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 bch_alloc_v4 a)
 283 {
 284         return a.dirty_sectors || a.stripe;
 285 }
 286
 287 static inline int bucket_sectors_fragmented(struct bch_dev *ca,
 288                                             struct bch_alloc_v4 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 bch_alloc_v4 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 bch_alloc_v4 old,
 315                                   struct bch_alloc_v4 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 bch_alloc_v4 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 bch_alloc_v4 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 bch_alloc_v4 old_a, new_a;
 510         struct bch_dev *ca = bch_dev_bkey_exists(c, new.k->p.inode);
 511         int ret = 0;
 512
 513         if (bch2_trans_inconsistent_on(new.k->p.offset < ca->mi.first_bucket ||
 514                                        new.k->p.offset >= ca->mi.nbuckets, trans,
 515                                        "alloc key outside range of device's buckets"))
 516                 return -EIO;
 517
 518         /*
 519          * alloc btree is read in by bch2_alloc_read, not gc:
 520          */
 521         if ((flags & BTREE_TRIGGER_GC) &&
 522             !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
 523                 return 0;
 524
 525         bch2_alloc_to_v4(old, &old_a);
 526         bch2_alloc_to_v4(new, &new_a);
 527
 528         if ((flags & BTREE_TRIGGER_INSERT) &&
 529             !old_a.data_type != !new_a.data_type &&
 530             new.k->type == KEY_TYPE_alloc_v4) {
 531                 struct bch_alloc_v4 *v = (struct bch_alloc_v4 *) new.v;
 532
 533                 BUG_ON(!journal_seq);
 534
 535                 /*
 536                  * If the btree updates referring to a bucket weren't flushed
 537                  * before the bucket became empty again, then the we don't have
 538                  * to wait on a journal flush before we can reuse the bucket:
 539                  */
 540                 new_a.journal_seq = !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                 v->journal_seq = new_a.journal_seq;
 545         }
 546
 547         if (old_a.data_type && !new_a.data_type && new_a.journal_seq) {
 548                 ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 549                                 c->journal.flushed_seq_ondisk,
 550                                 new.k->p.inode, new.k->p.offset,
 551                                 new_a.journal_seq);
 552                 if (ret) {
 553                         bch2_fs_fatal_error(c,
 554                                 "error setting bucket_needs_journal_commit: %i", ret);
 555                         return ret;
 556                 }
 557         }
 558
 559         if (!new_a.data_type &&
 560             (!new_a.journal_seq || new_a.journal_seq < c->journal.flushed_seq_ondisk))
 561                 closure_wake_up(&c->freelist_wait);
 562
 563         if ((flags & BTREE_TRIGGER_INSERT) &&
 564             BCH_ALLOC_V4_NEED_DISCARD(&new_a) &&
 565             !new_a.journal_seq)
 566                 bch2_do_discards(c);
 567
 568         if (!old_a.data_type &&
 569             new_a.data_type &&
 570             should_invalidate_buckets(ca))
 571                 bch2_do_invalidates(c);
 572
 573         if (bucket_state(new_a) == BUCKET_need_gc_gens) {
 574                 atomic_inc(&c->kick_gc);
 575                 wake_up_process(c->gc_thread);
 576         }
 577
 578         percpu_down_read(&c->mark_lock);
 579         if (!gc && new_a.gen != old_a.gen)
 580                 *bucket_gen(ca, new.k->p.offset) = new_a.gen;
 581
 582         bch2_dev_usage_update(c, ca, old_a, new_a, journal_seq, gc);
 583
 584         if (gc) {
 585                 struct bucket *g = gc_bucket(ca, new.k->p.offset);
 586
 587                 bucket_lock(g);
 588
 589                 g->gen_valid            = 1;
 590                 g->gen                  = new_a.gen;
 591                 g->data_type            = new_a.data_type;
 592                 g->stripe               = new_a.stripe;
 593                 g->stripe_redundancy    = new_a.stripe_redundancy;
 594                 g->dirty_sectors        = new_a.dirty_sectors;
 595                 g->cached_sectors       = new_a.cached_sectors;
 596
 597                 bucket_unlock(g);
 598         }
 599         percpu_up_read(&c->mark_lock);
 600
 601         /*
 602          * need to know if we're getting called from the invalidate path or
 603          * not:
 604          */
 605
 606         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 607             old_a.cached_sectors) {
 608                 ret = update_cached_sectors(c, new, ca->dev_idx,
 609                                             -old_a.cached_sectors,
 610                                             journal_seq, gc);
 611                 if (ret) {
 612                         bch2_fs_fatal_error(c, "bch2_mark_alloc(): no replicas entry while updating cached sectors");
 613                         return ret;
 614                 }
 615
 616                 trace_invalidate(ca, bucket_to_sector(ca, new.k->p.offset),
 617                                  old_a.cached_sectors);
 618         }
 619
 620         return 0;
 621 }
 622
 623 void bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 624                                size_t b, enum bch_data_type data_type,
 625                                unsigned sectors, struct gc_pos pos,
 626                                unsigned flags)
 627 {
 628         struct bucket old, new, *g;
 629         bool overflow;
 630
 631         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 632         BUG_ON(data_type != BCH_DATA_sb &&
 633                data_type != BCH_DATA_journal);
 634
 635         /*
 636          * Backup superblock might be past the end of our normal usable space:
 637          */
 638         if (b >= ca->mi.nbuckets)
 639                 return;
 640
 641         percpu_down_read(&c->mark_lock);
 642         g = gc_bucket(ca, b);
 643
 644         bucket_lock(g);
 645         old = *g;
 646
 647         g->data_type = data_type;
 648         g->dirty_sectors += sectors;
 649         overflow = g->dirty_sectors < sectors;
 650
 651         new = *g;
 652         bucket_unlock(g);
 653
 654         bch2_fs_inconsistent_on(old.data_type &&
 655                                 old.data_type != data_type, c,
 656                 "different types of data in same bucket: %s, %s",
 657                 bch2_data_types[old.data_type],
 658                 bch2_data_types[data_type]);
 659
 660         bch2_fs_inconsistent_on(overflow, c,
 661                 "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > U16_MAX",
 662                 ca->dev_idx, b, new.gen,
 663                 bch2_data_types[old.data_type ?: data_type],
 664                 old.dirty_sectors, sectors);
 665
 666         bch2_dev_usage_update_m(c, ca, old, new, 0, true);
 667         percpu_up_read(&c->mark_lock);
 668 }
 669
 670 static s64 ptr_disk_sectors(s64 sectors, struct extent_ptr_decoded p)
 671 {
 672         EBUG_ON(sectors < 0);
 673
 674         return p.crc.compression_type &&
 675                 p.crc.compression_type != BCH_COMPRESSION_TYPE_incompressible
 676                 ? DIV_ROUND_UP_ULL(sectors * p.crc.compressed_size,
 677                                p.crc.uncompressed_size)
 678                 : sectors;
 679 }
 680
 681 static int check_bucket_ref(struct bch_fs *c,
 682                             struct bkey_s_c k,
 683                             const struct bch_extent_ptr *ptr,
