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 "bset.h"
  11 #include "btree_gc.h"
  12 #include "btree_update.h"
  13 #include "buckets.h"
  14 #include "buckets_waiting_for_journal.h"
  15 #include "ec.h"
  16 #include "error.h"
  17 #include "inode.h"
  18 #include "movinggc.h"
  19 #include "recovery.h"
  20 #include "reflink.h"
  21 #include "replicas.h"
  22 #include "subvolume.h"
  23
  24 #include <linux/preempt.h>
  25 #include <trace/events/bcachefs.h>
  26
  27 static inline void fs_usage_data_type_to_base(struct bch_fs_usage *fs_usage,
  28                                               enum bch_data_type data_type,
  29                                               s64 sectors)
  30 {
  31         switch (data_type) {
  32         case BCH_DATA_btree:
  33                 fs_usage->btree         += sectors;
  34                 break;
  35         case BCH_DATA_user:
  36         case BCH_DATA_parity:
  37                 fs_usage->data          += sectors;
  38                 break;
  39         case BCH_DATA_cached:
  40                 fs_usage->cached        += sectors;
  41                 break;
  42         default:
  43                 break;
  44         }
  45 }
  46
  47 void bch2_fs_usage_initialize(struct bch_fs *c)
  48 {
  49         struct bch_fs_usage *usage;
  50         struct bch_dev *ca;
  51         unsigned i;
  52
  53         percpu_down_write(&c->mark_lock);
  54         usage = c->usage_base;
  55
  56         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
  57                 bch2_fs_usage_acc_to_base(c, i);
  58
  59         for (i = 0; i < BCH_REPLICAS_MAX; i++)
  60                 usage->reserved += usage->persistent_reserved[i];
  61
  62         for (i = 0; i < c->replicas.nr; i++) {
  63                 struct bch_replicas_entry *e =
  64                         cpu_replicas_entry(&c->replicas, i);
  65
  66                 fs_usage_data_type_to_base(usage, e->data_type, usage->replicas[i]);
  67         }
  68
  69         for_each_member_device(ca, c, i) {
  70                 struct bch_dev_usage dev = bch2_dev_usage_read(ca);
  71
  72                 usage->hidden += (dev.d[BCH_DATA_sb].buckets +
  73                                   dev.d[BCH_DATA_journal].buckets) *
  74                         ca->mi.bucket_size;
  75         }
  76
  77         percpu_up_write(&c->mark_lock);
  78 }
  79
  80 static inline struct bch_dev_usage *dev_usage_ptr(struct bch_dev *ca,
  81                                                   unsigned journal_seq,
  82                                                   bool gc)
  83 {
  84         BUG_ON(!gc && !journal_seq);
  85
  86         return this_cpu_ptr(gc
  87                             ? ca->usage_gc
  88                             : ca->usage[journal_seq & JOURNAL_BUF_MASK]);
  89 }
  90
  91 struct bch_dev_usage bch2_dev_usage_read(struct bch_dev *ca)
  92 {
  93         struct bch_fs *c = ca->fs;
  94         struct bch_dev_usage ret;
  95         unsigned seq, i, u64s = dev_usage_u64s();
  96
  97         do {
  98                 seq = read_seqcount_begin(&c->usage_lock);
  99                 memcpy(&ret, ca->usage_base, u64s * sizeof(u64));
 100                 for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
 101                         acc_u64s_percpu((u64 *) &ret, (u64 __percpu *) ca->usage[i], u64s);
 102         } while (read_seqcount_retry(&c->usage_lock, seq));
 103
 104         return ret;
 105 }
 106
 107 static inline struct bch_fs_usage *fs_usage_ptr(struct bch_fs *c,
 108                                                 unsigned journal_seq,
 109                                                 bool gc)
 110 {
 111         percpu_rwsem_assert_held(&c->mark_lock);
 112         BUG_ON(!gc && !journal_seq);
 113
 114         return this_cpu_ptr(gc
 115                             ? c->usage_gc
 116                             : c->usage[journal_seq & JOURNAL_BUF_MASK]);
 117 }
 118
 119 u64 bch2_fs_usage_read_one(struct bch_fs *c, u64 *v)
 120 {
 121         ssize_t offset = v - (u64 *) c->usage_base;
 122         unsigned i, seq;
 123         u64 ret;
 124
 125         BUG_ON(offset < 0 || offset >= fs_usage_u64s(c));
 126         percpu_rwsem_assert_held(&c->mark_lock);
 127
 128         do {
 129                 seq = read_seqcount_begin(&c->usage_lock);
 130                 ret = *v;
 131
 132                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 133                         ret += percpu_u64_get((u64 __percpu *) c->usage[i] + offset);
 134         } while (read_seqcount_retry(&c->usage_lock, seq));
 135
 136         return ret;
 137 }
 138
 139 struct bch_fs_usage_online *bch2_fs_usage_read(struct bch_fs *c)
 140 {
 141         struct bch_fs_usage_online *ret;
 142         unsigned seq, i, u64s;
 143
 144         percpu_down_read(&c->mark_lock);
 145
 146         ret = kmalloc(sizeof(struct bch_fs_usage_online) +
 147                       sizeof(u64) * c->replicas.nr, GFP_NOFS);
 148         if (unlikely(!ret)) {
 149                 percpu_up_read(&c->mark_lock);
 150                 return NULL;
 151         }
 152
 153         ret->online_reserved = percpu_u64_get(c->online_reserved);
 154
 155         u64s = fs_usage_u64s(c);
 156         do {
 157                 seq = read_seqcount_begin(&c->usage_lock);
 158                 memcpy(&ret->u, c->usage_base, u64s * sizeof(u64));
 159                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 160                         acc_u64s_percpu((u64 *) &ret->u, (u64 __percpu *) c->usage[i], u64s);
 161         } while (read_seqcount_retry(&c->usage_lock, seq));
 162
 163         return ret;
 164 }
 165
 166 void bch2_fs_usage_acc_to_base(struct bch_fs *c, unsigned idx)
 167 {
 168         struct bch_dev *ca;
 169         unsigned i, u64s = fs_usage_u64s(c);
 170
 171         BUG_ON(idx >= ARRAY_SIZE(c->usage));
 172
 173         preempt_disable();
 174         write_seqcount_begin(&c->usage_lock);
 175
 176         acc_u64s_percpu((u64 *) c->usage_base,
 177                         (u64 __percpu *) c->usage[idx], u64s);
 178         percpu_memset(c->usage[idx], 0, u64s * sizeof(u64));
 179
 180         rcu_read_lock();
 181         for_each_member_device_rcu(ca, c, i, NULL) {
 182                 u64s = dev_usage_u64s();
 183
 184                 acc_u64s_percpu((u64 *) ca->usage_base,
 185                                 (u64 __percpu *) ca->usage[idx], u64s);
 186                 percpu_memset(ca->usage[idx], 0, u64s * sizeof(u64));
 187         }
 188         rcu_read_unlock();
 189
 190         write_seqcount_end(&c->usage_lock);
 191         preempt_enable();
 192 }
 193
 194 void bch2_fs_usage_to_text(struct printbuf *out,
 195                            struct bch_fs *c,
 196                            struct bch_fs_usage_online *fs_usage)
 197 {
 198         unsigned i;
 199
 200         pr_buf(out, "capacity:\t\t\t%llu\n", c->capacity);
 201
 202         pr_buf(out, "hidden:\t\t\t\t%llu\n",
 203                fs_usage->u.hidden);
 204         pr_buf(out, "data:\t\t\t\t%llu\n",
 205                fs_usage->u.data);
 206         pr_buf(out, "cached:\t\t\t\t%llu\n",
 207                fs_usage->u.cached);
 208         pr_buf(out, "reserved:\t\t\t%llu\n",
 209                fs_usage->u.reserved);
 210         pr_buf(out, "nr_inodes:\t\t\t%llu\n",
 211                fs_usage->u.nr_inodes);
 212         pr_buf(out, "online reserved:\t\t%llu\n",
 213                fs_usage->online_reserved);
 214
 215         for (i = 0;
 216              i < ARRAY_SIZE(fs_usage->u.persistent_reserved);
 217              i++) {
 218                 pr_buf(out, "%u replicas:\n", i + 1);
 219                 pr_buf(out, "\treserved:\t\t%llu\n",
 220                        fs_usage->u.persistent_reserved[i]);
 221         }
 222
 223         for (i = 0; i < c->replicas.nr; i++) {
 224                 struct bch_replicas_entry *e =
 225                         cpu_replicas_entry(&c->replicas, i);
 226
 227                 pr_buf(out, "\t");
 228                 bch2_replicas_entry_to_text(out, e);
 229                 pr_buf(out, ":\t%llu\n", fs_usage->u.replicas[i]);
 230         }
 231 }
 232
 233 static u64 reserve_factor(u64 r)
 234 {
 235         return r + (round_up(r, (1 << RESERVE_FACTOR)) >> RESERVE_FACTOR);
 236 }
 237
 238 u64 bch2_fs_sectors_used(struct bch_fs *c, struct bch_fs_usage_online *fs_usage)
 239 {
 240         return min(fs_usage->u.hidden +
 241                    fs_usage->u.btree +
 242                    fs_usage->u.data +
 243                    reserve_factor(fs_usage->u.reserved +
 244                                   fs_usage->online_reserved),
 245                    c->capacity);
 246 }
 247
 248 static struct bch_fs_usage_short
 249 __bch2_fs_usage_read_short(struct bch_fs *c)
 250 {
 251         struct bch_fs_usage_short ret;
 252         u64 data, reserved;
 253
 254         ret.capacity = c->capacity -
 255                 bch2_fs_usage_read_one(c, &c->usage_base->hidden);
 256
 257         data            = bch2_fs_usage_read_one(c, &c->usage_base->data) +
 258                 bch2_fs_usage_read_one(c, &c->usage_base->btree);
 259         reserved        = bch2_fs_usage_read_one(c, &c->usage_base->reserved) +
 260                 percpu_u64_get(c->online_reserved);
 261
 262         ret.used        = min(ret.capacity, data + reserve_factor(reserved));
 263         ret.free        = ret.capacity - ret.used;
 264
 265         ret.nr_inodes   = bch2_fs_usage_read_one(c, &c->usage_base->nr_inodes);
 266
 267         return ret;
 268 }
 269
 270 struct bch_fs_usage_short
 271 bch2_fs_usage_read_short(struct bch_fs *c)
 272 {
 273         struct bch_fs_usage_short ret;
 274
 275         percpu_down_read(&c->mark_lock);
 276         ret = __bch2_fs_usage_read_short(c);
 277         percpu_up_read(&c->mark_lock);
 278
 279         return ret;
 280 }
 281
 282 static inline int is_unavailable_bucket(struct bkey_alloc_unpacked a)
 283 {
 284         return a.dirty_sectors || a.stripe;
 285 }
 286
 287 static inline int bucket_sectors_fragmented(struct bch_dev *ca,
 288                                             struct bkey_alloc_unpacked a)
 289 {
