git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/buckets.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Code for manipulating bucket marks for garbage collection.
   4  *
   5  * Copyright 2014 Datera, Inc.
   6  */
   7
   8 #include "bcachefs.h"
   9 #include "alloc_background.h"
  10 #include "backpointers.h"
  11 #include "bset.h"
  12 #include "btree_gc.h"
  13 #include "btree_update.h"
  14 #include "buckets.h"
  15 #include "buckets_waiting_for_journal.h"
  16 #include "ec.h"
  17 #include "error.h"
  18 #include "inode.h"
  19 #include "movinggc.h"
  20 #include "recovery.h"
  21 #include "reflink.h"
  22 #include "replicas.h"
  23 #include "subvolume.h"
  24
  25 #include <linux/preempt.h>
  26 #include <trace/events/bcachefs.h>
  27
  28 static inline void fs_usage_data_type_to_base(struct bch_fs_usage *fs_usage,
  29                                               enum bch_data_type data_type,
  30                                               s64 sectors)
  31 {
  32         switch (data_type) {
  33         case BCH_DATA_btree:
  34                 fs_usage->btree         += sectors;
  35                 break;
  36         case BCH_DATA_user:
  37         case BCH_DATA_parity:
  38                 fs_usage->data          += sectors;
  39                 break;
  40         case BCH_DATA_cached:
  41                 fs_usage->cached        += sectors;
  42                 break;
  43         default:
  44                 break;
  45         }
  46 }
  47
  48 void bch2_fs_usage_initialize(struct bch_fs *c)
  49 {
  50         struct bch_fs_usage *usage;
  51         struct bch_dev *ca;
  52         unsigned i;
  53
  54         percpu_down_write(&c->mark_lock);
  55         usage = c->usage_base;
  56
  57         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
  58                 bch2_fs_usage_acc_to_base(c, i);
  59
  60         for (i = 0; i < BCH_REPLICAS_MAX; i++)
  61                 usage->reserved += usage->persistent_reserved[i];
  62
  63         for (i = 0; i < c->replicas.nr; i++) {
  64                 struct bch_replicas_entry *e =
  65                         cpu_replicas_entry(&c->replicas, i);
  66
  67                 fs_usage_data_type_to_base(usage, e->data_type, usage->replicas[i]);
  68         }
  69
  70         for_each_member_device(ca, c, i) {
  71                 struct bch_dev_usage dev = bch2_dev_usage_read(ca);
  72
  73                 usage->hidden += (dev.d[BCH_DATA_sb].buckets +
  74                                   dev.d[BCH_DATA_journal].buckets) *
  75                         ca->mi.bucket_size;
  76         }
  77
  78         percpu_up_write(&c->mark_lock);
  79 }
  80
  81 static inline struct bch_dev_usage *dev_usage_ptr(struct bch_dev *ca,
  82                                                   unsigned journal_seq,
  83                                                   bool gc)
  84 {
  85         BUG_ON(!gc && !journal_seq);
  86
  87         return this_cpu_ptr(gc
  88                             ? ca->usage_gc
  89                             : ca->usage[journal_seq & JOURNAL_BUF_MASK]);
  90 }
  91
  92 struct bch_dev_usage bch2_dev_usage_read(struct bch_dev *ca)
  93 {
  94         struct bch_fs *c = ca->fs;
  95         struct bch_dev_usage ret;
  96         unsigned seq, i, u64s = dev_usage_u64s();
  97
  98         do {
  99                 seq = read_seqcount_begin(&c->usage_lock);
 100                 memcpy(&ret, ca->usage_base, u64s * sizeof(u64));
 101                 for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
 102                         acc_u64s_percpu((u64 *) &ret, (u64 __percpu *) ca->usage[i], u64s);
 103         } while (read_seqcount_retry(&c->usage_lock, seq));
 104
 105         return ret;
 106 }
 107
 108 static inline struct bch_fs_usage *fs_usage_ptr(struct bch_fs *c,
 109                                                 unsigned journal_seq,
 110                                                 bool gc)
 111 {
 112         percpu_rwsem_assert_held(&c->mark_lock);
 113         BUG_ON(!gc && !journal_seq);
 114
 115         return this_cpu_ptr(gc
 116                             ? c->usage_gc
 117                             : c->usage[journal_seq & JOURNAL_BUF_MASK]);
 118 }
 119
 120 u64 bch2_fs_usage_read_one(struct bch_fs *c, u64 *v)
 121 {
 122         ssize_t offset = v - (u64 *) c->usage_base;
 123         unsigned i, seq;
 124         u64 ret;
 125
 126         BUG_ON(offset < 0 || offset >= fs_usage_u64s(c));
 127         percpu_rwsem_assert_held(&c->mark_lock);
 128
 129         do {
 130                 seq = read_seqcount_begin(&c->usage_lock);
 131                 ret = *v;
 132
 133                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 134                         ret += percpu_u64_get((u64 __percpu *) c->usage[i] + offset);
 135         } while (read_seqcount_retry(&c->usage_lock, seq));
 136
 137         return ret;
 138 }
 139
 140 struct bch_fs_usage_online *bch2_fs_usage_read(struct bch_fs *c)
 141 {
 142         struct bch_fs_usage_online *ret;
 143         unsigned seq, i, u64s;
 144
 145         percpu_down_read(&c->mark_lock);
 146
 147         ret = kmalloc(sizeof(struct bch_fs_usage_online) +
 148                       sizeof(u64) * c->replicas.nr, GFP_NOFS);
 149         if (unlikely(!ret)) {
 150                 percpu_up_read(&c->mark_lock);
 151                 return NULL;
 152         }
 153
 154         ret->online_reserved = percpu_u64_get(c->online_reserved);
 155
 156         u64s = fs_usage_u64s(c);
 157         do {
 158                 seq = read_seqcount_begin(&c->usage_lock);
 159                 memcpy(&ret->u, c->usage_base, u64s * sizeof(u64));
 160                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 161                         acc_u64s_percpu((u64 *) &ret->u, (u64 __percpu *) c->usage[i], u64s);
 162         } while (read_seqcount_retry(&c->usage_lock, seq));
 163
 164         return ret;
 165 }
 166
 167 void bch2_fs_usage_acc_to_base(struct bch_fs *c, unsigned idx)
 168 {
 169         struct bch_dev *ca;
 170         unsigned i, u64s = fs_usage_u64s(c);
 171
 172         BUG_ON(idx >= ARRAY_SIZE(c->usage));
 173
 174         preempt_disable();
 175         write_seqcount_begin(&c->usage_lock);
 176
 177         acc_u64s_percpu((u64 *) c->usage_base,
 178                         (u64 __percpu *) c->usage[idx], u64s);
 179         percpu_memset(c->usage[idx], 0, u64s * sizeof(u64));
 180
 181         rcu_read_lock();
 182         for_each_member_device_rcu(ca, c, i, NULL) {
 183                 u64s = dev_usage_u64s();
 184
 185                 acc_u64s_percpu((u64 *) ca->usage_base,
 186                                 (u64 __percpu *) ca->usage[idx], u64s);
 187                 percpu_memset(ca->usage[idx], 0, u64s * sizeof(u64));
 188         }
 189         rcu_read_unlock();
 190
 191         write_seqcount_end(&c->usage_lock);
 192         preempt_enable();
 193 }
 194
 195 void bch2_fs_usage_to_text(struct printbuf *out,
 196                            struct bch_fs *c,
 197                            struct bch_fs_usage_online *fs_usage)
 198 {
 199         unsigned i;
 200
 201         prt_printf(out, "capacity:\t\t\t%llu\n", c->capacity);
 202
 203         prt_printf(out, "hidden:\t\t\t\t%llu\n",
 204                fs_usage->u.hidden);
 205         prt_printf(out, "data:\t\t\t\t%llu\n",
 206                fs_usage->u.data);
 207         prt_printf(out, "cached:\t\t\t\t%llu\n",
 208                fs_usage->u.cached);
 209         prt_printf(out, "reserved:\t\t\t%llu\n",
 210                fs_usage->u.reserved);
 211         prt_printf(out, "nr_inodes:\t\t\t%llu\n",
 212                fs_usage->u.nr_inodes);
 213         prt_printf(out, "online reserved:\t\t%llu\n",
 214                fs_usage->online_reserved);
 215
 216         for (i = 0;
 217              i < ARRAY_SIZE(fs_usage->u.persistent_reserved);
 218              i++) {
 219                 prt_printf(out, "%u replicas:\n", i + 1);
 220                 prt_printf(out, "\treserved:\t\t%llu\n",
 221                        fs_usage->u.persistent_reserved[i]);
 222         }
 223
 224         for (i = 0; i < c->replicas.nr; i++) {
 225                 struct bch_replicas_entry *e =
 226                         cpu_replicas_entry(&c->replicas, i);
 227
 228                 prt_printf(out, "\t");
 229                 bch2_replicas_entry_to_text(out, e);
 230                 prt_printf(out, ":\t%llu\n", fs_usage->u.replicas[i]);
 231         }
 232 }
 233
 234 static u64 reserve_factor(u64 r)
 235 {
 236         return r + (round_up(r, (1 << RESERVE_FACTOR)) >> RESERVE_FACTOR);
 237 }
 238
 239 u64 bch2_fs_sectors_used(struct bch_fs *c, struct bch_fs_usage_online *fs_usage)
 240 {
 241         return min(fs_usage->u.hidden +
 242                    fs_usage->u.btree +
 243                    fs_usage->u.data +
 244                    reserve_factor(fs_usage->u.reserved +
 245                                   fs_usage->online_reserved),
 246                    c->capacity);
 247 }
 248
 249 static struct bch_fs_usage_short
 250 __bch2_fs_usage_read_short(struct bch_fs *c)
 251 {
 252         struct bch_fs_usage_short ret;
 253         u64 data, reserved;
 254
 255         ret.capacity = c->capacity -
 256                 bch2_fs_usage_read_one(c, &c->usage_base->hidden);
 257
 258         data            = bch2_fs_usage_read_one(c, &c->usage_base->data) +
 259                 bch2_fs_usage_read_one(c, &c->usage_base->btree);
 260         reserved        = bch2_fs_usage_read_one(c, &c->usage_base->reserved) +
 261                 percpu_u64_get(c->online_reserved);
 262
 263         ret.used        = min(ret.capacity, data + reserve_factor(reserved));
 264         ret.free        = ret.capacity - ret.used;
 265
 266         ret.nr_inodes   = bch2_fs_usage_read_one(c, &c->usage_base->nr_inodes);
 267
 268         return ret;
 269 }
 270
 271 struct bch_fs_usage_short
 272 bch2_fs_usage_read_short(struct bch_fs *c)
 273 {
 274         struct bch_fs_usage_short ret;
 275
 276         percpu_down_read(&c->mark_lock);
 277         ret = __bch2_fs_usage_read_short(c);
 278         percpu_up_read(&c->mark_lock);
 279
 280         return ret;
 281 }
 282
 283 void bch2_dev_usage_init(struct bch_dev *ca)
 284 {
 285         ca->usage_base->d[BCH_DATA_free].buckets = ca->mi.nbuckets - ca->mi.first_bucket;
 286 }
 287