 684                             s64 sectors, enum bch_data_type ptr_data_type,
 685                             u8 b_gen, u8 bucket_data_type,
 686                             u32 dirty_sectors, u32 cached_sectors)
 687 {
 688         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 689         size_t bucket_nr = PTR_BUCKET_NR(ca, ptr);
 690         u16 bucket_sectors = !ptr->cached
 691                 ? dirty_sectors
 692                 : cached_sectors;
 693         struct printbuf buf = PRINTBUF;
 694         int ret = 0;
 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 err;
 738         }
 739
 740         if (bucket_data_type && ptr_data_type &&
 741             bucket_data_type != ptr_data_type) {
 742                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 743                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 744                         "while marking %s",
 745                         ptr->dev, bucket_nr, b_gen,
 746                         bch2_data_types[bucket_data_type],
 747                         bch2_data_types[ptr_data_type],
 748                         (printbuf_reset(&buf),
 749                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 750                 ret = -EIO;
 751                 goto err;
 752         }
 753
 754         if ((unsigned) (bucket_sectors + sectors) > U32_MAX) {
 755                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 756                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U16_MAX\n"
 757                         "while marking %s",
 758                         ptr->dev, bucket_nr, b_gen,
 759                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 760                         bucket_sectors, sectors,
 761                         (printbuf_reset(&buf),
 762                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 763                 ret = -EIO;
 764                 goto err;
 765         }
 766 err:
 767         printbuf_exit(&buf);
 768         return ret;
 769 }
 770
 771 static int mark_stripe_bucket(struct btree_trans *trans,
 772                               struct bkey_s_c k,
 773                               unsigned ptr_idx,
 774                               unsigned flags)
 775 {
 776         struct bch_fs *c = trans->c;
 777         u64 journal_seq = trans->journal_res.seq;
 778         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 779         unsigned nr_data = s->nr_blocks - s->nr_redundant;
 780         bool parity = ptr_idx >= nr_data;
 781         enum bch_data_type data_type = parity ? BCH_DATA_parity : 0;
 782         s64 sectors = parity ? le16_to_cpu(s->sectors) : 0;
 783         const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
 784         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 785         struct bucket old, new, *g;
 786         struct printbuf buf = PRINTBUF;
 787         int ret = 0;
 788
 789         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 790
 791         /* * XXX doesn't handle deletion */
 792
 793         percpu_down_read(&c->mark_lock);
 794         buf.atomic++;
 795         g = PTR_GC_BUCKET(ca, ptr);
 796
 797         if (g->dirty_sectors ||
 798             (g->stripe && g->stripe != k.k->p.offset)) {
 799                 bch2_fs_inconsistent(c,
 800                               "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
 801                               ptr->dev, PTR_BUCKET_NR(ca, ptr), g->gen,
 802                               (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 803                 ret = -EINVAL;
 804                 goto err;
 805         }
 806
 807         bucket_lock(g);
 808         old = *g;
 809
 810         ret = check_bucket_ref(c, k, ptr, sectors, data_type,
 811                                new.gen, new.data_type,
 812                                new.dirty_sectors, new.cached_sectors);
 813         if (ret) {
 814                 bucket_unlock(g);
 815                 goto err;
 816         }
 817
 818         new.dirty_sectors += sectors;
 819         if (data_type)
 820                 new.data_type = data_type;
 821
 822         g->stripe               = k.k->p.offset;
 823         g->stripe_redundancy    = s->nr_redundant;
 824
 825         new = *g;
 826         bucket_unlock(g);
 827
 828         bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 829 err:
 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->c, 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                              struct bkey_s_c k,
 860                              struct extent_ptr_decoded p,
 861                              s64 sectors, enum bch_data_type data_type,
 862                              unsigned flags)
 863 {
 864         u64 journal_seq = trans->journal_res.seq;
 865         struct bch_fs *c = trans->c;
 866         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 867         struct bucket old, new, *g;
 868         u8 bucket_data_type;
 869         int ret = 0;
 870
 871         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 872
 873         percpu_down_read(&c->mark_lock);
 874         g = PTR_GC_BUCKET(ca, &p.ptr);
 875
 876         bucket_lock(g);
 877         old = *g;
 878
 879         bucket_data_type = g->data_type;
 880
 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                 bucket_unlock(g);
 888                 goto err;
 889         }
 890
 891         g->data_type = bucket_data_type;
 892
 893         new = *g;
 894         bucket_unlock(g);
 895
 896         bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 897 err:
 898         percpu_up_read(&c->mark_lock);
 899
 900         return ret;
 901 }
 902
 903 static int bch2_mark_stripe_ptr(struct btree_trans *trans,
 904                                 struct bkey_s_c k,
 905                                 struct bch_extent_stripe_ptr p,
 906                                 enum bch_data_type data_type,
 907                                 s64 sectors,
 908                                 unsigned flags)
 909 {
 910         struct bch_fs *c = trans->c;
 911         struct bch_replicas_padded r;
 912         struct gc_stripe *m;
 913
 914         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 915
 916         m = genradix_ptr_alloc(&c->gc_stripes, p.idx, GFP_KERNEL);
 917         if (!m) {
 918                 bch_err(c, "error allocating memory for gc_stripes, idx %llu",
 919                         (u64) p.idx);
 920                 return -ENOMEM;
 921         }
 922
 923         spin_lock(&c->ec_stripes_heap_lock);
 924
 925         if (!m || !m->alive) {
 926                 spin_unlock(&c->ec_stripes_heap_lock);
 927                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
 928                                     (u64) p.idx);
 929                 bch2_inconsistent_error(c);
 930                 return -EIO;
 931         }
 932
 933         m->block_sectors[p.block] += sectors;
 934
 935         r = m->r;
 936         spin_unlock(&c->ec_stripes_heap_lock);
 937
 938         r.e.data_type = data_type;
 939         update_replicas(c, k, &r.e, sectors, trans->journal_res.seq, true);
 940
 941         return 0;
 942 }
 943
 944 int bch2_mark_extent(struct btree_trans *trans,
 945                      struct bkey_s_c old, struct bkey_s_c new,
 946                      unsigned flags)
 947 {
 948         u64 journal_seq = trans->journal_res.seq;
 949         struct bch_fs *c = trans->c;
 950         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
 951         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 952         const union bch_extent_entry *entry;