 290         return a.dirty_sectors
 291                 ? max(0, (int) ca->mi.bucket_size - (int) a.dirty_sectors)
 292                 : 0;
 293 }
 294
 295 static inline enum bch_data_type bucket_type(struct bkey_alloc_unpacked a)
 296 {
 297         return a.cached_sectors && !a.dirty_sectors
 298                 ? BCH_DATA_cached
 299                 : a.data_type;
 300 }
 301
 302 static inline void account_bucket(struct bch_fs_usage *fs_usage,
 303                                   struct bch_dev_usage *dev_usage,
 304                                   enum bch_data_type type,
 305                                   int nr, s64 size)
 306 {
 307         if (type == BCH_DATA_sb || type == BCH_DATA_journal)
 308                 fs_usage->hidden        += size;
 309
 310         dev_usage->d[type].buckets      += nr;
 311 }
 312
 313 static void bch2_dev_usage_update(struct bch_fs *c, struct bch_dev *ca,
 314                                   struct bkey_alloc_unpacked old,
 315                                   struct bkey_alloc_unpacked new,
 316                                   u64 journal_seq, bool gc)
 317 {
 318         struct bch_fs_usage *fs_usage;
 319         struct bch_dev_usage *u;
 320
 321         preempt_disable();
 322         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 323         u = dev_usage_ptr(ca, journal_seq, gc);
 324
 325         if (bucket_type(old))
 326                 account_bucket(fs_usage, u, bucket_type(old),
 327                                -1, -ca->mi.bucket_size);
 328
 329         if (bucket_type(new))
 330                 account_bucket(fs_usage, u, bucket_type(new),
 331                                1, ca->mi.bucket_size);
 332
 333         u->buckets_ec += (int) new.stripe - (int) old.stripe;
 334         u->buckets_unavailable +=
 335                 is_unavailable_bucket(new) - is_unavailable_bucket(old);
 336
 337         u->d[old.data_type].sectors -= old.dirty_sectors;
 338         u->d[new.data_type].sectors += new.dirty_sectors;
 339         u->d[BCH_DATA_cached].sectors +=
 340                 (int) new.cached_sectors - (int) old.cached_sectors;
 341
 342         u->d[old.data_type].fragmented -= bucket_sectors_fragmented(ca, old);
 343         u->d[new.data_type].fragmented += bucket_sectors_fragmented(ca, new);
 344
 345         preempt_enable();
 346 }
 347
 348 static void bch2_dev_usage_update_m(struct bch_fs *c, struct bch_dev *ca,
 349                                     struct bucket old, struct bucket new,
 350                                     u64 journal_seq, bool gc)
 351 {
 352         struct bkey_alloc_unpacked old_a = {
 353                 .gen            = old.gen,
 354                 .data_type      = old.data_type,
 355                 .dirty_sectors  = old.dirty_sectors,
 356                 .cached_sectors = old.cached_sectors,
 357                 .stripe         = old.stripe,
 358         };
 359         struct bkey_alloc_unpacked new_a = {
 360                 .gen            = new.gen,
 361                 .data_type      = new.data_type,
 362                 .dirty_sectors  = new.dirty_sectors,
 363                 .cached_sectors = new.cached_sectors,
 364                 .stripe         = new.stripe,
 365         };
 366
 367         bch2_dev_usage_update(c, ca, old_a, new_a, journal_seq, gc);
 368 }
 369
 370 static inline int __update_replicas(struct bch_fs *c,
 371                                     struct bch_fs_usage *fs_usage,
 372                                     struct bch_replicas_entry *r,
 373                                     s64 sectors)
 374 {
 375         int idx = bch2_replicas_entry_idx(c, r);
 376
 377         if (idx < 0)
 378                 return -1;
 379
 380         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 381         fs_usage->replicas[idx]         += sectors;
 382         return 0;
 383 }
 384
 385 static inline int update_replicas(struct bch_fs *c, struct bkey_s_c k,
 386                         struct bch_replicas_entry *r, s64 sectors,
 387                         unsigned journal_seq, bool gc)
 388 {
 389         struct bch_fs_usage __percpu *fs_usage;
 390         int idx, ret = 0;
 391         struct printbuf buf = PRINTBUF;
 392
 393         percpu_down_read(&c->mark_lock);
 394         buf.atomic++;
 395
 396         idx = bch2_replicas_entry_idx(c, r);
 397         if (idx < 0 &&
 398             (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
 399              fsck_err(c, "no replicas entry\n"
 400                       "  while marking %s",
 401                       (bch2_bkey_val_to_text(&buf, c, k), buf.buf)))) {
 402                 percpu_up_read(&c->mark_lock);
 403                 ret = bch2_mark_replicas(c, r);
 404                 percpu_down_read(&c->mark_lock);
 405
 406                 if (ret)
 407                         goto err;
 408                 idx = bch2_replicas_entry_idx(c, r);
 409         }
 410         if (idx < 0) {
 411                 ret = -1;
 412                 goto err;
 413         }
 414
 415         preempt_disable();
 416         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 417         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 418         fs_usage->replicas[idx]         += sectors;
 419         preempt_enable();
 420 err:
 421 fsck_err:
 422         percpu_up_read(&c->mark_lock);
 423         printbuf_exit(&buf);
 424         return ret;
 425 }
 426
 427 static inline int update_cached_sectors(struct bch_fs *c,
 428                         struct bkey_s_c k,
 429                         unsigned dev, s64 sectors,
 430                         unsigned journal_seq, bool gc)
 431 {
 432         struct bch_replicas_padded r;
 433
 434         bch2_replicas_entry_cached(&r.e, dev);
 435
 436         return update_replicas(c, k, &r.e, sectors, journal_seq, gc);
 437 }
 438
 439 static struct replicas_delta_list *
 440 replicas_deltas_realloc(struct btree_trans *trans, unsigned more)
 441 {
 442         struct replicas_delta_list *d = trans->fs_usage_deltas;
 443         unsigned new_size = d ? (d->size + more) * 2 : 128;
 444         unsigned alloc_size = sizeof(*d) + new_size;
 445
 446         WARN_ON_ONCE(alloc_size > REPLICAS_DELTA_LIST_MAX);
 447
 448         if (!d || d->used + more > d->size) {
 449                 d = krealloc(d, alloc_size, GFP_NOIO|__GFP_ZERO);
 450
 451                 BUG_ON(!d && alloc_size > REPLICAS_DELTA_LIST_MAX);
 452
 453                 if (!d) {
 454                         d = mempool_alloc(&trans->c->replicas_delta_pool, GFP_NOIO);
 455                         memset(d, 0, REPLICAS_DELTA_LIST_MAX);
 456
 457                         if (trans->fs_usage_deltas)
 458                                 memcpy(d, trans->fs_usage_deltas,
 459                                        trans->fs_usage_deltas->size + sizeof(*d));
 460
 461                         new_size = REPLICAS_DELTA_LIST_MAX - sizeof(*d);
 462                         kfree(trans->fs_usage_deltas);
 463                 }
 464
 465                 d->size = new_size;
 466                 trans->fs_usage_deltas = d;
 467         }
 468         return d;
 469 }
 470
 471 static inline void update_replicas_list(struct btree_trans *trans,
 472                                         struct bch_replicas_entry *r,
 473                                         s64 sectors)
 474 {
 475         struct replicas_delta_list *d;
 476         struct replicas_delta *n;
 477         unsigned b;
 478
 479         if (!sectors)
 480                 return;
 481
 482         b = replicas_entry_bytes(r) + 8;
 483         d = replicas_deltas_realloc(trans, b);
 484
 485         n = (void *) d->d + d->used;
 486         n->delta = sectors;
 487         memcpy(&n->r, r, replicas_entry_bytes(r));
 488         bch2_replicas_entry_sort(&n->r);
 489         d->used += b;
 490 }
 491
 492 static inline void update_cached_sectors_list(struct btree_trans *trans,
 493                                               unsigned dev, s64 sectors)
 494 {
 495         struct bch_replicas_padded r;
 496
 497         bch2_replicas_entry_cached(&r.e, dev);
 498
 499         update_replicas_list(trans, &r.e, sectors);
 500 }
 501
 502 int bch2_mark_alloc(struct btree_trans *trans,
 503                     struct bkey_s_c old, struct bkey_s_c new,
 504                     unsigned flags)
 505 {
 506         bool gc = flags & BTREE_TRIGGER_GC;
 507         u64 journal_seq = trans->journal_res.seq;
 508         struct bch_fs *c = trans->c;
 509         struct bkey_alloc_unpacked old_u = bch2_alloc_unpack(old);
 510         struct bkey_alloc_unpacked new_u = bch2_alloc_unpack(new);
 511         struct bch_dev *ca = bch_dev_bkey_exists(c, new_u.dev);
 512         int ret = 0;
 513
 514         if (bch2_trans_inconsistent_on(new_u.bucket < ca->mi.first_bucket ||
 515                                        new_u.bucket >= ca->mi.nbuckets, trans,
 516                                        "alloc key outside range of device's buckets"))
 517                 return -EIO;
 518
 519         /*
 520          * alloc btree is read in by bch2_alloc_read, not gc:
 521          */
 522         if ((flags & BTREE_TRIGGER_GC) &&
 523             !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
 524                 return 0;
 525
 526         if ((flags & BTREE_TRIGGER_INSERT) &&
 527             !old_u.data_type != !new_u.data_type &&
 528             new.k->type == KEY_TYPE_alloc_v3) {
 529                 struct bch_alloc_v3 *v = (struct bch_alloc_v3 *) new.v;
 530                 u64 old_journal_seq = le64_to_cpu(v->journal_seq);
 531
 532                 BUG_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                 new_u.journal_seq = !new_u.data_type &&
 540                         (journal_seq == old_journal_seq ||
 541                          bch2_journal_noflush_seq(&c->journal, old_journal_seq))
 542                         ? 0 : journal_seq;
 543                 v->journal_seq = cpu_to_le64(new_u.journal_seq);
 544         }
 545
 546         if (old_u.data_type && !new_u.data_type && new_u.journal_seq) {
 547                 ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 548                                 c->journal.flushed_seq_ondisk,
