert(c, m, idx);
1070                         else
1071                                 bch2_stripes_heap_update(c, m, idx);
1072                 }
1073         } else {
1074                 struct gc_stripe *m =
1075                         genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
1076
1077                 if (!m) {
1078                         bch_err(c, "error allocating memory for gc_stripes, idx %llu",
1079                                 idx);
1080                         return -BCH_ERR_ENOMEM_mark_stripe;
1081                 }
1082                 /*
1083                  * This will be wrong when we bring back runtime gc: we should
1084                  * be unmarking the old key and then marking the new key
1085                  */
1086                 m->alive        = true;
1087                 m->sectors      = le16_to_cpu(new_s->sectors);
1088                 m->nr_blocks    = new_s->nr_blocks;
1089                 m->nr_redundant = new_s->nr_redundant;
1090
1091                 for (i = 0; i < new_s->nr_blocks; i++)
1092                         m->ptrs[i] = new_s->ptrs[i];
1093
1094                 bch2_bkey_to_replicas(&m->r.e, new);
1095
1096                 /*
1097                  * gc recalculates this field from stripe ptr
1098                  * references:
1099                  */
1100                 memset(m->block_sectors, 0, sizeof(m->block_sectors));
1101
1102                 for (i = 0; i < new_s->nr_blocks; i++) {
1103                         ret = mark_stripe_bucket(trans, new, i, flags);
1104                         if (ret)
1105                                 return ret;
1106                 }
1107
1108                 ret = update_replicas(c, new, &m->r.e,
1109                                       ((s64) m->sectors * m->nr_redundant),
1110                                       journal_seq, gc);
1111                 if (ret) {
1112                         struct printbuf buf = PRINTBUF;
1113
1114                         bch2_bkey_val_to_text(&buf, c, new);
1115                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
1116                         printbuf_exit(&buf);
1117                         return ret;
1118                 }
1119         }
1120
1121         return 0;
1122 }
1123
1124 int bch2_mark_inode(struct btree_trans *trans,
1125                     enum btree_id btree_id, unsigned level,
1126                     struct bkey_s_c old, struct bkey_s_c new,
1127                     unsigned flags)
1128 {
1129         struct bch_fs *c = trans->c;
1130         struct bch_fs_usage __percpu *fs_usage;
1131         u64 journal_seq = trans->journal_res.seq;
1132
1133         if (flags & BTREE_TRIGGER_INSERT) {
1134                 struct bch_inode_v3 *v = (struct bch_inode_v3 *) new.v;
1135
1136                 BUG_ON(!journal_seq);
1137                 BUG_ON(new.k->type != KEY_TYPE_inode_v3);
1138
1139                 v->bi_journal_seq = cpu_to_le64(journal_seq);
1140         }
1141
1142         if (flags & BTREE_TRIGGER_GC) {
1143                 percpu_down_read(&c->mark_lock);
1144                 preempt_disable();
1145
1146                 fs_usage = fs_usage_ptr(c, journal_seq, flags & BTREE_TRIGGER_GC);
1147                 fs_usage->nr_inodes += bkey_is_inode(new.k);
1148                 fs_usage->nr_inodes -= bkey_is_inode(old.k);
1149
1150                 preempt_enable();
1151                 percpu_up_read(&c->mark_lock);
1152         }
1153         return 0;
1154 }
1155
1156 int bch2_mark_reservation(struct btree_trans *trans,
1157                           enum btree_id btree_id, unsigned level,
1158                           struct bkey_s_c old, struct bkey_s_c new,
1159                           unsigned flags)
1160 {
1161         struct bch_fs *c = trans->c;
1162         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old : new;
1163         struct bch_fs_usage __percpu *fs_usage;
1164         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1165         s64 sectors = (s64) k.k->size;
1166
1167         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1168
1169         if (flags & BTREE_TRIGGER_OVERWRITE)
1170                 sectors = -sectors;
1171         sectors *= replicas;
1172
1173         percpu_down_read(&c->mark_lock);
1174         preempt_disable();
1175
1176         fs_usage = fs_usage_ptr(c, trans->journal_res.seq, flags & BTREE_TRIGGER_GC);
1177         replicas = clamp_t(unsigned, replicas, 1,
1178                            ARRAY_SIZE(fs_usage->persistent_reserved));
1179
1180         fs_usage->reserved                              += sectors;
1181         fs_usage->persistent_reserved[replicas - 1]     += sectors;
1182
1183         preempt_enable();
1184         percpu_up_read(&c->mark_lock);
1185
1186         return 0;
1187 }
1188
1189 static s64 __bch2_mark_reflink_p(struct btree_trans *trans,
1190                                  struct bkey_s_c_reflink_p p,
1191                                  u64 start, u64 end,
1192                                  u64 *idx, unsigned flags, size_t r_idx)
1193 {
1194         struct bch_fs *c = trans->c;
1195         struct reflink_gc *r;
1196         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1197         u64 next_idx = end;