 953         struct extent_ptr_decoded p;
 954         struct bch_replicas_padded r;
 955         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
 956                 ? BCH_DATA_btree
 957                 : BCH_DATA_user;
 958         s64 sectors = bkey_is_btree_ptr(k.k)
 959                 ? btree_sectors(c)
 960                 : k.k->size;
 961         s64 dirty_sectors = 0;
 962         bool stale;
 963         int ret;
 964
 965         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 966
 967         r.e.data_type   = data_type;
 968         r.e.nr_devs     = 0;
 969         r.e.nr_required = 1;
 970
 971         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 972                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
 973
 974                 if (flags & BTREE_TRIGGER_OVERWRITE)
 975                         disk_sectors = -disk_sectors;
 976
 977                 ret = bch2_mark_pointer(trans, k, p, disk_sectors,
 978                                         data_type, flags);
 979                 if (ret < 0)
 980                         return ret;
 981
 982                 stale = ret > 0;
 983
 984                 if (p.ptr.cached) {
 985                         if (!stale) {
 986                                 ret = update_cached_sectors(c, k, p.ptr.dev,
 987                                                 disk_sectors, journal_seq, true);
 988                                 if (ret) {
 989                                         bch2_fs_fatal_error(c, "bch2_mark_extent(): no replicas entry while updating cached sectors");
 990                                         return ret;
 991                                 }
 992                         }
 993                 } else if (!p.has_ec) {
 994                         dirty_sectors          += disk_sectors;
 995                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
 996                 } else {
 997                         ret = bch2_mark_stripe_ptr(trans, k, p.ec, data_type,
 998                                         disk_sectors, flags);
 999                         if (ret)
1000                                 return ret;
1001
1002                         /*
1003                          * There may be other dirty pointers in this extent, but
1004                          * if so they're not required for mounting if we have an
1005                          * erasure coded pointer in this extent:
1006                          */
1007                         r.e.nr_required = 0;
1008                 }
1009         }
1010
1011         if (r.e.nr_devs) {
1012                 ret = update_replicas(c, k, &r.e, dirty_sectors, journal_seq, true);
1013                 if (ret) {
1014                         struct printbuf buf = PRINTBUF;
1015
1016                         bch2_bkey_val_to_text(&buf, c, k);
1017                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
1018                         printbuf_exit(&buf);
1019                         return ret;
1020                 }
1021         }
1022
1023         return 0;
1024 }
1025
1026 int bch2_mark_stripe(struct btree_trans *trans,
1027                      struct bkey_s_c old, struct bkey_s_c new,
1028                      unsigned flags)
1029 {
1030         bool gc = flags & BTREE_TRIGGER_GC;
1031         u64 journal_seq = trans->journal_res.seq;
1032         struct bch_fs *c = trans->c;
1033         u64 idx = new.k->p.offset;
1034         const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
1035                 ? bkey_s_c_to_stripe(old).v : NULL;
1036         const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
1037                 ? bkey_s_c_to_stripe(new).v : NULL;
1038         unsigned i;
1039         int ret;
1040
1041         BUG_ON(gc && old_s);
1042
1043         if (!gc) {
1044                 struct stripe *m = genradix_ptr(&c->stripes, idx);
1045
1046                 if (!m || (old_s && !m->alive)) {
1047                         struct printbuf buf1 = PRINTBUF;
1048                         struct printbuf buf2 = PRINTBUF;
1049
1050                         bch2_bkey_val_to_text(&buf1, c, old);
1051                         bch2_bkey_val_to_text(&buf2, c, new);
1052                         bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
1053                                             "old %s\n"
1054                                             "new %s", idx, buf1.buf, buf2.buf);
1055                         printbuf_exit(&buf2);
1056                         printbuf_exit(&buf1);
1057                         bch2_inconsistent_error(c);
1058                         return -1;
1059                 }
1060
1061                 if (!new_s) {
1062                         spin_lock(&c->ec_stripes_heap_lock);
1063                         bch2_stripes_heap_del(c, m, idx);
1064                         spin_unlock(&c->ec_stripes_heap_lock);
1065
1066                         memset(m, 0, sizeof(*m));
1067                 } else {
1068                         m->alive        = true;
1069                         m->sectors      = le16_to_cpu(new_s->sectors);
1070                         m->algorithm    = new_s->algorithm;
1071                         m->nr_blocks    = new_s->nr_blocks;
1072                         m->nr_redundant = new_s->nr_redundant;
1073                         m->blocks_nonempty = 0;
1074
1075                         for (i = 0; i < new_s->nr_blocks; i++)
1076                                 m->blocks_nonempty += !!stripe_blockcount_get(new_s, i);
1077
1078                         spin_lock(&c->ec_stripes_heap_lock);
1079                         bch2_stripes_heap_update(c, m, idx);
1080                         spin_unlock(&c->ec_stripes_heap_lock);
1081                 }
1082         } else {
1083                 struct gc_stripe *m =
1084                         genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
1085
1086                 if (!m) {
1087                         bch_err(c, "error allocating memory for gc_stripes, idx %llu",
1088                                 idx);
1089                         return -ENOMEM;
1090                 }
1091                 /*
1092                  * This will be wrong when we bring back runtime gc: we should
1093                  * be unmarking the old key and then marking the new key
1094                  */
1095                 m->alive        = true;
1096                 m->sectors      = le16_to_cpu(new_s->sectors);
1097                 m->nr_blocks    = new_s->nr_blocks;
1098                 m->nr_redundant = new_s->nr_redundant;
1099
1100                 for (i = 0; i < new_s->nr_blocks; i++)
1101                         m->ptrs[i] = new_s->ptrs[i];
1102
1103                 bch2_bkey_to_replicas(&m->r.e, new);
1104
1105                 /*
1106                  * gc recalculates this field from stripe ptr
1107                  * references:
1108                  */
1109                 memset(m->block_sectors, 0, sizeof(m->block_sectors));
1110
1111                 for (i = 0; i < new_s->nr_blocks; i++) {
1112                         ret = mark_stripe_bucket(trans, new, i, flags);
1113                         if (ret)
1114                                 return ret;
1115                 }
1116
1117                 ret = update_replicas(c, new, &m->r.e,
1118                                       ((s64) m->sectors * m->nr_redundant),
1119                                       journal_seq, gc);
1120                 if (ret) {
1121                         struct printbuf buf = PRINTBUF;
1122
1123                         bch2_bkey_val_to_text(&buf, c, new);
1124                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
