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         buf.atomic++;
 371
 372         idx = bch2_replicas_entry_idx(c, r);
 373         if (idx < 0 &&
 374             fsck_err(c, "no replicas entry\n"
 375                      "  while marking %s",
 376                      (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 377                 percpu_up_read(&c->mark_lock);
 378                 ret = bch2_mark_replicas(c, r);
 379                 percpu_down_read(&c->mark_lock);
 380
 381                 if (ret)
 382                         goto err;
 383                 idx = bch2_replicas_entry_idx(c, r);
 384         }
 385         if (idx < 0) {
 386                 ret = -1;
 387                 goto err;
 388         }
 389
 390         preempt_disable();
 391         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 392         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 393         fs_usage->replicas[idx]         += sectors;
 394         preempt_enable();
 395 err:
 396 fsck_err:
 397         percpu_up_read(&c->mark_lock);
 398         printbuf_exit(&buf);
 399         return ret;
 400 }
 401
 402 static inline int update_cached_sectors(struct bch_fs *c,
 403                         struct bkey_s_c k,
 404                         unsigned dev, s64 sectors,
 405                         unsigned journal_seq, bool gc)
 406 {
 407         struct bch_replicas_padded r;
 408
 409         bch2_replicas_entry_cached(&r.e, dev);
 410
 411         return update_replicas(c, k, &r.e, sectors, journal_seq, gc);
 412 }
 413
 414 static int __replicas_deltas_realloc(struct btree_trans *trans, unsigned more,
 415                                      gfp_t gfp)
 416 {
 417         struct replicas_delta_list *d = trans->fs_usage_deltas;
 418         unsigned new_size = d ? (d->size + more) * 2 : 128;
 419         unsigned alloc_size = sizeof(*d) + new_size;
 420
 421         WARN_ON_ONCE(alloc_size > REPLICAS_DELTA_LIST_MAX);
 422
 423         if (!d || d->used + more > d->size) {
 424                 d = krealloc(d, alloc_size, gfp|__GFP_ZERO);
 425
 426                 if (unlikely(!d)) {
 427                         if (alloc_size > REPLICAS_DELTA_LIST_MAX)
 428                                 return -ENOMEM;
 429
 430                         d = mempool_alloc(&trans->c->replicas_delta_pool, gfp);
 431                         if (!d)
 432                                 return -ENOMEM;
 433
 434                         memset(d, 0, REPLICAS_DELTA_LIST_MAX);
 435
 436                         if (trans->fs_usage_deltas)
 437                                 memcpy(d, trans->fs_usage_deltas,
 438                                        trans->fs_usage_deltas->size + sizeof(*d));
 439
 440                         new_size = REPLICAS_DELTA_LIST_MAX - sizeof(*d);
 441                         kfree(trans->fs_usage_deltas);
 442                 }
 443
 444                 d->size = new_size;
 445                 trans->fs_usage_deltas = d;
 446         }
 447
 448         return 0;
 449 }
 450
 451 int bch2_replicas_deltas_realloc(struct btree_trans *trans, unsigned more)
 452 {
 453         return allocate_dropping_locks_errcode(trans,
 454                                 __replicas_deltas_realloc(trans, more, _gfp));
 455 }
 456
 457 static inline int update_replicas_list(struct btree_trans *trans,
 458                                         struct bch_replicas_entry *r,
 459                                         s64 sectors)
 460 {
 461         struct replicas_delta_list *d;
 462         struct replicas_delta *n;
 463         unsigned b;
 464         int ret;
 465
 466         if (!sectors)
 467                 return 0;
 468
 469         b = replicas_entry_bytes(r) + 8;
 470         ret = bch2_replicas_deltas_realloc(trans, b);
 471         if (ret)
 472                 return ret;
 473
 474         d = trans->fs_usage_deltas;
 475         n = (void *) d->d + d->used;
 476         n->delta = sectors;
 477         memcpy((void *) n + offsetof(struct replicas_delta, r),
 478                r, replicas_entry_bytes(r));
 479         bch2_replicas_entry_sort(&n->r);
 480         d->used += b;
 481         return 0;
 482 }
 483
 484 static inline int update_cached_sectors_list(struct btree_trans *trans,
 485                                               unsigned dev, s64 sectors)
 486 {
 487         struct bch_replicas_padded r;
 488
 489         bch2_replicas_entry_cached(&r.e, dev);
 490
 491         return update_replicas_list(trans, &r.e, sectors);
 492 }
 493
 494 int bch2_mark_alloc(struct btree_trans *trans,
 495                     enum btree_id btree, unsigned level,
 496                     struct bkey_s_c old, struct bkey_s_c new,
 497                     unsigned flags)
 498 {
 499         bool gc = flags & BTREE_TRIGGER_GC;
 500         u64 journal_seq = trans->journal_res.seq;
 501         u64 bucket_journal_seq;
 502         struct bch_fs *c = trans->c;
 503         struct bch_alloc_v4 old_a_convert, new_a_convert;
 504         const struct bch_alloc_v4 *old_a, *new_a;
 505         struct bch_dev *ca;
 506         int ret = 0;
 507
 508         /*
 509          * alloc btree is read in by bch2_alloc_read, not gc:
 510          */
 511         if ((flags & BTREE_TRIGGER_GC) &&
 512             !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
 513                 return 0;
 514
 515         if (bch2_trans_inconsistent_on(!bch2_dev_bucket_exists(c, new.k->p), trans,
 516                                        "alloc key for invalid device or bucket"))
 517                 return -EIO;
 518
 519         ca = bch_dev_bkey_exists(c, new.k->p.inode);
 520
 521         old_a = bch2_alloc_to_v4(old, &old_a_convert);
 522         new_a = bch2_alloc_to_v4(new, &new_a_convert);
 523
 524         bucket_journal_seq = new_a->journal_seq;
 525
 526         if ((flags & BTREE_TRIGGER_INSERT) &&
 527             data_type_is_empty(old_a->data_type) !=
 528             data_type_is_empty(new_a->data_type) &&
 529             new.k->type == KEY_TYPE_alloc_v4) {
 530                 struct bch_alloc_v4 *v = (struct bch_alloc_v4 *) new.v;
 531
 532                 EBUG_ON(!journal_seq);
 533
 534                 /*
 535                  * If the btree updates referring to a bucket weren't flushed
 536                  * before the bucket became empty again, then the we don't have
 537                  * to wait on a journal flush before we can reuse the bucket:
 538                  */
 539                 v->journal_seq = bucket_journal_seq =
 540                         data_type_is_empty(new_a->data_type) &&
 541                         (journal_seq == v->journal_seq ||
 542                          bch2_journal_noflush_seq(&c->journal, v->journal_seq))
 543                         ? 0 : journal_seq;
 544         }
 545
 546         if (!data_type_is_empty(old_a->data_type) &&
 547             data_type_is_empty(new_a->data_type) &&
 548             bucket_journal_seq) {
 549                 ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 550                                 c->journal.flushed_seq_ondisk,
 551                                 new.k->p.inode, new.k->p.offset,
 552                                 bucket_journal_seq);
 553                 if (ret) {
 554                         bch2_fs_fatal_error(c,
 555                                 "error setting bucket_needs_journal_commit: %i", ret);
 556                         return ret;
 557                 }
 558         }
 559
 560         percpu_down_read(&c->mark_lock);
 561         if (!gc && new_a->gen != old_a->gen)
 562                 *bucket_gen(ca, new.k->p.offset) = new_a->gen;
 563
 564         bch2_dev_usage_update(c, ca, *old_a, *new_a, journal_seq, gc);
 565
 566         if (gc) {
 567                 struct bucket *g = gc_bucket(ca, new.k->p.offset);
 568
 569                 bucket_lock(g);
 570
 571                 g->gen_valid            = 1;
 572                 g->gen                  = new_a->gen;
 573                 g->data_type            = new_a->data_type;
 574                 g->stripe               = new_a->stripe;
 575                 g->stripe_redundancy    = new_a->stripe_redundancy;
 576                 g->dirty_sectors        = new_a->dirty_sectors;
 577                 g->cached_sectors       = new_a->cached_sectors;
 578
 579                 bucket_unlock(g);
 580         }
 581         percpu_up_read(&c->mark_lock);
 582
 583         /*
 584          * need to know if we're getting called from the invalidate path or
 585          * not:
 586          */
 587
 588         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 589             old_a->cached_sectors) {
 590                 ret = update_cached_sectors(c, new, ca->dev_idx,