1198         s64 ret = 0;
1199         struct printbuf buf = PRINTBUF;
1200
1201         if (r_idx >= c->reflink_gc_nr)
1202                 goto not_found;
1203
1204         r = genradix_ptr(&c->reflink_gc_table, r_idx);
1205         next_idx = min(next_idx, r->offset - r->size);
1206         if (*idx < next_idx)
1207                 goto not_found;
1208
1209         BUG_ON((s64) r->refcount + add < 0);
1210
1211         r->refcount += add;
1212         *idx = r->offset;
1213         return 0;
1214 not_found:
1215         if (fsck_err(c, "pointer to missing indirect extent\n"
1216                      "  %s\n"
1217                      "  missing range %llu-%llu",
1218                      (bch2_bkey_val_to_text(&buf, c, p.s_c), buf.buf),
1219                      *idx, next_idx)) {
1220                 struct bkey_i_error *new;
1221
1222                 new = bch2_trans_kmalloc(trans, sizeof(*new));
1223                 ret = PTR_ERR_OR_ZERO(new);
1224                 if (ret)
1225                         goto err;
1226
1227                 bkey_init(&new->k);
1228                 new->k.type     = KEY_TYPE_error;
1229                 new->k.p                = bkey_start_pos(p.k);
1230                 new->k.p.offset += *idx - start;
1231                 bch2_key_resize(&new->k, next_idx - *idx);
1232                 ret = __bch2_btree_insert(trans, BTREE_ID_extents, &new->k_i,
1233                                           BTREE_TRIGGER_NORUN);
1234         }
1235
1236         *idx = next_idx;
1237 err:
1238 fsck_err:
1239         printbuf_exit(&buf);
1240         return ret;
1241 }
1242
1243 int bch2_mark_reflink_p(struct btree_trans *trans,
1244                         enum btree_id btree_id, unsigned level,
1245                         struct bkey_s_c old, struct bkey_s_c new,
1246                         unsigned flags)
1247 {
1248         struct bch_fs *c = trans->c;
1249         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old : new;
1250         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1251         struct reflink_gc *ref;
1252         size_t l, r, m;
1253         u64 idx = le64_to_cpu(p.v->idx), start = idx;
1254         u64 end = le64_to_cpu(p.v->idx) + p.k->size;
1255         int ret = 0;
1256
1257         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1258
1259         if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix) {
1260                 idx -= le32_to_cpu(p.v->front_pad);
1261                 end += le32_to_cpu(p.v->back_pad);
1262         }
1263
1264         l = 0;
1265         r = c->reflink_gc_nr;
1266         while (l < r) {
1267                 m = l + (r - l) / 2;
1268
1269                 ref = genradix_ptr(&c->reflink_gc_table, m);
1270                 if (ref->offset <= idx)
1271                         l = m + 1;
1272                 else
1273                         r = m;
1274         }
1275
1276         while (idx < end && !ret)
1277                 ret = __bch2_mark_reflink_p(trans, p, start, end,
1278                                             &idx, flags, l++);
1279
1280         return ret;
1281 }
1282
1283 void bch2_trans_fs_usage_revert(struct btree_trans *trans,
1284                                 struct replicas_delta_list *deltas)
1285 {
1286         struct bch_fs *c = trans->c;
1287         struct bch_fs_usage *dst;
1288         struct replicas_delta *d, *top = (void *) deltas->d + deltas->used;
1289         s64 added = 0;
1290         unsigned i;
1291
1292         percpu_down_read(&c->mark_lock);
1293         preempt_disable();
1294         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1295
1296         /* revert changes: */
1297         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1298                 switch (d->r.data_type) {
1299                 case BCH_DATA_btree:
1300                 case BCH_DATA_user:
1301                 case BCH_DATA_parity:
1302                         added += d->delta;
1303                 }
1304                 BUG_ON(__update_replicas(c, dst, &d->r, -d->delta));
1305         }
1306
1307         dst->nr_inodes -= deltas->nr_inodes;
1308
1309         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1310                 added                           -= deltas->persistent_reserved[i];
1311                 dst->reserved                   -= deltas->persistent_reserved[i];
1312                 dst->persistent_reserved[i]     -= deltas->persistent_reserved[i];
1313         }
1314
1315         if (added > 0) {
1316                 trans->disk_res->sectors += added;
1317                 this_cpu_add(*c->online_reserved, added);
1318         }
1319
1320         preempt_enable();
1321         percpu_up_read(&c->mark_lock);
1322 }
1323
1324 int bch2_trans_fs_usage_apply(struct btree_trans *trans,
1325                               struct replicas_delta_list *deltas)
1326 {
1327         struct bch_fs *c = trans->c;
1328         static int warned_disk_usage = 0;
1329         bool warn = false;
1330         unsigned disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1331         struct replicas_delta *d = deltas->d, *d2;
1332         struct replicas_delta *top = (void *) deltas->d + deltas->used;
1333         struct bch_fs_usage *dst;
1334         s64 added = 0, should_not_have_added;