1125                         printbuf_exit(&buf);
1126                         return ret;
1127                 }
1128         }
1129
1130         return 0;
1131 }
1132
1133 int bch2_mark_inode(struct btree_trans *trans,
1134                     struct bkey_s_c old, struct bkey_s_c new,
1135                     unsigned flags)
1136 {
1137         struct bch_fs *c = trans->c;
1138         struct bch_fs_usage __percpu *fs_usage;
1139         u64 journal_seq = trans->journal_res.seq;
1140
1141         if (flags & BTREE_TRIGGER_INSERT) {
1142                 struct bch_inode_v2 *v = (struct bch_inode_v2 *) new.v;
1143
1144                 BUG_ON(!journal_seq);
1145                 BUG_ON(new.k->type != KEY_TYPE_inode_v2);
1146
1147                 v->bi_journal_seq = cpu_to_le64(journal_seq);
1148         }
1149
1150         if (flags & BTREE_TRIGGER_GC) {
1151                 percpu_down_read(&c->mark_lock);
1152                 preempt_disable();
1153
1154                 fs_usage = fs_usage_ptr(c, journal_seq, flags & BTREE_TRIGGER_GC);
1155                 fs_usage->nr_inodes += bkey_is_inode(new.k);
1156                 fs_usage->nr_inodes -= bkey_is_inode(old.k);
1157
1158                 preempt_enable();
1159                 percpu_up_read(&c->mark_lock);
1160         }
1161         return 0;
1162 }
1163
1164 int bch2_mark_reservation(struct btree_trans *trans,
1165                           struct bkey_s_c old, struct bkey_s_c new,
1166                           unsigned flags)
1167 {
1168         struct bch_fs *c = trans->c;
1169         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1170         struct bch_fs_usage __percpu *fs_usage;
1171         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1172         s64 sectors = (s64) k.k->size;
1173
1174         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1175
1176         if (flags & BTREE_TRIGGER_OVERWRITE)
1177                 sectors = -sectors;
1178         sectors *= replicas;
1179
1180         percpu_down_read(&c->mark_lock);
1181         preempt_disable();
1182
1183         fs_usage = fs_usage_ptr(c, trans->journal_res.seq, flags & BTREE_TRIGGER_GC);
1184         replicas = clamp_t(unsigned, replicas, 1,
1185                            ARRAY_SIZE(fs_usage->persistent_reserved));
1186
1187         fs_usage->reserved                              += sectors;
1188         fs_usage->persistent_reserved[replicas - 1]     += sectors;
1189
1190         preempt_enable();
1191         percpu_up_read(&c->mark_lock);
1192
1193         return 0;
1194 }
1195
1196 static s64 __bch2_mark_reflink_p(struct btree_trans *trans,
1197                                  struct bkey_s_c_reflink_p p,
1198                                  u64 start, u64 end,
1199                                  u64 *idx, unsigned flags, size_t r_idx)
1200 {
1201         struct bch_fs *c = trans->c;
1202         struct reflink_gc *r;
1203         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1204         u64 next_idx = end;
1205         s64 ret = 0;
1206         struct printbuf buf = PRINTBUF;
1207
1208         if (r_idx >= c->reflink_gc_nr)
1209                 goto not_found;
1210
1211         r = genradix_ptr(&c->reflink_gc_table, r_idx);
1212         next_idx = min(next_idx, r->offset - r->size);
1213         if (*idx < next_idx)
1214                 goto not_found;
1215
1216         BUG_ON((s64) r->refcount + add < 0);
1217
1218         r->refcount += add;
1219         *idx = r->offset;
1220         return 0;
1221 not_found:
1222         if (fsck_err(c, "pointer to missing indirect extent\n"
1223                      "  %s\n"
1224                      "  missing range %llu-%llu",
1225                      (bch2_bkey_val_to_text(&buf, c, p.s_c), buf.buf),
1226                      *idx, next_idx)) {
1227                 struct bkey_i_error new;
1228
1229                 bkey_init(&new.k);
1230                 new.k.type      = KEY_TYPE_error;
1231                 new.k.p         = bkey_start_pos(p.k);
1232                 new.k.p.offset += *idx - start;
1233                 bch2_key_resize(&new.k, next_idx - *idx);
1234                 ret = __bch2_btree_insert(trans, BTREE_ID_extents, &new.k_i);
1235         }
1236
1237         *idx = next_idx;
1238 fsck_err:
1239         printbuf_exit(&buf);
1240         return ret;
1241 }
1242
1243 int bch2_mark_reflink_p(struct btree_trans *trans,
1244                         struct bkey_s_c old, struct bkey_s_c new,
1245                         unsigned flags)
1246 {
1247         struct bch_fs *c = trans->c;
1248         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1249         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1250         struct reflink_gc *ref;
1251         size_t l, r, m;
1252         u64 idx = le64_to_cpu(p.v->idx), start = idx;
1253         u64 end = le64_to_cpu(p.v->idx) + p.k->size;
1254         int ret = 0;
1255
1256         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1257
1258         if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix) {
1259                 idx -= le32_to_cpu(p.v->front_pad);
1260                 end += le32_to_cpu(p.v->back_pad);
1261         }
1262
1263         l = 0;
1264         r = c->reflink_gc_nr;
1265         while (l < r) {
1266                 m = l + (r - l) / 2;
1267
1268                 ref = genradix_ptr(&c->reflink_gc_table, m);
1269                 if (ref->offset <= idx)
1270                         l = m + 1;
1271                 else
1272                         r = m;
1273         }
1274
1275         while (idx < end && !ret)
1276                 ret = __bch2_mark_reflink_p(trans, p, start, end,
1277                                             &idx, flags, l++);
1278
1279         return ret;
1280 }
1281
1282 static noinline __cold
1283 void fs_usage_apply_warn(struct btree_trans *trans,
1284                          unsigned disk_res_sectors,
1285                          s64 should_not_have_added)
1286 {
1287         struct bch_fs *c = trans->c;
1288         struct btree_insert_entry *i;
1289         struct printbuf buf = PRINTBUF;
1290
1291         bch_err(c, "disk usage increased %lli more than %u sectors reserved",
1292                 should_not_have_added, disk_res_sectors);
1293
1294         trans_for_each_update(trans, i) {
1295                 struct bkey_s_c old = { &i->old_k, i->old_v };
1296
1297                 pr_err("while inserting");
1298                 printbuf_reset(&buf);
1299                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(i->k));
1300                 pr_err("  %s", buf.buf);
1301                 pr_err("overlapping with");
1302                 printbuf_reset(&buf);
1303                 bch2_bkey_val_to_text(&buf, c, old);
1304                 pr_err("  %s", buf.buf);
1305         }
1306
1307         __WARN();
1308         printbuf_exit(&buf);
1309 }
1310
1311 int bch2_trans_fs_usage_apply(struct btree_trans *trans,
1312                               struct replicas_delta_list *deltas)
1313 {
1314         struct bch_fs *c = trans->c;
1315         static int warned_disk_usage = 0;
1316         bool warn = false;
1317         unsigned disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1318         struct replicas_delta *d = deltas->d, *d2;
1319         struct replicas_delta *top = (void *) deltas->d + deltas->used;
1320         struct bch_fs_usage *dst;
1321         s64 added = 0, should_not_have_added;
1322         unsigned i;
1323
1324         percpu_down_read(&c->mark_lock);
1325         preempt_disable();
1326         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1327
1328         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1329                 switch (d->r.data_type) {