 549                                 new_u.dev, new_u.bucket,
 550                                 new_u.journal_seq);
 551                 if (ret) {
 552                         bch2_fs_fatal_error(c,
 553                                 "error setting bucket_needs_journal_commit: %i", ret);
 554                         return ret;
 555                 }
 556         }
 557
 558         if (!new_u.data_type &&
 559             (!new_u.journal_seq || new_u.journal_seq < c->journal.flushed_seq_ondisk))
 560                 closure_wake_up(&c->freelist_wait);
 561
 562         if ((flags & BTREE_TRIGGER_INSERT) &&
 563             new_u.need_discard &&
 564             !new_u.journal_seq)
 565                 bch2_do_discards(c);
 566
 567         if (!old_u.data_type &&
 568             new_u.data_type &&
 569             should_invalidate_buckets(ca))
 570                 bch2_do_invalidates(c);
 571
 572         if (bucket_state(new_u) == BUCKET_need_gc_gens) {
 573                 atomic_inc(&c->kick_gc);
 574                 wake_up_process(c->gc_thread);
 575         }
 576
 577         percpu_down_read(&c->mark_lock);
 578         if (!gc && new_u.gen != old_u.gen)
 579                 *bucket_gen(ca, new_u.bucket) = new_u.gen;
 580
 581         bch2_dev_usage_update(c, ca, old_u, new_u, journal_seq, gc);
 582
 583         if (gc) {
 584                 struct bucket *g = gc_bucket(ca, new_u.bucket);
 585
 586                 bucket_lock(g);
 587
 588                 g->gen_valid            = 1;
 589                 g->gen                  = new_u.gen;
 590                 g->data_type            = new_u.data_type;
 591                 g->stripe               = new_u.stripe;
 592                 g->stripe_redundancy    = new_u.stripe_redundancy;
 593                 g->dirty_sectors        = new_u.dirty_sectors;
 594                 g->cached_sectors       = new_u.cached_sectors;
 595
 596                 bucket_unlock(g);
 597         }
 598         percpu_up_read(&c->mark_lock);
 599
 600         /*
 601          * need to know if we're getting called from the invalidate path or
 602          * not:
 603          */
 604
 605         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 606             old_u.cached_sectors) {
 607                 ret = update_cached_sectors(c, new, ca->dev_idx,
 608                                             -old_u.cached_sectors,
 609                                             journal_seq, gc);
 610                 if (ret) {
 611                         bch2_fs_fatal_error(c, "bch2_mark_alloc(): no replicas entry while updating cached sectors");
 612                         return ret;
 613                 }
 614
 615                 trace_invalidate(ca, bucket_to_sector(ca, new_u.bucket),
 616                                  old_u.cached_sectors);
 617         }
 618
 619         return 0;
 620 }
 621
 622 void bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 623                                size_t b, enum bch_data_type data_type,
 624                                unsigned sectors, struct gc_pos pos,
 625                                unsigned flags)
 626 {
 627         struct bucket old, new, *g;
 628         bool overflow;
 629
 630         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 631         BUG_ON(data_type != BCH_DATA_sb &&
 632                data_type != BCH_DATA_journal);
 633
 634         /*
 635          * Backup superblock might be past the end of our normal usable space:
 636          */
 637         if (b >= ca->mi.nbuckets)
 638                 return;
 639
 640         percpu_down_read(&c->mark_lock);
 641         g = gc_bucket(ca, b);
 642
 643         bucket_lock(g);
 644         old = *g;
 645
 646         g->data_type = data_type;
 647         g->dirty_sectors += sectors;
 648         overflow = g->dirty_sectors < sectors;
 649
 650         new = *g;
 651         bucket_unlock(g);
 652
 653         bch2_fs_inconsistent_on(old.data_type &&
 654                                 old.data_type != data_type, c,
 655                 "different types of data in same bucket: %s, %s",
 656                 bch2_data_types[old.data_type],
 657                 bch2_data_types[data_type]);
 658
 659         bch2_fs_inconsistent_on(overflow, c,
 660                 "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > U16_MAX",
 661                 ca->dev_idx, b, new.gen,
 662                 bch2_data_types[old.data_type ?: data_type],
 663                 old.dirty_sectors, sectors);
 664
 665         bch2_dev_usage_update_m(c, ca, old, new, 0, true);
 666         percpu_up_read(&c->mark_lock);
 667 }
 668
 669 static s64 ptr_disk_sectors(s64 sectors, struct extent_ptr_decoded p)
 670 {
 671         EBUG_ON(sectors < 0);
 672
 673         return p.crc.compression_type &&
 674                 p.crc.compression_type != BCH_COMPRESSION_TYPE_incompressible
 675                 ? DIV_ROUND_UP_ULL(sectors * p.crc.compressed_size,
 676                                p.crc.uncompressed_size)
 677                 : sectors;
 678 }
 679
 680 static int check_bucket_ref(struct bch_fs *c,
 681                             struct bkey_s_c k,
 682                             const struct bch_extent_ptr *ptr,
 683                             s64 sectors, enum bch_data_type ptr_data_type,
 684                             u8 b_gen, u8 bucket_data_type,
 685                             u32 dirty_sectors, u32 cached_sectors)
 686 {
 687         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 688         size_t bucket_nr = PTR_BUCKET_NR(ca, ptr);
 689         u16 bucket_sectors = !ptr->cached
 690                 ? dirty_sectors
 691                 : cached_sectors;
 692         struct printbuf buf = PRINTBUF;
 693         int ret = 0;
 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 err;
 737         }
 738
 739         if (bucket_data_type && ptr_data_type &&
 740             bucket_data_type != ptr_data_type) {
 741                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 742                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 743                         "while marking %s",
 744                         ptr->dev, bucket_nr, b_gen,
 745                         bch2_data_types[bucket_data_type],
 746                         bch2_data_types[ptr_data_type],
 747                         (printbuf_reset(&buf),
 748                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 749                 ret = -EIO;
 750                 goto err;
 751         }
 752
 753         if ((unsigned) (bucket_sectors + sectors) > U32_MAX) {
 754                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 755                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U16_MAX\n"
 756                         "while marking %s",
 757                         ptr->dev, bucket_nr, b_gen,
 758                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 759                         bucket_sectors, sectors,
 760                         (printbuf_reset(&buf),
 761                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 762                 ret = -EIO;
 763                 goto err;
 764         }
 765 err:
 766         printbuf_exit(&buf);
 767         return ret;
 768 }
 769
 770 static int mark_stripe_bucket(struct btree_trans *trans,
 771                               struct bkey_s_c k,
 772                               unsigned ptr_idx,
 773                               unsigned flags)
 774 {
 775         struct bch_fs *c = trans->c;
 776         u64 journal_seq = trans->journal_res.seq;
 777         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 778         unsigned nr_data = s->nr_blocks - s->nr_redundant;
 779         bool parity = ptr_idx >= nr_data;
 780         enum bch_data_type data_type = parity ? BCH_DATA_parity : 0;
 781         s64 sectors = parity ? le16_to_cpu(s->sectors) : 0;
 782         const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
 783         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 784         struct bucket old, new, *g;
 785         struct printbuf buf = PRINTBUF;
 786         int ret = 0;
 787
 788         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 789
 790         /* * XXX doesn't handle deletion */
 791
 792         percpu_down_read(&c->mark_lock);
 793         buf.atomic++;
 794         g = PTR_GC_BUCKET(ca, ptr);
 795
 796         if (g->dirty_sectors ||
 797             (g->stripe && g->stripe != k.k->p.offset)) {
 798                 bch2_fs_inconsistent(c,
 799                               "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
 800                               ptr->dev, PTR_BUCKET_NR(ca, ptr), g->gen,
 801                               (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 802                 ret = -EINVAL;
 803                 goto err;
 804         }
 805
 806         bucket_lock(g);
 807         old = *g;
 808
 809         ret = check_bucket_ref(c, k, ptr, sectors, data_type,
 810                                new.gen, new.data_type,
 811                                new.dirty_sectors, new.cached_sectors);
 812         if (ret) {
 813                 bucket_unlock(g);
 814                 goto err;
 815         }
 816
 817         new.dirty_sectors += sectors;
 818         if (data_type)
 819                 new.data_type = data_type;
 820
 821         g->stripe               = k.k->p.offset;
 822         g->stripe_redundancy    = s->nr_redundant;
 823
 824         new = *g;
 825         bucket_unlock(g);
 826
 827         bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 828 err:
 829         percpu_up_read(&c->mark_lock);
 830         printbuf_exit(&buf);
 831         return ret;
 832 }
 833
 834 static int __mark_pointer(struct btree_trans *trans,
 835                           struct bkey_s_c k,
 836                           const struct bch_extent_ptr *ptr,
 837                           s64 sectors, enum bch_data_type ptr_data_type,
 838                           u8 bucket_gen, u8 *bucket_data_type,
 839                           u32 *dirty_sectors, u32 *cached_sectors)
 840 {
