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 u64 bch2_fs_usage_read_one(struct bch_fs *c, u64 *v)
 106 {
 107         ssize_t offset = v - (u64 *) c->usage_base;
 108         unsigned i, seq;
 109         u64 ret;
 110
 111         BUG_ON(offset < 0 || offset >= fs_usage_u64s(c));
 112         percpu_rwsem_assert_held(&c->mark_lock);
 113
 114         do {
 115                 seq = read_seqcount_begin(&c->usage_lock);
 116                 ret = *v;
 117
 118                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 119                         ret += percpu_u64_get((u64 __percpu *) c->usage[i] + offset);
 120         } while (read_seqcount_retry(&c->usage_lock, seq));
 121
 122         return ret;
 123 }
 124
 125 struct bch_fs_usage_online *bch2_fs_usage_read(struct bch_fs *c)
 126 {
 127         struct bch_fs_usage_online *ret;
 128         unsigned nr_replicas = READ_ONCE(c->replicas.nr);
 129         unsigned seq, i;
 130 retry:
 131         ret = kmalloc(__fs_usage_online_u64s(nr_replicas) * sizeof(u64), GFP_KERNEL);
 132         if (unlikely(!ret))
 133                 return NULL;
 134
 135         percpu_down_read(&c->mark_lock);
 136
 137         if (nr_replicas != c->replicas.nr) {
 138                 nr_replicas = c->replicas.nr;
 139                 percpu_up_read(&c->mark_lock);
 140                 kfree(ret);
 141                 goto retry;
 142         }
 143
 144         ret->online_reserved = percpu_u64_get(c->online_reserved);
 145
 146         do {
 147                 seq = read_seqcount_begin(&c->usage_lock);
 148                 unsafe_memcpy(&ret->u, c->usage_base,
 149                               __fs_usage_u64s(nr_replicas) * sizeof(u64),
 150                               "embedded variable length struct");
 151                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 152                         acc_u64s_percpu((u64 *) &ret->u, (u64 __percpu *) c->usage[i],
 153                                         __fs_usage_u64s(nr_replicas));
 154         } while (read_seqcount_retry(&c->usage_lock, seq));
 155
 156         return ret;
 157 }
 158
 159 void bch2_fs_usage_acc_to_base(struct bch_fs *c, unsigned idx)
 160 {
 161         struct bch_dev *ca;
 162         unsigned i, u64s = fs_usage_u64s(c);
 163
 164         BUG_ON(idx >= ARRAY_SIZE(c->usage));
 165
 166         preempt_disable();
 167         write_seqcount_begin(&c->usage_lock);
 168
 169         acc_u64s_percpu((u64 *) c->usage_base,
 170                         (u64 __percpu *) c->usage[idx], u64s);
 171         percpu_memset(c->usage[idx], 0, u64s * sizeof(u64));
 172
 173         rcu_read_lock();
 174         for_each_member_device_rcu(ca, c, i, NULL) {
 175                 u64s = dev_usage_u64s();
 176
 177                 acc_u64s_percpu((u64 *) ca->usage_base,
 178                                 (u64 __percpu *) ca->usage[idx], u64s);
 179                 percpu_memset(ca->usage[idx], 0, u64s * sizeof(u64));
 180         }
 181         rcu_read_unlock();
 182
 183         write_seqcount_end(&c->usage_lock);
 184         preempt_enable();
 185 }
 186
 187 void bch2_fs_usage_to_text(struct printbuf *out,
 188                            struct bch_fs *c,
 189                            struct bch_fs_usage_online *fs_usage)
 190 {
 191         unsigned i;
 192
 193         prt_printf(out, "capacity:\t\t\t%llu\n", c->capacity);
 194
 195         prt_printf(out, "hidden:\t\t\t\t%llu\n",
 196                fs_usage->u.hidden);
 197         prt_printf(out, "data:\t\t\t\t%llu\n",
 198                fs_usage->u.data);
 199         prt_printf(out, "cached:\t\t\t\t%llu\n",
 200                fs_usage->u.cached);
 201         prt_printf(out, "reserved:\t\t\t%llu\n",
 202                fs_usage->u.reserved);
 203         prt_printf(out, "nr_inodes:\t\t\t%llu\n",
 204                fs_usage->u.nr_inodes);
 205         prt_printf(out, "online reserved:\t\t%llu\n",
 206                fs_usage->online_reserved);
 207
 208         for (i = 0;
 209              i < ARRAY_SIZE(fs_usage->u.persistent_reserved);
 210              i++) {
 211                 prt_printf(out, "%u replicas:\n", i + 1);
 212                 prt_printf(out, "\treserved:\t\t%llu\n",
 213                        fs_usage->u.persistent_reserved[i]);
 214         }
 215
 216         for (i = 0; i < c->replicas.nr; i++) {
 217                 struct bch_replicas_entry *e =
 218                         cpu_replicas_entry(&c->replicas, i);
 219
 220                 prt_printf(out, "\t");
 221                 bch2_replicas_entry_to_text(out, e);
 222                 prt_printf(out, ":\t%llu\n", fs_usage->u.replicas[i]);
 223         }
 224 }
 225
 226 static u64 reserve_factor(u64 r)
 227 {
 228         return r + (round_up(r, (1 << RESERVE_FACTOR)) >> RESERVE_FACTOR);
 229 }
 230
 231 u64 bch2_fs_sectors_used(struct bch_fs *c, struct bch_fs_usage_online *fs_usage)
 232 {
 233         return min(fs_usage->u.hidden +
 234                    fs_usage->u.btree +
 235                    fs_usage->u.data +
 236                    reserve_factor(fs_usage->u.reserved +
 237                                   fs_usage->online_reserved),
 238                    c->capacity);
 239 }
 240
 241 static struct bch_fs_usage_short
 242 __bch2_fs_usage_read_short(struct bch_fs *c)
 243 {
 244         struct bch_fs_usage_short ret;
 245         u64 data, reserved;
 246
 247         ret.capacity = c->capacity -
 248                 bch2_fs_usage_read_one(c, &c->usage_base->hidden);
 249
 250         data            = bch2_fs_usage_read_one(c, &c->usage_base->data) +
 251                 bch2_fs_usage_read_one(c, &c->usage_base->btree);
 252         reserved        = bch2_fs_usage_read_one(c, &c->usage_base->reserved) +
 253                 percpu_u64_get(c->online_reserved);
 254
 255         ret.used        = min(ret.capacity, data + reserve_factor(reserved));
 256         ret.free        = ret.capacity - ret.used;
 257
 258         ret.nr_inodes   = bch2_fs_usage_read_one(c, &c->usage_base->nr_inodes);
 259
 260         return ret;
 261 }
 262
 263 struct bch_fs_usage_short
 264 bch2_fs_usage_read_short(struct bch_fs *c)
 265 {
 266         struct bch_fs_usage_short ret;
 267
 268         percpu_down_read(&c->mark_lock);
 269         ret = __bch2_fs_usage_read_short(c);
 270         percpu_up_read(&c->mark_lock);
 271
 272         return ret;
 273 }
 274
 275 void bch2_dev_usage_init(struct bch_dev *ca)
 276 {
 277         ca->usage_base->d[BCH_DATA_free].buckets = ca->mi.nbuckets - ca->mi.first_bucket;
 278 }
 279
 280 static inline int bucket_sectors_fragmented(struct bch_dev *ca,
 281                                             struct bch_alloc_v4 a)
 282 {
 283         return a.dirty_sectors
 284                 ? max(0, (int) ca->mi.bucket_size - (int) a.dirty_sectors)
 285                 : 0;
 286 }
 287
 288 static void bch2_dev_usage_update(struct bch_fs *c, struct bch_dev *ca,
 289                                   struct bch_alloc_v4 old,
 290                                   struct bch_alloc_v4 new,
 291                                   u64 journal_seq, bool gc)
 292 {
 293         struct bch_fs_usage *fs_usage;
 294         struct bch_dev_usage *u;
 295
 296         preempt_disable();
 297         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 298
 299         if (data_type_is_hidden(old.data_type))
 300                 fs_usage->hidden -= ca->mi.bucket_size;
 301         if (data_type_is_hidden(new.data_type))
 302                 fs_usage->hidden += ca->mi.bucket_size;
 303
 304         u = dev_usage_ptr(ca, journal_seq, gc);
 305
 306         u->d[old.data_type].buckets--;
 307         u->d[new.data_type].buckets++;
 308
 309         u->buckets_ec -= (int) !!old.stripe;
 310         u->buckets_ec += (int) !!new.stripe;
 311
 312         u->d[old.data_type].sectors -= old.dirty_sectors;
 313         u->d[new.data_type].sectors += new.dirty_sectors;
 314
 315         u->d[BCH_DATA_cached].sectors += new.cached_sectors;
 316         u->d[BCH_DATA_cached].sectors -= old.cached_sectors;
 317
 318         u->d[old.data_type].fragmented -= bucket_sectors_fragmented(ca, old);
 319         u->d[new.data_type].fragmented += bucket_sectors_fragmented(ca, new);
 320
 321         preempt_enable();
 322 }
 323
 324 static void bch2_dev_usage_update_m(struct bch_fs *c, struct bch_dev *ca,
 325                                     struct bucket old, struct bucket new,
 326                                     u64 journal_seq, bool gc)
 327 {
 328         struct bch_alloc_v4 old_a = {
 329                 .gen            = old.gen,
 330                 .data_type      = old.data_type,
 331                 .dirty_sectors  = old.dirty_sectors,
 332                 .cached_sectors = old.cached_sectors,
 333                 .stripe         = old.stripe,
 334         };
 335         struct bch_alloc_v4 new_a = {
 336                 .gen            = new.gen,
 337                 .data_type      = new.data_type,
 338                 .dirty_sectors  = new.dirty_sectors,
 339                 .cached_sectors = new.cached_sectors,
 340                 .stripe         = new.stripe,
 341         };
 342
 343         bch2_dev_usage_update(c, ca, old_a, new_a, journal_seq, gc);
 344 }
 345
 346 static inline int __update_replicas(struct bch_fs *c,
 347                                     struct bch_fs_usage *fs_usage,
 348                                     struct bch_replicas_entry *r,
 349                                     s64 sectors)
 350 {
 351         int idx = bch2_replicas_entry_idx(c, r);
 352
 353         if (idx < 0)
 354                 return -1;
 355
 356         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 357         fs_usage->replicas[idx]         += sectors;
 358         return 0;
 359 }
 360
 361 static inline int update_replicas(struct bch_fs *c, struct bkey_s_c k,
 362                         struct bch_replicas_entry *r, s64 sectors,
 363                         unsigned journal_seq, bool gc)