1330                 case BCH_DATA_btree:
1331                 case BCH_DATA_user:
1332                 case BCH_DATA_parity:
1333                         added += d->delta;
1334                 }
1335
1336                 if (__update_replicas(c, dst, &d->r, d->delta))
1337                         goto need_mark;
1338         }
1339
1340         dst->nr_inodes += deltas->nr_inodes;
1341
1342         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1343                 added                           += deltas->persistent_reserved[i];
1344                 dst->reserved                   += deltas->persistent_reserved[i];
1345                 dst->persistent_reserved[i]     += deltas->persistent_reserved[i];
1346         }
1347
1348         /*
1349          * Not allowed to reduce sectors_available except by getting a
1350          * reservation:
1351          */
1352         should_not_have_added = added - (s64) disk_res_sectors;
1353         if (unlikely(should_not_have_added > 0)) {
1354                 u64 old, new, v = atomic64_read(&c->sectors_available);
1355
1356                 do {
1357                         old = v;
1358                         new = max_t(s64, 0, old - should_not_have_added);
1359                 } while ((v = atomic64_cmpxchg(&c->sectors_available,
1360                                                old, new)) != old);
1361
1362                 added -= should_not_have_added;
1363                 warn = true;
1364         }
1365
1366         if (added > 0) {
1367                 trans->disk_res->sectors -= added;
1368                 this_cpu_sub(*c->online_reserved, added);
1369         }
1370
1371         preempt_enable();
1372         percpu_up_read(&c->mark_lock);
1373
1374         if (unlikely(warn) && !xchg(&warned_disk_usage, 1))
1375                 fs_usage_apply_warn(trans, disk_res_sectors, should_not_have_added);
1376         return 0;
1377 need_mark:
1378         /* revert changes: */
1379         for (d2 = deltas->d; d2 != d; d2 = replicas_delta_next(d2))
1380                 BUG_ON(__update_replicas(c, dst, &d2->r, -d2->delta));
1381
1382         preempt_enable();
1383         percpu_up_read(&c->mark_lock);
1384         return -1;
1385 }
1386
1387 /* trans_mark: */
1388
1389 static int bch2_trans_mark_pointer(struct btree_trans *trans,
1390                         struct bkey_s_c k, struct extent_ptr_decoded p,
1391                         s64 sectors, enum bch_data_type data_type)
1392 {
1393         struct btree_iter iter;
1394         struct bkey_i_alloc_v4 *a;
1395         int ret;
1396
1397         a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(trans->c, &p.ptr));
1398         if (IS_ERR(a))
1399                 return PTR_ERR(a);
1400
1401         ret = __mark_pointer(trans, k, &p.ptr, sectors, data_type,
1402                              a->v.gen, &a->v.data_type,
1403                              &a->v.dirty_sectors, &a->v.cached_sectors);
1404         if (ret)
1405                 goto out;
1406
1407         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1408         if (ret)
1409                 goto out;
1410 out:
1411         bch2_trans_iter_exit(trans, &iter);
1412         return ret;
1413 }
1414
1415 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1416                         struct extent_ptr_decoded p,
1417                         s64 sectors, enum bch_data_type data_type)
1418 {
1419         struct btree_iter iter;
1420         struct bkey_s_c k;
1421         struct bkey_i_stripe *s;
1422         struct bch_replicas_padded r;
1423         int ret = 0;
1424
1425         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes, POS(0, p.ec.idx),
1426                              BTREE_ITER_INTENT|
1427                              BTREE_ITER_WITH_UPDATES);
1428         k = bch2_btree_iter_peek_slot(&iter);
1429         ret = bkey_err(k);
1430         if (ret)
1431                 goto err;
1432
1433         if (k.k->type != KEY_TYPE_stripe) {
1434                 bch2_trans_inconsistent(trans,
1435                         "pointer to nonexistent stripe %llu",
1436                         (u64) p.ec.idx);
1437                 ret = -EIO;
1438                 goto err;
1439         }
1440
1441         if (!bch2_ptr_matches_stripe(bkey_s_c_to_stripe(k).v, p)) {
1442                 bch2_trans_inconsistent(trans,
1443                         "stripe pointer doesn't match stripe %llu",
1444                         (u64) p.ec.idx);
1445                 ret = -EIO;
1446                 goto err;
1447         }
1448
1449         s = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1450         ret = PTR_ERR_OR_ZERO(s);
1451         if (ret)
1452                 goto err;
1453
1454         bkey_reassemble(&s->k_i, k);
1455         stripe_blockcount_set(&s->v, p.ec.block,
1456                 stripe_blockcount_get(&s->v, p.ec.block) +
1457                 sectors);
1458
1459         ret = bch2_trans_update(trans, &iter, &s->k_i, 0);
1460         if (ret)
1461                 goto err;
1462
1463         bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1464         r.e.data_type = data_type;
1465         update_replicas_list(trans, &r.e, sectors);
1466 err:
1467         bch2_trans_iter_exit(trans, &iter);
1468         return ret;
1469 }
1470
1471 int bch2_trans_mark_extent(struct btree_trans *trans,
1472                            struct bkey_s_c old, struct bkey_i *new,
1473                            unsigned flags)
1474 {
1475         struct bch_fs *c = trans->c;
1476         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1477                 ? old
1478                 : bkey_i_to_s_c(new);
1479         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1480         const union bch_extent_entry *entry;
1481         struct extent_ptr_decoded p;
1482         struct bch_replicas_padded r;
1483         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
1484                 ? BCH_DATA_btree
1485                 : BCH_DATA_user;
1486         s64 sectors = bkey_is_btree_ptr(k.k)
1487                 ? btree_sectors(c)
1488                 : k.k->size;
1489         s64 dirty_sectors = 0;
1490         bool stale;
1491         int ret;
1492
1493         r.e.data_type   = data_type;
1494         r.e.nr_devs     = 0;
1495         r.e.nr_required = 1;
1496
1497         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1498                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
1499
1500                 if (flags & BTREE_TRIGGER_OVERWRITE)
1501                         disk_sectors = -disk_sectors;
1502
1503                 ret = bch2_trans_mark_pointer(trans, k, p,
1504                                         disk_sectors, data_type);
1505                 if (ret < 0)
1506                         return ret;
1507
1508                 stale = ret > 0;
1509
1510                 if (p.ptr.cached) {
1511                         if (!stale)
1512                                 update_cached_sectors_list(trans, p.ptr.dev,
1513                                                            disk_sectors);
1514                 } else if (!p.has_ec) {
1515                         dirty_sectors          += disk_sectors;
1516                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1517                 } else {
1518                         ret = bch2_trans_mark_stripe_ptr(trans, p,
1519                                         disk_sectors, data_type);
1520                         if (ret)
1521                                 return ret;
1522
1523                         r.e.nr_required = 0;
1524                 }
1525         }
1526
1527         if (r.e.nr_devs)