 591                                             -((s64) old_a->cached_sectors),
 592                                             journal_seq, gc);
 593                 if (ret) {
 594                         bch2_fs_fatal_error(c, "%s(): no replicas entry while updating cached sectors",
 595                                             __func__);
 596                         return ret;
 597                 }
 598         }
 599
 600         if (new_a->data_type == BCH_DATA_free &&
 601             (!new_a->journal_seq || new_a->journal_seq < c->journal.flushed_seq_ondisk))
 602                 closure_wake_up(&c->freelist_wait);
 603
 604         if (new_a->data_type == BCH_DATA_need_discard &&
 605             (!bucket_journal_seq || bucket_journal_seq < c->journal.flushed_seq_ondisk))
 606                 bch2_do_discards(c);
 607
 608         if (old_a->data_type != BCH_DATA_cached &&
 609             new_a->data_type == BCH_DATA_cached &&
 610             should_invalidate_buckets(ca, bch2_dev_usage_read(ca)))
 611                 bch2_do_invalidates(c);
 612
 613         if (new_a->data_type == BCH_DATA_need_gc_gens)
 614                 bch2_do_gc_gens(c);
 615
 616         return 0;
 617 }
 618
 619 int bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 620                               size_t b, enum bch_data_type data_type,
 621                               unsigned sectors, struct gc_pos pos,
 622                               unsigned flags)
 623 {
 624         struct bucket old, new, *g;
 625         int ret = 0;
 626
 627         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 628         BUG_ON(data_type != BCH_DATA_sb &&
 629                data_type != BCH_DATA_journal);
 630
 631         /*
 632          * Backup superblock might be past the end of our normal usable space:
 633          */
 634         if (b >= ca->mi.nbuckets)
 635                 return 0;
 636
 637         percpu_down_read(&c->mark_lock);
 638         g = gc_bucket(ca, b);
 639
 640         bucket_lock(g);
 641         old = *g;
 642
 643         if (bch2_fs_inconsistent_on(g->data_type &&
 644                         g->data_type != data_type, c,
 645                         "different types of data in same bucket: %s, %s",
 646                         bch2_data_types[g->data_type],
 647                         bch2_data_types[data_type])) {
 648                 ret = -EIO;
 649                 goto err;
 650         }
 651
 652         if (bch2_fs_inconsistent_on((u64) g->dirty_sectors + sectors > ca->mi.bucket_size, c,
 653                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > bucket size",
 654                         ca->dev_idx, b, g->gen,
 655                         bch2_data_types[g->data_type ?: data_type],
 656                         g->dirty_sectors, sectors)) {
 657                 ret = -EIO;
 658                 goto err;
 659         }
 660
 661
 662         g->data_type = data_type;
 663         g->dirty_sectors += sectors;
 664         new = *g;
 665 err:
 666         bucket_unlock(g);
 667         if (!ret)
 668                 bch2_dev_usage_update_m(c, ca, old, new, 0, true);
 669         percpu_up_read(&c->mark_lock);
 670         return ret;
 671 }
 672
 673 static int check_bucket_ref(struct btree_trans *trans,
 674                             struct bkey_s_c k,
 675                             const struct bch_extent_ptr *ptr,
 676                             s64 sectors, enum bch_data_type ptr_data_type,
 677                             u8 b_gen, u8 bucket_data_type,
 678                             u32 dirty_sectors, u32 cached_sectors)
 679 {
 680         struct bch_fs *c = trans->c;
 681         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 682         size_t bucket_nr = PTR_BUCKET_NR(ca, ptr);
 683         u16 bucket_sectors = !ptr->cached
 684                 ? dirty_sectors
 685                 : cached_sectors;
 686         struct printbuf buf = PRINTBUF;
 687         int ret = 0;
 688
 689         if (bucket_data_type == BCH_DATA_cached)
 690                 bucket_data_type = BCH_DATA_user;
 691
 692         if ((bucket_data_type == BCH_DATA_stripe && ptr_data_type == BCH_DATA_user) ||
 693             (bucket_data_type == BCH_DATA_user   && ptr_data_type == BCH_DATA_stripe))
 694                 bucket_data_type = ptr_data_type = BCH_DATA_stripe;
 695
 696         if (gen_after(ptr->gen, b_gen)) {
 697                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 698                         "bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\n"
 699                         "while marking %s",
 700                         ptr->dev, bucket_nr, b_gen,
 701                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 702                         ptr->gen,
 703                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 704                 ret = -EIO;
 705                 goto err;
 706         }
 707
 708         if (gen_cmp(b_gen, ptr->gen) > BUCKET_GC_GEN_MAX) {
 709                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 710                         "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 711                         "while marking %s",
 712                         ptr->dev, bucket_nr, b_gen,
 713                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 714                         ptr->gen,
 715                         (printbuf_reset(&buf),
 716                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 717                 ret = -EIO;
 718                 goto err;
 719         }
 720
 721         if (b_gen != ptr->gen && !ptr->cached) {
 722                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 723                         "bucket %u:%zu gen %u (mem gen %u) data type %s: stale dirty ptr (gen %u)\n"
 724                         "while marking %s",
 725                         ptr->dev, bucket_nr, b_gen,
 726                         *bucket_gen(ca, bucket_nr),
 727                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 728                         ptr->gen,
 729                         (printbuf_reset(&buf),
 730                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 731                 ret = -EIO;
 732                 goto err;
 733         }
 734
 735         if (b_gen != ptr->gen) {
 736                 ret = 1;
 737                 goto out;
 738         }
 739
 740         if (!data_type_is_empty(bucket_data_type) &&
 741             ptr_data_type &&
 742             bucket_data_type != ptr_data_type) {
 743                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 744                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 745                         "while marking %s",
 746                         ptr->dev, bucket_nr, b_gen,
 747                         bch2_data_types[bucket_data_type],
 748                         bch2_data_types[ptr_data_type],
 749                         (printbuf_reset(&buf),
 750                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 751                 ret = -EIO;
 752                 goto err;
 753         }
 754
 755         if ((unsigned) (bucket_sectors + sectors) > U32_MAX) {
 756                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 757                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U16_MAX\n"
 758                         "while marking %s",
 759                         ptr->dev, bucket_nr, b_gen,
 760                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 761                         bucket_sectors, sectors,
 762                         (printbuf_reset(&buf),
 763                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 764                 ret = -EIO;
 765                 goto err;
 766         }
 767 out:
 768         printbuf_exit(&buf);
 769         return ret;
 770 err:
 771         bch2_dump_trans_updates(trans);
 772         goto out;
 773 }
 774
 775 static int mark_stripe_bucket(struct btree_trans *trans,
 776                               struct bkey_s_c k,
 777                               unsigned ptr_idx,
 778                               unsigned flags)
 779 {
 780         struct bch_fs *c = trans->c;
 781         u64 journal_seq = trans->journal_res.seq;
 782         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 783         unsigned nr_data = s->nr_blocks - s->nr_redundant;
 784         bool parity = ptr_idx >= nr_data;
 785         enum bch_data_type data_type = parity ? BCH_DATA_parity : BCH_DATA_stripe;
 786         s64 sectors = parity ? le16_to_cpu(s->sectors) : 0;
 787         const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
 788         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 789         struct bucket old, new, *g;