 841         u32 *dst_sectors = !ptr->cached
 842                 ? dirty_sectors
 843                 : cached_sectors;
 844         int ret = check_bucket_ref(trans->c, k, ptr, sectors, ptr_data_type,
 845                                    bucket_gen, *bucket_data_type,
 846                                    *dirty_sectors, *cached_sectors);
 847
 848         if (ret)
 849                 return ret;
 850
 851         *dst_sectors += sectors;
 852         *bucket_data_type = *dirty_sectors || *cached_sectors
 853                 ? ptr_data_type : 0;
 854         return 0;
 855 }
 856
 857 static int bch2_mark_pointer(struct btree_trans *trans,
 858                              struct bkey_s_c k,
 859                              struct extent_ptr_decoded p,
 860                              s64 sectors, enum bch_data_type data_type,
 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         u8 bucket_data_type;
 868         int ret = 0;
 869
 870         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 871
 872         percpu_down_read(&c->mark_lock);
 873         g = PTR_GC_BUCKET(ca, &p.ptr);
 874
 875         bucket_lock(g);
 876         old = *g;
 877
 878         bucket_data_type = g->data_type;
 879
 880         ret = __mark_pointer(trans, k, &p.ptr, sectors,
 881                              data_type, g->gen,
 882                              &bucket_data_type,
 883                              &g->dirty_sectors,
 884                              &g->cached_sectors);
 885         if (ret) {
 886                 bucket_unlock(g);
 887                 goto err;
 888         }
 889
 890         g->data_type = bucket_data_type;
 891
 892         new = *g;
 893         bucket_unlock(g);
 894
 895         bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 896 err:
 897         percpu_up_read(&c->mark_lock);
 898
 899         return ret;
 900 }
 901
 902 static int bch2_mark_stripe_ptr(struct btree_trans *trans,
 903                                 struct bkey_s_c k,
 904                                 struct bch_extent_stripe_ptr p,
 905                                 enum bch_data_type data_type,
 906                                 s64 sectors,
 907                                 unsigned flags)
 908 {
 909         struct bch_fs *c = trans->c;
 910         struct bch_replicas_padded r;
 911         struct gc_stripe *m;
 912
 913         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 914
 915         m = genradix_ptr_alloc(&c->gc_stripes, p.idx, GFP_KERNEL);
 916         if (!m) {
 917                 bch_err(c, "error allocating memory for gc_stripes, idx %llu",
 918                         (u64) p.idx);
 919                 return -ENOMEM;
 920         }
 921
 922         spin_lock(&c->ec_stripes_heap_lock);
 923
 924         if (!m || !m->alive) {
 925                 spin_unlock(&c->ec_stripes_heap_lock);
 926                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
 927                                     (u64) p.idx);
 928                 bch2_inconsistent_error(c);
 929                 return -EIO;
 930         }
 931
 932         m->block_sectors[p.block] += sectors;
 933
 934         r = m->r;
 935         spin_unlock(&c->ec_stripes_heap_lock);
 936
 937         r.e.data_type = data_type;
 938         update_replicas(c, k, &r.e, sectors, trans->journal_res.seq, true);
 939
 940         return 0;
 941 }
 942
 943 int bch2_mark_extent(struct btree_trans *trans,
 944                      struct bkey_s_c old, struct bkey_s_c new,
 945                      unsigned flags)
 946 {
 947         u64 journal_seq = trans->journal_res.seq;
 948         struct bch_fs *c = trans->c;
 949         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
 950         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 951         const union bch_extent_entry *entry;
 952         struct extent_ptr_decoded p;
 953         struct bch_replicas_padded r;
 954         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
 955                 ? BCH_DATA_btree
 956                 : BCH_DATA_user;
 957         s64 sectors = bkey_is_btree_ptr(k.k)
 958                 ? btree_sectors(c)
 959                 : k.k->size;
 960         s64 dirty_sectors = 0;
 961         bool stale;
 962         int ret;
 963
 964         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 965
 966         r.e.data_type   = data_type;
 967         r.e.nr_devs     = 0;
 968         r.e.nr_required = 1;
 969
 970         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 971                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
 972
 973                 if (flags & BTREE_TRIGGER_OVERWRITE)
 974                         disk_sectors = -disk_sectors;
 975
 976                 ret = bch2_mark_pointer(trans, k, p, disk_sectors,
 977                                         data_type, flags);
 978                 if (ret < 0)
 979                         return ret;
 980
 981                 stale = ret > 0;
 982
 983                 if (p.ptr.cached) {
 984                         if (!stale) {
 985                                 ret = update_cached_sectors(c, k, p.ptr.dev,
 986                                                 disk_sectors, journal_seq, true);
 987                                 if (ret) {
 988                                         bch2_fs_fatal_error(c, "bch2_mark_extent(): no replicas entry while updating cached sectors");
 989                                         return ret;
 990                                 }
 991                         }
 992                 } else if (!p.has_ec) {
 993                         dirty_sectors          += disk_sectors;
 994                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
 995                 } else {
 996                         ret = bch2_mark_stripe_ptr(trans, k, p.ec, data_type,
 997                                         disk_sectors, flags);
 998                         if (ret)
 999                                 return ret;
1000
1001                         /*
1002                          * There may be other dirty pointers in this extent, but
1003                          * if so they're not required for mounting if we have an
1004                          * erasure coded pointer in this extent:
1005                          */
1006                         r.e.nr_required = 0;
1007                 }
1008         }
1009
1010         if (r.e.nr_devs) {
1011                 ret = update_replicas(c, k, &r.e, dirty_sectors, journal_seq, true);
1012                 if (ret) {
1013                         struct printbuf buf = PRINTBUF;
1014
1015                         bch2_bkey_val_to_text(&buf, c, k);
1016                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
1017                         printbuf_exit(&buf);
1018                         return ret;
1019                 }
1020         }
1021
1022         return 0;
1023 }
1024
1025 int bch2_mark_stripe(struct btree_trans *trans,
1026                      struct bkey_s_c old, struct bkey_s_c new,
1027                      unsigned flags)
1028 {
1029         bool gc = flags & BTREE_TRIGGER_GC;
1030         u64 journal_seq = trans->journal_res.seq;
1031         struct bch_fs *c = trans->c;
1032         u64 idx = new.k->p.offset;
1033         const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
1034                 ? bkey_s_c_to_stripe(old).v : NULL;
1035         const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
1036                 ? bkey_s_c_to_stripe(new).v : NULL;
1037         unsigned i;
1038         int ret;
1039
1040         BUG_ON(gc && old_s);
1041
1042         if (!gc) {
1043                 struct stripe *m = genradix_ptr(&c->stripes, idx);
1044
1045                 if (!m || (old_s && !m->alive)) {
1046                         struct printbuf buf1 = PRINTBUF;
1047                         struct printbuf buf2 = PRINTBUF;
1048
1049                         bch2_bkey_val_to_text(&buf1, c, old);
1050                         bch2_bkey_val_to_text(&buf2, c, new);
1051                         bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
1052                                             "old %s\n"
1053                                             "new %s", idx, buf1.buf, buf2.buf);
1054                         printbuf_exit(&buf2);
1055                         printbuf_exit(&buf1);
1056                         bch2_inconsistent_error(c);
1057                         return -1;
1058                 }
1059
1060                 if (!new_s) {
1061                         spin_lock(&c->ec_stripes_heap_lock);
1062                         bch2_stripes_heap_del(c, m, idx);
1063                         spin_unlock(&c->ec_stripes_heap_lock);
1064
1065                         memset(m, 0, sizeof(*m));
1066                 } else {
1067                         m->alive        = true;
1068                         m->sectors      = le16_to_cpu(new_s->sectors);
1069                         m->algorithm    = new_s->algorithm;
1070                         m->nr_blocks    = new_s->nr_blocks;
1071                         m->nr_redundant = new_s->nr_redundant;
1072                         m->blocks_nonempty = 0;
1073
1074                         for (i = 0; i < new_s->nr_blocks; i++)
1075                                 m->blocks_nonempty += !!stripe_blockcount_get(new_s, i);
1076
1077                         spin_lock(&c->ec_stripes_heap_lock);
1078                         bch2_stripes_heap_update(c, m, idx);
1079                         spin_unlock(&c->ec_stripes_heap_lock);
1080                 }
1081         } else {
1082                 struct gc_stripe *m =
1083                         genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
1084
1085                 if (!m) {
1086                         bch_err(c, "error allocating memory for gc_stripes, idx %llu",
1087                                 idx);
1088                         return -ENOMEM;
1089                 }
1090                 /*
1091                  * This will be wrong when we bring back runtime gc: we should
1092                  * be unmarking the old key and then marking the new key
1093                  */
1094                 m->alive        = true;