 364 {
 365         struct bch_fs_usage *fs_usage;
 366         int idx, ret = 0;
 367         struct printbuf buf = PRINTBUF;
 368
 369         percpu_down_read(&c->mark_lock);
 370
 371         idx = bch2_replicas_entry_idx(c, r);
 372         if (idx < 0 &&
 373             fsck_err(c, "no replicas entry\n"
 374                      "  while marking %s",
 375                      (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 376                 percpu_up_read(&c->mark_lock);
 377                 ret = bch2_mark_replicas(c, r);
 378                 percpu_down_read(&c->mark_lock);
 379
 380                 if (ret)
 381                         goto err;
 382                 idx = bch2_replicas_entry_idx(c, r);
 383         }
 384         if (idx < 0) {
 385                 ret = -1;
 386                 goto err;
 387         }
 388
 389         preempt_disable();
 390         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 391         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 392         fs_usage->replicas[idx]         += sectors;
 393         preempt_enable();
 394 err:
 395 fsck_err:
 396         percpu_up_read(&c->mark_lock);
 397         printbuf_exit(&buf);
 398         return ret;
 399 }
 400
 401 static inline int update_cached_sectors(struct bch_fs *c,
 402                         struct bkey_s_c k,
 403                         unsigned dev, s64 sectors,
 404                         unsigned journal_seq, bool gc)
 405 {
 406         struct bch_replicas_padded r;
 407
 408         bch2_replicas_entry_cached(&r.e, dev);
 409
 410         return update_replicas(c, k, &r.e, sectors, journal_seq, gc);
 411 }
 412
 413 static int __replicas_deltas_realloc(struct btree_trans *trans, unsigned more,
 414                                      gfp_t gfp)
 415 {
 416         struct replicas_delta_list *d = trans->fs_usage_deltas;
 417         unsigned new_size = d ? (d->size + more) * 2 : 128;
 418         unsigned alloc_size = sizeof(*d) + new_size;
 419
 420         WARN_ON_ONCE(alloc_size > REPLICAS_DELTA_LIST_MAX);
 421
 422         if (!d || d->used + more > d->size) {
 423                 d = krealloc(d, alloc_size, gfp|__GFP_ZERO);
 424
 425                 if (unlikely(!d)) {
 426                         if (alloc_size > REPLICAS_DELTA_LIST_MAX)
 427                                 return -ENOMEM;
 428
 429                         d = mempool_alloc(&trans->c->replicas_delta_pool, gfp);
 430                         if (!d)
 431                                 return -ENOMEM;
 432
 433                         memset(d, 0, REPLICAS_DELTA_LIST_MAX);
 434
 435                         if (trans->fs_usage_deltas)
 436                                 memcpy(d, trans->fs_usage_deltas,
 437                                        trans->fs_usage_deltas->size + sizeof(*d));
 438
 439                         new_size = REPLICAS_DELTA_LIST_MAX - sizeof(*d);
 440                         kfree(trans->fs_usage_deltas);
 441                 }
 442
 443                 d->size = new_size;
 444                 trans->fs_usage_deltas = d;
 445         }
 446
 447         return 0;
 448 }
 449
 450 int bch2_replicas_deltas_realloc(struct btree_trans *trans, unsigned more)
 451 {
 452         return allocate_dropping_locks_errcode(trans,
 453                                 __replicas_deltas_realloc(trans, more, _gfp));
 454 }
 455
 456 static inline int 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         int ret;
 464
 465         if (!sectors)
 466                 return 0;
 467
 468         b = replicas_entry_bytes(r) + 8;
 469         ret = bch2_replicas_deltas_realloc(trans, b);
 470         if (ret)
 471                 return ret;
 472
 473         d = trans->fs_usage_deltas;
 474         n = (void *) d->d + d->used;
 475         n->delta = sectors;
 476         memcpy((void *) n + offsetof(struct replicas_delta, r),
 477                r, replicas_entry_bytes(r));
 478         bch2_replicas_entry_sort(&n->r);
 479         d->used += b;
 480         return 0;
 481 }
 482
 483 static inline int update_cached_sectors_list(struct btree_trans *trans,
 484                                               unsigned dev, s64 sectors)
 485 {
 486         struct bch_replicas_padded r;
 487
 488         bch2_replicas_entry_cached(&r.e, dev);
 489
 490         return update_replicas_list(trans, &r.e, sectors);
 491 }
 492
 493 int bch2_mark_alloc(struct btree_trans *trans,
 494                     enum btree_id btree, unsigned level,
 495                     struct bkey_s_c old, struct bkey_s_c new,
 496                     unsigned flags)
 497 {
 498         bool gc = flags & BTREE_TRIGGER_GC;
 499         u64 journal_seq = trans->journal_res.seq;
 500         u64 bucket_journal_seq;
 501         struct bch_fs *c = trans->c;
 502         struct bch_alloc_v4 old_a_convert, new_a_convert;
 503         const struct bch_alloc_v4 *old_a, *new_a;
 504         struct bch_dev *ca;
 505         int ret = 0;
 506
 507         /*
 508          * alloc btree is read in by bch2_alloc_read, not gc:
 509          */
 510         if ((flags & BTREE_TRIGGER_GC) &&
 511             !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
 512                 return 0;
 513
 514         if (bch2_trans_inconsistent_on(!bch2_dev_bucket_exists(c, new.k->p), trans,
 515                                        "alloc key for invalid device or bucket"))
 516                 return -EIO;
 517
 518         ca = bch_dev_bkey_exists(c, new.k->p.inode);
 519
 520         old_a = bch2_alloc_to_v4(old, &old_a_convert);
 521         new_a = bch2_alloc_to_v4(new, &new_a_convert);
 522
 523         bucket_journal_seq = new_a->journal_seq;
 524
 525         if ((flags & BTREE_TRIGGER_INSERT) &&
 526             data_type_is_empty(old_a->data_type) !=
 527             data_type_is_empty(new_a->data_type) &&
 528             new.k->type == KEY_TYPE_alloc_v4) {
 529                 struct bch_alloc_v4 *v = (struct bch_alloc_v4 *) new.v;
 530
 531                 EBUG_ON(!journal_seq);
 532
 533                 /*
 534                  * If the btree updates referring to a bucket weren't flushed
 535                  * before the bucket became empty again, then the we don't have
 536                  * to wait on a journal flush before we can reuse the bucket:
 537                  */
 538                 v->journal_seq = bucket_journal_seq =
 539                         data_type_is_empty(new_a->data_type) &&
 540                         (journal_seq == v->journal_seq ||
 541                          bch2_journal_noflush_seq(&c->journal, v->journal_seq))
 542                         ? 0 : journal_seq;
 543         }
 544
 545         if (!data_type_is_empty(old_a->data_type) &&
 546             data_type_is_empty(new_a->data_type) &&
 547             bucket_journal_seq) {
 548                 ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 549                                 c->journal.flushed_seq_ondisk,
 550                                 new.k->p.inode, new.k->p.offset,
 551                                 bucket_journal_seq);
 552                 if (ret) {
 553                         bch2_fs_fatal_error(c,
 554                                 "error setting bucket_needs_journal_commit: %i", ret);
 555                         return ret;
 556                 }
 557         }
 558
 559         percpu_down_read(&c->mark_lock);
 560         if (!gc && new_a->gen != old_a->gen)
 561                 *bucket_gen(ca, new.k->p.offset) = new_a->gen;
 562
 563         bch2_dev_usage_update(c, ca, *old_a, *new_a, journal_seq, gc);
 564
 565         if (gc) {
 566                 struct bucket *g = gc_bucket(ca, new.k->p.offset);
 567
 568                 bucket_lock(g);
 569
 570                 g->gen_valid            = 1;
 571                 g->gen                  = new_a->gen;
 572                 g->data_type            = new_a->data_type;
 573                 g->stripe               = new_a->stripe;
 574                 g->stripe_redundancy    = new_a->stripe_redundancy;
 575                 g->dirty_sectors        = new_a->dirty_sectors;
 576                 g->cached_sectors       = new_a->cached_sectors;
 577
 578                 bucket_unlock(g);
 579         }
 580         percpu_up_read(&c->mark_lock);
 581
 582         /*
 583          * need to know if we're getting called from the invalidate path or
 584          * not:
 585          */
 586
 587         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 588             old_a->cached_sectors) {
 589                 ret = update_cached_sectors(c, new, ca->dev_idx,
 590                                             -((s64) old_a->cached_sectors),
 591                                             journal_seq, gc);
 592                 if (ret) {
 593                         bch2_fs_fatal_error(c, "%s(): no replicas entry while updating cached sectors",
 594                                             __func__);
 595                         return ret;
 596                 }
 597         }
 598
 599         if (new_a->data_type == BCH_DATA_free &&
 600             (!new_a->journal_seq || new_a->journal_seq < c->journal.flushed_seq_ondisk))
 601                 closure_wake_up(&c->freelist_wait);
 602
 603         if (new_a->data_type == BCH_DATA_need_discard &&
 604             (!bucket_journal_seq || bucket_journal_seq < c->journal.flushed_seq_ondisk))
 605                 bch2_do_discards(c);
 606
 607         if (old_a->data_type != BCH_DATA_cached &&
 608             new_a->data_type == BCH_DATA_cached &&
 609             should_invalidate_buckets(ca, bch2_dev_usage_read(ca)))
 610                 bch2_do_invalidates(c);
 611
 612         if (new_a->data_type == BCH_DATA_need_gc_gens)
 613                 bch2_do_gc_gens(c);
 614
 615         return 0;
 616 }
 617
 618 int bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 619                               size_t b, enum bch_data_type data_type,