 790         struct printbuf buf = PRINTBUF;
 791         int ret = 0;
 792
 793         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 794
 795         /* * XXX doesn't handle deletion */
 796
 797         percpu_down_read(&c->mark_lock);
 798         buf.atomic++;
 799         g = PTR_GC_BUCKET(ca, ptr);
 800
 801         if (g->dirty_sectors ||
 802             (g->stripe && g->stripe != k.k->p.offset)) {
 803                 bch2_fs_inconsistent(c,
 804                               "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
 805                               ptr->dev, PTR_BUCKET_NR(ca, ptr), g->gen,
 806                               (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 807                 ret = -EINVAL;
 808                 goto err;
 809         }
 810
 811         bucket_lock(g);
 812         old = *g;
 813
 814         ret = check_bucket_ref(trans, k, ptr, sectors, data_type,
 815                                g->gen, g->data_type,
 816                                g->dirty_sectors, g->cached_sectors);
 817         if (ret)
 818                 goto err;
 819
 820         g->data_type = data_type;
 821         g->dirty_sectors += sectors;
 822
 823         g->stripe               = k.k->p.offset;
 824         g->stripe_redundancy    = s->nr_redundant;
 825         new = *g;
 826 err:
 827         bucket_unlock(g);
 828         if (!ret)
 829                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 830         percpu_up_read(&c->mark_lock);
 831         printbuf_exit(&buf);
 832         return ret;
 833 }
 834
 835 static int __mark_pointer(struct btree_trans *trans,
 836                           struct bkey_s_c k,
 837                           const struct bch_extent_ptr *ptr,
 838                           s64 sectors, enum bch_data_type ptr_data_type,
 839                           u8 bucket_gen, u8 *bucket_data_type,
 840                           u32 *dirty_sectors, u32 *cached_sectors)
 841 {
 842         u32 *dst_sectors = !ptr->cached
 843                 ? dirty_sectors
 844                 : cached_sectors;
 845         int ret = check_bucket_ref(trans, k, ptr, sectors, ptr_data_type,
 846                                    bucket_gen, *bucket_data_type,
 847                                    *dirty_sectors, *cached_sectors);
 848
 849         if (ret)
 850                 return ret;
 851
 852         *dst_sectors += sectors;
 853         *bucket_data_type = *dirty_sectors || *cached_sectors
 854                 ? ptr_data_type : 0;
 855         return 0;
 856 }
 857
 858 static int bch2_mark_pointer(struct btree_trans *trans,
 859                              enum btree_id btree_id, unsigned level,
 860                              struct bkey_s_c k,
 861                              struct extent_ptr_decoded p,
 862                              s64 sectors,
 863                              unsigned flags)
 864 {
 865         u64 journal_seq = trans->journal_res.seq;
 866         struct bch_fs *c = trans->c;
 867         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 868         struct bucket old, new, *g;
 869         enum bch_data_type data_type = bkey_ptr_data_type(btree_id, level, k, p);
 870         u8 bucket_data_type;
 871         int ret = 0;
 872
 873         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 874
 875         percpu_down_read(&c->mark_lock);
 876         g = PTR_GC_BUCKET(ca, &p.ptr);
 877         bucket_lock(g);
 878         old = *g;
 879
 880         bucket_data_type = g->data_type;
 881         ret = __mark_pointer(trans, k, &p.ptr, sectors,
 882                              data_type, g->gen,
 883                              &bucket_data_type,
 884                              &g->dirty_sectors,
 885                              &g->cached_sectors);
 886         if (!ret)
 887                 g->data_type = bucket_data_type;
 888
 889         new = *g;
 890         bucket_unlock(g);
 891         if (!ret)
 892                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 893         percpu_up_read(&c->mark_lock);
 894
 895         return ret;
 896 }
 897
 898 static int bch2_mark_stripe_ptr(struct btree_trans *trans,
 899                                 struct bkey_s_c k,
 900                                 struct bch_extent_stripe_ptr p,
 901                                 enum bch_data_type data_type,
 902                                 s64 sectors,
 903                                 unsigned flags)
 904 {
 905         struct bch_fs *c = trans->c;
 906         struct bch_replicas_padded r;
 907         struct gc_stripe *m;
 908
 909         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 910
 911         m = genradix_ptr_alloc(&c->gc_stripes, p.idx, GFP_KERNEL);
 912         if (!m) {
 913                 bch_err(c, "error allocating memory for gc_stripes, idx %llu",
 914                         (u64) p.idx);
 915                 return -BCH_ERR_ENOMEM_mark_stripe_ptr;
 916         }
 917
 918         mutex_lock(&c->ec_stripes_heap_lock);
 919
 920         if (!m || !m->alive) {
 921                 mutex_unlock(&c->ec_stripes_heap_lock);
 922                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
 923                                     (u64) p.idx);
 924                 bch2_inconsistent_error(c);
 925                 return -EIO;
 926         }
 927
 928         m->block_sectors[p.block] += sectors;
 929
 930         r = m->r;
 931         mutex_unlock(&c->ec_stripes_heap_lock);
 932
 933         r.e.data_type = data_type;
 934         update_replicas(c, k, &r.e, sectors, trans->journal_res.seq, true);
 935
 936         return 0;
 937 }
 938
 939 int bch2_mark_extent(struct btree_trans *trans,
 940                      enum btree_id btree_id, unsigned level,
 941                      struct bkey_s_c old, struct bkey_s_c new,
 942                      unsigned flags)
 943 {
 944         u64 journal_seq = trans->journal_res.seq;
 945         struct bch_fs *c = trans->c;
 946         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old : new;
 947         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 948         const union bch_extent_entry *entry;
 949         struct extent_ptr_decoded p;
 950         struct bch_replicas_padded r;
 951         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
 952                 ? BCH_DATA_btree
 953                 : BCH_DATA_user;
 954         s64 sectors = bkey_is_btree_ptr(k.k)
 955                 ? btree_sectors(c)
 956                 : k.k->size;
 957         s64 dirty_sectors = 0;
 958         bool stale;
 959         int ret;
 960
 961         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 962
 963         r.e.data_type   = data_type;
 964         r.e.nr_devs     = 0;
 965         r.e.nr_required = 1;
 966
 967         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 968                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
 969
 970                 if (flags & BTREE_TRIGGER_OVERWRITE)
 971                         disk_sectors = -disk_sectors;
 972
 973                 ret = bch2_mark_pointer(trans, btree_id, level, k, p, disk_sectors, flags);
 974                 if (ret < 0)
 975                         return ret;
 976
 977                 stale = ret > 0;
 978
 979                 if (p.ptr.cached) {
 980                         if (!stale) {
 981                                 ret = update_cached_sectors(c, k, p.ptr.dev,
 982                                                 disk_sectors, journal_seq, true);
 983                                 if (ret) {
 984                                         bch2_fs_fatal_error(c, "%s(): no replicas entry while updating cached sectors",
 985                                                             __func__);
 986                                         return ret;
 987                                 }
 988                         }
 989                 } else if (!p.has_ec) {
 990                         dirty_sectors          += disk_sectors;
 991                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
 992                 } else {
 993                         ret = bch2_mark_stripe_ptr(trans, k, p.ec, data_type,
 994                                         disk_sectors, flags);
 995                         if (ret)
 996                                 return ret;
 997
 998                         /*
 999                          * There may be other dirty pointers in this extent, but
1000                          * if so they're not required for mounting if we have an
1001                          * erasure coded pointer in this extent:
1002                          */
1003                         r.e.nr_required = 0;