1095                 m->sectors      = le16_to_cpu(new_s->sectors);
1096                 m->nr_blocks    = new_s->nr_blocks;
1097                 m->nr_redundant = new_s->nr_redundant;
1098
1099                 for (i = 0; i < new_s->nr_blocks; i++)
1100                         m->ptrs[i] = new_s->ptrs[i];
1101
1102                 bch2_bkey_to_replicas(&m->r.e, new);
1103
1104                 /*
1105                  * gc recalculates this field from stripe ptr
1106                  * references:
1107                  */
1108                 memset(m->block_sectors, 0, sizeof(m->block_sectors));
1109
1110                 for (i = 0; i < new_s->nr_blocks; i++) {
1111                         ret = mark_stripe_bucket(trans, new, i, flags);
1112                         if (ret)
1113                                 return ret;
1114                 }
1115
1116                 ret = update_replicas(c, new, &m->r.e,
1117                                       ((s64) m->sectors * m->nr_redundant),
1118                                       journal_seq, gc);
1119                 if (ret) {
1120                         struct printbuf buf = PRINTBUF;
1121
1122                         bch2_bkey_val_to_text(&buf, c, new);
1123                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
1124                         printbuf_exit(&buf);
1125                         return ret;
1126                 }
1127         }
1128
1129         return 0;
1130 }
1131
1132 int bch2_mark_inode(struct btree_trans *trans,
1133                     struct bkey_s_c old, struct bkey_s_c new,
1134                     unsigned flags)
1135 {
1136         struct bch_fs *c = trans->c;
1137         struct bch_fs_usage __percpu *fs_usage;
1138         u64 journal_seq = trans->journal_res.seq;
1139
1140         if (flags & BTREE_TRIGGER_INSERT) {
1141                 struct bch_inode_v2 *v = (struct bch_inode_v2 *) new.v;
1142
1143                 BUG_ON(!journal_seq);
1144                 BUG_ON(new.k->type != KEY_TYPE_inode_v2);
1145
1146                 v->bi_journal_seq = cpu_to_le64(journal_seq);
1147         }
1148
1149         if (flags & BTREE_TRIGGER_GC) {
1150                 percpu_down_read(&c->mark_lock);
1151                 preempt_disable();
1152
1153                 fs_usage = fs_usage_ptr(c, journal_seq, flags & BTREE_TRIGGER_GC);
1154                 fs_usage->nr_inodes += bkey_is_inode(new.k);
1155                 fs_usage->nr_inodes -= bkey_is_inode(old.k);
1156
1157                 preempt_enable();
1158                 percpu_up_read(&c->mark_lock);
1159         }
1160         return 0;
1161 }
1162
1163 int bch2_mark_reservation(struct btree_trans *trans,
1164                           struct bkey_s_c old, struct bkey_s_c new,
1165                           unsigned flags)
1166 {
1167         struct bch_fs *c = trans->c;
1168         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1169         struct bch_fs_usage __percpu *fs_usage;
1170         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1171         s64 sectors = (s64) k.k->size;
1172
1173         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1174
1175         if (flags & BTREE_TRIGGER_OVERWRITE)
1176                 sectors = -sectors;
1177         sectors *= replicas;
1178
1179         percpu_down_read(&c->mark_lock);
1180         preempt_disable();
1181
1182         fs_usage = fs_usage_ptr(c, trans->journal_res.seq, flags & BTREE_TRIGGER_GC);
1183         replicas = clamp_t(unsigned, replicas, 1,
1184                            ARRAY_SIZE(fs_usage->persistent_reserved));
1185
1186         fs_usage->reserved                              += sectors;
1187         fs_usage->persistent_reserved[replicas - 1]     += sectors;
1188
1189         preempt_enable();
1190         percpu_up_read(&c->mark_lock);
1191
1192         return 0;
1193 }
1194
1195 static s64 __bch2_mark_reflink_p(struct btree_trans *trans,
1196                                  struct bkey_s_c_reflink_p p,
1197                                  u64 start, u64 end,
1198                                  u64 *idx, unsigned flags, size_t r_idx)
1199 {
1200         struct bch_fs *c = trans->c;
1201         struct reflink_gc *r;
1202         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1203         u64 next_idx = end;
1204         s64 ret = 0;
1205         struct printbuf buf = PRINTBUF;
1206
1207         if (r_idx >= c->reflink_gc_nr)
1208                 goto not_found;
1209
1210         r = genradix_ptr(&c->reflink_gc_table, r_idx);
1211         next_idx = min(next_idx, r->offset - r->size);
1212         if (*idx < next_idx)
1213                 goto not_found;
1214
1215         BUG_ON((s64) r->refcount + add < 0);
1216
1217         r->refcount += add;
1218         *idx = r->offset;
1219         return 0;
1220 not_found:
1221         if (fsck_err(c, "pointer to missing indirect extent\n"
1222                      "  %s\n"
1223                      "  missing range %llu-%llu",
1224                      (bch2_bkey_val_to_text(&buf, c, p.s_c), buf.buf),
1225                      *idx, next_idx)) {
1226                 struct bkey_i_error new;
1227
1228                 bkey_init(&new.k);
1229                 new.k.type      = KEY_TYPE_error;
1230                 new.k.p         = bkey_start_pos(p.k);
1231                 new.k.p.offset += *idx - start;
1232                 bch2_key_resize(&new.k, next_idx - *idx);
1233                 ret = __bch2_btree_insert(trans, BTREE_ID_extents, &new.k_i);
1234         }
1235
1236         *idx = next_idx;
1237 fsck_err:
1238         printbuf_exit(&buf);
1239         return ret;
1240 }
1241
1242 int bch2_mark_reflink_p(struct btree_trans *trans,
1243                         struct bkey_s_c old, struct bkey_s_c new,
1244                         unsigned flags)
1245 {
1246         struct bch_fs *c = trans->c;
1247         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1248         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1249         struct reflink_gc *ref;
1250         size_t l, r, m;
1251         u64 idx = le64_to_cpu(p.v->idx), start = idx;
1252         u64 end = le64_to_cpu(p.v->idx) + p.k->size;
1253         int ret = 0;
1254
1255         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1256
1257         if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix) {
1258                 idx -= le32_to_cpu(p.v->front_pad);
1259                 end += le32_to_cpu(p.v->back_pad);
1260         }
1261
1262         l = 0;
1263         r = c->reflink_gc_nr;
1264         while (l < r) {
1265                 m = l + (r - l) / 2;
1266
1267                 ref = genradix_ptr(&c->reflink_gc_table, m);
1268                 if (ref->offset <= idx)
1269                         l = m + 1;
1270                 else
1271                         r = m;
1272         }
1273
1274         while (idx < end && !ret)
1275                 ret = __bch2_mark_reflink_p(trans, p, start, end,
1276                                             &idx, flags, l++);
1277
1278         return ret;
1279 }
1280
1281 static noinline __cold
1282 void fs_usage_apply_warn(struct btree_trans *trans,
1283                          unsigned disk_res_sectors,
1284                          s64 should_not_have_added)
1285 {
1286         struct bch_fs *c = trans->c;
1287         struct btree_insert_entry *i;
1288         struct printbuf buf = PRINTBUF;
1289
1290         bch_err(c, "disk usage increased %lli more than %u sectors reserved",
1291                 should_not_have_added, disk_res_sectors);
1292
1293         trans_for_each_update(trans, i) {
1294                 struct bkey_s_c old = { &i->old_k, i->old_v };
1295
1296                 pr_err("while inserting");
1297                 printbuf_reset(&buf);
1298                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(i->k));
1299                 pr_err("  %s", buf.buf);
1300                 pr_err("overlapping with");
1301                 printbuf_reset(&buf);
1302                 bch2_bkey_val_to_text(&buf, c, old);
1303                 pr_err("  %s", buf.buf);
1304         }
1305
1306         __WARN();
1307         printbuf_exit(&buf);
1308 }
1309
1310 int bch2_trans_fs_usage_apply(struct btree_trans *trans,
1311                               struct replicas_delta_list *deltas)
1312 {
1313         struct bch_fs *c = trans->c;
1314         static int warned_disk_usage = 0;
1315         bool warn = false;
1316         unsigned disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1317         struct replicas_delta *d = deltas->d, *d2;
1318         struct replicas_delta *top = (void *) deltas->d + deltas->used;
1319         struct bch_fs_usage *dst;
1320         s64 added = 0, should_not_have_added;
1321         unsigned i;
1322
1323         percpu_down_read(&c->mark_lock);
1324         preempt_disable();
1325         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1326
1327         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1328                 switch (d->r.data_type) {
1329                 case BCH_DATA_btree:
1330                 case BCH_DATA_user:
1331                 case BCH_DATA_parity:
1332                         added += d->delta;
1333                 }
1334
1335                 if (__update_replicas(c, dst, &d->r, d->delta))
1336                         goto need_mark;
1337         }
1338
1339         dst->nr_inodes += deltas->nr_inodes;
1340
1341         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1342                 added                           += deltas->persistent_reserved[i];
1343                 dst->reserved                   += deltas->persistent_reserved[i];
1344                 dst->persistent_reserved[i]     += deltas->persistent_reserved[i];
1345         }
1346
1347         /*
1348          * Not allowed to reduce sectors_available except by getting a
1349          * reservation:
1350          */
1351         should_not_have_added = added - (s64) disk_res_sectors;
1352         if (unlikely(should_not_have_added > 0)) {
1353                 u64 old, new, v = atomic64_read(&c->sectors_available);
1354
1355                 do {
1356                         old = v;
1357                         new = max_t(s64, 0, old - should_not_have_added);