1335         unsigned i;
1336
1337         percpu_down_read(&c->mark_lock);
1338         preempt_disable();
1339         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1340
1341         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1342                 switch (d->r.data_type) {
1343                 case BCH_DATA_btree:
1344                 case BCH_DATA_user:
1345                 case BCH_DATA_parity:
1346                         added += d->delta;
1347                 }
1348
1349                 if (__update_replicas(c, dst, &d->r, d->delta))
1350                         goto need_mark;
1351         }
1352
1353         dst->nr_inodes += deltas->nr_inodes;
1354
1355         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1356                 added                           += deltas->persistent_reserved[i];
1357                 dst->reserved                   += deltas->persistent_reserved[i];
1358                 dst->persistent_reserved[i]     += deltas->persistent_reserved[i];
1359         }
1360
1361         /*
1362          * Not allowed to reduce sectors_available except by getting a
1363          * reservation:
1364          */
1365         should_not_have_added = added - (s64) disk_res_sectors;
1366         if (unlikely(should_not_have_added > 0)) {
1367                 u64 old, new, v = atomic64_read(&c->sectors_available);
1368
1369                 do {
1370                         old = v;
1371                         new = max_t(s64, 0, old - should_not_have_added);
1372                 } while ((v = atomic64_cmpxchg(&c->sectors_available,
1373                                                old, new)) != old);
1374
1375                 added -= should_not_have_added;
1376                 warn = true;
1377         }
1378
1379         if (added > 0) {
1380                 trans->disk_res->sectors -= added;
1381                 this_cpu_sub(*c->online_reserved, added);
1382         }
1383
1384         preempt_enable();
1385         percpu_up_read(&c->mark_lock);
1386
1387         if (unlikely(warn) && !xchg(&warned_disk_usage, 1))
1388                 bch2_trans_inconsistent(trans,
1389                                         "disk usage increased %lli more than %u sectors reserved)",
1390                                         should_not_have_added, disk_res_sectors);
1391         return 0;
1392 need_mark:
1393         /* revert changes: */
1394         for (d2 = deltas->d; d2 != d; d2 = replicas_delta_next(d2))
1395                 BUG_ON(__update_replicas(c, dst, &d2->r, -d2->delta));
1396
1397         preempt_enable();
1398         percpu_up_read(&c->mark_lock);
1399         return -1;
1400 }
1401
1402 /* trans_mark: */
1403
1404 static inline int bch2_trans_mark_pointer(struct btree_trans *trans,
1405                                    enum btree_id btree_id, unsigned level,
1406                                    struct bkey_s_c k, struct extent_ptr_decoded p,
1407                                    unsigned flags)
1408 {
1409         bool insert = !(flags & BTREE_TRIGGER_OVERWRITE);
1410         struct btree_iter iter;
1411         struct bkey_i_alloc_v4 *a;
1412         struct bpos bucket;
1413         struct bch_backpointer bp;
1414         s64 sectors;
1415         int ret;
1416
1417         bch2_extent_ptr_to_bp(trans->c, btree_id, level, k, p, &bucket, &bp);
1418         sectors = bp.bucket_len;
1419         if (!insert)
1420                 sectors = -sectors;
1421
1422         a = bch2_trans_start_alloc_update(trans, &iter, bucket);
1423         if (IS_ERR(a))
1424                 return PTR_ERR(a);
1425
1426         ret = __mark_pointer(trans, k, &p.ptr, sectors, bp.data_type,
1427                              a->v.gen, &a->v.data_type,
1428                              &a->v.dirty_sectors, &a->v.cached_sectors);
1429         if (ret)
1430                 goto err;
1431
1432         if (!p.ptr.cached) {
1433                 ret = bch2_bucket_backpointer_mod(trans, bucket, bp, k, insert);
1434                 if (ret)
1435                         goto err;
1436         }
1437
1438         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1439 err:
1440         bch2_trans_iter_exit(trans, &iter);
1441         return ret;
1442 }
1443
1444 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1445                         struct extent_ptr_decoded p,
1446                         s64 sectors, enum bch_data_type data_type)
1447 {
1448         struct btree_iter iter;
1449         struct bkey_i_stripe *s;
1450         struct bch_replicas_padded r;
1451         int ret = 0;
1452
1453         s = bch2_bkey_get_mut_typed(trans, &iter,
1454                         BTREE_ID_stripes, POS(0, p.ec.idx),
1455                         BTREE_ITER_WITH_UPDATES, stripe);
1456         ret = PTR_ERR_OR_ZERO(s);
1457         if (unlikely(ret)) {
1458                 bch2_trans_inconsistent_on(ret == -ENOENT, trans,
1459                         "pointer to nonexistent stripe %llu",
1460                         (u64) p.ec.idx);
1461                 goto err;
1462         }
1463
1464         if (!bch2_ptr_matches_stripe(&s->v, p)) {
1465                 bch2_trans_inconsistent(trans,
1466                         "stripe pointer doesn't match stripe %llu",