 620                               unsigned sectors, struct gc_pos pos,
 621                               unsigned flags)
 622 {
 623         struct bucket old, new, *g;
 624         int ret = 0;
 625
 626         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 627         BUG_ON(data_type != BCH_DATA_sb &&
 628                data_type != BCH_DATA_journal);
 629
 630         /*
 631          * Backup superblock might be past the end of our normal usable space:
 632          */
 633         if (b >= ca->mi.nbuckets)
 634                 return 0;
 635
 636         percpu_down_read(&c->mark_lock);
 637         g = gc_bucket(ca, b);
 638
 639         bucket_lock(g);
 640         old = *g;
 641
 642         if (bch2_fs_inconsistent_on(g->data_type &&
 643                         g->data_type != data_type, c,
 644                         "different types of data in same bucket: %s, %s",
 645                         bch2_data_types[g->data_type],
 646                         bch2_data_types[data_type])) {
 647                 ret = -EIO;
 648                 goto err;
 649         }
 650
 651         if (bch2_fs_inconsistent_on((u64) g->dirty_sectors + sectors > ca->mi.bucket_size, c,
 652                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > bucket size",
 653                         ca->dev_idx, b, g->gen,
 654                         bch2_data_types[g->data_type ?: data_type],
 655                         g->dirty_sectors, sectors)) {
 656                 ret = -EIO;
 657                 goto err;
 658         }
 659
 660
 661         g->data_type = data_type;
 662         g->dirty_sectors += sectors;
 663         new = *g;
 664 err:
 665         bucket_unlock(g);
 666         if (!ret)
 667                 bch2_dev_usage_update_m(c, ca, old, new, 0, true);
 668         percpu_up_read(&c->mark_lock);
 669         return ret;
 670 }
 671
 672 static int check_bucket_ref(struct btree_trans *trans,
 673                             struct bkey_s_c k,
 674                             const struct bch_extent_ptr *ptr,
 675                             s64 sectors, enum bch_data_type ptr_data_type,
 676                             u8 b_gen, u8 bucket_data_type,
 677                             u32 dirty_sectors, u32 cached_sectors)
 678 {
 679         struct bch_fs *c = trans->c;
 680         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 681         size_t bucket_nr = PTR_BUCKET_NR(ca, ptr);
 682         u32 bucket_sectors = !ptr->cached
 683                 ? dirty_sectors
 684                 : cached_sectors;
 685         struct printbuf buf = PRINTBUF;
 686         int ret = 0;
 687
 688         if (bucket_data_type == BCH_DATA_cached)
 689                 bucket_data_type = BCH_DATA_user;
 690
 691         if ((bucket_data_type == BCH_DATA_stripe && ptr_data_type == BCH_DATA_user) ||
 692             (bucket_data_type == BCH_DATA_user   && ptr_data_type == BCH_DATA_stripe))
 693                 bucket_data_type = ptr_data_type = BCH_DATA_stripe;
 694
 695         if (gen_after(ptr->gen, b_gen)) {
 696                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 697                         "bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\n"
 698                         "while marking %s",
 699                         ptr->dev, bucket_nr, b_gen,
 700                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 701                         ptr->gen,
 702                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 703                 ret = -EIO;
 704                 goto err;
 705         }
 706
 707         if (gen_cmp(b_gen, ptr->gen) > BUCKET_GC_GEN_MAX) {
 708                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 709                         "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 710                         "while marking %s",
 711                         ptr->dev, bucket_nr, b_gen,
 712                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 713                         ptr->gen,
 714                         (printbuf_reset(&buf),
 715                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 716                 ret = -EIO;
 717                 goto err;
 718         }
 719
 720         if (b_gen != ptr->gen && !ptr->cached) {
 721                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 722                         "bucket %u:%zu gen %u (mem gen %u) data type %s: stale dirty ptr (gen %u)\n"
 723                         "while marking %s",
 724                         ptr->dev, bucket_nr, b_gen,
 725                         *bucket_gen(ca, bucket_nr),
 726                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 727                         ptr->gen,
 728                         (printbuf_reset(&buf),
 729                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 730                 ret = -EIO;
 731                 goto err;
 732         }
 733
 734         if (b_gen != ptr->gen) {
 735                 ret = 1;
 736                 goto out;
 737         }
 738
 739         if (!data_type_is_empty(bucket_data_type) &&
 740             ptr_data_type &&
 741             bucket_data_type != ptr_data_type) {
 742                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 743                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 744                         "while marking %s",
 745                         ptr->dev, bucket_nr, b_gen,
 746                         bch2_data_types[bucket_data_type],
 747                         bch2_data_types[ptr_data_type],
 748                         (printbuf_reset(&buf),
 749                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 750                 ret = -EIO;
 751                 goto err;
 752         }
 753
 754         if ((u64) bucket_sectors + sectors > U32_MAX) {
 755                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 756                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U32_MAX\n"
 757                         "while marking %s",
 758                         ptr->dev, bucket_nr, b_gen,
 759                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 760                         bucket_sectors, sectors,
 761                         (printbuf_reset(&buf),
 762                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 763                 ret = -EIO;
 764                 goto err;
 765         }
 766 out:
 767         printbuf_exit(&buf);
 768         return ret;
 769 err:
 770         bch2_dump_trans_updates(trans);
 771         goto out;
 772 }
 773
 774 static int mark_stripe_bucket(struct btree_trans *trans,
 775                               struct bkey_s_c k,
 776                               unsigned ptr_idx,
 777                               unsigned flags)
 778 {
 779         struct bch_fs *c = trans->c;
 780         u64 journal_seq = trans->journal_res.seq;
 781         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 782         unsigned nr_data = s->nr_blocks - s->nr_redundant;
 783         bool parity = ptr_idx >= nr_data;
 784         enum bch_data_type data_type = parity ? BCH_DATA_parity : BCH_DATA_stripe;
 785         s64 sectors = parity ? le16_to_cpu(s->sectors) : 0;
 786         const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
 787         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 788         struct bucket old, new, *g;
 789         struct printbuf buf = PRINTBUF;
 790         int ret = 0;
 791
 792         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 793
 794         /* * XXX doesn't handle deletion */
 795
 796         percpu_down_read(&c->mark_lock);
 797         g = PTR_GC_BUCKET(ca, ptr);
 798
 799         if (g->dirty_sectors ||
 800             (g->stripe && g->stripe != k.k->p.offset)) {
 801                 bch2_fs_inconsistent(c,
 802                               "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
 803                               ptr->dev, PTR_BUCKET_NR(ca, ptr), g->gen,
 804                               (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 805                 ret = -EINVAL;
 806                 goto err;
 807         }
 808
 809         bucket_lock(g);
 810         old = *g;
 811
 812         ret = check_bucket_ref(trans, k, ptr, sectors, data_type,
 813                                g->gen, g->data_type,
 814                                g->dirty_sectors, g->cached_sectors);
 815         if (ret)
 816                 goto err;
 817
 818         g->data_type = data_type;
 819         g->dirty_sectors += sectors;
 820
 821         g->stripe               = k.k->p.offset;
 822         g->stripe_redundancy    = s->nr_redundant;
 823         new = *g;
 824 err:
 825         bucket_unlock(g);
 826         if (!ret)
 827                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 828         percpu_up_read(&c->mark_lock);
 829         printbuf_exit(&buf);
 830         return ret;
 831 }
 832
 833 static int __mark_pointer(struct btree_trans *trans,
 834                           struct bkey_s_c k,
 835                           const struct bch_extent_ptr *ptr,
 836                           s64 sectors, enum bch_data_type ptr_data_type,
 837                           u8 bucket_gen, u8 *bucket_data_type,
 838                           u32 *dirty_sectors, u32 *cached_sectors)
 839 {
 840         u32 *dst_sectors = !ptr->cached
 841                 ? dirty_sectors
 842                 : cached_sectors;
 843         int ret = check_bucket_ref(trans, k, ptr, sectors, ptr_data_type,
 844                                    bucket_gen, *bucket_data_type,
 845                                    *dirty_sectors, *cached_sectors);
 846
 847         if (ret)
 848                 return ret;
 849
 850         *dst_sectors += sectors;
 851         *bucket_data_type = *dirty_sectors || *cached_sectors
 852                 ? ptr_data_type : 0;
 853         return 0;
 854 }
 855