 288 static inline int bucket_sectors_fragmented(struct bch_dev *ca,
 289                                             struct bch_alloc_v4 a)
 290 {
 291         return a.dirty_sectors
 292                 ? max(0, (int) ca->mi.bucket_size - (int) a.dirty_sectors)
 293                 : 0;
 294 }
 295
 296 static void bch2_dev_usage_update(struct bch_fs *c, struct bch_dev *ca,
 297                                   struct bch_alloc_v4 old,
 298                                   struct bch_alloc_v4 new,
 299                                   u64 journal_seq, bool gc)
 300 {
 301         struct bch_fs_usage *fs_usage;
 302         struct bch_dev_usage *u;
 303
 304         preempt_disable();
 305         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 306
 307         if (data_type_is_hidden(old.data_type))
 308                 fs_usage->hidden -= ca->mi.bucket_size;
 309         if (data_type_is_hidden(new.data_type))
 310                 fs_usage->hidden += ca->mi.bucket_size;
 311
 312         u = dev_usage_ptr(ca, journal_seq, gc);
 313
 314         u->d[old.data_type].buckets--;
 315         u->d[new.data_type].buckets++;
 316
 317         u->buckets_ec -= (int) !!old.stripe;
 318         u->buckets_ec += (int) !!new.stripe;
 319
 320         u->d[old.data_type].sectors -= old.dirty_sectors;
 321         u->d[new.data_type].sectors += new.dirty_sectors;
 322
 323         u->d[BCH_DATA_cached].sectors += new.cached_sectors;
 324         u->d[BCH_DATA_cached].sectors -= old.cached_sectors;
 325
 326         u->d[old.data_type].fragmented -= bucket_sectors_fragmented(ca, old);
 327         u->d[new.data_type].fragmented += bucket_sectors_fragmented(ca, new);
 328
 329         preempt_enable();
 330 }
 331
 332 static void bch2_dev_usage_update_m(struct bch_fs *c, struct bch_dev *ca,
 333                                     struct bucket old, struct bucket new,
 334                                     u64 journal_seq, bool gc)
 335 {
 336         struct bch_alloc_v4 old_a = {
 337                 .gen            = old.gen,
 338                 .data_type      = old.data_type,
 339                 .dirty_sectors  = old.dirty_sectors,
 340                 .cached_sectors = old.cached_sectors,
 341                 .stripe         = old.stripe,
 342         };
 343         struct bch_alloc_v4 new_a = {
 344                 .gen            = new.gen,
 345                 .data_type      = new.data_type,
 346                 .dirty_sectors  = new.dirty_sectors,
 347                 .cached_sectors = new.cached_sectors,
 348                 .stripe         = new.stripe,
 349         };
 350
 351         bch2_dev_usage_update(c, ca, old_a, new_a, journal_seq, gc);
 352 }
 353
 354 static inline int __update_replicas(struct bch_fs *c,
 355                                     struct bch_fs_usage *fs_usage,
 356                                     struct bch_replicas_entry *r,
 357                                     s64 sectors)
 358 {
 359         int idx = bch2_replicas_entry_idx(c, r);
 360
 361         if (idx < 0)
 362                 return -1;
 363
 364         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 365         fs_usage->replicas[idx]         += sectors;
 366         return 0;
 367 }
 368
 369 static inline int update_replicas(struct bch_fs *c, struct bkey_s_c k,
 370                         struct bch_replicas_entry *r, s64 sectors,
 371                         unsigned journal_seq, bool gc)
 372 {
 373         struct bch_fs_usage __percpu *fs_usage;
 374         int idx, ret = 0;
 375         struct printbuf buf = PRINTBUF;
 376
 377         percpu_down_read(&c->mark_lock);
 378         buf.atomic++;
 379
 380         idx = bch2_replicas_entry_idx(c, r);
 381         if (idx < 0 &&
 382             fsck_err(c, "no replicas entry\n"
 383                      "  while marking %s",
 384                      (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 385                 percpu_up_read(&c->mark_lock);
 386                 ret = bch2_mark_replicas(c, r);
 387                 percpu_down_read(&c->mark_lock);
 388
 389                 if (ret)
 390                         goto err;
 391                 idx = bch2_replicas_entry_idx(c, r);
 392         }
 393         if (idx < 0) {
 394                 ret = -1;
 395                 goto err;
 396         }
 397
 398         preempt_disable();
 399         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 400         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 401         fs_usage->replicas[idx]         += sectors;
 402         preempt_enable();
 403 err:
 404 fsck_err:
 405         percpu_up_read(&c->mark_lock);
 406         printbuf_exit(&buf);
 407         return ret;
 408 }
 409
 410 static inline int update_cached_sectors(struct bch_fs *c,
 411                         struct bkey_s_c k,
 412                         unsigned dev, s64 sectors,
 413                         unsigned journal_seq, bool gc)
 414 {
 415         struct bch_replicas_padded r;
 416
 417         bch2_replicas_entry_cached(&r.e, dev);
 418
 419         return update_replicas(c, k, &r.e, sectors, journal_seq, gc);
 420 }
 421
 422 static struct replicas_delta_list *
 423 replicas_deltas_realloc(struct btree_trans *trans, unsigned more)
 424 {
 425         struct replicas_delta_list *d = trans->fs_usage_deltas;
 426         unsigned new_size = d ? (d->size + more) * 2 : 128;
 427         unsigned alloc_size = sizeof(*d) + new_size;
 428
 429         WARN_ON_ONCE(alloc_size > REPLICAS_DELTA_LIST_MAX);
 430
 431         if (!d || d->used + more > d->size) {
 432                 d = krealloc(d, alloc_size, GFP_NOIO|__GFP_ZERO);
 433
 434                 BUG_ON(!d && alloc_size > REPLICAS_DELTA_LIST_MAX);
 435
 436                 if (!d) {
 437                         d = mempool_alloc(&trans->c->replicas_delta_pool, GFP_NOIO);
 438                         memset(d, 0, REPLICAS_DELTA_LIST_MAX);
 439
 440                         if (trans->fs_usage_deltas)
 441                                 memcpy(d, trans->fs_usage_deltas,
 442                                        trans->fs_usage_deltas->size + sizeof(*d));
 443
 444                         new_size = REPLICAS_DELTA_LIST_MAX - sizeof(*d);
 445                         kfree(trans->fs_usage_deltas);
 446                 }
 447
 448                 d->size = new_size;
 449                 trans->fs_usage_deltas = d;
 450         }
 451         return d;
 452 }
 453
 454 static inline void update_replicas_list(struct btree_trans *trans,
 455                                         struct bch_replicas_entry *r,
 456                                         s64 sectors)
 457 {
 458         struct replicas_delta_list *d;
 459         struct replicas_delta *n;
 460         unsigned b;
 461
 462         if (!sectors)
 463                 return;
 464
 465         b = replicas_entry_bytes(r) + 8;
 466         d = replicas_deltas_realloc(trans, b);
 467
 468         n = (void *) d->d + d->used;
 469         n->delta = sectors;
 470         memcpy((void *) n + offsetof(struct replicas_delta, r),
 471                r, replicas_entry_bytes(r));
 472         bch2_replicas_entry_sort(&n->r);
 473         d->used += b;
 474 }
 475
 476 static inline void update_cached_sectors_list(struct btree_trans *trans,
 477                                               unsigned dev, s64 sectors)
 478 {
 479         struct bch_replicas_padded r;
 480
 481         bch2_replicas_entry_cached(&r.e, dev);
 482
 483         update_replicas_list(trans, &r.e, sectors);
 484 }
 485
 486 int bch2_mark_alloc(struct btree_trans *trans,
 487                     struct bkey_s_c old, struct bkey_s_c new,
 488                     unsigned flags)
 489 {
 490         bool gc = flags & BTREE_TRIGGER_GC;
 491         u64 journal_seq = trans->journal_res.seq;
 492         struct bch_fs *c = trans->c;
 493         struct bch_alloc_v4 old_a, new_a;
 494         struct bch_dev *ca;
 495         int ret = 0;
 496
 497         /*
 498          * alloc btree is read in by bch2_alloc_read, not gc:
 499          */
 500         if ((flags & BTREE_TRIGGER_GC) &&
 501             !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
 502                 return 0;
 503
 504         if (bch2_trans_inconsistent_on(!bch2_dev_bucket_exists(c, new.k->p), trans,
 505                                        "alloc key for invalid device or bucket"))
 506                 return -EIO;
 507
 508         ca = bch_dev_bkey_exists(c, new.k->p.inode);
 509
 510         bch2_alloc_to_v4(old, &old_a);
 511         bch2_alloc_to_v4(new, &new_a);
 512
 513         if ((flags & BTREE_TRIGGER_INSERT) &&
 514             data_type_is_empty(old_a.data_type) !=
 515             data_type_is_empty(new_a.data_type) &&
 516             new.k->type == KEY_TYPE_alloc_v4) {
 517                 struct bch_alloc_v4 *v = (struct bch_alloc_v4 *) new.v;
 518
 519                 BUG_ON(!journal_seq);
 520
 521                 /*
 522                  * If the btree updates referring to a bucket weren't flushed
 523                  * before the bucket became empty again, then the we don't have
 524                  * to wait on a journal flush before we can reuse the bucket:
 525                  */
 526                 new_a.journal_seq = data_type_is_empty(new_a.data_type) &&
 527                         (journal_seq == v->journal_seq ||
 528                          bch2_journal_noflush_seq(&c->journal, v->journal_seq))
 529                         ? 0 : journal_seq;
 530                 v->journal_seq = new_a.journal_seq;
 531         }
 532
 533         if (!data_type_is_empty(old_a.data_type) &&
 534             data_type_is_empty(new_a.data_type) &&
 535             new_a.journal_seq) {
 536                 ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 537                                 c->journal.flushed_seq_ondisk,
 538                                 new.k->p.inode, new.k->p.offset,
 539                                 new_a.journal_seq);
 540                 if (ret) {
 541                         bch2_fs_fatal_error(c,
 542                                 "error setting bucket_needs_journal_commit: %i", ret);
 543                         return ret;
 544                 }
 545         }
 546
 547         if (new_a.data_type == BCH_DATA_free &&
 548             (!new_a.journal_seq || new_a.journal_seq < c->journal.flushed_seq_ondisk))