1467                         (u64) p.ec.idx);
1468                 ret = -EIO;
1469                 goto err;
1470         }
1471
1472         stripe_blockcount_set(&s->v, p.ec.block,
1473                 stripe_blockcount_get(&s->v, p.ec.block) +
1474                 sectors);
1475
1476         bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1477         r.e.data_type = data_type;
1478         update_replicas_list(trans, &r.e, sectors);
1479 err:
1480         bch2_trans_iter_exit(trans, &iter);
1481         return ret;
1482 }
1483
1484 int bch2_trans_mark_extent(struct btree_trans *trans,
1485                            enum btree_id btree_id, unsigned level,
1486                            struct bkey_s_c old, struct bkey_i *new,
1487                            unsigned flags)
1488 {
1489         struct bch_fs *c = trans->c;
1490         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1491                 ? old
1492                 : bkey_i_to_s_c(new);
1493         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1494         const union bch_extent_entry *entry;
1495         struct extent_ptr_decoded p;
1496         struct bch_replicas_padded r;
1497         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
1498                 ? BCH_DATA_btree
1499                 : BCH_DATA_user;
1500         s64 sectors = bkey_is_btree_ptr(k.k)
1501                 ? btree_sectors(c)
1502                 : k.k->size;
1503         s64 dirty_sectors = 0;
1504         bool stale;
1505         int ret;
1506
1507         r.e.data_type   = data_type;
1508         r.e.nr_devs     = 0;
1509         r.e.nr_required = 1;
1510
1511         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1512                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
1513
1514                 if (flags & BTREE_TRIGGER_OVERWRITE)
1515                         disk_sectors = -disk_sectors;
1516
1517                 ret = bch2_trans_mark_pointer(trans, btree_id, level, k, p, flags);
1518                 if (ret < 0)
1519                         return ret;
1520
1521                 stale = ret > 0;
1522
1523                 if (p.ptr.cached) {
1524                         if (!stale)
1525                                 update_cached_sectors_list(trans, p.ptr.dev,
1526                                                            disk_sectors);
1527                 } else if (!p.has_ec) {
1528                         dirty_sectors          += disk_sectors;
1529                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1530                 } else {
1531                         ret = bch2_trans_mark_stripe_ptr(trans, p,
1532                                         disk_sectors, data_type);
1533                         if (ret)
1534                                 return ret;
1535
1536                         r.e.nr_required = 0;
1537                 }
1538         }
1539
1540         if (r.e.nr_devs)
1541                 update_replicas_list(trans, &r.e, dirty_sectors);
1542
1543         return 0;
1544 }
1545
1546 static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
1547                                          struct bkey_s_c_stripe s,
1548                                          unsigned idx, bool deleting)
1549 {
1550         struct bch_fs *c = trans->c;
1551         const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
1552         struct btree_iter iter;
1553         struct bkey_i_alloc_v4 *a;
1554         enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
1555                 ? BCH_DATA_parity : 0;
1556         s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
1557         int ret = 0;
1558
1559         if (deleting)
1560                 sectors = -sectors;
1561
1562         a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(c, ptr));
1563         if (IS_ERR(a))
1564                 return PTR_ERR(a);
1565
1566         ret = check_bucket_ref(trans, s.s_c, ptr, sectors, data_type,
1567                                a->v.gen, a->v.data_type,
1568                                a->v.dirty_sectors, a->v.cached_sectors);
1569         if (ret)
1570                 goto err;
1571
1572         if (!deleting) {
1573                 if (bch2_trans_inconsistent_on(a->v.stripe ||
1574                                                a->v.stripe_redundancy, trans,
1575                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
1576                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1577                                 bch2_data_types[a->v.data_type],
1578                                 a->v.dirty_sectors,
1579                                 a->v.stripe, s.k->p.offset)) {
1580                         ret = -EIO;
1581                         goto err;
1582                 }
1583
1584                 if (bch2_trans_inconsistent_on(data_type && a->v.dirty_sectors, trans,
1585                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
1586                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1587                                 bch2_data_types[a->v.data_type],
1588                                 a->v.dirty_sectors,
1589                                 s.k->p.offset)) {
1590                         ret = -EIO;
1591                         goto err;