1358                 } while ((v = atomic64_cmpxchg(&c->sectors_available,
1359                                                old, new)) != old);
1360
1361                 added -= should_not_have_added;
1362                 warn = true;
1363         }
1364
1365         if (added > 0) {
1366                 trans->disk_res->sectors -= added;
1367                 this_cpu_sub(*c->online_reserved, added);
1368         }
1369
1370         preempt_enable();
1371         percpu_up_read(&c->mark_lock);
1372
1373         if (unlikely(warn) && !xchg(&warned_disk_usage, 1))
1374                 fs_usage_apply_warn(trans, disk_res_sectors, should_not_have_added);
1375         return 0;
1376 need_mark:
1377         /* revert changes: */
1378         for (d2 = deltas->d; d2 != d; d2 = replicas_delta_next(d2))
1379                 BUG_ON(__update_replicas(c, dst, &d2->r, -d2->delta));
1380
1381         preempt_enable();
1382         percpu_up_read(&c->mark_lock);
1383         return -1;
1384 }
1385
1386 /* trans_mark: */
1387
1388 static int bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
1389                               const struct bch_extent_ptr *ptr,
1390                               struct bkey_alloc_unpacked *u)
1391 {
1392         struct bch_fs *c = trans->c;
1393         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1394         struct bkey_s_c k;
1395         int ret;
1396
1397         bch2_trans_iter_init(trans, iter, BTREE_ID_alloc,
1398                              POS(ptr->dev, PTR_BUCKET_NR(ca, ptr)),
1399                              BTREE_ITER_WITH_UPDATES|
1400                              BTREE_ITER_CACHED|
1401                              BTREE_ITER_INTENT);
1402         k = bch2_btree_iter_peek_slot(iter);
1403         ret = bkey_err(k);
1404         if (ret) {
1405                 bch2_trans_iter_exit(trans, iter);
1406                 return ret;
1407         }
1408
1409         *u = bch2_alloc_unpack(k);
1410         return 0;
1411 }
1412
1413 static int bch2_trans_mark_pointer(struct btree_trans *trans,
1414                         struct bkey_s_c k, struct extent_ptr_decoded p,
1415                         s64 sectors, enum bch_data_type data_type)
1416 {
1417         struct btree_iter iter;
1418         struct bkey_alloc_unpacked u;
1419         int ret;
1420
1421         ret = bch2_trans_start_alloc_update(trans, &iter, &p.ptr, &u);
1422         if (ret)
1423                 return ret;
1424
1425         ret = __mark_pointer(trans, k, &p.ptr, sectors, data_type,
1426                              u.gen, &u.data_type,
1427                              &u.dirty_sectors, &u.cached_sectors);
1428         if (ret)
1429                 goto out;
1430
1431         ret = bch2_alloc_write(trans, &iter, &u, 0);
1432         if (ret)
1433                 goto out;
1434 out:
1435         bch2_trans_iter_exit(trans, &iter);
1436         return ret;
1437 }
1438
1439 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1440                         struct extent_ptr_decoded p,
1441                         s64 sectors, enum bch_data_type data_type)
1442 {
1443         struct btree_iter iter;
1444         struct bkey_s_c k;
1445         struct bkey_i_stripe *s;
1446         struct bch_replicas_padded r;
1447         int ret = 0;
1448
1449         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes, POS(0, p.ec.idx),
1450                              BTREE_ITER_INTENT|
1451                              BTREE_ITER_WITH_UPDATES);
1452         k = bch2_btree_iter_peek_slot(&iter);
1453         ret = bkey_err(k);
1454         if (ret)
1455                 goto err;
1456
1457         if (k.k->type != KEY_TYPE_stripe) {
1458                 bch2_trans_inconsistent(trans,
1459                         "pointer to nonexistent stripe %llu",
1460                         (u64) p.ec.idx);
1461                 ret = -EIO;
1462                 goto err;
1463         }
1464
1465         if (!bch2_ptr_matches_stripe(bkey_s_c_to_stripe(k).v, p)) {
1466                 bch2_trans_inconsistent(trans,
1467                         "stripe pointer doesn't match stripe %llu",
1468                         (u64) p.ec.idx);
1469                 ret = -EIO;
1470                 goto err;
1471         }
1472
1473         s = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1474         ret = PTR_ERR_OR_ZERO(s);
1475         if (ret)
1476                 goto err;
1477
1478         bkey_reassemble(&s->k_i, k);
1479         stripe_blockcount_set(&s->v, p.ec.block,
1480                 stripe_blockcount_get(&s->v, p.ec.block) +
1481                 sectors);
1482
1483         ret = bch2_trans_update(trans, &iter, &s->k_i, 0);
1484         if (ret)
1485                 goto err;
1486
1487         bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1488         r.e.data_type = data_type;
1489         update_replicas_list(trans, &r.e, sectors);
1490 err:
1491         bch2_trans_iter_exit(trans, &iter);
1492         return ret;
1493 }
1494
1495 int bch2_trans_mark_extent(struct btree_trans *trans,
1496                            struct bkey_s_c old, struct bkey_i *new,
1497                            unsigned flags)
1498 {
1499         struct bch_fs *c = trans->c;
1500         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1501                 ? old
1502                 : bkey_i_to_s_c(new);
1503         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1504         const union bch_extent_entry *entry;
1505         struct extent_ptr_decoded p;
1506         struct bch_replicas_padded r;
1507         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
1508                 ? BCH_DATA_btree
1509                 : BCH_DATA_user;
1510         s64 sectors = bkey_is_btree_ptr(k.k)
1511                 ? btree_sectors(c)
1512                 : k.k->size;
1513         s64 dirty_sectors = 0;
1514         bool stale;
1515         int ret;
1516
1517         r.e.data_type   = data_type;
1518         r.e.nr_devs     = 0;
1519         r.e.nr_required = 1;
1520
1521         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1522                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
1523
1524                 if (flags & BTREE_TRIGGER_OVERWRITE)
1525                         disk_sectors = -disk_sectors;
1526
1527                 ret = bch2_trans_mark_pointer(trans, k, p,
1528                                         disk_sectors, data_type);
1529                 if (ret < 0)
1530                         return ret;
1531
1532                 stale = ret > 0;
1533
1534                 if (p.ptr.cached) {
1535                         if (!stale)
1536                                 update_cached_sectors_list(trans, p.ptr.dev,
1537                                                            disk_sectors);
1538                 } else if (!p.has_ec) {
1539                         dirty_sectors          += disk_sectors;
1540                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1541                 } else {
1542                         ret = bch2_trans_mark_stripe_ptr(trans, p,
1543                                         disk_sectors, data_type);
1544                         if (ret)
1545                                 return ret;
1546
1547                         r.e.nr_required = 0;
1548                 }
1549         }
1550
1551         if (r.e.nr_devs)
1552                 update_replicas_list(trans, &r.e, dirty_sectors);
1553
1554         return 0;
1555 }
1556
1557 static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
1558                                          struct bkey_s_c_stripe s,
1559                                          unsigned idx, bool deleting)
1560 {
1561         struct bch_fs *c = trans->c;
1562         const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
1563         struct btree_iter iter;
1564         struct bkey_alloc_unpacked u;
1565         enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
1566                 ? BCH_DATA_parity : 0;
1567         s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
1568         int ret = 0;
1569
1570         if (deleting)
1571                 sectors = -sectors;
1572
1573         ret = bch2_trans_start_alloc_update(trans, &iter, ptr, &u);
1574         if (ret)
1575                 return ret;
1576
1577         ret = check_bucket_ref(c, s.s_c, ptr, sectors, data_type,
1578                                u.gen, u.data_type,
1579                                u.dirty_sectors, u.cached_sectors);
1580         if (ret)
1581                 goto err;
1582
1583         if (!deleting) {
1584                 if (bch2_trans_inconsistent_on(u.stripe ||
1585                                             u.stripe_redundancy, trans,
1586                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
1587                                 iter.pos.inode, iter.pos.offset, u.gen,
1588                                 bch2_data_types[u.data_type],
1589                                 u.dirty_sectors,
1590                                 u.stripe, s.k->p.offset)) {
1591                         ret = -EIO;
1592                         goto err;
1593                 }
1594
1595                 if (bch2_trans_inconsistent_on(data_type && u.dirty_sectors, trans,
1596                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
1597                                 iter.pos.inode, iter.pos.offset, u.gen,
1598                                 bch2_data_types[u.data_type],
1599                                 u.dirty_sectors,
1600                                 s.k->p.offset)) {
1601                         ret = -EIO;
1602                         goto err;
1603                 }
1604
1605                 u.stripe                = s.k->p.offset;
1606                 u.stripe_redundancy     = s.v->nr_redundant;
1607         } else {
1608                 if (bch2_trans_inconsistent_on(u.stripe != s.k->p.offset ||
1609                                             u.stripe_redundancy != s.v->nr_redundant, trans,
1610                                 "bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
1611                                 iter.pos.inode, iter.pos.offset, u.gen,