 549                 closure_wake_up(&c->freelist_wait);
 550
 551         if (new_a.data_type == BCH_DATA_need_discard &&
 552             (!new_a.journal_seq || new_a.journal_seq < c->journal.flushed_seq_ondisk))
 553                 bch2_do_discards(c);
 554
 555         if (old_a.data_type != BCH_DATA_cached &&
 556             new_a.data_type == BCH_DATA_cached &&
 557             should_invalidate_buckets(ca, bch2_dev_usage_read(ca)))
 558                 bch2_do_invalidates(c);
 559
 560         if (new_a.data_type == BCH_DATA_need_gc_gens)
 561                 bch2_do_gc_gens(c);
 562
 563         percpu_down_read(&c->mark_lock);
 564         if (!gc && new_a.gen != old_a.gen)
 565                 *bucket_gen(ca, new.k->p.offset) = new_a.gen;
 566
 567         bch2_dev_usage_update(c, ca, old_a, new_a, journal_seq, gc);
 568
 569         if (gc) {
 570                 struct bucket *g = gc_bucket(ca, new.k->p.offset);
 571
 572                 bucket_lock(g);
 573
 574                 g->gen_valid            = 1;
 575                 g->gen                  = new_a.gen;
 576                 g->data_type            = new_a.data_type;
 577                 g->stripe               = new_a.stripe;
 578                 g->stripe_redundancy    = new_a.stripe_redundancy;
 579                 g->dirty_sectors        = new_a.dirty_sectors;
 580                 g->cached_sectors       = new_a.cached_sectors;
 581
 582                 bucket_unlock(g);
 583         }
 584         percpu_up_read(&c->mark_lock);
 585
 586         /*
 587          * need to know if we're getting called from the invalidate path or
 588          * not:
 589          */
 590
 591         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 592             old_a.cached_sectors) {
 593                 ret = update_cached_sectors(c, new, ca->dev_idx,
 594                                             -old_a.cached_sectors,
 595                                             journal_seq, gc);
 596                 if (ret) {
 597                         bch2_fs_fatal_error(c, "bch2_mark_alloc(): no replicas entry while updating cached sectors");
 598                         return ret;
 599                 }
 600         }
 601
 602         return 0;
 603 }
 604
 605 int bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 606                               size_t b, enum bch_data_type data_type,
 607                               unsigned sectors, struct gc_pos pos,
 608                               unsigned flags)
 609 {
 610         struct bucket old, new, *g;
 611         int ret = 0;
 612
 613         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 614         BUG_ON(data_type != BCH_DATA_sb &&
 615                data_type != BCH_DATA_journal);
 616
 617         /*
 618          * Backup superblock might be past the end of our normal usable space:
 619          */
 620         if (b >= ca->mi.nbuckets)
 621                 return 0;
 622
 623         percpu_down_read(&c->mark_lock);
 624         g = gc_bucket(ca, b);
 625
 626         bucket_lock(g);
 627         old = *g;
 628
 629         if (bch2_fs_inconsistent_on(g->data_type &&
 630                         g->data_type != data_type, c,
 631                         "different types of data in same bucket: %s, %s",
 632                         bch2_data_types[g->data_type],
 633                         bch2_data_types[data_type])) {
 634                 ret = -EIO;
 635                 goto err;
 636         }
 637
 638         if (bch2_fs_inconsistent_on((u64) g->dirty_sectors + sectors > ca->mi.bucket_size, c,
 639                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > bucket size",
 640                         ca->dev_idx, b, g->gen,
 641                         bch2_data_types[g->data_type ?: data_type],
 642                         g->dirty_sectors, sectors)) {
 643                 ret = -EIO;
 644                 goto err;
 645         }
 646
 647
 648         g->data_type = data_type;
 649         g->dirty_sectors += sectors;
 650         new = *g;
 651 err:
 652         bucket_unlock(g);
 653         if (!ret)
 654                 bch2_dev_usage_update_m(c, ca, old, new, 0, true);
 655         percpu_up_read(&c->mark_lock);
 656         return ret;
 657 }
 658
 659 static int check_bucket_ref(struct bch_fs *c,
 660                             struct bkey_s_c k,
 661                             const struct bch_extent_ptr *ptr,
 662                             s64 sectors, enum bch_data_type ptr_data_type,
 663                             u8 b_gen, u8 bucket_data_type,
 664                             u32 dirty_sectors, u32 cached_sectors)
 665 {
 666         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 667         size_t bucket_nr = PTR_BUCKET_NR(ca, ptr);
 668         u16 bucket_sectors = !ptr->cached
 669                 ? dirty_sectors
 670                 : cached_sectors;
 671         struct printbuf buf = PRINTBUF;
 672         int ret = 0;
 673
 674         if (bucket_data_type == BCH_DATA_cached)
 675                 bucket_data_type = BCH_DATA_user;
 676
 677         if (gen_after(ptr->gen, b_gen)) {
 678                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 679                         "bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\n"
 680                         "while marking %s",
 681                         ptr->dev, bucket_nr, b_gen,
 682                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 683                         ptr->gen,
 684                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 685                 ret = -EIO;
 686                 goto err;
 687         }
 688
 689         if (gen_cmp(b_gen, ptr->gen) > BUCKET_GC_GEN_MAX) {
 690                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 691                         "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 692                         "while marking %s",
 693                         ptr->dev, bucket_nr, b_gen,
 694                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 695                         ptr->gen,
 696                         (printbuf_reset(&buf),
 697                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 698                 ret = -EIO;
 699                 goto err;
 700         }
 701
 702         if (b_gen != ptr->gen && !ptr->cached) {
 703                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 704                         "bucket %u:%zu gen %u (mem gen %u) data type %s: stale dirty ptr (gen %u)\n"
 705                         "while marking %s",
 706                         ptr->dev, bucket_nr, b_gen,
 707                         *bucket_gen(ca, bucket_nr),
 708                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 709                         ptr->gen,
 710                         (printbuf_reset(&buf),
 711                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 712                 ret = -EIO;
 713                 goto err;
 714         }
 715
 716         if (b_gen != ptr->gen) {
 717                 ret = 1;
 718                 goto err;
 719         }
 720
 721         if (!data_type_is_empty(bucket_data_type) &&
 722             ptr_data_type &&
 723             bucket_data_type != ptr_data_type) {
 724                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 725                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 726                         "while marking %s",
 727                         ptr->dev, bucket_nr, b_gen,
 728                         bch2_data_types[bucket_data_type],
 729                         bch2_data_types[ptr_data_type],
 730                         (printbuf_reset(&buf),
 731                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 732                 ret = -EIO;
 733                 goto err;
 734         }
 735
 736         if ((unsigned) (bucket_sectors + sectors) > U32_MAX) {
 737                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 738                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U16_MAX\n"
 739                         "while marking %s",
 740                         ptr->dev, bucket_nr, b_gen,
 741                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 742                         bucket_sectors, sectors,
 743                         (printbuf_reset(&buf),
 744                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 745                 ret = -EIO;
 746                 goto err;
 747         }
 748 err:
 749         printbuf_exit(&buf);
 750         return ret;
 751 }
 752
 753 static int mark_stripe_bucket(struct btree_trans *trans,
 754                               struct bkey_s_c k,
 755                               unsigned ptr_idx,
 756                               unsigned flags)
 757 {
 758         struct bch_fs *c = trans->c;
 759         u64 journal_seq = trans->journal_res.seq;
 760         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 761         unsigned nr_data = s->nr_blocks - s->nr_redundant;
 762         bool parity = ptr_idx >= nr_data;
 763         enum bch_data_type data_type = parity ? BCH_DATA_parity : 0;
 764         s64 sectors = parity ? le16_to_cpu(s->sectors) : 0;
 765         const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
 766         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 767         struct bucket old, new, *g;
 768         struct printbuf buf = PRINTBUF;
 769         int ret = 0;
 770
 771         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 772
 773         /* * XXX doesn't handle deletion */
 774
 775         percpu_down_read(&c->mark_lock);
 776         buf.atomic++;
 777         g = PTR_GC_BUCKET(ca, ptr);
 778
 779         if (g->dirty_sectors ||
 780             (g->stripe && g->stripe != k.k->p.offset)) {
 781                 bch2_fs_inconsistent(c,
 782                               "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
 783                               ptr->dev, PTR_BUCKET_NR(ca, ptr), g->gen,
 784                               (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 785                 ret = -EINVAL;
 786                 goto err;
 787         }
 788
 789         bucket_lock(g);
 790         old = *g;
 791
 792         ret = check_bucket_ref(c, k, ptr, sectors, data_type,
 793                                g->gen, g->data_type,