1592                 }
1593
1594                 a->v.stripe             = s.k->p.offset;
1595                 a->v.stripe_redundancy  = s.v->nr_redundant;
1596                 a->v.data_type          = BCH_DATA_stripe;
1597         } else {
1598                 if (bch2_trans_inconsistent_on(a->v.stripe != s.k->p.offset ||
1599                                                a->v.stripe_redundancy != s.v->nr_redundant, trans,
1600                                 "bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
1601                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1602                                 s.k->p.offset, a->v.stripe)) {
1603                         ret = -EIO;
1604                         goto err;
1605                 }
1606
1607                 a->v.stripe             = 0;
1608                 a->v.stripe_redundancy  = 0;
1609                 a->v.data_type          = alloc_data_type(a->v, BCH_DATA_user);
1610         }
1611
1612         a->v.dirty_sectors += sectors;
1613         if (data_type)
1614                 a->v.data_type = !deleting ? data_type : 0;
1615
1616         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1617         if (ret)
1618                 goto err;
1619 err:
1620         bch2_trans_iter_exit(trans, &iter);
1621         return ret;
1622 }
1623
1624 int bch2_trans_mark_stripe(struct btree_trans *trans,
1625                            enum btree_id btree_id, unsigned level,
1626                            struct bkey_s_c old, struct bkey_i *new,
1627                            unsigned flags)
1628 {
1629         const struct bch_stripe *old_s = NULL;
1630         struct bch_stripe *new_s = NULL;
1631         struct bch_replicas_padded r;
1632         unsigned i, nr_blocks;
1633         int ret = 0;
1634
1635         if (old.k->type == KEY_TYPE_stripe)
1636                 old_s = bkey_s_c_to_stripe(old).v;
1637         if (new->k.type == KEY_TYPE_stripe)
1638                 new_s = &bkey_i_to_stripe(new)->v;
1639
1640         /*
1641          * If the pointers aren't changing, we don't need to do anything:
1642          */
1643         if (new_s && old_s &&
1644             new_s->nr_blocks    == old_s->nr_blocks &&
1645             new_s->nr_redundant == old_s->nr_redundant &&
1646             !memcmp(old_s->ptrs, new_s->ptrs,
1647                     new_s->nr_blocks * sizeof(struct bch_extent_ptr)))
1648                 return 0;
1649
1650         BUG_ON(new_s && old_s &&
1651                (new_s->nr_blocks        != old_s->nr_blocks ||
1652                 new_s->nr_redundant     != old_s->nr_redundant));
1653
1654         nr_blocks = new_s ? new_s->nr_blocks : old_s->nr_blocks;
1655
1656         if (new_s) {
1657                 s64 sectors = le16_to_cpu(new_s->sectors);
1658
1659                 bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(new));
1660                 update_replicas_list(trans, &r.e, sectors * new_s->nr_redundant);
1661         }
1662
1663         if (old_s) {
1664                 s64 sectors = -((s64) le16_to_cpu(old_s->sectors));
1665
1666                 bch2_bkey_to_replicas(&r.e, old);
1667                 update_replicas_list(trans, &r.e, sectors * old_s->nr_redundant);
1668         }
1669
1670         for (i = 0; i < nr_blocks; i++) {
1671                 if (new_s && old_s &&
1672                     !memcmp(&new_s->ptrs[i],
1673                             &old_s->ptrs[i],
1674                             sizeof(new_s->ptrs[i])))
1675                         continue;
1676
1677                 if (new_s) {
1678                         ret = bch2_trans_mark_stripe_bucket(trans,
1679                                         bkey_i_to_s_c_stripe(new), i, false);
1680                         if (ret)
1681                                 break;
1682                 }
1683
1684                 if (old_s) {
1685                         ret = bch2_trans_mark_stripe_bucket(trans,
1686                                         bkey_s_c_to_stripe(old), i, true);
1687                         if (ret)
1688                                 break;
1689                 }
1690         }
1691
1692         return ret;
1693 }
1694
1695 int bch2_trans_mark_inode(struct btree_trans *trans,
1696                           enum btree_id btree_id, unsigned level,
1697                           struct bkey_s_c old,
1698                           struct bkey_i *new,
1699                           unsigned flags)
1700 {
1701         int nr = bkey_is_inode(&new->k) - bkey_is_inode(old.k);
1702
1703         if (nr) {
1704                 struct replicas_delta_list *d =
1705                         replicas_deltas_realloc(trans, 0);
1706                 d->nr_inodes += nr;
1707         }
1708
1709         return 0;
1710 }
1711
1712 int bch2_trans_mark_reservation(struct btree_trans *trans,
1713                                 enum btree_id btree_id, unsigned level,
1714                                 struct bkey_s_c old,
1715                                 struct bkey_i *new,
1716                                 unsigned flags)
1717 {
1718         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1719                 ? old
1720                 : bkey_i_to_s_c(new);