1004                 }
1005         }
1006
1007         if (r.e.nr_devs) {
1008                 ret = update_replicas(c, k, &r.e, dirty_sectors, journal_seq, true);
1009                 if (ret) {
1010                         struct printbuf buf = PRINTBUF;
1011
1012                         bch2_bkey_val_to_text(&buf, c, k);
1013                         bch2_fs_fatal_error(c, "%s(): no replicas entry for %s", __func__, buf.buf);
1014                         printbuf_exit(&buf);
1015                         return ret;
1016                 }
1017         }
1018
1019         return 0;
1020 }
1021
1022 int bch2_mark_stripe(struct btree_trans *trans,
1023                      enum btree_id btree_id, unsigned level,
1024                      struct bkey_s_c old, struct bkey_s_c new,
1025                      unsigned flags)
1026 {
1027         bool gc = flags & BTREE_TRIGGER_GC;
1028         u64 journal_seq = trans->journal_res.seq;
1029         struct bch_fs *c = trans->c;
1030         u64 idx = new.k->p.offset;
1031         const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
1032                 ? bkey_s_c_to_stripe(old).v : NULL;
1033         const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
1034                 ? bkey_s_c_to_stripe(new).v : NULL;
1035         unsigned i;
1036         int ret;
1037
1038         BUG_ON(gc && old_s);
1039
1040         if (!gc) {
1041                 struct stripe *m = genradix_ptr(&c->stripes, idx);
1042
1043                 if (!m) {
1044                         struct printbuf buf1 = PRINTBUF;
1045                         struct printbuf buf2 = PRINTBUF;
1046
1047                         bch2_bkey_val_to_text(&buf1, c, old);
1048                         bch2_bkey_val_to_text(&buf2, c, new);
1049                         bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
1050                                             "old %s\n"
1051                                             "new %s", idx, buf1.buf, buf2.buf);
1052                         printbuf_exit(&buf2);
1053                         printbuf_exit(&buf1);
1054                         bch2_inconsistent_error(c);
1055                         return -1;
1056                 }
1057
1058                 if (!new_s) {
1059                         bch2_stripes_heap_del(c, m, idx);
1060
1061                         memset(m, 0, sizeof(*m));
1062                 } else {
1063                         m->sectors      = le16_to_cpu(new_s->sectors);
1064                         m->algorithm    = new_s->algorithm;
1065                         m->nr_blocks    = new_s->nr_blocks;
1066                         m->nr_redundant = new_s->nr_redundant;
1067                         m->blocks_nonempty = 0;
1068
1069                         for (i = 0; i < new_s->nr_blocks; i++)
1070                                 m->blocks_nonempty += !!stripe_blockcount_get(new_s, i);
1071
1072                         if (!old_s)
1073                                 bch2_stripes_heap_insert(c, m, idx);
1074                         else
1075                                 bch2_stripes_heap_update(c, m, idx);
1076                 }
1077         } else {
1078                 struct gc_stripe *m =
1079                         genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
1080
1081                 if (!m) {
1082                         bch_err(c, "error allocating memory for gc_stripes, idx %llu",
1083                                 idx);
1084                         return -BCH_ERR_ENOMEM_mark_stripe;
1085                 }
1086                 /*
1087                  * This will be wrong when we bring back runtime gc: we should
1088                  * be unmarking the old key and then marking the new key
1089                  */
1090                 m->alive        = true;
1091                 m->sectors      = le16_to_cpu(new_s->sectors);
1092                 m->nr_blocks    = new_s->nr_blocks;
1093                 m->nr_redundant = new_s->nr_redundant;
1094
1095                 for (i = 0; i < new_s->nr_blocks; i++)
1096                         m->ptrs[i] = new_s->ptrs[i];
1097
1098                 bch2_bkey_to_replicas(&m->r.e, new);
1099
1100                 /*
1101                  * gc recalculates this field from stripe ptr
1102                  * references:
1103                  */
1104                 memset(m->block_sectors, 0, sizeof(m->block_sectors));
1105
1106                 for (i = 0; i < new_s->nr_blocks; i++) {
1107                         ret = mark_stripe_bucket(trans, new, i, flags);
1108                         if (ret)
1109                                 return ret;
1110                 }
1111
1112                 ret = update_replicas(c, new, &m->r.e,
1113                                       ((s64) m->sectors * m->nr_redundant),
1114                                       journal_seq, gc);
1115                 if (ret) {
1116                         struct printbuf buf = PRINTBUF;
1117
1118                         bch2_bkey_val_to_text(&buf, c, new);
1119                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
1120                         printbuf_exit(&buf);
1121                         return ret;
1122                 }
1123         }
1124
1125         return 0;
1126 }
1127
1128 int bch2_mark_reservation(struct btree_trans *trans,
1129                           enum btree_id btree_id, unsigned level,
1130                           struct bkey_s_c old, struct bkey_s_c new,
1131                           unsigned flags)
1132 {
1133         struct bch_fs *c = trans->c;
1134         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old : new;
1135         struct bch_fs_usage *fs_usage;
1136         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1137         s64 sectors = (s64) k.k->size;
1138
1139         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1140
1141         if (flags & BTREE_TRIGGER_OVERWRITE)
1142                 sectors = -sectors;
1143         sectors *= replicas;
1144
1145         percpu_down_read(&c->mark_lock);
1146         preempt_disable();
1147
1148         fs_usage = fs_usage_ptr(c, trans->journal_res.seq, flags & BTREE_TRIGGER_GC);
1149         replicas = clamp_t(unsigned, replicas, 1,
1150                            ARRAY_SIZE(fs_usage->persistent_reserved));
1151
1152         fs_usage->reserved                              += sectors;
1153         fs_usage->persistent_reserved[replicas - 1]     += sectors;
1154
1155         preempt_enable();
1156         percpu_up_read(&c->mark_lock);
1157
1158         return 0;
1159 }
1160
1161 static s64 __bch2_mark_reflink_p(struct btree_trans *trans,
1162                                  struct bkey_s_c_reflink_p p,
1163                                  u64 start, u64 end,
1164                                  u64 *idx, unsigned flags, size_t r_idx)
1165 {
1166         struct bch_fs *c = trans->c;
1167         struct reflink_gc *r;
1168         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1169         u64 next_idx = end;
1170         s64 ret = 0;
1171         struct printbuf buf = PRINTBUF;
1172
1173         if (r_idx >= c->reflink_gc_nr)
1174                 goto not_found;
1175
1176         r = genradix_ptr(&c->reflink_gc_table, r_idx);
1177         next_idx = min(next_idx, r->offset - r->size);
1178         if (*idx < next_idx)
1179                 goto not_found;
1180
1181         BUG_ON((s64) r->refcount + add < 0);
1182
1183         r->refcount += add;
1184         *idx = r->offset;
1185         return 0;
1186 not_found:
1187         if (fsck_err(c, "pointer to missing indirect extent\n"
1188                      "  %s\n"
1189                      "  missing range %llu-%llu",
1190                      (bch2_bkey_val_to_text(&buf, c, p.s_c), buf.buf),
1191                      *idx, next_idx)) {
1192                 struct bkey_i_error *new;
1193
1194                 new = bch2_trans_kmalloc(trans, sizeof(*new));
1195                 ret = PTR_ERR_OR_ZERO(new);
1196                 if (ret)
1197                         goto err;
1198
1199                 bkey_init(&new->k);
1200                 new->k.type     = KEY_TYPE_error;
1201                 new->k.p                = bkey_start_pos(p.k);
1202                 new->k.p.offset += *idx - start;
1203                 bch2_key_resize(&new->k, next_idx - *idx);
1204                 ret = __bch2_btree_insert(trans, BTREE_ID_extents, &new->k_i,
1205                                           BTREE_TRIGGER_NORUN);
1206         }
1207
1208         *idx = next_idx;
1209 err:
1210 fsck_err:
1211         printbuf_exit(&buf);
1212         return ret;
1213 }
1214
1215 int bch2_mark_reflink_p(struct btree_trans *trans,
1216                         enum btree_id btree_id, unsigned level,