 794                                g->dirty_sectors, g->cached_sectors);
 795         if (ret)
 796                 goto err;
 797
 798         if (data_type)
 799                 g->data_type = data_type;
 800         g->dirty_sectors += sectors;
 801
 802         g->stripe               = k.k->p.offset;
 803         g->stripe_redundancy    = s->nr_redundant;
 804         new = *g;
 805 err:
 806         bucket_unlock(g);
 807         if (!ret)
 808                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 809         percpu_up_read(&c->mark_lock);
 810         printbuf_exit(&buf);
 811         return ret;
 812 }
 813
 814 static int __mark_pointer(struct btree_trans *trans,
 815                           struct bkey_s_c k,
 816                           const struct bch_extent_ptr *ptr,
 817                           s64 sectors, enum bch_data_type ptr_data_type,
 818                           u8 bucket_gen, u8 *bucket_data_type,
 819                           u32 *dirty_sectors, u32 *cached_sectors)
 820 {
 821         u32 *dst_sectors = !ptr->cached
 822                 ? dirty_sectors
 823                 : cached_sectors;
 824         int ret = check_bucket_ref(trans->c, k, ptr, sectors, ptr_data_type,
 825                                    bucket_gen, *bucket_data_type,
 826                                    *dirty_sectors, *cached_sectors);
 827
 828         if (ret)
 829                 return ret;
 830
 831         *dst_sectors += sectors;
 832         *bucket_data_type = *dirty_sectors || *cached_sectors
 833                 ? ptr_data_type : 0;
 834         return 0;
 835 }
 836
 837 static int bch2_mark_pointer(struct btree_trans *trans,
 838                              struct bkey_s_c k,
 839                              struct extent_ptr_decoded p,
 840                              s64 sectors, enum bch_data_type data_type,
 841                              unsigned flags)
 842 {
 843         u64 journal_seq = trans->journal_res.seq;
 844         struct bch_fs *c = trans->c;
 845         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 846         struct bucket old, new, *g;
 847         u8 bucket_data_type;
 848         int ret = 0;
 849
 850         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 851
 852         percpu_down_read(&c->mark_lock);
 853         g = PTR_GC_BUCKET(ca, &p.ptr);
 854         bucket_lock(g);
 855         old = *g;
 856
 857         bucket_data_type = g->data_type;
 858         ret = __mark_pointer(trans, k, &p.ptr, sectors,
 859                              data_type, g->gen,
 860                              &bucket_data_type,
 861                              &g->dirty_sectors,
 862                              &g->cached_sectors);
 863         if (!ret)
 864                 g->data_type = bucket_data_type;
 865
 866         new = *g;
 867         bucket_unlock(g);
 868         if (!ret)
 869                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 870         percpu_up_read(&c->mark_lock);
 871
 872         return ret;
 873 }
 874
 875 static int bch2_mark_stripe_ptr(struct btree_trans *trans,
 876                                 struct bkey_s_c k,
 877                                 struct bch_extent_stripe_ptr p,
 878                                 enum bch_data_type data_type,
 879                                 s64 sectors,
 880                                 unsigned flags)
 881 {
 882         struct bch_fs *c = trans->c;
 883         struct bch_replicas_padded r;
 884         struct gc_stripe *m;
 885
 886         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 887
 888         m = genradix_ptr_alloc(&c->gc_stripes, p.idx, GFP_KERNEL);
 889         if (!m) {
 890                 bch_err(c, "error allocating memory for gc_stripes, idx %llu",
 891                         (u64) p.idx);
 892                 return -ENOMEM;
 893         }
 894
 895         spin_lock(&c->ec_stripes_heap_lock);
 896
 897         if (!m || !m->alive) {
 898                 spin_unlock(&c->ec_stripes_heap_lock);
 899                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
 900                                     (u64) p.idx);
 901                 bch2_inconsistent_error(c);
 902                 return -EIO;
 903         }
 904
 905         m->block_sectors[p.block] += sectors;
 906
 907         r = m->r;
 908         spin_unlock(&c->ec_stripes_heap_lock);
 909
 910         r.e.data_type = data_type;
 911         update_replicas(c, k, &r.e, sectors, trans->journal_res.seq, true);
 912
 913         return 0;
 914 }
 915
 916 int bch2_mark_extent(struct btree_trans *trans,
 917                      struct bkey_s_c old, struct bkey_s_c new,
 918                      unsigned flags)
 919 {
 920         u64 journal_seq = trans->journal_res.seq;
 921         struct bch_fs *c = trans->c;
 922         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
 923         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 924         const union bch_extent_entry *entry;
 925         struct extent_ptr_decoded p;
 926         struct bch_replicas_padded r;
 927         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
 928                 ? BCH_DATA_btree
 929                 : BCH_DATA_user;
 930         s64 sectors = bkey_is_btree_ptr(k.k)
 931                 ? btree_sectors(c)
 932                 : k.k->size;
 933         s64 dirty_sectors = 0;
 934         bool stale;
 935         int ret;
 936
 937         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 938
 939         r.e.data_type   = data_type;
 940         r.e.nr_devs     = 0;
 941         r.e.nr_required = 1;
 942
 943         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 944                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
 945
 946                 if (flags & BTREE_TRIGGER_OVERWRITE)
 947                         disk_sectors = -disk_sectors;
 948
 949                 ret = bch2_mark_pointer(trans, k, p, disk_sectors,
 950                                         data_type, flags);
 951                 if (ret < 0)
 952                         return ret;
 953
 954                 stale = ret > 0;
 955
 956                 if (p.ptr.cached) {
 957                         if (!stale) {
 958                                 ret = update_cached_sectors(c, k, p.ptr.dev,
 959                                                 disk_sectors, journal_seq, true);
 960                                 if (ret) {
 961                                         bch2_fs_fatal_error(c, "bch2_mark_extent(): no replicas entry while updating cached sectors");
 962                                         return ret;
 963                                 }
 964                         }
 965                 } else if (!p.has_ec) {
 966                         dirty_sectors          += disk_sectors;
 967                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
 968                 } else {
 969                         ret = bch2_mark_stripe_ptr(trans, k, p.ec, data_type,
 970                                         disk_sectors, flags);
 971                         if (ret)
 972                                 return ret;
 973
 974                         /*
 975                          * There may be other dirty pointers in this extent, but
 976                          * if so they're not required for mounting if we have an
 977                          * erasure coded pointer in this extent:
 978                          */
 979                         r.e.nr_required = 0;
 980                 }
 981         }
 982
 983         if (r.e.nr_devs) {
 984                 ret = update_replicas(c, k, &r.e, dirty_sectors, journal_seq, true);
 985                 if (ret) {
 986                         struct printbuf buf = PRINTBUF;
 987
 988                         bch2_bkey_val_to_text(&buf, c, k);
 989                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
 990                         printbuf_exit(&buf);
 991                         return ret;
 992                 }
 993         }
 994
 995         return 0;
 996 }
 997
 998 int bch2_mark_stripe(struct btree_trans *trans,
 999                      struct bkey_s_c old, struct bkey_s_c new,
1000                      unsigned flags)
1001 {
1002         bool gc = flags & BTREE_TRIGGER_GC;
1003         u64 journal_seq = trans->journal_res.seq;
1004         struct bch_fs *c = trans->c;
1005         u64 idx = new.k->p.offset;
1006         const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
1007                 ? bkey_s_c_to_stripe(old).v : NULL;
1008         const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
1009                 ? bkey_s_c_to_stripe(new).v : NULL;
1010         unsigned i;
1011         int ret;
1012
1013         BUG_ON(gc && old_s);
1014
1015         if (!gc) {
1016                 struct stripe *m = genradix_ptr(&c->stripes, idx);
1017
1018                 if (!m || (old_s && !m->alive)) {
1019                         struct printbuf buf1 = PRINTBUF;
1020                         struct printbuf buf2 = PRINTBUF;
1021
1022                         bch2_bkey_val_to_text(&buf1, c, old);
1023                         bch2_bkey_val_to_text(&buf2, c, new);
1024                         bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
1025                                             "old %s\n"
1026                                             "new %s", idx, buf1.buf, buf2.buf);
1027                         printbuf_exit(&buf2);
1028                         printbuf_exit(&buf1);
1029                         bch2_inconsistent_error(c);
1030                         return -1;
1031                 }
1032
1033                 if (!new_s) {
1034                         spin_lock(&c->ec_stripes_heap_lock);
1035                         bch2_stripes_heap_del(c, m, idx);
1036                         spin_unlock(&c->ec_stripes_heap_lock);
1037
1038                         memset(m, 0, sizeof(*m));
1039                 } else {
1040                         m->alive        = true;
1041                         m->sectors      = le16_to_cpu(new_s->sectors);
1042                         m->algorithm    = new_s->algorithm;
1043                         m->nr_blocks    = new_s->nr_blocks;
1044                         m->nr_redundant = new_s->nr_redundant;
1045                         m->blocks_nonempty = 0;
1046
1047                         for (i = 0; i < new_s->nr_blocks; i++)