1721         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1722         s64 sectors = (s64) k.k->size;
1723         struct replicas_delta_list *d;
1724
1725         if (flags & BTREE_TRIGGER_OVERWRITE)
1726                 sectors = -sectors;
1727         sectors *= replicas;
1728
1729         d = replicas_deltas_realloc(trans, 0);
1730
1731         replicas = clamp_t(unsigned, replicas, 1,
1732                            ARRAY_SIZE(d->persistent_reserved));
1733
1734         d->persistent_reserved[replicas - 1] += sectors;
1735         return 0;
1736 }
1737
1738 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1739                         struct bkey_s_c_reflink_p p,
1740                         u64 *idx, unsigned flags)
1741 {
1742         struct bch_fs *c = trans->c;
1743         struct btree_iter iter;
1744         struct bkey_i *k;
1745         __le64 *refcount;
1746         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1747         struct printbuf buf = PRINTBUF;
1748         int ret;
1749
1750         k = bch2_bkey_get_mut_noupdate(trans, &iter,
1751                         BTREE_ID_reflink, POS(0, *idx),
1752                         BTREE_ITER_WITH_UPDATES);
1753         ret = PTR_ERR_OR_ZERO(k);
1754         if (ret)
1755                 goto err;
1756
1757         refcount = bkey_refcount(k);
1758         if (!refcount) {
1759                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1760                 bch2_trans_inconsistent(trans,
1761                         "nonexistent indirect extent at %llu while marking\n  %s",
1762                         *idx, buf.buf);
1763                 ret = -EIO;
1764                 goto err;
1765         }
1766
1767         if (!*refcount && (flags & BTREE_TRIGGER_OVERWRITE)) {
1768                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1769                 bch2_trans_inconsistent(trans,
1770                         "indirect extent refcount underflow at %llu while marking\n  %s",
1771                         *idx, buf.buf);
1772                 ret = -EIO;
1773                 goto err;
1774         }
1775
1776         if (flags & BTREE_TRIGGER_INSERT) {
1777                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1778                 u64 pad;
1779
1780                 pad = max_t(s64, le32_to_cpu(v->front_pad),
1781                             le64_to_cpu(v->idx) - bkey_start_offset(&k->k));
1782                 BUG_ON(pad > U32_MAX);
1783                 v->front_pad = cpu_to_le32(pad);
1784
1785                 pad = max_t(s64, le32_to_cpu(v->back_pad),
1786                             k->k.p.offset - p.k->size - le64_to_cpu(v->idx));
1787                 BUG_ON(pad > U32_MAX);
1788                 v->back_pad = cpu_to_le32(pad);
1789         }
1790
1791         le64_add_cpu(refcount, add);
1792
1793         bch2_btree_iter_set_pos_to_extent_start(&iter);
1794         ret = bch2_trans_update(trans, &iter, k, 0);
1795         if (ret)
1796                 goto err;
1797
1798         *idx = k->k.p.offset;
1799 err:
1800         bch2_trans_iter_exit(trans, &iter);
1801         printbuf_exit(&buf);
1802         return ret;
1803 }
1804
1805 int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1806                               enum btree_id btree_id, unsigned level,
1807                               struct bkey_s_c old,
1808                               struct bkey_i *new,
1809                               unsigned flags)
1810 {
1811         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1812                 ? old
1813                 : bkey_i_to_s_c(new);
1814         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1815         u64 idx, end_idx;
1816         int ret = 0;
1817
1818         if (flags & BTREE_TRIGGER_INSERT) {
1819                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1820
1821                 v->front_pad = v->back_pad = 0;
1822         }
1823
1824         idx     = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
1825         end_idx = le64_to_cpu(p.v->idx) + p.k->size +
1826                 le32_to_cpu(p.v->back_pad);
1827
1828         while (idx < end_idx && !ret)
1829                 ret = __bch2_trans_mark_reflink_p(trans, p, &idx, flags);
1830
1831         return ret;
1832 }
1833
1834 static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1835                                     struct bch_dev *ca, size_t b,
1836                                     enum bch_data_type type,
1837                                     unsigned sectors)
1838 {
1839         struct bch_fs *c = trans->c;
1840         struct btree_iter iter;
1841         struct bkey_i_alloc_v4 *a;
1842         int ret = 0;
1843
1844         /*
1845          * Backup superblock might be past the end of our normal usable space:
1846          */
1847         if (b >= ca->mi.nbuckets)
1848                 return 0;
1849
1850         a = bch2_trans_start_alloc_update(trans, &iter, POS(ca->dev_idx, b));
1851         if (IS_ERR(a))
1852                 return PTR_ERR(a);
1853
1854         if (a->v.data_type && type && a->v.data_type != type) {
1855                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1856                         "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
1857                         "while marking %s",