 856 static int bch2_mark_pointer(struct btree_trans *trans,
 857                              enum btree_id btree_id, unsigned level,
 858                              struct bkey_s_c k,
 859                              struct extent_ptr_decoded p,
 860                              s64 sectors,
 861                              unsigned flags)
 862 {
 863         u64 journal_seq = trans->journal_res.seq;
 864         struct bch_fs *c = trans->c;
 865         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 866         struct bucket old, new, *g;
 867         enum bch_data_type data_type = bkey_ptr_data_type(btree_id, level, k, p);
 868         u8 bucket_data_type;
 869         int ret = 0;
 870
 871         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 872
 873         percpu_down_read(&c->mark_lock);
 874         g = PTR_GC_BUCKET(ca, &p.ptr);
 875         bucket_lock(g);
 876         old = *g;
 877
 878         bucket_data_type = g->data_type;
 879         ret = __mark_pointer(trans, k, &p.ptr, sectors,
 880                              data_type, g->gen,
 881                              &bucket_data_type,
 882                              &g->dirty_sectors,
 883                              &g->cached_sectors);
 884         if (!ret)
 885                 g->data_type = bucket_data_type;
 886
 887         new = *g;
 888         bucket_unlock(g);
 889         if (!ret)
 890                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 891         percpu_up_read(&c->mark_lock);
 892
 893         return ret;
 894 }
 895
 896 static int bch2_mark_stripe_ptr(struct btree_trans *trans,
 897                                 struct bkey_s_c k,
 898                                 struct bch_extent_stripe_ptr p,
 899                                 enum bch_data_type data_type,
 900                                 s64 sectors,
 901                                 unsigned flags)
 902 {
 903         struct bch_fs *c = trans->c;
 904         struct bch_replicas_padded r;
 905         struct gc_stripe *m;
 906
 907         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 908
 909         m = genradix_ptr_alloc(&c->gc_stripes, p.idx, GFP_KERNEL);
 910         if (!m) {
 911                 bch_err(c, "error allocating memory for gc_stripes, idx %llu",
 912                         (u64) p.idx);
 913                 return -BCH_ERR_ENOMEM_mark_stripe_ptr;
 914         }
 915
 916         mutex_lock(&c->ec_stripes_heap_lock);
 917
 918         if (!m || !m->alive) {
 919                 mutex_unlock(&c->ec_stripes_heap_lock);
 920                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
 921                                     (u64) p.idx);
 922                 bch2_inconsistent_error(c);
 923                 return -EIO;
 924         }
 925
 926         m->block_sectors[p.block] += sectors;
 927
 928         r = m->r;
 929         mutex_unlock(&c->ec_stripes_heap_lock);
 930
 931         r.e.data_type = data_type;
 932         update_replicas(c, k, &r.e, sectors, trans->journal_res.seq, true);
 933
 934         return 0;
 935 }
 936
 937 int bch2_mark_extent(struct btree_trans *trans,
 938                      enum btree_id btree_id, unsigned level,
 939                      struct bkey_s_c old, struct bkey_s_c new,
 940                      unsigned flags)
 941 {
 942         u64 journal_seq = trans->journal_res.seq;
 943         struct bch_fs *c = trans->c;
 944         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old : new;
 945         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 946         const union bch_extent_entry *entry;
 947         struct extent_ptr_decoded p;
 948         struct bch_replicas_padded r;
 949         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
 950                 ? BCH_DATA_btree
 951                 : BCH_DATA_user;
 952         s64 sectors = bkey_is_btree_ptr(k.k)
 953                 ? btree_sectors(c)
 954                 : k.k->size;
 955         s64 dirty_sectors = 0;
 956         bool stale;
 957         int ret;
 958
 959         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 960
 961         r.e.data_type   = data_type;
 962         r.e.nr_devs     = 0;
 963         r.e.nr_required = 1;
 964
 965         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 966                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
 967
 968                 if (flags & BTREE_TRIGGER_OVERWRITE)
 969                         disk_sectors = -disk_sectors;
 970
 971                 ret = bch2_mark_pointer(trans, btree_id, level, k, p, disk_sectors, flags);
 972                 if (ret < 0)
 973                         return ret;
 974
 975                 stale = ret > 0;
 976
 977                 if (p.ptr.cached) {
 978                         if (!stale) {
 979                                 ret = update_cached_sectors(c, k, p.ptr.dev,
 980                                                 disk_sectors, journal_seq, true);
 981                                 if (ret) {
 982                                         bch2_fs_fatal_error(c, "%s(): no replicas entry while updating cached sectors",
 983                                                             __func__);
 984                                         return ret;
 985                                 }
 986                         }
 987                 } else if (!p.has_ec) {
 988                         dirty_sectors          += disk_sectors;
 989                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
 990                 } else {
 991                         ret = bch2_mark_stripe_ptr(trans, k, p.ec, data_type,
 992                                         disk_sectors, flags);
 993                         if (ret)
 994                                 return ret;
 995
 996                         /*
 997                          * There may be other dirty pointers in this extent, but
 998                          * if so they're not required for mounting if we have an
 999                          * erasure coded pointer in this extent:
1000                          */
1001                         r.e.nr_required = 0;
1002                 }
1003         }
1004
1005         if (r.e.nr_devs) {
1006                 ret = update_replicas(c, k, &r.e, dirty_sectors, journal_seq, true);
1007                 if (ret) {
1008                         struct printbuf buf = PRINTBUF;
1009
1010                         bch2_bkey_val_to_text(&buf, c, k);
1011                         bch2_fs_fatal_error(c, "%s(): no replicas entry for %s", __func__, buf.buf);
1012                         printbuf_exit(&buf);
1013                         return ret;
1014                 }
1015         }
1016
1017         return 0;
1018 }
1019
1020 int bch2_mark_stripe(struct btree_trans *trans,
1021                      enum btree_id btree_id, unsigned level,
1022                      struct bkey_s_c old, struct bkey_s_c new,
1023                      unsigned flags)
1024 {
1025         bool gc = flags & BTREE_TRIGGER_GC;
1026         u64 journal_seq = trans->journal_res.seq;
1027         struct bch_fs *c = trans->c;
1028         u64 idx = new.k->p.offset;
1029         const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
1030                 ? bkey_s_c_to_stripe(old).v : NULL;
1031         const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
1032                 ? bkey_s_c_to_stripe(new).v : NULL;
1033         unsigned i;
1034         int ret;
1035
1036         BUG_ON(gc && old_s);
1037
1038         if (!gc) {
1039                 struct stripe *m = genradix_ptr(&c->stripes, idx);
1040
1041                 if (!m) {
1042                         struct printbuf buf1 = PRINTBUF;
1043                         struct printbuf buf2 = PRINTBUF;
1044
1045                         bch2_bkey_val_to_text(&buf1, c, old);
1046                         bch2_bkey_val_to_text(&buf2, c, new);
1047                         bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
1048                                             "old %s\n"
1049                                             "new %s", idx, buf1.buf, buf2.buf);
1050                         printbuf_exit(&buf2);
1051                         printbuf_exit(&buf1);
1052                         bch2_inconsistent_error(c);
1053                         return -1;
1054                 }
1055
1056                 if (!new_s) {
1057                         bch2_stripes_heap_del(c, m, idx);
1058
1059                         memset(m, 0, sizeof(*m));
1060                 } else {
1061                         m->sectors      = le16_to_cpu(new_s->sectors);
1062                         m->algorithm    = new_s->algorithm;
1063                         m->nr_blocks    = new_s->nr_blocks;
1064                         m->nr_redundant = new_s->nr_redundant;
1065                         m->blocks_nonempty = 0;
1066
1067                         for (i = 0; i < new_s->nr_blocks; i++)
1068                                 m->blocks_nonempty += !!stripe_blockcount_get(new_s, i);
1069
1070                         if (!old_s)
1071                                 bch2_stripes_heap_insert(c, m, idx);
1072                         else
1073                                 bch2_stripes_heap_update(c, m, idx);
1074                 }
1075         } else {
1076                 struct gc_stripe *m =
1077                         genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
1078
1079                 if (!m) {
1080                         bch_err(c, "error allocating memory for gc_stripes, idx %llu",
1081                                 idx);
1082                         return -BCH_ERR_ENOMEM_mark_stripe;
1083                 }
1084                 /*
1085                  * This will be wrong when we bring back runtime gc: we should
1086                  * be unmarking the old key and then marking the new key
1087                  */
1088                 m->alive        = true;
1089                 m->sectors      = le16_to_cpu(new_s->sectors);