1612                                 s.k->p.offset, u.stripe)) {
1613                         ret = -EIO;
1614                         goto err;
1615                 }
1616
1617                 u.stripe                = 0;
1618                 u.stripe_redundancy     = 0;
1619         }
1620
1621         u.dirty_sectors += sectors;
1622         if (data_type)
1623                 u.data_type = !deleting ? data_type : 0;
1624
1625         ret = bch2_alloc_write(trans, &iter, &u, 0);
1626         if (ret)
1627                 goto err;
1628 err:
1629         bch2_trans_iter_exit(trans, &iter);
1630         return ret;
1631 }
1632
1633 int bch2_trans_mark_stripe(struct btree_trans *trans,
1634                            struct bkey_s_c old, struct bkey_i *new,
1635                            unsigned flags)
1636 {
1637         const struct bch_stripe *old_s = NULL;
1638         struct bch_stripe *new_s = NULL;
1639         struct bch_replicas_padded r;
1640         unsigned i, nr_blocks;
1641         int ret = 0;
1642
1643         if (old.k->type == KEY_TYPE_stripe)
1644                 old_s = bkey_s_c_to_stripe(old).v;
1645         if (new->k.type == KEY_TYPE_stripe)
1646                 new_s = &bkey_i_to_stripe(new)->v;
1647
1648         /*
1649          * If the pointers aren't changing, we don't need to do anything:
1650          */
1651         if (new_s && old_s &&
1652             new_s->nr_blocks    == old_s->nr_blocks &&
1653             new_s->nr_redundant == old_s->nr_redundant &&
1654             !memcmp(old_s->ptrs, new_s->ptrs,
1655                     new_s->nr_blocks * sizeof(struct bch_extent_ptr)))
1656                 return 0;
1657
1658         BUG_ON(new_s && old_s &&
1659                (new_s->nr_blocks        != old_s->nr_blocks ||
1660                 new_s->nr_redundant     != old_s->nr_redundant));
1661
1662         nr_blocks = new_s ? new_s->nr_blocks : old_s->nr_blocks;
1663
1664         if (new_s) {
1665                 s64 sectors = le16_to_cpu(new_s->sectors);
1666
1667                 bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(new));
1668                 update_replicas_list(trans, &r.e, sectors * new_s->nr_redundant);
1669         }
1670
1671         if (old_s) {
1672                 s64 sectors = -((s64) le16_to_cpu(old_s->sectors));
1673
1674                 bch2_bkey_to_replicas(&r.e, old);
1675                 update_replicas_list(trans, &r.e, sectors * old_s->nr_redundant);
1676         }
1677
1678         for (i = 0; i < nr_blocks; i++) {
1679                 if (new_s && old_s &&
1680                     !memcmp(&new_s->ptrs[i],
1681                             &old_s->ptrs[i],
1682                             sizeof(new_s->ptrs[i])))
1683                         continue;
1684
1685                 if (new_s) {
1686                         ret = bch2_trans_mark_stripe_bucket(trans,
1687                                         bkey_i_to_s_c_stripe(new), i, false);
1688                         if (ret)
1689                                 break;
1690                 }
1691
1692                 if (old_s) {
1693                         ret = bch2_trans_mark_stripe_bucket(trans,
1694                                         bkey_s_c_to_stripe(old), i, true);
1695                         if (ret)
1696                                 break;
1697                 }
1698         }
1699
1700         return ret;
1701 }
1702
1703 int bch2_trans_mark_inode(struct btree_trans *trans,
1704                           struct bkey_s_c old,
1705                           struct bkey_i *new,
1706                           unsigned flags)
1707 {
1708         int nr = bkey_is_inode(&new->k) - bkey_is_inode(old.k);
1709
1710         if (nr) {
1711                 struct replicas_delta_list *d =
1712                         replicas_deltas_realloc(trans, 0);
1713                 d->nr_inodes += nr;
1714         }
1715
1716         return 0;
1717 }
1718
1719 int bch2_trans_mark_reservation(struct btree_trans *trans,
1720                                 struct bkey_s_c old,
1721                                 struct bkey_i *new,
1722                                 unsigned flags)
1723 {
1724         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1725                 ? old
1726                 : bkey_i_to_s_c(new);
1727         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1728         s64 sectors = (s64) k.k->size;
1729         struct replicas_delta_list *d;
1730
1731         if (flags & BTREE_TRIGGER_OVERWRITE)
1732                 sectors = -sectors;
1733         sectors *= replicas;
1734
1735         d = replicas_deltas_realloc(trans, 0);
1736
1737         replicas = clamp_t(unsigned, replicas, 1,
1738                            ARRAY_SIZE(d->persistent_reserved));
1739
1740         d->persistent_reserved[replicas - 1] += sectors;
1741         return 0;
1742 }
1743
1744 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1745                         struct bkey_s_c_reflink_p p,
1746                         u64 *idx, unsigned flags)
1747 {
1748         struct bch_fs *c = trans->c;
1749         struct btree_iter iter;
1750         struct bkey_s_c k;
1751         struct bkey_i *n;
1752         __le64 *refcount;
1753         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1754         struct printbuf buf = PRINTBUF;
1755         int ret;
1756
1757         bch2_trans_iter_init(trans, &iter, BTREE_ID_reflink, POS(0, *idx),
1758                              BTREE_ITER_INTENT|
1759                              BTREE_ITER_WITH_UPDATES);
1760         k = bch2_btree_iter_peek_slot(&iter);
1761         ret = bkey_err(k);
1762         if (ret)
1763                 goto err;
1764
1765         n = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1766         ret = PTR_ERR_OR_ZERO(n);
1767         if (ret)
1768                 goto err;
1769
1770         bkey_reassemble(n, k);
1771
1772         refcount = bkey_refcount(n);
1773         if (!refcount) {
1774                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1775                 bch2_trans_inconsistent(trans,
1776                         "nonexistent indirect extent at %llu while marking\n  %s",
1777                         *idx, buf.buf);
1778                 ret = -EIO;
1779                 goto err;
1780         }
1781
1782         if (!*refcount && (flags & BTREE_TRIGGER_OVERWRITE)) {
1783                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1784                 bch2_trans_inconsistent(trans,
1785                         "indirect extent refcount underflow at %llu while marking\n  %s",
1786                         *idx, buf.buf);
1787                 ret = -EIO;
1788                 goto err;
1789         }
1790
1791         if (flags & BTREE_TRIGGER_INSERT) {
1792                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1793                 u64 pad;
1794
1795                 pad = max_t(s64, le32_to_cpu(v->front_pad),
1796                             le64_to_cpu(v->idx) - bkey_start_offset(k.k));
1797                 BUG_ON(pad > U32_MAX);
1798                 v->front_pad = cpu_to_le32(pad);
1799
1800                 pad = max_t(s64, le32_to_cpu(v->back_pad),
1801                             k.k->p.offset - p.k->size - le64_to_cpu(v->idx));
1802                 BUG_ON(pad > U32_MAX);
1803                 v->back_pad = cpu_to_le32(pad);
1804         }
1805
1806         le64_add_cpu(refcount, add);
1807
1808         if (!*refcount) {
1809                 n->k.type = KEY_TYPE_deleted;
1810                 set_bkey_val_u64s(&n->k, 0);
1811         }
1812
1813         bch2_btree_iter_set_pos_to_extent_start(&iter);
1814         ret = bch2_trans_update(trans, &iter, n, 0);
1815         if (ret)
1816                 goto err;
1817
1818         *idx = k.k->p.offset;
1819 err:
1820         bch2_trans_iter_exit(trans, &iter);
1821         printbuf_exit(&buf);
1822         return ret;
1823 }
1824
1825 int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1826                               struct bkey_s_c old,
1827                               struct bkey_i *new,
1828                               unsigned flags)
1829 {
1830         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1831                 ? old
1832                 : bkey_i_to_s_c(new);
1833         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1834         u64 idx, end_idx;
1835         int ret = 0;
1836
1837         if (flags & BTREE_TRIGGER_INSERT) {
1838                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1839
1840                 v->front_pad = v->back_pad = 0;
1841         }
1842
1843         idx     = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
1844         end_idx = le64_to_cpu(p.v->idx) + p.k->size +
1845                 le32_to_cpu(p.v->back_pad);
1846
1847         while (idx < end_idx && !ret)
1848                 ret = __bch2_trans_mark_reflink_p(trans, p, &idx, flags);
1849
1850         return ret;
1851 }
1852
1853 static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1854                                     struct bch_dev *ca, size_t b,
1855                                     enum bch_data_type type,
1856                                     unsigned sectors)
1857 {
1858         struct bch_fs *c = trans->c;
1859         struct btree_iter iter;
1860         struct bkey_alloc_unpacked u;
1861         struct bch_extent_ptr ptr = {
1862                 .dev = ca->dev_idx,
1863                 .offset = bucket_to_sector(ca, b),
1864         };
1865         int ret = 0;
1866
1867         /*
1868          * Backup superblock might be past the end of our normal usable space:
1869          */
1870         if (b >= ca->mi.nbuckets)
1871                 return 0;
1872
1873         ret = bch2_trans_start_alloc_update(trans, &iter, &ptr, &u);
1874         if (ret)
1875                 return ret;
1876
1877         if (u.data_type && u.data_type != type) {
1878                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1879                         "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
1880                         "while marking %s",
1881                         iter.pos.inode, iter.pos.offset, u.gen,
1882                         bch2_data_types[u.data_type],