1858                         iter.pos.inode, iter.pos.offset, a->v.gen,
1859                         bch2_data_types[a->v.data_type],
1860                         bch2_data_types[type],
1861                         bch2_data_types[type]);
1862                 ret = -EIO;
1863                 goto out;
1864         }
1865
1866         a->v.data_type          = type;
1867         a->v.dirty_sectors      = sectors;
1868
1869         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1870         if (ret)
1871                 goto out;
1872 out:
1873         bch2_trans_iter_exit(trans, &iter);
1874         return ret;
1875 }
1876
1877 int bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1878                                     struct bch_dev *ca, size_t b,
1879                                     enum bch_data_type type,
1880                                     unsigned sectors)
1881 {
1882         return commit_do(trans, NULL, NULL, 0,
1883                         __bch2_trans_mark_metadata_bucket(trans, ca, b, type, sectors));
1884 }
1885
1886 static int bch2_trans_mark_metadata_sectors(struct btree_trans *trans,
1887                                             struct bch_dev *ca,
1888                                             u64 start, u64 end,
1889                                             enum bch_data_type type,
1890                                             u64 *bucket, unsigned *bucket_sectors)
1891 {
1892         do {
1893                 u64 b = sector_to_bucket(ca, start);
1894                 unsigned sectors =
1895                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1896
1897                 if (b != *bucket && *bucket_sectors) {
1898                         int ret = bch2_trans_mark_metadata_bucket(trans, ca, *bucket,
1899                                                                   type, *bucket_sectors);
1900                         if (ret)
1901                                 return ret;
1902
1903                         *bucket_sectors = 0;
1904                 }
1905
1906                 *bucket         = b;
1907                 *bucket_sectors += sectors;
1908                 start += sectors;
1909         } while (start < end);
1910
1911         return 0;
1912 }
1913
1914 static int __bch2_trans_mark_dev_sb(struct btree_trans *trans,
1915                                     struct bch_dev *ca)
1916 {
1917         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1918         u64 bucket = 0;
1919         unsigned i, bucket_sectors = 0;
1920         int ret;
1921
1922         for (i = 0; i < layout->nr_superblocks; i++) {
1923                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1924
1925                 if (offset == BCH_SB_SECTOR) {
1926                         ret = bch2_trans_mark_metadata_sectors(trans, ca,
1927                                                 0, BCH_SB_SECTOR,
1928                                                 BCH_DATA_sb, &bucket, &bucket_sectors);
1929                         if (ret)
1930                                 return ret;
1931                 }
1932
1933                 ret = bch2_trans_mark_metadata_sectors(trans, ca, offset,
1934                                       offset + (1 << layout->sb_max_size_bits),
1935                                       BCH_DATA_sb, &bucket, &bucket_sectors);
1936                 if (ret)
1937                         return ret;
1938         }
1939
1940         if (bucket_sectors) {
1941                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1942                                 bucket, BCH_DATA_sb, bucket_sectors);
1943                 if (ret)
1944                         return ret;
1945         }
1946
1947         for (i = 0; i < ca->journal.nr; i++) {
1948                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1949                                 ca->journal.buckets[i],
1950                                 BCH_DATA_journal, ca->mi.bucket_size);
1951                 if (ret)
1952                         return ret;
1953         }
1954
1955         return 0;
1956 }
1957
1958 int bch2_trans_mark_dev_sb(struct bch_fs *c, struct bch_dev *ca)
1959 {
1960         return bch2_trans_run(c, __bch2_trans_mark_dev_sb(&trans, ca));
1961 }
1962
1963 /* Disk reservations: */
1964
1965 #define SECTORS_CACHE   1024
1966
1967 int __bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
1968                               u64 sectors, int flags)
1969 {
1970         struct bch_fs_pcpu *pcpu;
1971         u64 old, v, get;
1972         s64 sectors_available;
1973         int ret;
1974
1975         percpu_down_read(&c->mark_lock);
1976         preempt_disable();
1977         pcpu = this_cpu_ptr(c->pcpu);
1978
1979         if (sectors <= pcpu->sectors_available)
1980                 goto out;
1981
1982         v = atomic64_read(&c->sectors_available);
1983         do {
1984                 old = v;
1985                 get = min((u64) sectors + SECTORS_CACHE, old);
1986
1987                 if (get < sectors) {
1988                         preempt_enable();
1989                         goto recalculate;
1990                 }
1991         } while ((v = atomic64_cmpxchg(&c->sectors_available,
1992                                        old, old - get)) != old);