1048                                 m->blocks_nonempty += !!stripe_blockcount_get(new_s, i);
1049
1050                         spin_lock(&c->ec_stripes_heap_lock);
1051                         bch2_stripes_heap_update(c, m, idx);
1052                         spin_unlock(&c->ec_stripes_heap_lock);
1053                 }
1054         } else {
1055                 struct gc_stripe *m =
1056                         genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
1057
1058                 if (!m) {
1059                         bch_err(c, "error allocating memory for gc_stripes, idx %llu",
1060                                 idx);
1061                         return -ENOMEM;
1062                 }
1063                 /*
1064                  * This will be wrong when we bring back runtime gc: we should
1065                  * be unmarking the old key and then marking the new key
1066                  */
1067                 m->alive        = true;
1068                 m->sectors      = le16_to_cpu(new_s->sectors);
1069                 m->nr_blocks    = new_s->nr_blocks;
1070                 m->nr_redundant = new_s->nr_redundant;
1071
1072                 for (i = 0; i < new_s->nr_blocks; i++)
1073                         m->ptrs[i] = new_s->ptrs[i];
1074
1075                 bch2_bkey_to_replicas(&m->r.e, new);
1076
1077                 /*
1078                  * gc recalculates this field from stripe ptr
1079                  * references:
1080                  */
1081                 memset(m->block_sectors, 0, sizeof(m->block_sectors));
1082
1083                 for (i = 0; i < new_s->nr_blocks; i++) {
1084                         ret = mark_stripe_bucket(trans, new, i, flags);
1085                         if (ret)
1086                                 return ret;
1087                 }
1088
1089                 ret = update_replicas(c, new, &m->r.e,
1090                                       ((s64) m->sectors * m->nr_redundant),
1091                                       journal_seq, gc);
1092                 if (ret) {
1093                         struct printbuf buf = PRINTBUF;
1094
1095                         bch2_bkey_val_to_text(&buf, c, new);
1096                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
1097                         printbuf_exit(&buf);
1098                         return ret;
1099                 }
1100         }
1101
1102         return 0;
1103 }
1104
1105 int bch2_mark_inode(struct btree_trans *trans,
1106                     struct bkey_s_c old, struct bkey_s_c new,
1107                     unsigned flags)
1108 {
1109         struct bch_fs *c = trans->c;
1110         struct bch_fs_usage __percpu *fs_usage;
1111         u64 journal_seq = trans->journal_res.seq;
1112
1113         if (flags & BTREE_TRIGGER_INSERT) {
1114                 struct bch_inode_v2 *v = (struct bch_inode_v2 *) new.v;
1115
1116                 BUG_ON(!journal_seq);
1117                 BUG_ON(new.k->type != KEY_TYPE_inode_v2);
1118
1119                 v->bi_journal_seq = cpu_to_le64(journal_seq);
1120         }
1121
1122         if (flags & BTREE_TRIGGER_GC) {
1123                 percpu_down_read(&c->mark_lock);
1124                 preempt_disable();
1125
1126                 fs_usage = fs_usage_ptr(c, journal_seq, flags & BTREE_TRIGGER_GC);
1127                 fs_usage->nr_inodes += bkey_is_inode(new.k);
1128                 fs_usage->nr_inodes -= bkey_is_inode(old.k);
1129
1130                 preempt_enable();
1131                 percpu_up_read(&c->mark_lock);
1132         }
1133         return 0;
1134 }
1135
1136 int bch2_mark_reservation(struct btree_trans *trans,
1137                           struct bkey_s_c old, struct bkey_s_c new,
1138                           unsigned flags)
1139 {
1140         struct bch_fs *c = trans->c;
1141         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1142         struct bch_fs_usage __percpu *fs_usage;
1143         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1144         s64 sectors = (s64) k.k->size;
1145
1146         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1147
1148         if (flags & BTREE_TRIGGER_OVERWRITE)
1149                 sectors = -sectors;
1150         sectors *= replicas;
1151
1152         percpu_down_read(&c->mark_lock);
1153         preempt_disable();
1154
1155         fs_usage = fs_usage_ptr(c, trans->journal_res.seq, flags & BTREE_TRIGGER_GC);
1156         replicas = clamp_t(unsigned, replicas, 1,
1157                            ARRAY_SIZE(fs_usage->persistent_reserved));
1158
1159         fs_usage->reserved                              += sectors;
1160         fs_usage->persistent_reserved[replicas - 1]     += sectors;
1161
1162         preempt_enable();
1163         percpu_up_read(&c->mark_lock);
1164
1165         return 0;
1166 }
1167
1168 static s64 __bch2_mark_reflink_p(struct btree_trans *trans,
1169                                  struct bkey_s_c_reflink_p p,
1170                                  u64 start, u64 end,
1171                                  u64 *idx, unsigned flags, size_t r_idx)
1172 {
1173         struct bch_fs *c = trans->c;
1174         struct reflink_gc *r;
1175         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1176         u64 next_idx = end;
1177         s64 ret = 0;
1178         struct printbuf buf = PRINTBUF;
1179
1180         if (r_idx >= c->reflink_gc_nr)
1181                 goto not_found;
1182
1183         r = genradix_ptr(&c->reflink_gc_table, r_idx);
1184         next_idx = min(next_idx, r->offset - r->size);
1185         if (*idx < next_idx)
1186                 goto not_found;
1187
1188         BUG_ON((s64) r->refcount + add < 0);
1189
1190         r->refcount += add;
1191         *idx = r->offset;
1192         return 0;
1193 not_found:
1194         if (fsck_err(c, "pointer to missing indirect extent\n"
1195                      "  %s\n"
1196                      "  missing range %llu-%llu",
1197                      (bch2_bkey_val_to_text(&buf, c, p.s_c), buf.buf),
1198                      *idx, next_idx)) {
1199                 struct bkey_i_error new;
1200
1201                 bkey_init(&new.k);
1202                 new.k.type      = KEY_TYPE_error;
1203                 new.k.p         = bkey_start_pos(p.k);
1204                 new.k.p.offset += *idx - start;
1205                 bch2_key_resize(&new.k, next_idx - *idx);
1206                 ret = __bch2_btree_insert(trans, BTREE_ID_extents, &new.k_i);
1207         }
1208
1209         *idx = next_idx;
1210 fsck_err:
1211         printbuf_exit(&buf);
1212         return ret;
1213 }
1214
1215 int bch2_mark_reflink_p(struct btree_trans *trans,
1216                         struct bkey_s_c old, struct bkey_s_c new,
1217                         unsigned flags)
1218 {
1219         struct bch_fs *c = trans->c;
1220         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1221         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1222         struct reflink_gc *ref;
1223         size_t l, r, m;
1224         u64 idx = le64_to_cpu(p.v->idx), start = idx;
1225         u64 end = le64_to_cpu(p.v->idx) + p.k->size;
1226         int ret = 0;
1227
1228         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1229
1230         if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix) {
1231                 idx -= le32_to_cpu(p.v->front_pad);
1232                 end += le32_to_cpu(p.v->back_pad);
1233         }
1234
1235         l = 0;
1236         r = c->reflink_gc_nr;
1237         while (l < r) {
1238                 m = l + (r - l) / 2;
1239
1240                 ref = genradix_ptr(&c->reflink_gc_table, m);
1241                 if (ref->offset <= idx)
1242                         l = m + 1;
1243                 else
1244                         r = m;
1245         }
1246
1247         while (idx < end && !ret)
1248                 ret = __bch2_mark_reflink_p(trans, p, start, end,
1249                                             &idx, flags, l++);
1250
1251         return ret;
1252 }
1253
1254 static noinline __cold
1255 void fs_usage_apply_warn(struct btree_trans *trans,
1256                          unsigned disk_res_sectors,
1257                          s64 should_not_have_added)
1258 {
1259         struct bch_fs *c = trans->c;
1260         struct btree_insert_entry *i;
1261         struct printbuf buf = PRINTBUF;
1262
1263         bch_err(c, "disk usage increased %lli more than %u sectors reserved",
1264                 should_not_have_added, disk_res_sectors);
1265
1266         trans_for_each_update(trans, i) {
1267                 struct bkey_s_c old = { &i->old_k, i->old_v };
1268
1269                 pr_err("while inserting");
1270                 printbuf_reset(&buf);
1271                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(i->k));
1272                 pr_err("  %s", buf.buf);
1273                 pr_err("overlapping with");
1274                 printbuf_reset(&buf);
1275                 bch2_bkey_val_to_text(&buf, c, old);
1276                 pr_err("  %s", buf.buf);
1277         }
1278
1279         __WARN();
1280         printbuf_exit(&buf);
1281 }
1282
1283 int bch2_trans_fs_usage_apply(struct btree_trans *trans,
1284                               struct replicas_delta_list *deltas)
1285 {
1286         struct bch_fs *c = trans->c;
1287         static int warned_disk_usage = 0;
1288         bool warn = false;
1289         unsigned disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1290         struct replicas_delta *d = deltas->d, *d2;
1291         struct replicas_delta *top = (void *) deltas->d + deltas->used;
1292         struct bch_fs_usage *dst;
1293         s64 added = 0, should_not_have_added;
1294         unsigned i;
1295
1296         percpu_down_read(&c->mark_lock);
1297         preempt_disable();
1298         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1299
1300         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1301                 switch (d->r.data_type) {
1302                 case BCH_DATA_btree:
1303                 case BCH_DATA_user:
1304                 case BCH_DATA_parity:
1305                         added += d->delta;
1306                 }
1307
1308                 if (__update_replicas(c, dst, &d->r, d->delta))
1309                         goto need_mark;
1310         }
1311
1312         dst->nr_inodes += deltas->nr_inodes;
1313