1528                 update_replicas_list(trans, &r.e, dirty_sectors);
1529
1530         return 0;
1531 }
1532
1533 static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
1534                                          struct bkey_s_c_stripe s,
1535                                          unsigned idx, bool deleting)
1536 {
1537         struct bch_fs *c = trans->c;
1538         const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
1539         struct btree_iter iter;
1540         struct bkey_i_alloc_v4 *a;
1541         enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
1542                 ? BCH_DATA_parity : 0;
1543         s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
1544         int ret = 0;
1545
1546         if (deleting)
1547                 sectors = -sectors;
1548
1549         a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(c, ptr));
1550         if (IS_ERR(a))
1551                 return PTR_ERR(a);
1552
1553         ret = check_bucket_ref(c, s.s_c, ptr, sectors, data_type,
1554                                a->v.gen, a->v.data_type,
1555                                a->v.dirty_sectors, a->v.cached_sectors);
1556         if (ret)
1557                 goto err;
1558
1559         if (!deleting) {
1560                 if (bch2_trans_inconsistent_on(a->v.stripe ||
1561                                                a->v.stripe_redundancy, trans,
1562                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
1563                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1564                                 bch2_data_types[a->v.data_type],
1565                                 a->v.dirty_sectors,
1566                                 a->v.stripe, s.k->p.offset)) {
1567                         ret = -EIO;
1568                         goto err;
1569                 }
1570
1571                 if (bch2_trans_inconsistent_on(data_type && a->v.dirty_sectors, trans,
1572                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
1573                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1574                                 bch2_data_types[a->v.data_type],
1575                                 a->v.dirty_sectors,
1576                                 s.k->p.offset)) {
1577                         ret = -EIO;
1578                         goto err;
1579                 }
1580
1581                 a->v.stripe             = s.k->p.offset;
1582                 a->v.stripe_redundancy  = s.v->nr_redundant;
1583         } else {
1584                 if (bch2_trans_inconsistent_on(a->v.stripe != s.k->p.offset ||
1585                                                a->v.stripe_redundancy != s.v->nr_redundant, trans,
1586                                 "bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
1587                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1588                                 s.k->p.offset, a->v.stripe)) {
1589                         ret = -EIO;
1590                         goto err;
1591                 }
1592
1593                 a->v.stripe             = 0;
1594                 a->v.stripe_redundancy  = 0;
1595         }
1596
1597         a->v.dirty_sectors += sectors;
1598         if (data_type)
1599                 a->v.data_type = !deleting ? data_type : 0;
1600
1601         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1602         if (ret)
1603                 goto err;
1604 err:
1605         bch2_trans_iter_exit(trans, &iter);
1606         return ret;
1607 }
1608
1609 int bch2_trans_mark_stripe(struct btree_trans *trans,
1610                            struct bkey_s_c old, struct bkey_i *new,
1611                            unsigned flags)
1612 {
1613         const struct bch_stripe *old_s = NULL;
1614         struct bch_stripe *new_s = NULL;
1615         struct bch_replicas_padded r;
1616         unsigned i, nr_blocks;
1617         int ret = 0;
1618
1619         if (old.k->type == KEY_TYPE_stripe)
1620                 old_s = bkey_s_c_to_stripe(old).v;
1621         if (new->k.type == KEY_TYPE_stripe)
1622                 new_s = &bkey_i_to_stripe(new)->v;
1623
1624         /*
1625          * If the pointers aren't changing, we don't need to do anything:
1626          */
1627         if (new_s && old_s &&
1628             new_s->nr_blocks    == old_s->nr_blocks &&
1629             new_s->nr_redundant == old_s->nr_redundant &&
1630             !memcmp(old_s->ptrs, new_s->ptrs,
1631                     new_s->nr_blocks * sizeof(struct bch_extent_ptr)))
1632                 return 0;
1633
1634         BUG_ON(new_s && old_s &&
1635                (new_s->nr_blocks        != old_s->nr_blocks ||
1636                 new_s->nr_redundant     != old_s->nr_redundant));
1637
1638         nr_blocks = new_s ? new_s->nr_blocks : old_s->nr_blocks;
1639
1640         if (new_s) {
1641                 s64 sectors = le16_to_cpu(new_s->sectors);
1642
1643                 bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(new));
1644                 update_replicas_list(trans, &r.e, sectors * new_s->nr_redundant);
1645         }
1646
1647         if (old_s) {
1648                 s64 sectors = -((s64) le16_to_cpu(old_s->sectors));
1649
1650                 bch2_bkey_to_replicas(&r.e, old);
1651                 update_replicas_list(trans, &r.e, sectors * old_s->nr_redundant);
1652         }
1653
1654         for (i = 0; i < nr_blocks; i++) {
1655                 if (new_s && old_s &&
1656                     !memcmp(&new_s->ptrs[i],
1657                             &old_s->ptrs[i],
1658                             sizeof(new_s->ptrs[i])))
1659                         continue;
1660
1661                 if (new_s) {
1662                         ret = bch2_trans_mark_stripe_bucket(trans,
1663                                         bkey_i_to_s_c_stripe(new), i, false);
1664                         if (ret)
1665                                 break;
1666                 }
1667
1668                 if (old_s) {
1669                         ret = bch2_trans_mark_stripe_bucket(trans,
1670                                         bkey_s_c_to_stripe(old), i, true);
1671                         if (ret)
1672                                 break;
1673                 }
1674         }
1675
1676         return ret;
1677 }
1678
1679 int bch2_trans_mark_inode(struct btree_trans *trans,
1680                           struct bkey_s_c old,
1681                           struct bkey_i *new,
1682                           unsigned flags)
1683 {
1684         int nr = bkey_is_inode(&new->k) - bkey_is_inode(old.k);
1685
1686         if (nr) {
1687                 struct replicas_delta_list *d =
1688                         replicas_deltas_realloc(trans, 0);
1689                 d->nr_inodes += nr;
1690         }
1691
1692         return 0;
1693 }
1694
1695 int bch2_trans_mark_reservation(struct btree_trans *trans,
1696                                 struct bkey_s_c old,
1697                                 struct bkey_i *new,
1698                                 unsigned flags)
1699 {
1700         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1701                 ? old
1702                 : bkey_i_to_s_c(new);
1703         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1704         s64 sectors = (s64) k.k->size;
1705         struct replicas_delta_list *d;
1706
1707         if (flags & BTREE_TRIGGER_OVERWRITE)
1708                 sectors = -sectors;
1709         sectors *= replicas;
1710
1711         d = replicas_deltas_realloc(trans, 0);
1712
1713         replicas = clamp_t(unsigned, replicas, 1,
1714                            ARRAY_SIZE(d->persistent_reserved));
1715
1716         d->persistent_reserved[replicas - 1] += sectors;