1090                 m->nr_blocks    = new_s->nr_blocks;
1091                 m->nr_redundant = new_s->nr_redundant;
1092
1093                 for (i = 0; i < new_s->nr_blocks; i++)
1094                         m->ptrs[i] = new_s->ptrs[i];
1095
1096                 bch2_bkey_to_replicas(&m->r.e, new);
1097
1098                 /*
1099                  * gc recalculates this field from stripe ptr
1100                  * references:
1101                  */
1102                 memset(m->block_sectors, 0, sizeof(m->block_sectors));
1103
1104                 for (i = 0; i < new_s->nr_blocks; i++) {
1105                         ret = mark_stripe_bucket(trans, new, i, flags);
1106                         if (ret)
1107                                 return ret;
1108                 }
1109
1110                 ret = update_replicas(c, new, &m->r.e,
1111                                       ((s64) m->sectors * m->nr_redundant),
1112                                       journal_seq, gc);
1113                 if (ret) {
1114                         struct printbuf buf = PRINTBUF;
1115
1116                         bch2_bkey_val_to_text(&buf, c, new);
1117                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
1118                         printbuf_exit(&buf);
1119                         return ret;
1120                 }
1121         }
1122
1123         return 0;
1124 }
1125
1126 int bch2_mark_reservation(struct btree_trans *trans,
1127                           enum btree_id btree_id, unsigned level,
1128                           struct bkey_s_c old, struct bkey_s_c new,
1129                           unsigned flags)
1130 {
1131         struct bch_fs *c = trans->c;
1132         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old : new;
1133         struct bch_fs_usage *fs_usage;
1134         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1135         s64 sectors = (s64) k.k->size;
1136
1137         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1138
1139         if (flags & BTREE_TRIGGER_OVERWRITE)
1140                 sectors = -sectors;
1141         sectors *= replicas;
1142
1143         percpu_down_read(&c->mark_lock);
1144         preempt_disable();
1145
1146         fs_usage = fs_usage_ptr(c, trans->journal_res.seq, flags & BTREE_TRIGGER_GC);
1147         replicas = clamp_t(unsigned, replicas, 1,
1148                            ARRAY_SIZE(fs_usage->persistent_reserved));
1149
1150         fs_usage->reserved                              += sectors;
1151         fs_usage->persistent_reserved[replicas - 1]     += sectors;
1152
1153         preempt_enable();
1154         percpu_up_read(&c->mark_lock);
1155
1156         return 0;
1157 }
1158
1159 static s64 __bch2_mark_reflink_p(struct btree_trans *trans,
1160                                  struct bkey_s_c_reflink_p p,
1161                                  u64 start, u64 end,
1162                                  u64 *idx, unsigned flags, size_t r_idx)
1163 {
1164         struct bch_fs *c = trans->c;
1165         struct reflink_gc *r;
1166         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1167         u64 next_idx = end;
1168         s64 ret = 0;
1169         struct printbuf buf = PRINTBUF;
1170
1171         if (r_idx >= c->reflink_gc_nr)
1172                 goto not_found;
1173
1174         r = genradix_ptr(&c->reflink_gc_table, r_idx);
1175         next_idx = min(next_idx, r->offset - r->size);
1176         if (*idx < next_idx)
1177                 goto not_found;
1178
1179         BUG_ON((s64) r->refcount + add < 0);
1180
1181         r->refcount += add;
1182         *idx = r->offset;
1183         return 0;
1184 not_found:
1185         if (fsck_err(c, "pointer to missing indirect extent\n"
1186                      "  %s\n"
1187                      "  missing range %llu-%llu",
1188                      (bch2_bkey_val_to_text(&buf, c, p.s_c), buf.buf),
1189                      *idx, next_idx)) {
1190                 struct bkey_i_error *new;
1191
1192                 new = bch2_trans_kmalloc(trans, sizeof(*new));
1193                 ret = PTR_ERR_OR_ZERO(new);
1194                 if (ret)
1195                         goto err;
1196
1197                 bkey_init(&new->k);
1198                 new->k.type     = KEY_TYPE_error;
1199                 new->k.p                = bkey_start_pos(p.k);
1200                 new->k.p.offset += *idx - start;
1201                 bch2_key_resize(&new->k, next_idx - *idx);
1202                 ret = bch2_btree_insert_trans(trans, BTREE_ID_extents, &new->k_i,
1203                                           BTREE_TRIGGER_NORUN);
1204         }
1205
1206         *idx = next_idx;
1207 err:
1208 fsck_err:
1209         printbuf_exit(&buf);
1210         return ret;
1211 }
1212
1213 int bch2_mark_reflink_p(struct btree_trans *trans,
1214                         enum btree_id btree_id, unsigned level,
1215                         struct bkey_s_c old, struct bkey_s_c new,
1216                         unsigned flags)
1217 {
1218         struct bch_fs *c = trans->c;
1219         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old : new;
1220         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1221         struct reflink_gc *ref;
1222         size_t l, r, m;
1223         u64 idx = le64_to_cpu(p.v->idx), start = idx;
1224         u64 end = le64_to_cpu(p.v->idx) + p.k->size;
1225         int ret = 0;
1226
1227         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1228
1229         if (c->sb.version_upgrade_complete >= bcachefs_metadata_version_reflink_p_fix) {
1230                 idx -= le32_to_cpu(p.v->front_pad);
1231                 end += le32_to_cpu(p.v->back_pad);
1232         }
1233
1234         l = 0;
1235         r = c->reflink_gc_nr;
1236         while (l < r) {
1237                 m = l + (r - l) / 2;
1238
1239                 ref = genradix_ptr(&c->reflink_gc_table, m);
1240                 if (ref->offset <= idx)
1241                         l = m + 1;
1242                 else
1243                         r = m;
1244         }
1245
1246         while (idx < end && !ret)
1247                 ret = __bch2_mark_reflink_p(trans, p, start, end,
1248                                             &idx, flags, l++);
1249
1250         return ret;
1251 }
1252
1253 void bch2_trans_fs_usage_revert(struct btree_trans *trans,
1254                                 struct replicas_delta_list *deltas)
1255 {
1256         struct bch_fs *c = trans->c;
1257         struct bch_fs_usage *dst;
1258         struct replicas_delta *d, *top = (void *) deltas->d + deltas->used;
1259         s64 added = 0;
1260         unsigned i;
1261
1262         percpu_down_read(&c->mark_lock);
1263         preempt_disable();
1264         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1265
1266         /* revert changes: */
1267         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1268                 switch (d->r.data_type) {
1269                 case BCH_DATA_btree:
1270                 case BCH_DATA_user:
1271                 case BCH_DATA_parity:
1272                         added += d->delta;
1273                 }
1274                 BUG_ON(__update_replicas(c, dst, &d->r, -d->delta));
1275         }
1276
1277         dst->nr_inodes -= deltas->nr_inodes;
1278
1279         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1280                 added                           -= deltas->persistent_reserved[i];
1281                 dst->reserved                   -= deltas->persistent_reserved[i];
1282                 dst->persistent_reserved[i]     -= deltas->persistent_reserved[i];
1283         }
1284
1285         if (added > 0) {
1286                 trans->disk_res->sectors += added;
1287                 this_cpu_add(*c->online_reserved, added);
1288         }
1289
1290         preempt_enable();
1291         percpu_up_read(&c->mark_lock);
1292 }
1293
1294 int bch2_trans_fs_usage_apply(struct btree_trans *trans,
1295                               struct replicas_delta_list *deltas)
1296 {
1297         struct bch_fs *c = trans->c;
1298         static int warned_disk_usage = 0;
1299         bool warn = false;
1300         unsigned disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1301         struct replicas_delta *d, *d2;
1302         struct replicas_delta *top = (void *) deltas->d + deltas->used;
1303         struct bch_fs_usage *dst;
1304         s64 added = 0, should_not_have_added;
1305         unsigned i;
1306
1307         percpu_down_read(&c->mark_lock);
1308         preempt_disable();
1309         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1310
1311         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1312                 switch (d->r.data_type) {
1313                 case BCH_DATA_btree:
1314                 case BCH_DATA_user:
1315                 case BCH_DATA_parity:
1316                         added += d->delta;
1317                 }
1318
1319                 if (__update_replicas(c, dst, &d->r, d->delta))
1320                         goto need_mark;
1321         }
1322
1323         dst->nr_inodes += deltas->nr_inodes;
1324
1325         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1326                 added                           += deltas->persistent_reserved[i];
1327                 dst->reserved                   += deltas->persistent_reserved[i];
1328                 dst->persistent_reserved[i]     += deltas->persistent_reserved[i];
1329         }
1330
1331         /*
1332          * Not allowed to reduce sectors_available except by getting a
1333          * reservation:
1334          */
1335         should_not_have_added = added - (s64) disk_res_sectors;
1336         if (unlikely(should_not_have_added > 0)) {
1337                 u64 old, new, v = atomic64_read(&c->sectors_available);
1338
1339                 do {
1340                         old = v;
1341                         new = max_t(s64, 0, old - should_not_have_added);