1217                         struct bkey_s_c old, struct bkey_s_c new,
1218                         unsigned flags)
1219 {
1220         struct bch_fs *c = trans->c;
1221         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old : new;
1222         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1223         struct reflink_gc *ref;
1224         size_t l, r, m;
1225         u64 idx = le64_to_cpu(p.v->idx), start = idx;
1226         u64 end = le64_to_cpu(p.v->idx) + p.k->size;
1227         int ret = 0;
1228
1229         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1230
1231         if (c->sb.version_upgrade_complete >= bcachefs_metadata_version_reflink_p_fix) {
1232                 idx -= le32_to_cpu(p.v->front_pad);
1233                 end += le32_to_cpu(p.v->back_pad);
1234         }
1235
1236         l = 0;
1237         r = c->reflink_gc_nr;
1238         while (l < r) {
1239                 m = l + (r - l) / 2;
1240
1241                 ref = genradix_ptr(&c->reflink_gc_table, m);
1242                 if (ref->offset <= idx)
1243                         l = m + 1;
1244                 else
1245                         r = m;
1246         }
1247
1248         while (idx < end && !ret)
1249                 ret = __bch2_mark_reflink_p(trans, p, start, end,
1250                                             &idx, flags, l++);
1251
1252         return ret;
1253 }
1254
1255 void bch2_trans_fs_usage_revert(struct btree_trans *trans,
1256                                 struct replicas_delta_list *deltas)
1257 {
1258         struct bch_fs *c = trans->c;
1259         struct bch_fs_usage *dst;
1260         struct replicas_delta *d, *top = (void *) deltas->d + deltas->used;
1261         s64 added = 0;
1262         unsigned i;
1263
1264         percpu_down_read(&c->mark_lock);
1265         preempt_disable();
1266         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1267
1268         /* revert changes: */
1269         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1270                 switch (d->r.data_type) {
1271                 case BCH_DATA_btree:
1272                 case BCH_DATA_user:
1273                 case BCH_DATA_parity:
1274                         added += d->delta;
1275                 }
1276                 BUG_ON(__update_replicas(c, dst, &d->r, -d->delta));
1277         }
1278
1279         dst->nr_inodes -= deltas->nr_inodes;
1280
1281         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1282                 added                           -= deltas->persistent_reserved[i];
1283                 dst->reserved                   -= deltas->persistent_reserved[i];
1284                 dst->persistent_reserved[i]     -= deltas->persistent_reserved[i];
1285         }
1286
1287         if (added > 0) {
1288                 trans->disk_res->sectors += added;
1289                 this_cpu_add(*c->online_reserved, added);
1290         }
1291
1292         preempt_enable();
1293         percpu_up_read(&c->mark_lock);
1294 }
1295
1296 int bch2_trans_fs_usage_apply(struct btree_trans *trans,
1297                               struct replicas_delta_list *deltas)
1298 {
1299         struct bch_fs *c = trans->c;
1300         static int warned_disk_usage = 0;
1301         bool warn = false;
1302         unsigned disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1303         struct replicas_delta *d = deltas->d, *d2;
1304         struct replicas_delta *top = (void *) deltas->d + deltas->used;
1305         struct bch_fs_usage *dst;
1306         s64 added = 0, should_not_have_added;
1307         unsigned i;
1308
1309         percpu_down_read(&c->mark_lock);
1310         preempt_disable();
1311         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1312
1313         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1314                 switch (d->r.data_type) {
1315                 case BCH_DATA_btree:
1316                 case BCH_DATA_user:
1317                 case BCH_DATA_parity:
1318                         added += d->delta;
1319                 }
1320
1321                 if (__update_replicas(c, dst, &d->r, d->delta))
1322                         goto need_mark;
1323         }
1324
1325         dst->nr_inodes += deltas->nr_inodes;
1326
1327         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1328                 added                           += deltas->persistent_reserved[i];
1329                 dst->reserved                   += deltas->persistent_reserved[i];
1330                 dst->persistent_reserved[i]     += deltas->persistent_reserved[i];
1331         }
1332
1333         /*
1334          * Not allowed to reduce sectors_available except by getting a
1335          * reservation:
1336          */
1337         should_not_have_added = added - (s64) disk_res_sectors;
1338         if (unlikely(should_not_have_added > 0)) {
1339                 u64 old, new, v = atomic64_read(&c->sectors_available);
1340
1341                 do {
1342                         old = v;
1343                         new = max_t(s64, 0, old - should_not_have_added);
1344                 } while ((v = atomic64_cmpxchg(&c->sectors_available,
1345                                                old, new)) != old);
1346
1347                 added -= should_not_have_added;
1348                 warn = true;
1349         }
1350
1351         if (added > 0) {
1352                 trans->disk_res->sectors -= added;
1353                 this_cpu_sub(*c->online_reserved, added);
1354         }
1355
1356         preempt_enable();
1357         percpu_up_read(&c->mark_lock);
1358
1359         if (unlikely(warn) && !xchg(&warned_disk_usage, 1))
1360                 bch2_trans_inconsistent(trans,
1361                                         "disk usage increased %lli more than %u sectors reserved)",
1362                                         should_not_have_added, disk_res_sectors);
1363         return 0;
1364 need_mark:
1365         /* revert changes: */
1366         for (d2 = deltas->d; d2 != d; d2 = replicas_delta_next(d2))
1367                 BUG_ON(__update_replicas(c, dst, &d2->r, -d2->delta));
1368
1369         preempt_enable();
1370         percpu_up_read(&c->mark_lock);
1371         return -1;
1372 }
1373
1374 /* trans_mark: */
1375
1376 static inline int bch2_trans_mark_pointer(struct btree_trans *trans,
1377                                    enum btree_id btree_id, unsigned level,
1378                                    struct bkey_s_c k, struct extent_ptr_decoded p,
1379                                    unsigned flags)
1380 {
1381         bool insert = !(flags & BTREE_TRIGGER_OVERWRITE);
1382         struct btree_iter iter;
1383         struct bkey_i_alloc_v4 *a;
1384         struct bpos bucket;
1385         struct bch_backpointer bp;
1386         s64 sectors;
1387         int ret;
1388
1389         bch2_extent_ptr_to_bp(trans->c, btree_id, level, k, p, &bucket, &bp);
1390         sectors = bp.bucket_len;
1391         if (!insert)
1392                 sectors = -sectors;
1393
1394         a = bch2_trans_start_alloc_update(trans, &iter, bucket);
1395         if (IS_ERR(a))
1396                 return PTR_ERR(a);
1397
1398         ret = __mark_pointer(trans, k, &p.ptr, sectors, bp.data_type,
1399                              a->v.gen, &a->v.data_type,
1400                              &a->v.dirty_sectors, &a->v.cached_sectors) ?:
1401                 bch2_trans_update(trans, &iter, &a->k_i, 0);
1402         bch2_trans_iter_exit(trans, &iter);
1403
1404         if (ret)
1405                 return ret;
1406
1407         if (!p.ptr.cached) {
1408                 ret = bch2_bucket_backpointer_mod(trans, bucket, bp, k, insert);
1409                 if (ret)
1410                         return ret;
1411         }
1412
1413         return 0;
1414 }
1415
1416 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1417                         struct extent_ptr_decoded p,
1418                         s64 sectors, enum bch_data_type data_type)
1419 {
1420         struct btree_iter iter;
1421         struct bkey_i_stripe *s;
1422         struct bch_replicas_padded r;
1423         int ret = 0;
1424
1425         s = bch2_bkey_get_mut_typed(trans, &iter,
1426                         BTREE_ID_stripes, POS(0, p.ec.idx),
1427                         BTREE_ITER_WITH_UPDATES, stripe);
1428         ret = PTR_ERR_OR_ZERO(s);
1429         if (unlikely(ret)) {
1430                 bch2_trans_inconsistent_on(bch2_err_matches(ret, ENOENT), trans,
1431                         "pointer to nonexistent stripe %llu",
1432                         (u64) p.ec.idx);
1433                 goto err;
1434         }
1435
1436         if (!bch2_ptr_matches_stripe(&s->v, p)) {
1437                 bch2_trans_inconsistent(trans,
1438                         "stripe pointer doesn't match stripe %llu",