1314         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1315                 added                           += deltas->persistent_reserved[i];
1316                 dst->reserved                   += deltas->persistent_reserved[i];
1317                 dst->persistent_reserved[i]     += deltas->persistent_reserved[i];
1318         }
1319
1320         /*
1321          * Not allowed to reduce sectors_available except by getting a
1322          * reservation:
1323          */
1324         should_not_have_added = added - (s64) disk_res_sectors;
1325         if (unlikely(should_not_have_added > 0)) {
1326                 u64 old, new, v = atomic64_read(&c->sectors_available);
1327
1328                 do {
1329                         old = v;
1330                         new = max_t(s64, 0, old - should_not_have_added);
1331                 } while ((v = atomic64_cmpxchg(&c->sectors_available,
1332                                                old, new)) != old);
1333
1334                 added -= should_not_have_added;
1335                 warn = true;
1336         }
1337
1338         if (added > 0) {
1339                 trans->disk_res->sectors -= added;
1340                 this_cpu_sub(*c->online_reserved, added);
1341         }
1342
1343         preempt_enable();
1344         percpu_up_read(&c->mark_lock);
1345
1346         if (unlikely(warn) && !xchg(&warned_disk_usage, 1))
1347                 fs_usage_apply_warn(trans, disk_res_sectors, should_not_have_added);
1348         return 0;
1349 need_mark:
1350         /* revert changes: */
1351         for (d2 = deltas->d; d2 != d; d2 = replicas_delta_next(d2))
1352                 BUG_ON(__update_replicas(c, dst, &d2->r, -d2->delta));
1353
1354         preempt_enable();
1355         percpu_up_read(&c->mark_lock);
1356         return -1;
1357 }
1358
1359 /* trans_mark: */
1360
1361 static int bch2_trans_mark_pointer(struct btree_trans *trans,
1362                                    enum btree_id btree_id, unsigned level,
1363                                    struct bkey_s_c k, struct extent_ptr_decoded p,
1364                                    unsigned flags)
1365 {
1366         bool insert = !(flags & BTREE_TRIGGER_OVERWRITE);
1367         struct btree_iter iter;
1368         struct bkey_i_alloc_v4 *a;
1369         struct bpos bucket_pos;
1370         struct bch_backpointer bp;
1371         s64 sectors;
1372         int ret;
1373
1374         bch2_extent_ptr_to_bp(trans->c, btree_id, level, k, p, &bucket_pos, &bp);
1375         sectors = bp.bucket_len;
1376         if (!insert)
1377                 sectors = -sectors;
1378
1379         a = bch2_trans_start_alloc_update(trans, &iter, bucket_pos);
1380         if (IS_ERR(a))
1381                 return PTR_ERR(a);
1382
1383         ret = __mark_pointer(trans, k, &p.ptr, sectors, bp.data_type,
1384                              a->v.gen, &a->v.data_type,
1385                              &a->v.dirty_sectors, &a->v.cached_sectors);
1386         if (ret)
1387                 goto err;
1388
1389         if (!p.ptr.cached) {
1390                 ret = insert
1391                         ? bch2_bucket_backpointer_add(trans, a, bp, k)
1392                         : bch2_bucket_backpointer_del(trans, a, bp, k);
1393                 if (ret)
1394                         goto err;
1395         }
1396
1397         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1398 err:
1399         bch2_trans_iter_exit(trans, &iter);
1400         return ret;
1401 }
1402
1403 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1404                         struct extent_ptr_decoded p,
1405                         s64 sectors, enum bch_data_type data_type)
1406 {
1407         struct btree_iter iter;
1408         struct bkey_s_c k;
1409         struct bkey_i_stripe *s;
1410         struct bch_replicas_padded r;
1411         int ret = 0;
1412
1413         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes, POS(0, p.ec.idx),
1414                              BTREE_ITER_INTENT|
1415                              BTREE_ITER_WITH_UPDATES);
1416         k = bch2_btree_iter_peek_slot(&iter);
1417         ret = bkey_err(k);
1418         if (ret)
1419                 goto err;
1420
1421         if (k.k->type != KEY_TYPE_stripe) {
1422                 bch2_trans_inconsistent(trans,
1423                         "pointer to nonexistent stripe %llu",
1424                         (u64) p.ec.idx);
1425                 ret = -EIO;
1426                 goto err;
1427         }
1428
1429         if (!bch2_ptr_matches_stripe(bkey_s_c_to_stripe(k).v, p)) {
1430                 bch2_trans_inconsistent(trans,
1431                         "stripe pointer doesn't match stripe %llu",
1432                         (u64) p.ec.idx);
1433                 ret = -EIO;
1434                 goto err;
1435         }
1436
1437         s = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1438         ret = PTR_ERR_OR_ZERO(s);
1439         if (ret)
1440                 goto err;
1441
1442         bkey_reassemble(&s->k_i, k);
1443         stripe_blockcount_set(&s->v, p.ec.block,
1444                 stripe_blockcount_get(&s->v, p.ec.block) +
1445                 sectors);
1446
1447         ret = bch2_trans_update(trans, &iter, &s->k_i, 0);
1448         if (ret)
1449                 goto err;
1450
1451         bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1452         r.e.data_type = data_type;
1453         update_replicas_list(trans, &r.e, sectors);
1454 err:
1455         bch2_trans_iter_exit(trans, &iter);
1456         return ret;
1457 }
1458
1459 int bch2_trans_mark_extent(struct btree_trans *trans,
1460                            enum btree_id btree_id, unsigned level,
1461                            struct bkey_s_c old, struct bkey_i *new,
1462                            unsigned flags)
1463 {
1464         struct bch_fs *c = trans->c;
1465         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1466                 ? old
1467                 : bkey_i_to_s_c(new);
1468         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1469         const union bch_extent_entry *entry;
1470         struct extent_ptr_decoded p;
1471         struct bch_replicas_padded r;
1472         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
1473                 ? BCH_DATA_btree
1474                 : BCH_DATA_user;
1475         s64 sectors = bkey_is_btree_ptr(k.k)
1476                 ? btree_sectors(c)
1477                 : k.k->size;
1478         s64 dirty_sectors = 0;
1479         bool stale;
1480         int ret;
1481
1482         r.e.data_type   = data_type;
1483         r.e.nr_devs     = 0;
1484         r.e.nr_required = 1;
1485
1486         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1487                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
1488
1489                 if (flags & BTREE_TRIGGER_OVERWRITE)
1490                         disk_sectors = -disk_sectors;
1491
1492                 ret = bch2_trans_mark_pointer(trans, btree_id, level, k, p, flags);
1493                 if (ret < 0)
1494                         return ret;
1495
1496                 stale = ret > 0;
1497
1498                 if (p.ptr.cached) {
1499                         if (!stale)
1500                                 update_cached_sectors_list(trans, p.ptr.dev,
1501                                                            disk_sectors);
1502                 } else if (!p.has_ec) {
1503                         dirty_sectors          += disk_sectors;
1504                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1505                 } else {
1506                         ret = bch2_trans_mark_stripe_ptr(trans, p,
1507                                         disk_sectors, data_type);
1508                         if (ret)
1509                                 return ret;
1510
1511                         r.e.nr_required = 0;
1512                 }
1513         }
1514
1515         if (r.e.nr_devs)
1516                 update_replicas_list(trans, &r.e, dirty_sectors);
1517
1518         return 0;
1519 }
1520
1521 static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
1522                                          struct bkey_s_c_stripe s,
1523                                          unsigned idx, bool deleting)
1524 {
1525         struct bch_fs *c = trans->c;
1526         const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
1527         struct btree_iter iter;
1528         struct bkey_i_alloc_v4 *a;
1529         enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
1530                 ? BCH_DATA_parity : 0;
1531         s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
1532         int ret = 0;
1533
1534         if (deleting)
1535                 sectors = -sectors;
1536
1537         a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(c, ptr));
1538         if (IS_ERR(a))
1539                 return PTR_ERR(a);
1540
1541         ret = check_bucket_ref(c, s.s_c, ptr, sectors, data_type,
1542                                a->v.gen, a->v.data_type,
1543                                a->v.dirty_sectors, a->v.cached_sectors);
1544         if (ret)
1545                 goto err;
1546
1547         if (!deleting) {
1548                 if (bch2_trans_inconsistent_on(a->v.stripe ||
1549                                                a->v.stripe_redundancy, trans,
1550                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
1551                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1552                                 bch2_data_types[a->v.data_type],
1553                                 a->v.dirty_sectors,
1554                                 a->v.stripe, s.k->p.offset)) {
1555                         ret = -EIO;
1556                         goto err;
1557                 }
1558
1559                 if (bch2_trans_inconsistent_on(data_type && a->v.dirty_sectors, trans,
1560                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
1561                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1562                                 bch2_data_types[a->v.data_type],
1563                                 a->v.dirty_sectors,
1564                                 s.k->p.offset)) {
1565                         ret = -EIO;
1566                         goto err;
1567                 }
1568
1569                 a->v.stripe             = s.k->p.offset;
1570                 a->v.stripe_redundancy  = s.v->nr_redundant;
1571         } else {