1342                 } while ((v = atomic64_cmpxchg(&c->sectors_available,
1343                                                old, new)) != old);
1344
1345                 added -= should_not_have_added;
1346                 warn = true;
1347         }
1348
1349         if (added > 0) {
1350                 trans->disk_res->sectors -= added;
1351                 this_cpu_sub(*c->online_reserved, added);
1352         }
1353
1354         preempt_enable();
1355         percpu_up_read(&c->mark_lock);
1356
1357         if (unlikely(warn) && !xchg(&warned_disk_usage, 1))
1358                 bch2_trans_inconsistent(trans,
1359                                         "disk usage increased %lli more than %u sectors reserved)",
1360                                         should_not_have_added, disk_res_sectors);
1361         return 0;
1362 need_mark:
1363         /* revert changes: */
1364         for (d2 = deltas->d; d2 != d; d2 = replicas_delta_next(d2))
1365                 BUG_ON(__update_replicas(c, dst, &d2->r, -d2->delta));
1366
1367         preempt_enable();
1368         percpu_up_read(&c->mark_lock);
1369         return -1;
1370 }
1371
1372 /* trans_mark: */
1373
1374 static inline int bch2_trans_mark_pointer(struct btree_trans *trans,
1375                                    enum btree_id btree_id, unsigned level,
1376                                    struct bkey_s_c k, struct extent_ptr_decoded p,
1377                                    unsigned flags)
1378 {
1379         bool insert = !(flags & BTREE_TRIGGER_OVERWRITE);
1380         struct btree_iter iter;
1381         struct bkey_i_alloc_v4 *a;
1382         struct bpos bucket;
1383         struct bch_backpointer bp;
1384         s64 sectors;
1385         int ret;
1386
1387         bch2_extent_ptr_to_bp(trans->c, btree_id, level, k, p, &bucket, &bp);
1388         sectors = bp.bucket_len;
1389         if (!insert)
1390                 sectors = -sectors;
1391
1392         a = bch2_trans_start_alloc_update(trans, &iter, bucket);
1393         if (IS_ERR(a))
1394                 return PTR_ERR(a);
1395
1396         ret = __mark_pointer(trans, k, &p.ptr, sectors, bp.data_type,
1397                              a->v.gen, &a->v.data_type,
1398                              &a->v.dirty_sectors, &a->v.cached_sectors) ?:
1399                 bch2_trans_update(trans, &iter, &a->k_i, 0);
1400         bch2_trans_iter_exit(trans, &iter);
1401
1402         if (ret)
1403                 return ret;
1404
1405         if (!p.ptr.cached) {
1406                 ret = bch2_bucket_backpointer_mod(trans, bucket, bp, k, insert);
1407                 if (ret)
1408                         return ret;
1409         }
1410
1411         return 0;
1412 }
1413
1414 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1415                         struct extent_ptr_decoded p,
1416                         s64 sectors, enum bch_data_type data_type)
1417 {
1418         struct btree_iter iter;
1419         struct bkey_i_stripe *s;
1420         struct bch_replicas_padded r;
1421         int ret = 0;
1422
1423         s = bch2_bkey_get_mut_typed(trans, &iter,
1424                         BTREE_ID_stripes, POS(0, p.ec.idx),
1425                         BTREE_ITER_WITH_UPDATES, stripe);
1426         ret = PTR_ERR_OR_ZERO(s);
1427         if (unlikely(ret)) {
1428                 bch2_trans_inconsistent_on(bch2_err_matches(ret, ENOENT), trans,
1429                         "pointer to nonexistent stripe %llu",
1430                         (u64) p.ec.idx);
1431                 goto err;
1432         }
1433
1434         if (!bch2_ptr_matches_stripe(&s->v, p)) {
1435                 bch2_trans_inconsistent(trans,
1436                         "stripe pointer doesn't match stripe %llu",
1437                         (u64) p.ec.idx);
1438                 ret = -EIO;
1439                 goto err;
1440         }
1441
1442         stripe_blockcount_set(&s->v, p.ec.block,
1443                 stripe_blockcount_get(&s->v, p.ec.block) +
1444                 sectors);
1445
1446         bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1447         r.e.data_type = data_type;
1448         ret = update_replicas_list(trans, &r.e, sectors);
1449 err:
1450         bch2_trans_iter_exit(trans, &iter);
1451         return ret;
1452 }
1453
1454 int bch2_trans_mark_extent(struct btree_trans *trans,
1455                            enum btree_id btree_id, unsigned level,
1456                            struct bkey_s_c old, struct bkey_i *new,
1457                            unsigned flags)
1458 {
1459         struct bch_fs *c = trans->c;
1460         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1461                 ? old
1462                 : bkey_i_to_s_c(new);
1463         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1464         const union bch_extent_entry *entry;
1465         struct extent_ptr_decoded p;
1466         struct bch_replicas_padded r;
1467         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
1468                 ? BCH_DATA_btree
1469                 : BCH_DATA_user;
1470         s64 sectors = bkey_is_btree_ptr(k.k)
1471                 ? btree_sectors(c)
1472                 : k.k->size;
1473         s64 dirty_sectors = 0;
1474         bool stale;
1475         int ret = 0;
1476
1477         r.e.data_type   = data_type;
1478         r.e.nr_devs     = 0;
1479         r.e.nr_required = 1;
1480
1481         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1482                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
1483
1484                 if (flags & BTREE_TRIGGER_OVERWRITE)
1485                         disk_sectors = -disk_sectors;
1486
1487                 ret = bch2_trans_mark_pointer(trans, btree_id, level, k, p, flags);
1488                 if (ret < 0)
1489                         return ret;
1490
1491                 stale = ret > 0;
1492
1493                 if (p.ptr.cached) {
1494                         if (!stale) {
1495                                 ret = update_cached_sectors_list(trans, p.ptr.dev,
1496                                                                  disk_sectors);
1497                                 if (ret)
1498                                         return ret;
1499                         }
1500                 } else if (!p.has_ec) {
1501                         dirty_sectors          += disk_sectors;
1502                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1503                 } else {
1504                         ret = bch2_trans_mark_stripe_ptr(trans, p,
1505                                         disk_sectors, data_type);
1506                         if (ret)
1507                                 return ret;
1508
1509                         r.e.nr_required = 0;
1510                 }
1511         }
1512
1513         if (r.e.nr_devs)
1514                 ret = update_replicas_list(trans, &r.e, dirty_sectors);
1515
1516         return ret;
1517 }
1518
1519 static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
1520                                          struct bkey_s_c_stripe s,
1521                                          unsigned idx, bool deleting)
1522 {
1523         struct bch_fs *c = trans->c;
1524         const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
1525         struct btree_iter iter;
1526         struct bkey_i_alloc_v4 *a;
1527         enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
1528                 ? BCH_DATA_parity : 0;
1529         s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
1530         int ret = 0;
1531
1532         if (deleting)
1533                 sectors = -sectors;
1534
1535         a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(c, ptr));
1536         if (IS_ERR(a))
1537                 return PTR_ERR(a);
1538
1539         ret = check_bucket_ref(trans, s.s_c, ptr, sectors, data_type,
1540                                a->v.gen, a->v.data_type,
1541                                a->v.dirty_sectors, a->v.cached_sectors);
1542         if (ret)
1543                 goto err;
1544
1545         if (!deleting) {
1546                 if (bch2_trans_inconsistent_on(a->v.stripe ||
1547                                                a->v.stripe_redundancy, trans,
1548                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
1549                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1550                                 bch2_data_types[a->v.data_type],
1551                                 a->v.dirty_sectors,
1552                                 a->v.stripe, s.k->p.offset)) {
1553                         ret = -EIO;
1554                         goto err;
1555                 }
1556
1557                 if (bch2_trans_inconsistent_on(data_type && a->v.dirty_sectors, trans,
1558                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
1559                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1560                                 bch2_data_types[a->v.data_type],
1561                                 a->v.dirty_sectors,
1562                                 s.k->p.offset)) {
1563                         ret = -EIO;
1564                         goto err;
1565                 }
1566
1567                 a->v.stripe             = s.k->p.offset;
1568                 a->v.stripe_redundancy  = s.v->nr_redundant;
1569                 a->v.data_type          = BCH_DATA_stripe;
1570         } else {
1571                 if (bch2_trans_inconsistent_on(a->v.stripe != s.k->p.offset ||
1572                                                a->v.stripe_redundancy != s.v->nr_redundant, trans,
1573                                 "bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
1574                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1575                                 s.k->p.offset, a->v.stripe)) {
1576                         ret = -EIO;
1577                         goto err;
1578                 }
1579
1580                 a->v.stripe             = 0;