1439                         (u64) p.ec.idx);
1440                 ret = -EIO;
1441                 goto err;
1442         }
1443
1444         stripe_blockcount_set(&s->v, p.ec.block,
1445                 stripe_blockcount_get(&s->v, p.ec.block) +
1446                 sectors);
1447
1448         bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1449         r.e.data_type = data_type;
1450         ret = update_replicas_list(trans, &r.e, sectors);
1451 err:
1452         bch2_trans_iter_exit(trans, &iter);
1453         return ret;
1454 }
1455
1456 int bch2_trans_mark_extent(struct btree_trans *trans,
1457                            enum btree_id btree_id, unsigned level,
1458                            struct bkey_s_c old, struct bkey_i *new,
1459                            unsigned flags)
1460 {
1461         struct bch_fs *c = trans->c;
1462         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1463                 ? old
1464                 : bkey_i_to_s_c(new);
1465         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1466         const union bch_extent_entry *entry;
1467         struct extent_ptr_decoded p;
1468         struct bch_replicas_padded r;
1469         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
1470                 ? BCH_DATA_btree
1471                 : BCH_DATA_user;
1472         s64 sectors = bkey_is_btree_ptr(k.k)
1473                 ? btree_sectors(c)
1474                 : k.k->size;
1475         s64 dirty_sectors = 0;
1476         bool stale;
1477         int ret = 0;
1478
1479         r.e.data_type   = data_type;
1480         r.e.nr_devs     = 0;
1481         r.e.nr_required = 1;
1482
1483         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1484                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
1485
1486                 if (flags & BTREE_TRIGGER_OVERWRITE)
1487                         disk_sectors = -disk_sectors;
1488
1489                 ret = bch2_trans_mark_pointer(trans, btree_id, level, k, p, flags);
1490                 if (ret < 0)
1491                         return ret;
1492
1493                 stale = ret > 0;
1494
1495                 if (p.ptr.cached) {
1496                         if (!stale) {
1497                                 ret = update_cached_sectors_list(trans, p.ptr.dev,
1498                                                                  disk_sectors);
1499                                 if (ret)
1500                                         return ret;
1501                         }
1502                 } else if (!p.has_ec) {
1503                         dirty_sectors          += disk_sectors;
1504                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1505                 } else {
1506                         ret = bch2_trans_mark_stripe_ptr(trans, p,
1507                                         disk_sectors, data_type);
1508                         if (ret)
1509                                 return ret;
1510
1511                         r.e.nr_required = 0;
1512                 }
1513         }
1514
1515         if (r.e.nr_devs)
1516                 ret = update_replicas_list(trans, &r.e, dirty_sectors);
1517
1518         return ret;
1519 }
1520
1521 static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
1522                                          struct bkey_s_c_stripe s,
1523                                          unsigned idx, bool deleting)
1524 {
1525         struct bch_fs *c = trans->c;
1526         const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
1527         struct btree_iter iter;
1528         struct bkey_i_alloc_v4 *a;
1529         enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
1530                 ? BCH_DATA_parity : 0;
1531         s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
1532         int ret = 0;
1533
1534         if (deleting)
1535                 sectors = -sectors;
1536
1537         a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(c, ptr));
1538         if (IS_ERR(a))
1539                 return PTR_ERR(a);
1540
1541         ret = check_bucket_ref(trans, s.s_c, ptr, sectors, data_type,
1542                                a->v.gen, a->v.data_type,
1543                                a->v.dirty_sectors, a->v.cached_sectors);
1544         if (ret)
1545                 goto err;
1546
1547         if (!deleting) {
1548                 if (bch2_trans_inconsistent_on(a->v.stripe ||
1549                                                a->v.stripe_redundancy, trans,
1550                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
1551                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1552                                 bch2_data_types[a->v.data_type],
1553                                 a->v.dirty_sectors,
1554                                 a->v.stripe, s.k->p.offset)) {
1555                         ret = -EIO;
1556                         goto err;
1557                 }
1558
1559                 if (bch2_trans_inconsistent_on(data_type && a->v.dirty_sectors, trans,
1560                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
1561                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1562                                 bch2_data_types[a->v.data_type],
1563                                 a->v.dirty_sectors,
1564                                 s.k->p.offset)) {
1565                         ret = -EIO;
1566                         goto err;
1567                 }
1568
1569                 a->v.stripe             = s.k->p.offset;
1570                 a->v.stripe_redundancy  = s.v->nr_redundant;
1571                 a->v.data_type          = BCH_DATA_stripe;
1572         } else {
1573                 if (bch2_trans_inconsistent_on(a->v.stripe != s.k->p.offset ||
1574                                                a->v.stripe_redundancy != s.v->nr_redundant, trans,
1575                                 "bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
1576                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1577                                 s.k->p.offset, a->v.stripe)) {
1578                         ret = -EIO;
1579                         goto err;
1580                 }
1581
1582                 a->v.stripe             = 0;
1583                 a->v.stripe_redundancy  = 0;
1584                 a->v.data_type          = alloc_data_type(a->v, BCH_DATA_user);
1585         }
1586
1587         a->v.dirty_sectors += sectors;
1588         if (data_type)
1589                 a->v.data_type = !deleting ? data_type : 0;
1590
1591         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1592         if (ret)
1593                 goto err;
1594 err:
1595         bch2_trans_iter_exit(trans, &iter);
1596         return ret;
1597 }
1598
1599 int bch2_trans_mark_stripe(struct btree_trans *trans,
1600                            enum btree_id btree_id, unsigned level,
1601                            struct bkey_s_c old, struct bkey_i *new,
1602                            unsigned flags)
1603 {
1604         const struct bch_stripe *old_s = NULL;
1605         struct bch_stripe *new_s = NULL;
1606         struct bch_replicas_padded r;
1607         unsigned i, nr_blocks;
1608         int ret = 0;
1609
1610         if (old.k->type == KEY_TYPE_stripe)
1611                 old_s = bkey_s_c_to_stripe(old).v;
1612         if (new->k.type == KEY_TYPE_stripe)
1613                 new_s = &bkey_i_to_stripe(new)->v;
1614
1615         /*
1616          * If the pointers aren't changing, we don't need to do anything:
1617          */
1618         if (new_s && old_s &&
1619             new_s->nr_blocks    == old_s->nr_blocks &&
1620             new_s->nr_redundant == old_s->nr_redundant &&
1621             !memcmp(old_s->ptrs, new_s->ptrs,
1622                     new_s->nr_blocks * sizeof(struct bch_extent_ptr)))
1623                 return 0;
1624
1625         BUG_ON(new_s && old_s &&
1626                (new_s->nr_blocks        != old_s->nr_blocks ||
1627                 new_s->nr_redundant     != old_s->nr_redundant));
1628
1629         nr_blocks = new_s ? new_s->nr_blocks : old_s->nr_blocks;
1630
1631         if (new_s) {
1632                 s64 sectors = le16_to_cpu(new_s->sectors);
1633
1634                 bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(new));
1635                 ret = update_replicas_list(trans, &r.e, sectors * new_s->nr_redundant);
1636                 if (ret)
1637                         return ret;
1638         }
1639
1640         if (old_s) {
1641                 s64 sectors = -((s64) le16_to_cpu(old_s->sectors));
1642
1643                 bch2_bkey_to_replicas(&r.e, old);
1644                 ret = update_replicas_list(trans, &r.e, sectors * old_s->nr_redundant);