1572                 if (bch2_trans_inconsistent_on(a->v.stripe != s.k->p.offset ||
1573                                                a->v.stripe_redundancy != s.v->nr_redundant, trans,
1574                                 "bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
1575                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1576                                 s.k->p.offset, a->v.stripe)) {
1577                         ret = -EIO;
1578                         goto err;
1579                 }
1580
1581                 a->v.stripe             = 0;
1582                 a->v.stripe_redundancy  = 0;
1583         }
1584
1585         a->v.dirty_sectors += sectors;
1586         if (data_type)
1587                 a->v.data_type = !deleting ? data_type : 0;
1588
1589         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1590         if (ret)
1591                 goto err;
1592 err:
1593         bch2_trans_iter_exit(trans, &iter);
1594         return ret;
1595 }
1596
1597 int bch2_trans_mark_stripe(struct btree_trans *trans,
1598                            enum btree_id btree_id, unsigned level,
1599                            struct bkey_s_c old, struct bkey_i *new,
1600                            unsigned flags)
1601 {
1602         const struct bch_stripe *old_s = NULL;
1603         struct bch_stripe *new_s = NULL;
1604         struct bch_replicas_padded r;
1605         unsigned i, nr_blocks;
1606         int ret = 0;
1607
1608         if (old.k->type == KEY_TYPE_stripe)
1609                 old_s = bkey_s_c_to_stripe(old).v;
1610         if (new->k.type == KEY_TYPE_stripe)
1611                 new_s = &bkey_i_to_stripe(new)->v;
1612
1613         /*
1614          * If the pointers aren't changing, we don't need to do anything:
1615          */
1616         if (new_s && old_s &&
1617             new_s->nr_blocks    == old_s->nr_blocks &&
1618             new_s->nr_redundant == old_s->nr_redundant &&
1619             !memcmp(old_s->ptrs, new_s->ptrs,
1620                     new_s->nr_blocks * sizeof(struct bch_extent_ptr)))
1621                 return 0;
1622
1623         BUG_ON(new_s && old_s &&
1624                (new_s->nr_blocks        != old_s->nr_blocks ||
1625                 new_s->nr_redundant     != old_s->nr_redundant));
1626
1627         nr_blocks = new_s ? new_s->nr_blocks : old_s->nr_blocks;
1628
1629         if (new_s) {
1630                 s64 sectors = le16_to_cpu(new_s->sectors);
1631
1632                 bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(new));
1633                 update_replicas_list(trans, &r.e, sectors * new_s->nr_redundant);
1634         }
1635
1636         if (old_s) {
1637                 s64 sectors = -((s64) le16_to_cpu(old_s->sectors));
1638
1639                 bch2_bkey_to_replicas(&r.e, old);
1640                 update_replicas_list(trans, &r.e, sectors * old_s->nr_redundant);
1641         }
1642
1643         for (i = 0; i < nr_blocks; i++) {
1644                 if (new_s && old_s &&
1645                     !memcmp(&new_s->ptrs[i],
1646                             &old_s->ptrs[i],
1647                             sizeof(new_s->ptrs[i])))
1648                         continue;
1649
1650                 if (new_s) {
1651                         ret = bch2_trans_mark_stripe_bucket(trans,
1652                                         bkey_i_to_s_c_stripe(new), i, false);
1653                         if (ret)
1654                                 break;
1655                 }
1656
1657                 if (old_s) {
1658                         ret = bch2_trans_mark_stripe_bucket(trans,
1659                                         bkey_s_c_to_stripe(old), i, true);
1660                         if (ret)
1661                                 break;
1662                 }
1663         }
1664
1665         return ret;
1666 }
1667
1668 int bch2_trans_mark_inode(struct btree_trans *trans,
1669                           enum btree_id btree_id, unsigned level,
1670                           struct bkey_s_c old,
1671                           struct bkey_i *new,
1672                           unsigned flags)
1673 {
1674         int nr = bkey_is_inode(&new->k) - bkey_is_inode(old.k);
1675
1676         if (nr) {
1677                 struct replicas_delta_list *d =
1678                         replicas_deltas_realloc(trans, 0);
1679                 d->nr_inodes += nr;
1680         }
1681
1682         return 0;
1683 }
1684
1685 int bch2_trans_mark_reservation(struct btree_trans *trans,
1686                                 enum btree_id btree_id, unsigned level,
1687                                 struct bkey_s_c old,
1688                                 struct bkey_i *new,
1689                                 unsigned flags)
1690 {
1691         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1692                 ? old
1693                 : bkey_i_to_s_c(new);
1694         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1695         s64 sectors = (s64) k.k->size;
1696         struct replicas_delta_list *d;
1697
1698         if (flags & BTREE_TRIGGER_OVERWRITE)
1699                 sectors = -sectors;
1700         sectors *= replicas;
1701
1702         d = replicas_deltas_realloc(trans, 0);
1703
1704         replicas = clamp_t(unsigned, replicas, 1,
1705                            ARRAY_SIZE(d->persistent_reserved));
1706
1707         d->persistent_reserved[replicas - 1] += sectors;
1708         return 0;
1709 }
1710
1711 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1712                         struct bkey_s_c_reflink_p p,
1713                         u64 *idx, unsigned flags)
1714 {
1715         struct bch_fs *c = trans->c;
1716         struct btree_iter iter;
1717         struct bkey_s_c k;
1718         struct bkey_i *n;
1719         __le64 *refcount;
1720         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1721         struct printbuf buf = PRINTBUF;
1722         int ret;
1723
1724         bch2_trans_iter_init(trans, &iter, BTREE_ID_reflink, POS(0, *idx),
1725                              BTREE_ITER_INTENT|
1726                              BTREE_ITER_WITH_UPDATES);
1727         k = bch2_btree_iter_peek_slot(&iter);
1728         ret = bkey_err(k);
1729         if (ret)
1730                 goto err;
1731
1732         n = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1733         ret = PTR_ERR_OR_ZERO(n);
1734         if (ret)
1735                 goto err;
1736
1737         bkey_reassemble(n, k);
1738
1739         refcount = bkey_refcount(n);
1740         if (!refcount) {
1741                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1742                 bch2_trans_inconsistent(trans,
1743                         "nonexistent indirect extent at %llu while marking\n  %s",
1744                         *idx, buf.buf);
1745                 ret = -EIO;
1746                 goto err;
1747         }
1748
1749         if (!*refcount && (flags & BTREE_TRIGGER_OVERWRITE)) {
1750                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1751                 bch2_trans_inconsistent(trans,
1752                         "indirect extent refcount underflow at %llu while marking\n  %s",
1753                         *idx, buf.buf);
1754                 ret = -EIO;
1755                 goto err;
1756         }
1757
1758         if (flags & BTREE_TRIGGER_INSERT) {
1759                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1760                 u64 pad;
1761
1762                 pad = max_t(s64, le32_to_cpu(v->front_pad),
1763                             le64_to_cpu(v->idx) - bkey_start_offset(k.k));
1764                 BUG_ON(pad > U32_MAX);
1765                 v->front_pad = cpu_to_le32(pad);
1766
1767                 pad = max_t(s64, le32_to_cpu(v->back_pad),
1768                             k.k->p.offset - p.k->size - le64_to_cpu(v->idx));
1769                 BUG_ON(pad > U32_MAX);
1770                 v->back_pad = cpu_to_le32(pad);
1771         }
1772
1773         le64_add_cpu(refcount, add);
1774
1775         bch2_btree_iter_set_pos_to_extent_start(&iter);
1776         ret = bch2_trans_update(trans, &iter, n, 0);
1777         if (ret)
1778                 goto err;
1779
1780         *idx = k.k->p.offset;
1781 err:
1782         bch2_trans_iter_exit(trans, &iter);
1783         printbuf_exit(&buf);
1784         return ret;
1785 }
1786
1787 int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1788                               enum btree_id btree_id, unsigned level,
1789                               struct bkey_s_c old,
1790                               struct bkey_i *new,
1791                               unsigned flags)
1792 {
1793         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1794                 ? old
1795                 : bkey_i_to_s_c(new);
1796         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1797         u64 idx, end_idx;
1798         int ret = 0;
1799
1800         if (flags & BTREE_TRIGGER_INSERT) {
1801                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1802
1803                 v->front_pad = v->back_pad = 0;
1804         }
1805
1806         idx     = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
1807         end_idx = le64_to_cpu(p.v->idx) + p.k->size +
1808                 le32_to_cpu(p.v->back_pad);
1809
1810         while (idx < end_idx && !ret)
1811                 ret = __bch2_trans_mark_reflink_p(trans, p, &idx, flags);
1812
1813         return ret;
1814 }
1815
1816 static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1817                                     struct bch_dev *ca, size_t b,
1818                                     enum bch_data_type type,
1819                                     unsigned sectors)
1820 {
1821         struct bch_fs *c = trans->c;
1822         struct btree_iter iter;
1823         struct bkey_i_alloc_v4 *a;
1824         int ret = 0;
1825
1826         /*
1827          * Backup superblock might be past the end of our normal usable space:
1828          */
1829         if (b >= ca->mi.nbuckets)
1830                 return 0;
1831
1832         a = bch2_trans_start_alloc_update(trans, &iter, POS(ca->dev_idx, b));
1833         if (IS_ERR(a))
1834                 return PTR_ERR(a);
1835
1836         if (a->v.data_type && a->v.data_type != type) {
1837                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1838                         "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