1993
1994         pcpu->sectors_available         += get;
1995
1996 out:
1997         pcpu->sectors_available         -= sectors;
1998         this_cpu_add(*c->online_reserved, sectors);
1999         res->sectors                    += sectors;
2000
2001         preempt_enable();
2002         percpu_up_read(&c->mark_lock);
2003         return 0;
2004
2005 recalculate:
2006         mutex_lock(&c->sectors_available_lock);
2007
2008         percpu_u64_set(&c->pcpu->sectors_available, 0);
2009         sectors_available = avail_factor(__bch2_fs_usage_read_short(c).free);
2010
2011         if (sectors <= sectors_available ||
2012             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
2013                 atomic64_set(&c->sectors_available,
2014                              max_t(s64, 0, sectors_available - sectors));
2015                 this_cpu_add(*c->online_reserved, sectors);
2016                 res->sectors                    += sectors;
2017                 ret = 0;
2018         } else {
2019                 atomic64_set(&c->sectors_available, sectors_available);
2020                 ret = -BCH_ERR_ENOSPC_disk_reservation;
2021         }
2022
2023         mutex_unlock(&c->sectors_available_lock);
2024         percpu_up_read(&c->mark_lock);
2025
2026         return ret;
2027 }
2028
2029 /* Startup/shutdown: */
2030
2031 static void bucket_gens_free_rcu(struct rcu_head *rcu)
2032 {
2033         struct bucket_gens *buckets =
2034                 container_of(rcu, struct bucket_gens, rcu);
2035
2036         kvpfree(buckets, sizeof(*buckets) + buckets->nbuckets);
2037 }
2038
2039 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
2040 {
2041         struct bucket_gens *bucket_gens = NULL, *old_bucket_gens = NULL;
2042         unsigned long *buckets_nouse = NULL;
2043         bool resize = ca->bucket_gens != NULL;
2044         int ret;
2045
2046         if (!(bucket_gens       = kvpmalloc(sizeof(struct bucket_gens) + nbuckets,
2047                                             GFP_KERNEL|__GFP_ZERO))) {
2048                 ret = -BCH_ERR_ENOMEM_bucket_gens;
2049                 goto err;
2050         }
2051
2052         if ((c->opts.buckets_nouse &&
2053              !(buckets_nouse    = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2054                                             sizeof(unsigned long),
2055                                             GFP_KERNEL|__GFP_ZERO)))) {
2056                 ret = -BCH_ERR_ENOMEM_buckets_nouse;
2057                 goto err;
2058         }
2059
2060         bucket_gens->first_bucket = ca->mi.first_bucket;
2061         bucket_gens->nbuckets   = nbuckets;
2062
2063         bch2_copygc_stop(c);
2064
2065         if (resize) {
2066                 down_write(&c->gc_lock);
2067                 down_write(&ca->bucket_lock);
2068                 percpu_down_write(&c->mark_lock);
2069         }
2070
2071         old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
2072
2073         if (resize) {
2074                 size_t n = min(bucket_gens->nbuckets, old_bucket_gens->nbuckets);
2075
2076                 memcpy(bucket_gens->b,
2077                        old_bucket_gens->b,
2078                        n);
2079                 if (buckets_nouse)
2080                         memcpy(buckets_nouse,
2081                                ca->buckets_nouse,
2082                                BITS_TO_LONGS(n) * sizeof(unsigned long));
2083         }
2084
2085         rcu_assign_pointer(ca->bucket_gens, bucket_gens);
2086         bucket_gens     = old_bucket_gens;
2087
2088         swap(ca->buckets_nouse, buckets_nouse);
2089
2090         nbuckets = ca->mi.nbuckets;
2091
2092         if (resize) {
2093                 percpu_up_write(&c->mark_lock);
2094                 up_write(&ca->bucket_lock);
2095                 up_write(&c->gc_lock);
2096         }
2097
2098         ret = 0;
2099 err:
2100         kvpfree(buckets_nouse,
2101                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2102         if (bucket_gens)
2103                 call_rcu(&bucket_gens->rcu, bucket_gens_free_rcu);
2104
2105         return ret;
2106 }
2107
2108 void bch2_dev_buckets_free(struct bch_dev *ca)
2109 {
2110         unsigned i;
2111
2112         kvpfree(ca->buckets_nouse,
2113                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2114         kvpfree(rcu_dereference_protected(ca->bucket_gens, 1),
2115                 sizeof(struct bucket_gens) + ca->mi.nbuckets);
2116
2117         for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
2118                 free_percpu(ca->usage[i]);
2119         kfree(ca->usage_base);
2120 }
2121
2122 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2123 {
2124         unsigned i;
2125
2126         ca->usage_base = kzalloc(sizeof(struct bch_dev_usage), GFP_KERNEL);
2127         if (!ca->usage_base)
2128                 return -BCH_ERR_ENOMEM_usage_init;
2129
2130         for (i = 0; i < ARRAY_SIZE(ca->usage); i++) {
2131                 ca->usage[i] = alloc_percpu(struct bch_dev_usage);
2132                 if (!ca->usage[i])
2133                         return -BCH_ERR_ENOMEM_usage_init;
2134         }
2135
2136         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);
2137 }