1717         return 0;
1718 }
1719
1720 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1721                         struct bkey_s_c_reflink_p p,
1722                         u64 *idx, unsigned flags)
1723 {
1724         struct bch_fs *c = trans->c;
1725         struct btree_iter iter;
1726         struct bkey_s_c k;
1727         struct bkey_i *n;
1728         __le64 *refcount;
1729         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1730         struct printbuf buf = PRINTBUF;
1731         int ret;
1732
1733         bch2_trans_iter_init(trans, &iter, BTREE_ID_reflink, POS(0, *idx),
1734                              BTREE_ITER_INTENT|
1735                              BTREE_ITER_WITH_UPDATES);
1736         k = bch2_btree_iter_peek_slot(&iter);
1737         ret = bkey_err(k);
1738         if (ret)
1739                 goto err;
1740
1741         n = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1742         ret = PTR_ERR_OR_ZERO(n);
1743         if (ret)
1744                 goto err;
1745
1746         bkey_reassemble(n, k);
1747
1748         refcount = bkey_refcount(n);
1749         if (!refcount) {
1750                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1751                 bch2_trans_inconsistent(trans,
1752                         "nonexistent indirect extent at %llu while marking\n  %s",
1753                         *idx, buf.buf);
1754                 ret = -EIO;
1755                 goto err;
1756         }
1757
1758         if (!*refcount && (flags & BTREE_TRIGGER_OVERWRITE)) {
1759                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1760                 bch2_trans_inconsistent(trans,
1761                         "indirect extent refcount underflow at %llu while marking\n  %s",
1762                         *idx, buf.buf);
1763                 ret = -EIO;
1764                 goto err;
1765         }
1766
1767         if (flags & BTREE_TRIGGER_INSERT) {
1768                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1769                 u64 pad;
1770
1771                 pad = max_t(s64, le32_to_cpu(v->front_pad),
1772                             le64_to_cpu(v->idx) - bkey_start_offset(k.k));
1773                 BUG_ON(pad > U32_MAX);
1774                 v->front_pad = cpu_to_le32(pad);
1775
1776                 pad = max_t(s64, le32_to_cpu(v->back_pad),
1777                             k.k->p.offset - p.k->size - le64_to_cpu(v->idx));
1778                 BUG_ON(pad > U32_MAX);
1779                 v->back_pad = cpu_to_le32(pad);
1780         }
1781
1782         le64_add_cpu(refcount, add);
1783
1784         bch2_btree_iter_set_pos_to_extent_start(&iter);
1785         ret = bch2_trans_update(trans, &iter, n, 0);
1786         if (ret)
1787                 goto err;
1788
1789         *idx = k.k->p.offset;
1790 err:
1791         bch2_trans_iter_exit(trans, &iter);
1792         printbuf_exit(&buf);
1793         return ret;
1794 }
1795
1796 int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1797                               struct bkey_s_c old,
1798                               struct bkey_i *new,
1799                               unsigned flags)
1800 {
1801         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1802                 ? old
1803                 : bkey_i_to_s_c(new);
1804         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1805         u64 idx, end_idx;
1806         int ret = 0;
1807
1808         if (flags & BTREE_TRIGGER_INSERT) {
1809                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1810
1811                 v->front_pad = v->back_pad = 0;
1812         }
1813
1814         idx     = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
1815         end_idx = le64_to_cpu(p.v->idx) + p.k->size +
1816                 le32_to_cpu(p.v->back_pad);
1817
1818         while (idx < end_idx && !ret)
1819                 ret = __bch2_trans_mark_reflink_p(trans, p, &idx, flags);
1820
1821         return ret;
1822 }
1823
1824 static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1825                                     struct bch_dev *ca, size_t b,
1826                                     enum bch_data_type type,
1827                                     unsigned sectors)
1828 {
1829         struct bch_fs *c = trans->c;
1830         struct btree_iter iter;
1831         struct bkey_i_alloc_v4 *a;
1832         int ret = 0;
1833
1834         /*
1835          * Backup superblock might be past the end of our normal usable space:
1836          */
1837         if (b >= ca->mi.nbuckets)
1838                 return 0;
1839
1840         a = bch2_trans_start_alloc_update(trans, &iter, POS(ca->dev_idx, b));
1841         if (IS_ERR(a))
1842                 return PTR_ERR(a);
1843
1844         if (a->v.data_type && a->v.data_type != type) {
1845                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1846                         "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
1847                         "while marking %s",
1848                         iter.pos.inode, iter.pos.offset, a->v.gen,
1849                         bch2_data_types[a->v.data_type],
1850                         bch2_data_types[type],
1851                         bch2_data_types[type]);
1852                 ret = -EIO;
1853                 goto out;
1854         }
1855
1856         a->v.data_type          = type;
1857         a->v.dirty_sectors      = sectors;
1858
1859         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1860         if (ret)
1861                 goto out;
1862 out:
1863         bch2_trans_iter_exit(trans, &iter);
1864         return ret;
1865 }
1866
1867 int bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1868                                     struct bch_dev *ca, size_t b,
1869                                     enum bch_data_type type,
1870                                     unsigned sectors)
1871 {
1872         return __bch2_trans_do(trans, NULL, NULL, 0,
1873                         __bch2_trans_mark_metadata_bucket(trans, ca, b, type, sectors));
1874 }
1875
1876 static int bch2_trans_mark_metadata_sectors(struct btree_trans *trans,
1877                                             struct bch_dev *ca,
1878                                             u64 start, u64 end,
1879                                             enum bch_data_type type,
1880                                             u64 *bucket, unsigned *bucket_sectors)
1881 {
1882         do {
1883                 u64 b = sector_to_bucket(ca, start);
1884                 unsigned sectors =
1885                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1886
1887                 if (b != *bucket && *bucket_sectors) {
1888                         int ret = bch2_trans_mark_metadata_bucket(trans, ca, *bucket,
1889                                                                   type, *bucket_sectors);
1890                         if (ret)
1891                                 return ret;
1892
1893                         *bucket_sectors = 0;
1894                 }
1895
1896                 *bucket         = b;
1897                 *bucket_sectors += sectors;
1898                 start += sectors;
1899         } while (start < end);
1900
1901         return 0;
1902 }
1903
1904 static int __bch2_trans_mark_dev_sb(struct btree_trans *trans,
1905                                     struct bch_dev *ca)
1906 {
1907         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1908         u64 bucket = 0;
1909         unsigned i, bucket_sectors = 0;
1910         int ret;
1911
1912         for (i = 0; i < layout->nr_superblocks; i++) {
1913                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1914