1839                         "while marking %s",
1840                         iter.pos.inode, iter.pos.offset, a->v.gen,
1841                         bch2_data_types[a->v.data_type],
1842                         bch2_data_types[type],
1843                         bch2_data_types[type]);
1844                 ret = -EIO;
1845                 goto out;
1846         }
1847
1848         a->v.data_type          = type;
1849         a->v.dirty_sectors      = sectors;
1850
1851         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1852         if (ret)
1853                 goto out;
1854 out:
1855         bch2_trans_iter_exit(trans, &iter);
1856         return ret;
1857 }
1858
1859 int bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1860                                     struct bch_dev *ca, size_t b,
1861                                     enum bch_data_type type,
1862                                     unsigned sectors)
1863 {
1864         return commit_do(trans, NULL, NULL, 0,
1865                         __bch2_trans_mark_metadata_bucket(trans, ca, b, type, sectors));
1866 }
1867
1868 static int bch2_trans_mark_metadata_sectors(struct btree_trans *trans,
1869                                             struct bch_dev *ca,
1870                                             u64 start, u64 end,
1871                                             enum bch_data_type type,
1872                                             u64 *bucket, unsigned *bucket_sectors)
1873 {
1874         do {
1875                 u64 b = sector_to_bucket(ca, start);
1876                 unsigned sectors =
1877                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1878
1879                 if (b != *bucket && *bucket_sectors) {
1880                         int ret = bch2_trans_mark_metadata_bucket(trans, ca, *bucket,
1881                                                                   type, *bucket_sectors);
1882                         if (ret)
1883                                 return ret;
1884
1885                         *bucket_sectors = 0;
1886                 }
1887
1888                 *bucket         = b;
1889                 *bucket_sectors += sectors;
1890                 start += sectors;
1891         } while (start < end);
1892
1893         return 0;
1894 }
1895
1896 static int __bch2_trans_mark_dev_sb(struct btree_trans *trans,
1897                                     struct bch_dev *ca)
1898 {
1899         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1900         u64 bucket = 0;
1901         unsigned i, bucket_sectors = 0;
1902         int ret;
1903
1904         for (i = 0; i < layout->nr_superblocks; i++) {
1905                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1906
1907                 if (offset == BCH_SB_SECTOR) {
1908                         ret = bch2_trans_mark_metadata_sectors(trans, ca,
1909                                                 0, BCH_SB_SECTOR,
1910                                                 BCH_DATA_sb, &bucket, &bucket_sectors);
1911                         if (ret)
1912                                 return ret;
1913                 }
1914
1915                 ret = bch2_trans_mark_metadata_sectors(trans, ca, offset,
1916                                       offset + (1 << layout->sb_max_size_bits),
1917                                       BCH_DATA_sb, &bucket, &bucket_sectors);
1918                 if (ret)
1919                         return ret;
1920         }
1921
1922         if (bucket_sectors) {
1923                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1924                                 bucket, BCH_DATA_sb, bucket_sectors);
1925                 if (ret)
1926                         return ret;
1927         }
1928
1929         for (i = 0; i < ca->journal.nr; i++) {
1930                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1931                                 ca->journal.buckets[i],
1932                                 BCH_DATA_journal, ca->mi.bucket_size);
1933                 if (ret)
1934                         return ret;
1935         }
1936
1937         return 0;
1938 }
1939
1940 int bch2_trans_mark_dev_sb(struct bch_fs *c, struct bch_dev *ca)
1941 {
1942         return bch2_trans_do(c, NULL, NULL, BTREE_INSERT_LAZY_RW,
1943                         __bch2_trans_mark_dev_sb(&trans, ca));
1944 }
1945
1946 /* Disk reservations: */
1947
1948 #define SECTORS_CACHE   1024
1949
1950 int bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
1951                               u64 sectors, int flags)
1952 {
1953         struct bch_fs_pcpu *pcpu;
1954         u64 old, v, get;
1955         s64 sectors_available;
1956         int ret;
1957
1958         percpu_down_read(&c->mark_lock);
1959         preempt_disable();
1960         pcpu = this_cpu_ptr(c->pcpu);
1961
1962         if (sectors <= pcpu->sectors_available)
1963                 goto out;
1964
1965         v = atomic64_read(&c->sectors_available);
1966         do {
1967                 old = v;
1968                 get = min((u64) sectors + SECTORS_CACHE, old);
1969
1970                 if (get < sectors) {
1971                         preempt_enable();
1972                         goto recalculate;
1973                 }
1974         } while ((v = atomic64_cmpxchg(&c->sectors_available,
1975                                        old, old - get)) != old);
1976
1977         pcpu->sectors_available         += get;
1978
1979 out:
1980         pcpu->sectors_available         -= sectors;
1981         this_cpu_add(*c->online_reserved, sectors);
1982         res->sectors                    += sectors;
1983
1984         preempt_enable();
1985         percpu_up_read(&c->mark_lock);
1986         return 0;
1987
1988 recalculate:
1989         mutex_lock(&c->sectors_available_lock);
1990
1991         percpu_u64_set(&c->pcpu->sectors_available, 0);
1992         sectors_available = avail_factor(__bch2_fs_usage_read_short(c).free);
1993
1994         if (sectors <= sectors_available ||
1995             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
1996                 atomic64_set(&c->sectors_available,
1997                              max_t(s64, 0, sectors_available - sectors));
1998                 this_cpu_add(*c->online_reserved, sectors);
1999                 res->sectors                    += sectors;
2000                 ret = 0;
2001         } else {
2002                 atomic64_set(&c->sectors_available, sectors_available);
2003                 ret = -ENOSPC;
2004         }
2005
2006         mutex_unlock(&c->sectors_available_lock);
2007         percpu_up_read(&c->mark_lock);
2008
2009         return ret;
2010 }
2011
2012 /* Startup/shutdown: */
2013
2014 static void bucket_gens_free_rcu(struct rcu_head *rcu)
2015 {
2016         struct bucket_gens *buckets =
2017                 container_of(rcu, struct bucket_gens, rcu);
2018
2019         kvpfree(buckets, sizeof(*buckets) + buckets->nbuckets);
2020 }
2021
2022 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
2023 {
2024         struct bucket_gens *bucket_gens = NULL, *old_bucket_gens = NULL;
2025         unsigned long *buckets_nouse = NULL;
2026         bool resize = ca->bucket_gens != NULL;
2027         int ret = -ENOMEM;
2028
2029         if (!(bucket_gens       = kvpmalloc(sizeof(struct bucket_gens) + nbuckets,
2030                                             GFP_KERNEL|__GFP_ZERO)) ||
2031             (c->opts.buckets_nouse &&
2032              !(buckets_nouse    = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2033                                             sizeof(unsigned long),
2034                                             GFP_KERNEL|__GFP_ZERO))))
2035                 goto err;
2036
2037         bucket_gens->first_bucket = ca->mi.first_bucket;
2038         bucket_gens->nbuckets   = nbuckets;
2039
2040         bch2_copygc_stop(c);
2041
2042         if (resize) {
2043                 down_write(&c->gc_lock);
2044                 down_write(&ca->bucket_lock);
2045                 percpu_down_write(&c->mark_lock);
2046         }
2047
2048         old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
2049
2050         if (resize) {
2051                 size_t n = min(bucket_gens->nbuckets, old_bucket_gens->nbuckets);
2052
2053                 memcpy(bucket_gens->b,
2054                        old_bucket_gens->b,
2055                        n);
2056                 if (buckets_nouse)
2057                         memcpy(buckets_nouse,
2058                                ca->buckets_nouse,
2059                                BITS_TO_LONGS(n) * sizeof(unsigned long));
2060         }
2061
2062         rcu_assign_pointer(ca->bucket_gens, bucket_gens);
2063         bucket_gens     = old_bucket_gens;
2064
2065         swap(ca->buckets_nouse, buckets_nouse);
2066
2067         nbuckets = ca->mi.nbuckets;
2068
2069         if (resize) {
2070                 percpu_up_write(&c->mark_lock);
2071                 up_write(&ca->bucket_lock);
2072                 up_write(&c->gc_lock);
2073         }
2074
2075         ret = 0;
2076 err:
2077         kvpfree(buckets_nouse,
2078                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2079         if (bucket_gens)
2080                 call_rcu(&bucket_gens->rcu, bucket_gens_free_rcu);
2081
2082         return ret;
2083 }
2084
2085 void bch2_dev_buckets_free(struct bch_dev *ca)
2086 {
2087         unsigned i;
2088
2089         kvpfree(ca->buckets_nouse,
2090                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2091         kvpfree(rcu_dereference_protected(ca->bucket_gens, 1),
2092                 sizeof(struct bucket_gens) + ca->mi.nbuckets);
2093
2094         for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
2095                 free_percpu(ca->usage[i]);
2096         kfree(ca->usage_base);
2097 }
2098
2099 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2100 {
2101         unsigned i;
2102
2103         ca->usage_base = kzalloc(sizeof(struct bch_dev_usage), GFP_KERNEL);
2104         if (!ca->usage_base)
2105                 return -ENOMEM;
2106
2107         for (i = 0; i < ARRAY_SIZE(ca->usage); i++) {
2108                 ca->usage[i] = alloc_percpu(struct bch_dev_usage);
2109                 if (!ca->usage[i])
2110                         return -ENOMEM;
2111         }
2112
2113         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);;
2114 }