1581                 a->v.stripe_redundancy  = 0;
1582                 a->v.data_type          = alloc_data_type(a->v, BCH_DATA_user);
1583         }
1584
1585         a->v.dirty_sectors += sectors;
1586         if (data_type)
1587                 a->v.data_type = !deleting ? data_type : 0;
1588
1589         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1590         if (ret)
1591                 goto err;
1592 err:
1593         bch2_trans_iter_exit(trans, &iter);
1594         return ret;
1595 }
1596
1597 int bch2_trans_mark_stripe(struct btree_trans *trans,
1598                            enum btree_id btree_id, unsigned level,
1599                            struct bkey_s_c old, struct bkey_i *new,
1600                            unsigned flags)
1601 {
1602         const struct bch_stripe *old_s = NULL;
1603         struct bch_stripe *new_s = NULL;
1604         struct bch_replicas_padded r;
1605         unsigned i, nr_blocks;
1606         int ret = 0;
1607
1608         if (old.k->type == KEY_TYPE_stripe)
1609                 old_s = bkey_s_c_to_stripe(old).v;
1610         if (new->k.type == KEY_TYPE_stripe)
1611                 new_s = &bkey_i_to_stripe(new)->v;
1612
1613         /*
1614          * If the pointers aren't changing, we don't need to do anything:
1615          */
1616         if (new_s && old_s &&
1617             new_s->nr_blocks    == old_s->nr_blocks &&
1618             new_s->nr_redundant == old_s->nr_redundant &&
1619             !memcmp(old_s->ptrs, new_s->ptrs,
1620                     new_s->nr_blocks * sizeof(struct bch_extent_ptr)))
1621                 return 0;
1622
1623         BUG_ON(new_s && old_s &&
1624                (new_s->nr_blocks        != old_s->nr_blocks ||
1625                 new_s->nr_redundant     != old_s->nr_redundant));
1626
1627         nr_blocks = new_s ? new_s->nr_blocks : old_s->nr_blocks;
1628
1629         if (new_s) {
1630                 s64 sectors = le16_to_cpu(new_s->sectors);
1631
1632                 bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(new));
1633                 ret = update_replicas_list(trans, &r.e, sectors * new_s->nr_redundant);
1634                 if (ret)
1635                         return ret;
1636         }
1637
1638         if (old_s) {
1639                 s64 sectors = -((s64) le16_to_cpu(old_s->sectors));
1640
1641                 bch2_bkey_to_replicas(&r.e, old);
1642                 ret = update_replicas_list(trans, &r.e, sectors * old_s->nr_redundant);
1643                 if (ret)
1644                         return ret;
1645         }
1646
1647         for (i = 0; i < nr_blocks; i++) {
1648                 if (new_s && old_s &&
1649                     !memcmp(&new_s->ptrs[i],
1650                             &old_s->ptrs[i],
1651                             sizeof(new_s->ptrs[i])))
1652                         continue;
1653
1654                 if (new_s) {
1655                         ret = bch2_trans_mark_stripe_bucket(trans,
1656                                         bkey_i_to_s_c_stripe(new), i, false);
1657                         if (ret)
1658                                 break;
1659                 }
1660
1661                 if (old_s) {
1662                         ret = bch2_trans_mark_stripe_bucket(trans,
1663                                         bkey_s_c_to_stripe(old), i, true);
1664                         if (ret)
1665                                 break;
1666                 }
1667         }
1668
1669         return ret;
1670 }
1671
1672 int bch2_trans_mark_reservation(struct btree_trans *trans,
1673                                 enum btree_id btree_id, unsigned level,
1674                                 struct bkey_s_c old,
1675                                 struct bkey_i *new,
1676                                 unsigned flags)
1677 {
1678         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1679                 ? old
1680                 : bkey_i_to_s_c(new);
1681         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1682         s64 sectors = (s64) k.k->size;
1683         struct replicas_delta_list *d;
1684         int ret;
1685
1686         if (flags & BTREE_TRIGGER_OVERWRITE)
1687                 sectors = -sectors;
1688         sectors *= replicas;
1689
1690         ret = bch2_replicas_deltas_realloc(trans, 0);
1691         if (ret)
1692                 return ret;
1693
1694         d = trans->fs_usage_deltas;
1695         replicas = clamp_t(unsigned, replicas, 1,
1696                            ARRAY_SIZE(d->persistent_reserved));
1697
1698         d->persistent_reserved[replicas - 1] += sectors;
1699         return 0;
1700 }
1701
1702 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1703                         struct bkey_s_c_reflink_p p,
1704                         u64 *idx, unsigned flags)
1705 {
1706         struct bch_fs *c = trans->c;
1707         struct btree_iter iter;
1708         struct bkey_i *k;
1709         __le64 *refcount;
1710         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1711         struct printbuf buf = PRINTBUF;
1712         int ret;
1713
1714         k = bch2_bkey_get_mut_noupdate(trans, &iter,
1715                         BTREE_ID_reflink, POS(0, *idx),
1716                         BTREE_ITER_WITH_UPDATES);
1717         ret = PTR_ERR_OR_ZERO(k);
1718         if (ret)
1719                 goto err;
1720
1721         refcount = bkey_refcount(k);
1722         if (!refcount) {
1723                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1724                 bch2_trans_inconsistent(trans,
1725                         "nonexistent indirect extent at %llu while marking\n  %s",
1726                         *idx, buf.buf);
1727                 ret = -EIO;
1728                 goto err;
1729         }
1730
1731         if (!*refcount && (flags & BTREE_TRIGGER_OVERWRITE)) {
1732                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1733                 bch2_trans_inconsistent(trans,
1734                         "indirect extent refcount underflow at %llu while marking\n  %s",
1735                         *idx, buf.buf);
1736                 ret = -EIO;
1737                 goto err;
1738         }
1739
1740         if (flags & BTREE_TRIGGER_INSERT) {
1741                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1742                 u64 pad;
1743
1744                 pad = max_t(s64, le32_to_cpu(v->front_pad),
1745                             le64_to_cpu(v->idx) - bkey_start_offset(&k->k));
1746                 BUG_ON(pad > U32_MAX);
1747                 v->front_pad = cpu_to_le32(pad);
1748
1749                 pad = max_t(s64, le32_to_cpu(v->back_pad),
1750                             k->k.p.offset - p.k->size - le64_to_cpu(v->idx));
1751                 BUG_ON(pad > U32_MAX);
1752                 v->back_pad = cpu_to_le32(pad);
1753         }
1754
1755         le64_add_cpu(refcount, add);
1756
1757         bch2_btree_iter_set_pos_to_extent_start(&iter);
1758         ret = bch2_trans_update(trans, &iter, k, 0);
1759         if (ret)
1760                 goto err;
1761
1762         *idx = k->k.p.offset;
1763 err:
1764         bch2_trans_iter_exit(trans, &iter);
1765         printbuf_exit(&buf);
1766         return ret;
1767 }
1768
1769 int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1770                               enum btree_id btree_id, unsigned level,
1771                               struct bkey_s_c old,
1772                               struct bkey_i *new,
1773                               unsigned flags)
1774 {
1775         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1776                 ? old
1777                 : bkey_i_to_s_c(new);
1778         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1779         u64 idx, end_idx;
1780         int ret = 0;
1781
1782         if (flags & BTREE_TRIGGER_INSERT) {
1783                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1784
1785                 v->front_pad = v->back_pad = 0;
1786         }
1787
1788         idx     = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
1789         end_idx = le64_to_cpu(p.v->idx) + p.k->size +
1790                 le32_to_cpu(p.v->back_pad);
1791
1792         while (idx < end_idx && !ret)
1793                 ret = __bch2_trans_mark_reflink_p(trans, p, &idx, flags);
1794
1795         return ret;
1796 }
1797
1798 static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1799                                     struct bch_dev *ca, size_t b,
1800                                     enum bch_data_type type,
1801                                     unsigned sectors)
1802 {
1803         struct bch_fs *c = trans->c;
1804         struct btree_iter iter;
1805         struct bkey_i_alloc_v4 *a;
1806         int ret = 0;
1807
1808         /*
1809          * Backup superblock might be past the end of our normal usable space:
1810          */
1811         if (b >= ca->mi.nbuckets)
1812                 return 0;
1813
1814         a = bch2_trans_start_alloc_update(trans, &iter, POS(ca->dev_idx, b));
1815         if (IS_ERR(a))
1816                 return PTR_ERR(a);
1817
1818         if (a->v.data_type && type && a->v.data_type != type) {
1819                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1820                         "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
1821                         "while marking %s",
1822                         iter.pos.inode, iter.pos.offset, a->v.gen,
1823                         bch2_data_types[a->v.data_type],
1824                         bch2_data_types[type],
1825                         bch2_data_types[type]);
1826                 ret = -EIO;
1827                 goto out;
1828         }
1829
1830         a->v.data_type          = type;
1831         a->v.dirty_sectors      = sectors;
1832
1833         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1834         if (ret)
1835                 goto out;
1836 out:
1837         bch2_trans_iter_exit(trans, &iter);
1838         return ret;