1915                 if (offset == BCH_SB_SECTOR) {
1916                         ret = bch2_trans_mark_metadata_sectors(trans, ca,
1917                                                 0, BCH_SB_SECTOR,
1918                                                 BCH_DATA_sb, &bucket, &bucket_sectors);
1919                         if (ret)
1920                                 return ret;
1921                 }
1922
1923                 ret = bch2_trans_mark_metadata_sectors(trans, ca, offset,
1924                                       offset + (1 << layout->sb_max_size_bits),
1925                                       BCH_DATA_sb, &bucket, &bucket_sectors);
1926                 if (ret)
1927                         return ret;
1928         }
1929
1930         if (bucket_sectors) {
1931                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1932                                 bucket, BCH_DATA_sb, bucket_sectors);
1933                 if (ret)
1934                         return ret;
1935         }
1936
1937         for (i = 0; i < ca->journal.nr; i++) {
1938                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1939                                 ca->journal.buckets[i],
1940                                 BCH_DATA_journal, ca->mi.bucket_size);
1941                 if (ret)
1942                         return ret;
1943         }
1944
1945         return 0;
1946 }
1947
1948 int bch2_trans_mark_dev_sb(struct bch_fs *c, struct bch_dev *ca)
1949 {
1950         return bch2_trans_do(c, NULL, NULL, BTREE_INSERT_LAZY_RW,
1951                         __bch2_trans_mark_dev_sb(&trans, ca));
1952 }
1953
1954 /* Disk reservations: */
1955
1956 #define SECTORS_CACHE   1024
1957
1958 int bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
1959                               u64 sectors, int flags)
1960 {
1961         struct bch_fs_pcpu *pcpu;
1962         u64 old, v, get;
1963         s64 sectors_available;
1964         int ret;
1965
1966         percpu_down_read(&c->mark_lock);
1967         preempt_disable();
1968         pcpu = this_cpu_ptr(c->pcpu);
1969
1970         if (sectors <= pcpu->sectors_available)
1971                 goto out;
1972
1973         v = atomic64_read(&c->sectors_available);
1974         do {
1975                 old = v;
1976                 get = min((u64) sectors + SECTORS_CACHE, old);
1977
1978                 if (get < sectors) {
1979                         preempt_enable();
1980                         goto recalculate;
1981                 }
1982         } while ((v = atomic64_cmpxchg(&c->sectors_available,
1983                                        old, old - get)) != old);
1984
1985         pcpu->sectors_available         += get;
1986
1987 out:
1988         pcpu->sectors_available         -= sectors;
1989         this_cpu_add(*c->online_reserved, sectors);
1990         res->sectors                    += sectors;
1991
1992         preempt_enable();
1993         percpu_up_read(&c->mark_lock);
1994         return 0;
1995
1996 recalculate:
1997         mutex_lock(&c->sectors_available_lock);
1998
1999         percpu_u64_set(&c->pcpu->sectors_available, 0);
2000         sectors_available = avail_factor(__bch2_fs_usage_read_short(c).free);
2001
2002         if (sectors <= sectors_available ||
2003             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
2004                 atomic64_set(&c->sectors_available,
2005                              max_t(s64, 0, sectors_available - sectors));
2006                 this_cpu_add(*c->online_reserved, sectors);
2007                 res->sectors                    += sectors;
2008                 ret = 0;
2009         } else {
2010                 atomic64_set(&c->sectors_available, sectors_available);
2011                 ret = -ENOSPC;
2012         }
2013
2014         mutex_unlock(&c->sectors_available_lock);
2015         percpu_up_read(&c->mark_lock);
2016
2017         return ret;
2018 }
2019
2020 /* Startup/shutdown: */
2021
2022 static void bucket_gens_free_rcu(struct rcu_head *rcu)
2023 {
2024         struct bucket_gens *buckets =
2025                 container_of(rcu, struct bucket_gens, rcu);
2026
2027         kvpfree(buckets, sizeof(*buckets) + buckets->nbuckets);
2028 }
2029
2030 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
2031 {
2032         struct bucket_gens *bucket_gens = NULL, *old_bucket_gens = NULL;
2033         unsigned long *buckets_nouse = NULL;
2034         bool resize = ca->bucket_gens != NULL;
2035         int ret = -ENOMEM;
2036
2037         if (!(bucket_gens       = kvpmalloc(sizeof(struct bucket_gens) + nbuckets,
2038                                             GFP_KERNEL|__GFP_ZERO)) ||
2039             (c->opts.buckets_nouse &&
2040              !(buckets_nouse    = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2041                                             sizeof(unsigned long),
2042                                             GFP_KERNEL|__GFP_ZERO))))
2043                 goto err;
2044
2045         bucket_gens->first_bucket = ca->mi.first_bucket;
2046         bucket_gens->nbuckets   = nbuckets;
2047
2048         bch2_copygc_stop(c);
2049
2050         if (resize) {
2051                 down_write(&c->gc_lock);
2052                 down_write(&ca->bucket_lock);
2053                 percpu_down_write(&c->mark_lock);
2054         }
2055
2056         old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
2057
2058         if (resize) {
2059                 size_t n = min(bucket_gens->nbuckets, old_bucket_gens->nbuckets);
2060
2061                 memcpy(bucket_gens->b,
2062                        old_bucket_gens->b,
2063                        n);
2064                 if (buckets_nouse)
2065                         memcpy(buckets_nouse,
2066                                ca->buckets_nouse,
2067                                BITS_TO_LONGS(n) * sizeof(unsigned long));
2068         }
2069
2070         rcu_assign_pointer(ca->bucket_gens, bucket_gens);
2071         bucket_gens     = old_bucket_gens;
2072
2073         swap(ca->buckets_nouse, buckets_nouse);
2074
2075         nbuckets = ca->mi.nbuckets;
2076
2077         if (resize) {
2078                 percpu_up_write(&c->mark_lock);
2079                 up_write(&ca->bucket_lock);
2080                 up_write(&c->gc_lock);
2081         }
2082
2083         ret = 0;
2084 err:
2085         kvpfree(buckets_nouse,
2086                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2087         if (bucket_gens)
2088                 call_rcu(&bucket_gens->rcu, bucket_gens_free_rcu);
2089
2090         return ret;
2091 }
2092
2093 void bch2_dev_buckets_free(struct bch_dev *ca)
2094 {
2095         unsigned i;
2096
2097         kvpfree(ca->buckets_nouse,
2098                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2099         kvpfree(rcu_dereference_protected(ca->bucket_gens, 1),
2100                 sizeof(struct bucket_gens) + ca->mi.nbuckets);
2101
2102         for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
2103                 free_percpu(ca->usage[i]);
2104         kfree(ca->usage_base);
2105 }
2106
2107 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2108 {
2109         unsigned i;
2110
2111         ca->usage_base = kzalloc(sizeof(struct bch_dev_usage), GFP_KERNEL);
2112         if (!ca->usage_base)
2113                 return -ENOMEM;
2114
2115         for (i = 0; i < ARRAY_SIZE(ca->usage); i++) {
2116                 ca->usage[i] = alloc_percpu(struct bch_dev_usage);
2117                 if (!ca->usage[i])
2118                         return -ENOMEM;
2119         }
2120
2121         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);;
2122 }