1645                 if (ret)
1646                         return ret;
1647         }
1648
1649         for (i = 0; i < nr_blocks; i++) {
1650                 if (new_s && old_s &&
1651                     !memcmp(&new_s->ptrs[i],
1652                             &old_s->ptrs[i],
1653                             sizeof(new_s->ptrs[i])))
1654                         continue;
1655
1656                 if (new_s) {
1657                         ret = bch2_trans_mark_stripe_bucket(trans,
1658                                         bkey_i_to_s_c_stripe(new), i, false);
1659                         if (ret)
1660                                 break;
1661                 }
1662
1663                 if (old_s) {
1664                         ret = bch2_trans_mark_stripe_bucket(trans,
1665                                         bkey_s_c_to_stripe(old), i, true);
1666                         if (ret)
1667                                 break;
1668                 }
1669         }
1670
1671         return ret;
1672 }
1673
1674 int bch2_trans_mark_reservation(struct btree_trans *trans,
1675                                 enum btree_id btree_id, unsigned level,
1676                                 struct bkey_s_c old,
1677                                 struct bkey_i *new,
1678                                 unsigned flags)
1679 {
1680         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1681                 ? old
1682                 : bkey_i_to_s_c(new);
1683         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1684         s64 sectors = (s64) k.k->size;
1685         struct replicas_delta_list *d;
1686         int ret;
1687
1688         if (flags & BTREE_TRIGGER_OVERWRITE)
1689                 sectors = -sectors;
1690         sectors *= replicas;
1691
1692         ret = bch2_replicas_deltas_realloc(trans, 0);
1693         if (ret)
1694                 return ret;
1695
1696         d = trans->fs_usage_deltas;
1697         replicas = clamp_t(unsigned, replicas, 1,
1698                            ARRAY_SIZE(d->persistent_reserved));
1699
1700         d->persistent_reserved[replicas - 1] += sectors;
1701         return 0;
1702 }
1703
1704 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1705                         struct bkey_s_c_reflink_p p,
1706                         u64 *idx, unsigned flags)
1707 {
1708         struct bch_fs *c = trans->c;
1709         struct btree_iter iter;
1710         struct bkey_i *k;
1711         __le64 *refcount;
1712         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1713         struct printbuf buf = PRINTBUF;
1714         int ret;
1715
1716         k = bch2_bkey_get_mut_noupdate(trans, &iter,
1717                         BTREE_ID_reflink, POS(0, *idx),
1718                         BTREE_ITER_WITH_UPDATES);
1719         ret = PTR_ERR_OR_ZERO(k);
1720         if (ret)
1721                 goto err;
1722
1723         refcount = bkey_refcount(k);
1724         if (!refcount) {
1725                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1726                 bch2_trans_inconsistent(trans,
1727                         "nonexistent indirect extent at %llu while marking\n  %s",
1728                         *idx, buf.buf);
1729                 ret = -EIO;
1730                 goto err;
1731         }
1732
1733         if (!*refcount && (flags & BTREE_TRIGGER_OVERWRITE)) {
1734                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1735                 bch2_trans_inconsistent(trans,
1736                         "indirect extent refcount underflow at %llu while marking\n  %s",
1737                         *idx, buf.buf);
1738                 ret = -EIO;
1739                 goto err;
1740         }
1741
1742         if (flags & BTREE_TRIGGER_INSERT) {
1743                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1744                 u64 pad;
1745
1746                 pad = max_t(s64, le32_to_cpu(v->front_pad),
1747                             le64_to_cpu(v->idx) - bkey_start_offset(&k->k));
1748                 BUG_ON(pad > U32_MAX);
1749                 v->front_pad = cpu_to_le32(pad);
1750
1751                 pad = max_t(s64, le32_to_cpu(v->back_pad),
1752                             k->k.p.offset - p.k->size - le64_to_cpu(v->idx));
1753                 BUG_ON(pad > U32_MAX);
1754                 v->back_pad = cpu_to_le32(pad);
1755         }
1756
1757         le64_add_cpu(refcount, add);
1758
1759         bch2_btree_iter_set_pos_to_extent_start(&iter);
1760         ret = bch2_trans_update(trans, &iter, k, 0);
1761         if (ret)
1762                 goto err;
1763
1764         *idx = k->k.p.offset;
1765 err:
1766         bch2_trans_iter_exit(trans, &iter);
1767         printbuf_exit(&buf);
1768         return ret;
1769 }
1770
1771 int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1772                               enum btree_id btree_id, unsigned level,
1773                               struct bkey_s_c old,
1774                               struct bkey_i *new,
1775                               unsigned flags)
1776 {
1777         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1778                 ? old
1779                 : bkey_i_to_s_c(new);
1780         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1781         u64 idx, end_idx;
1782         int ret = 0;
1783
1784         if (flags & BTREE_TRIGGER_INSERT) {
1785                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1786
1787                 v->front_pad = v->back_pad = 0;
1788         }
1789
1790         idx     = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
1791         end_idx = le64_to_cpu(p.v->idx) + p.k->size +
1792                 le32_to_cpu(p.v->back_pad);
1793
1794         while (idx < end_idx && !ret)
1795                 ret = __bch2_trans_mark_reflink_p(trans, p, &idx, flags);
1796
1797         return ret;
1798 }
1799
1800 static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1801                                     struct bch_dev *ca, size_t b,
1802                                     enum bch_data_type type,
1803                                     unsigned sectors)
1804 {
1805         struct bch_fs *c = trans->c;
1806         struct btree_iter iter;
1807         struct bkey_i_alloc_v4 *a;
1808         int ret = 0;
1809
1810         /*
1811          * Backup superblock might be past the end of our normal usable space:
1812          */
1813         if (b >= ca->mi.nbuckets)
1814                 return 0;
1815
1816         a = bch2_trans_start_alloc_update(trans, &iter, POS(ca->dev_idx, b));
1817         if (IS_ERR(a))
1818                 return PTR_ERR(a);
1819
1820         if (a->v.data_type && type && a->v.data_type != type) {
1821                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1822                         "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
1823                         "while marking %s",
1824                         iter.pos.inode, iter.pos.offset, a->v.gen,
1825                         bch2_data_types[a->v.data_type],
1826                         bch2_data_types[type],
1827                         bch2_data_types[type]);
1828                 ret = -EIO;
1829                 goto out;
1830         }
1831
1832         a->v.data_type          = type;
1833         a->v.dirty_sectors      = sectors;
1834
1835         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1836         if (ret)
1837                 goto out;
1838 out:
1839         bch2_trans_iter_exit(trans, &iter);
1840         return ret;
1841 }
1842
1843 int bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1844                                     struct bch_dev *ca, size_t b,
1845                                     enum bch_data_type type,
1846                                     unsigned sectors)
1847 {
1848         return commit_do(trans, NULL, NULL, 0,
1849                         __bch2_trans_mark_metadata_bucket(trans, ca, b, type, sectors));
1850 }
1851
1852 static int bch2_trans_mark_metadata_sectors(struct btree_trans *trans,
1853                                             struct bch_dev *ca,
1854                                             u64 start, u64 end,
1855                                             enum bch_data_type type,
1856                                             u64 *bucket, unsigned *bucket_sectors)
1857 {
1858         do {
1859                 u64 b = sector_to_bucket(ca, start);
1860                 unsigned sectors =
1861                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1862
1863                 if (b != *bucket && *bucket_sectors) {
1864                         int ret = bch2_trans_mark_metadata_bucket(trans, ca, *bucket,
1865                                                                   type, *bucket_sectors);