1883                         bch2_data_types[type],
1884                         bch2_data_types[type]);
1885                 ret = -EIO;
1886                 goto out;
1887         }
1888
1889         u.data_type     = type;
1890         u.dirty_sectors = sectors;
1891
1892         ret = bch2_alloc_write(trans, &iter, &u, 0);
1893         if (ret)
1894                 goto out;
1895 out:
1896         bch2_trans_iter_exit(trans, &iter);
1897         return ret;
1898 }
1899
1900 int bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1901                                     struct bch_dev *ca, size_t b,
1902                                     enum bch_data_type type,
1903                                     unsigned sectors)
1904 {
1905         return __bch2_trans_do(trans, NULL, NULL, 0,
1906                         __bch2_trans_mark_metadata_bucket(trans, ca, b, type, sectors));
1907 }
1908
1909 static int bch2_trans_mark_metadata_sectors(struct btree_trans *trans,
1910                                             struct bch_dev *ca,
1911                                             u64 start, u64 end,
1912                                             enum bch_data_type type,
1913                                             u64 *bucket, unsigned *bucket_sectors)
1914 {
1915         do {
1916                 u64 b = sector_to_bucket(ca, start);
1917                 unsigned sectors =
1918                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1919
1920                 if (b != *bucket && *bucket_sectors) {
1921                         int ret = bch2_trans_mark_metadata_bucket(trans, ca, *bucket,
1922                                                                   type, *bucket_sectors);
1923                         if (ret)
1924                                 return ret;
1925
1926                         *bucket_sectors = 0;
1927                 }
1928
1929                 *bucket         = b;
1930                 *bucket_sectors += sectors;
1931                 start += sectors;
1932         } while (start < end);
1933
1934         return 0;
1935 }
1936
1937 static int __bch2_trans_mark_dev_sb(struct btree_trans *trans,
1938                                     struct bch_dev *ca)
1939 {
1940         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1941         u64 bucket = 0;
1942         unsigned i, bucket_sectors = 0;
1943         int ret;
1944
1945         for (i = 0; i < layout->nr_superblocks; i++) {
1946                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1947
1948                 if (offset == BCH_SB_SECTOR) {
1949                         ret = bch2_trans_mark_metadata_sectors(trans, ca,
1950                                                 0, BCH_SB_SECTOR,
1951                                                 BCH_DATA_sb, &bucket, &bucket_sectors);
1952                         if (ret)
1953                                 return ret;
1954                 }
1955
1956                 ret = bch2_trans_mark_metadata_sectors(trans, ca, offset,
1957                                       offset + (1 << layout->sb_max_size_bits),
1958                                       BCH_DATA_sb, &bucket, &bucket_sectors);
1959                 if (ret)
1960                         return ret;
1961         }
1962
1963         if (bucket_sectors) {
1964                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1965                                 bucket, BCH_DATA_sb, bucket_sectors);
1966                 if (ret)
1967                         return ret;
1968         }
1969
1970         for (i = 0; i < ca->journal.nr; i++) {
1971                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1972                                 ca->journal.buckets[i],
1973                                 BCH_DATA_journal, ca->mi.bucket_size);
1974                 if (ret)
1975                         return ret;
1976         }
1977
1978         return 0;
1979 }
1980
1981 int bch2_trans_mark_dev_sb(struct bch_fs *c, struct bch_dev *ca)
1982 {
1983         return bch2_trans_do(c, NULL, NULL, BTREE_INSERT_LAZY_RW,
1984                         __bch2_trans_mark_dev_sb(&trans, ca));
1985 }
1986
1987 /* Disk reservations: */
1988
1989 #define SECTORS_CACHE   1024
1990
1991 int bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
1992                               u64 sectors, int flags)
1993 {
1994         struct bch_fs_pcpu *pcpu;
1995         u64 old, v, get;
1996         s64 sectors_available;
1997         int ret;
1998
1999         percpu_down_read(&c->mark_lock);
2000         preempt_disable();
2001         pcpu = this_cpu_ptr(c->pcpu);
2002
2003         if (sectors <= pcpu->sectors_available)
2004                 goto out;
2005
2006         v = atomic64_read(&c->sectors_available);
2007         do {
2008                 old = v;
2009                 get = min((u64) sectors + SECTORS_CACHE, old);
2010
2011                 if (get < sectors) {
2012                         preempt_enable();
2013                         goto recalculate;
2014                 }
2015         } while ((v = atomic64_cmpxchg(&c->sectors_available,
2016                                        old, old - get)) != old);
2017
2018         pcpu->sectors_available         += get;
2019
2020 out:
2021         pcpu->sectors_available         -= sectors;
2022         this_cpu_add(*c->online_reserved, sectors);
2023         res->sectors                    += sectors;
2024
2025         preempt_enable();
2026         percpu_up_read(&c->mark_lock);
2027         return 0;
2028
2029 recalculate:
2030         mutex_lock(&c->sectors_available_lock);
2031
2032         percpu_u64_set(&c->pcpu->sectors_available, 0);
2033         sectors_available = avail_factor(__bch2_fs_usage_read_short(c).free);
2034
2035         if (sectors <= sectors_available ||
2036             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
2037                 atomic64_set(&c->sectors_available,
2038                              max_t(s64, 0, sectors_available - sectors));
2039                 this_cpu_add(*c->online_reserved, sectors);
2040                 res->sectors                    += sectors;
2041                 ret = 0;
2042         } else {
2043                 atomic64_set(&c->sectors_available, sectors_available);
2044                 ret = -ENOSPC;
2045         }
2046
2047         mutex_unlock(&c->sectors_available_lock);
2048         percpu_up_read(&c->mark_lock);
2049
2050         return ret;
2051 }
2052
2053 /* Startup/shutdown: */
2054
2055 static void bucket_gens_free_rcu(struct rcu_head *rcu)
2056 {
2057         struct bucket_gens *buckets =
2058                 container_of(rcu, struct bucket_gens, rcu);
2059
2060         kvpfree(buckets, sizeof(*buckets) + buckets->nbuckets);
2061 }
2062
2063 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
2064 {
2065         struct bucket_gens *bucket_gens = NULL, *old_bucket_gens = NULL;
2066         unsigned long *buckets_nouse = NULL;
2067         bool resize = ca->bucket_gens != NULL;
2068         int ret = -ENOMEM;
2069
2070         if (!(bucket_gens       = kvpmalloc(sizeof(struct bucket_gens) + nbuckets,
2071                                             GFP_KERNEL|__GFP_ZERO)) ||
2072             (c->opts.buckets_nouse &&
2073              !(buckets_nouse    = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2074                                             sizeof(unsigned long),
2075                                             GFP_KERNEL|__GFP_ZERO))))
2076                 goto err;
2077
2078         bucket_gens->first_bucket = ca->mi.first_bucket;
2079         bucket_gens->nbuckets   = nbuckets;
2080
2081         bch2_copygc_stop(c);
2082
2083         if (resize) {
2084                 down_write(&c->gc_lock);
2085                 down_write(&ca->bucket_lock);
2086                 percpu_down_write(&c->mark_lock);
2087         }
2088
2089         old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
2090
2091         if (resize) {
2092                 size_t n = min(bucket_gens->nbuckets, old_bucket_gens->nbuckets);
2093
2094                 memcpy(bucket_gens->b,
2095                        old_bucket_gens->b,
2096                        n);
2097                 if (buckets_nouse)
2098                         memcpy(buckets_nouse,
2099                                ca->buckets_nouse,
2100                                BITS_TO_LONGS(n) * sizeof(unsigned long));
2101         }
2102
2103         rcu_assign_pointer(ca->bucket_gens, bucket_gens);
2104         bucket_gens     = old_bucket_gens;
2105
2106         swap(ca->buckets_nouse, buckets_nouse);
2107
2108         nbuckets = ca->mi.nbuckets;
2109
2110         if (resize) {
2111                 percpu_up_write(&c->mark_lock);
2112                 up_write(&ca->bucket_lock);
2113                 up_write(&c->gc_lock);
2114         }
2115
2116         ret = 0;
2117 err:
2118         kvpfree(buckets_nouse,
2119                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2120         if (bucket_gens)
2121                 call_rcu(&bucket_gens->rcu, bucket_gens_free_rcu);
2122
2123         return ret;
2124 }
2125
2126 void bch2_dev_buckets_free(struct bch_dev *ca)
2127 {
2128         unsigned i;
2129
2130         kvpfree(ca->buckets_nouse,
2131                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2132         kvpfree(rcu_dereference_protected(ca->bucket_gens, 1),
2133                 sizeof(struct bucket_gens) + ca->mi.nbuckets);
2134
2135         for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
2136                 free_percpu(ca->usage[i]);
2137         kfree(ca->usage_base);
2138 }
2139
2140 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2141 {
2142         unsigned i;
2143
2144         ca->usage_base = kzalloc(sizeof(struct bch_dev_usage), GFP_KERNEL);
2145         if (!ca->usage_base)
2146                 return -ENOMEM;
2147
2148         for (i = 0; i < ARRAY_SIZE(ca->usage); i++) {
2149                 ca->usage[i] = alloc_percpu(struct bch_dev_usage);
2150                 if (!ca->usage[i])
2151                         return -ENOMEM;
2152         }
2153
2154         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);;
2155 }