1839 }
1840
1841 int bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1842                                     struct bch_dev *ca, size_t b,
1843                                     enum bch_data_type type,
1844                                     unsigned sectors)
1845 {
1846         return commit_do(trans, NULL, NULL, 0,
1847                         __bch2_trans_mark_metadata_bucket(trans, ca, b, type, sectors));
1848 }
1849
1850 static int bch2_trans_mark_metadata_sectors(struct btree_trans *trans,
1851                                             struct bch_dev *ca,
1852                                             u64 start, u64 end,
1853                                             enum bch_data_type type,
1854                                             u64 *bucket, unsigned *bucket_sectors)
1855 {
1856         do {
1857                 u64 b = sector_to_bucket(ca, start);
1858                 unsigned sectors =
1859                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1860
1861                 if (b != *bucket && *bucket_sectors) {
1862                         int ret = bch2_trans_mark_metadata_bucket(trans, ca, *bucket,
1863                                                                   type, *bucket_sectors);
1864                         if (ret)
1865                                 return ret;
1866
1867                         *bucket_sectors = 0;
1868                 }
1869
1870                 *bucket         = b;
1871                 *bucket_sectors += sectors;
1872                 start += sectors;
1873         } while (start < end);
1874
1875         return 0;
1876 }
1877
1878 static int __bch2_trans_mark_dev_sb(struct btree_trans *trans,
1879                                     struct bch_dev *ca)
1880 {
1881         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1882         u64 bucket = 0;
1883         unsigned i, bucket_sectors = 0;
1884         int ret;
1885
1886         for (i = 0; i < layout->nr_superblocks; i++) {
1887                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1888
1889                 if (offset == BCH_SB_SECTOR) {
1890                         ret = bch2_trans_mark_metadata_sectors(trans, ca,
1891                                                 0, BCH_SB_SECTOR,
1892                                                 BCH_DATA_sb, &bucket, &bucket_sectors);
1893                         if (ret)
1894                                 return ret;
1895                 }
1896
1897                 ret = bch2_trans_mark_metadata_sectors(trans, ca, offset,
1898                                       offset + (1 << layout->sb_max_size_bits),
1899                                       BCH_DATA_sb, &bucket, &bucket_sectors);
1900                 if (ret)
1901                         return ret;
1902         }
1903
1904         if (bucket_sectors) {
1905                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1906                                 bucket, BCH_DATA_sb, bucket_sectors);
1907                 if (ret)
1908                         return ret;
1909         }
1910
1911         for (i = 0; i < ca->journal.nr; i++) {
1912                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1913                                 ca->journal.buckets[i],
1914                                 BCH_DATA_journal, ca->mi.bucket_size);
1915                 if (ret)
1916                         return ret;
1917         }
1918
1919         return 0;
1920 }
1921
1922 int bch2_trans_mark_dev_sb(struct bch_fs *c, struct bch_dev *ca)
1923 {
1924         int ret = bch2_trans_run(c, __bch2_trans_mark_dev_sb(trans, ca));
1925
1926         if (ret)
1927                 bch_err_fn(c, ret);
1928         return ret;
1929 }
1930
1931 /* Disk reservations: */
1932
1933 #define SECTORS_CACHE   1024
1934
1935 int __bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
1936                               u64 sectors, int flags)
1937 {
1938         struct bch_fs_pcpu *pcpu;
1939         u64 old, v, get;
1940         s64 sectors_available;
1941         int ret;
1942
1943         percpu_down_read(&c->mark_lock);
1944         preempt_disable();
1945         pcpu = this_cpu_ptr(c->pcpu);
1946
1947         if (sectors <= pcpu->sectors_available)
1948                 goto out;
1949
1950         v = atomic64_read(&c->sectors_available);
1951         do {
1952                 old = v;
1953                 get = min((u64) sectors + SECTORS_CACHE, old);
1954
1955                 if (get < sectors) {
1956                         preempt_enable();
1957                         goto recalculate;
1958                 }
1959         } while ((v = atomic64_cmpxchg(&c->sectors_available,
1960                                        old, old - get)) != old);
1961
1962         pcpu->sectors_available         += get;
1963
1964 out:
1965         pcpu->sectors_available         -= sectors;
1966         this_cpu_add(*c->online_reserved, sectors);
1967         res->sectors                    += sectors;
1968
1969         preempt_enable();
1970         percpu_up_read(&c->mark_lock);
1971         return 0;
1972
1973 recalculate:
1974         mutex_lock(&c->sectors_available_lock);
1975
1976         percpu_u64_set(&c->pcpu->sectors_available, 0);
1977         sectors_available = avail_factor(__bch2_fs_usage_read_short(c).free);
1978
1979         if (sectors <= sectors_available ||
1980             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
1981                 atomic64_set(&c->sectors_available,
1982                              max_t(s64, 0, sectors_available - sectors));
1983                 this_cpu_add(*c->online_reserved, sectors);
1984                 res->sectors                    += sectors;
1985                 ret = 0;
1986         } else {
1987                 atomic64_set(&c->sectors_available, sectors_available);
1988                 ret = -BCH_ERR_ENOSPC_disk_reservation;
1989         }
1990
1991         mutex_unlock(&c->sectors_available_lock);
1992         percpu_up_read(&c->mark_lock);
1993
1994         return ret;
1995 }
1996
1997 /* Startup/shutdown: */
1998
1999 static void bucket_gens_free_rcu(struct rcu_head *rcu)
2000 {
2001         struct bucket_gens *buckets =
2002                 container_of(rcu, struct bucket_gens, rcu);
2003
2004         kvpfree(buckets, sizeof(*buckets) + buckets->nbuckets);
2005 }
2006
2007 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
2008 {
2009         struct bucket_gens *bucket_gens = NULL, *old_bucket_gens = NULL;
2010         unsigned long *buckets_nouse = NULL;
2011         bool resize = ca->bucket_gens != NULL;
2012         int ret;
2013
2014         if (!(bucket_gens       = kvpmalloc(sizeof(struct bucket_gens) + nbuckets,
2015                                             GFP_KERNEL|__GFP_ZERO))) {
2016                 ret = -BCH_ERR_ENOMEM_bucket_gens;
2017                 goto err;
2018         }
2019
2020         if ((c->opts.buckets_nouse &&
2021              !(buckets_nouse    = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2022                                             sizeof(unsigned long),
2023                                             GFP_KERNEL|__GFP_ZERO)))) {
2024                 ret = -BCH_ERR_ENOMEM_buckets_nouse;
2025                 goto err;
2026         }
2027
2028         bucket_gens->first_bucket = ca->mi.first_bucket;
2029         bucket_gens->nbuckets   = nbuckets;
2030
2031         bch2_copygc_stop(c);
2032
2033         if (resize) {
2034                 down_write(&c->gc_lock);
2035                 down_write(&ca->bucket_lock);
2036                 percpu_down_write(&c->mark_lock);
2037         }
2038
2039         old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
2040
2041         if (resize) {
2042                 size_t n = min(bucket_gens->nbuckets, old_bucket_gens->nbuckets);
2043
2044                 memcpy(bucket_gens->b,
2045                        old_bucket_gens->b,
2046                        n);
2047                 if (buckets_nouse)
2048                         memcpy(buckets_nouse,
2049                                ca->buckets_nouse,
2050                                BITS_TO_LONGS(n) * sizeof(unsigned long));
2051         }
2052
2053         rcu_assign_pointer(ca->bucket_gens, bucket_gens);
2054         bucket_gens     = old_bucket_gens;
2055
2056         swap(ca->buckets_nouse, buckets_nouse);
2057
2058         nbuckets = ca->mi.nbuckets;
2059
2060         if (resize) {
2061                 percpu_up_write(&c->mark_lock);
2062                 up_write(&ca->bucket_lock);
2063                 up_write(&c->gc_lock);
2064         }
2065
2066         ret = 0;
2067 err:
2068         kvpfree(buckets_nouse,
2069                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2070         if (bucket_gens)
2071                 call_rcu(&bucket_gens->rcu, bucket_gens_free_rcu);
2072
2073         return ret;
2074 }
2075
2076 void bch2_dev_buckets_free(struct bch_dev *ca)
2077 {
2078         unsigned i;
2079
2080         kvpfree(ca->buckets_nouse,
2081                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2082         kvpfree(rcu_dereference_protected(ca->bucket_gens, 1),
2083                 sizeof(struct bucket_gens) + ca->mi.nbuckets);
2084
2085         for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
2086                 free_percpu(ca->usage[i]);
2087         kfree(ca->usage_base);
2088 }
2089
2090 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2091 {
2092         unsigned i;
2093
2094         ca->usage_base = kzalloc(sizeof(struct bch_dev_usage), GFP_KERNEL);
2095         if (!ca->usage_base)
2096                 return -BCH_ERR_ENOMEM_usage_init;
2097
2098         for (i = 0; i < ARRAY_SIZE(ca->usage); i++) {
2099                 ca->usage[i] = alloc_percpu(struct bch_dev_usage);
2100                 if (!ca->usage[i])
2101                         return -BCH_ERR_ENOMEM_usage_init;
2102         }
2103
2104         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);
2105 }