1866                         if (ret)
1867                                 return ret;
1868
1869                         *bucket_sectors = 0;
1870                 }
1871
1872                 *bucket         = b;
1873                 *bucket_sectors += sectors;
1874                 start += sectors;
1875         } while (start < end);
1876
1877         return 0;
1878 }
1879
1880 static int __bch2_trans_mark_dev_sb(struct btree_trans *trans,
1881                                     struct bch_dev *ca)
1882 {
1883         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1884         u64 bucket = 0;
1885         unsigned i, bucket_sectors = 0;
1886         int ret;
1887
1888         for (i = 0; i < layout->nr_superblocks; i++) {
1889                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1890
1891                 if (offset == BCH_SB_SECTOR) {
1892                         ret = bch2_trans_mark_metadata_sectors(trans, ca,
1893                                                 0, BCH_SB_SECTOR,
1894                                                 BCH_DATA_sb, &bucket, &bucket_sectors);
1895                         if (ret)
1896                                 return ret;
1897                 }
1898
1899                 ret = bch2_trans_mark_metadata_sectors(trans, ca, offset,
1900                                       offset + (1 << layout->sb_max_size_bits),
1901                                       BCH_DATA_sb, &bucket, &bucket_sectors);
1902                 if (ret)
1903                         return ret;
1904         }
1905
1906         if (bucket_sectors) {
1907                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1908                                 bucket, BCH_DATA_sb, bucket_sectors);
1909                 if (ret)
1910                         return ret;
1911         }
1912
1913         for (i = 0; i < ca->journal.nr; i++) {
1914                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1915                                 ca->journal.buckets[i],
1916                                 BCH_DATA_journal, ca->mi.bucket_size);
1917                 if (ret)
1918                         return ret;
1919         }
1920
1921         return 0;
1922 }
1923
1924 int bch2_trans_mark_dev_sb(struct bch_fs *c, struct bch_dev *ca)
1925 {
1926         int ret = bch2_trans_run(c, __bch2_trans_mark_dev_sb(&trans, ca));
1927         if (ret)
1928                 bch_err_fn(c, ret);
1929         return ret;
1930 }
1931
1932 /* Disk reservations: */
1933
1934 #define SECTORS_CACHE   1024
1935
1936 int __bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
1937                               u64 sectors, int flags)
1938 {
1939         struct bch_fs_pcpu *pcpu;
1940         u64 old, v, get;
1941         s64 sectors_available;
1942         int ret;
1943
1944         percpu_down_read(&c->mark_lock);
1945         preempt_disable();
1946         pcpu = this_cpu_ptr(c->pcpu);
1947
1948         if (sectors <= pcpu->sectors_available)
1949                 goto out;
1950
1951         v = atomic64_read(&c->sectors_available);
1952         do {
1953                 old = v;
1954                 get = min((u64) sectors + SECTORS_CACHE, old);
1955
1956                 if (get < sectors) {
1957                         preempt_enable();
1958                         goto recalculate;
1959                 }
1960         } while ((v = atomic64_cmpxchg(&c->sectors_available,
1961                                        old, old - get)) != old);
1962
1963         pcpu->sectors_available         += get;
1964
1965 out:
1966         pcpu->sectors_available         -= sectors;
1967         this_cpu_add(*c->online_reserved, sectors);
1968         res->sectors                    += sectors;
1969
1970         preempt_enable();
1971         percpu_up_read(&c->mark_lock);
1972         return 0;
1973
1974 recalculate:
1975         mutex_lock(&c->sectors_available_lock);
1976
1977         percpu_u64_set(&c->pcpu->sectors_available, 0);
1978         sectors_available = avail_factor(__bch2_fs_usage_read_short(c).free);
1979
1980         if (sectors <= sectors_available ||
1981             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
1982                 atomic64_set(&c->sectors_available,
1983                              max_t(s64, 0, sectors_available - sectors));
1984                 this_cpu_add(*c->online_reserved, sectors);
1985                 res->sectors                    += sectors;
1986                 ret = 0;
1987         } else {
1988                 atomic64_set(&c->sectors_available, sectors_available);
1989                 ret = -BCH_ERR_ENOSPC_disk_reservation;
1990         }
1991
1992         mutex_unlock(&c->sectors_available_lock);
1993         percpu_up_read(&c->mark_lock);
1994
1995         return ret;
1996 }
1997
1998 /* Startup/shutdown: */
1999
2000 static void bucket_gens_free_rcu(struct rcu_head *rcu)
2001 {
2002         struct bucket_gens *buckets =
2003                 container_of(rcu, struct bucket_gens, rcu);
2004
2005         kvpfree(buckets, sizeof(*buckets) + buckets->nbuckets);
2006 }
2007
2008 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
2009 {
2010         struct bucket_gens *bucket_gens = NULL, *old_bucket_gens = NULL;
2011         unsigned long *buckets_nouse = NULL;
2012         bool resize = ca->bucket_gens != NULL;
2013         int ret;
2014
2015         if (!(bucket_gens       = kvpmalloc(sizeof(struct bucket_gens) + nbuckets,
2016                                             GFP_KERNEL|__GFP_ZERO))) {
2017                 ret = -BCH_ERR_ENOMEM_bucket_gens;
2018                 goto err;
2019         }
2020
2021         if ((c->opts.buckets_nouse &&
2022              !(buckets_nouse    = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2023                                             sizeof(unsigned long),
2024                                             GFP_KERNEL|__GFP_ZERO)))) {
2025                 ret = -BCH_ERR_ENOMEM_buckets_nouse;
2026                 goto err;
2027         }
2028
2029         bucket_gens->first_bucket = ca->mi.first_bucket;
2030         bucket_gens->nbuckets   = nbuckets;
2031
2032         bch2_copygc_stop(c);
2033
2034         if (resize) {
2035                 down_write(&c->gc_lock);
2036                 down_write(&ca->bucket_lock);
2037                 percpu_down_write(&c->mark_lock);
2038         }
2039
2040         old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
2041
2042         if (resize) {
2043                 size_t n = min(bucket_gens->nbuckets, old_bucket_gens->nbuckets);
2044
2045                 memcpy(bucket_gens->b,
2046                        old_bucket_gens->b,
2047                        n);
2048                 if (buckets_nouse)
2049                         memcpy(buckets_nouse,
2050                                ca->buckets_nouse,
2051                                BITS_TO_LONGS(n) * sizeof(unsigned long));
2052         }
2053
2054         rcu_assign_pointer(ca->bucket_gens, bucket_gens);
2055         bucket_gens     = old_bucket_gens;
2056
2057         swap(ca->buckets_nouse, buckets_nouse);
2058
2059         nbuckets = ca->mi.nbuckets;
2060
2061         if (resize) {
2062                 percpu_up_write(&c->mark_lock);
2063                 up_write(&ca->bucket_lock);
2064                 up_write(&c->gc_lock);
2065         }
2066
2067         ret = 0;
2068 err:
2069         kvpfree(buckets_nouse,
2070                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2071         if (bucket_gens)
2072                 call_rcu(&bucket_gens->rcu, bucket_gens_free_rcu);
2073
2074         return ret;
2075 }
2076
2077 void bch2_dev_buckets_free(struct bch_dev *ca)
2078 {
2079         unsigned i;
2080
2081         kvpfree(ca->buckets_nouse,
2082                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2083         kvpfree(rcu_dereference_protected(ca->bucket_gens, 1),
2084                 sizeof(struct bucket_gens) + ca->mi.nbuckets);
2085
2086         for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
2087                 free_percpu(ca->usage[i]);
2088         kfree(ca->usage_base);
2089 }
2090
2091 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2092 {
2093         unsigned i;
2094
2095         ca->usage_base = kzalloc(sizeof(struct bch_dev_usage), GFP_KERNEL);
2096         if (!ca->usage_base)
2097                 return -BCH_ERR_ENOMEM_usage_init;
2098
2099         for (i = 0; i < ARRAY_SIZE(ca->usage); i++) {
2100                 ca->usage[i] = alloc_percpu(struct bch_dev_usage);
2101                 if (!ca->usage[i])
2102                         return -BCH_ERR_ENOMEM_usage_init;
2103         }
2104
2105         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);
2106 }