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         percpu_down_read(&c->mark_lock);
 548         if (!gc && new_a.gen != old_a.gen)
 549                 *bucket_gen(ca, new.k->p.offset) = new_a.gen;
 550
 551         bch2_dev_usage_update(c, ca, old_a, new_a, journal_seq, gc);
 552
 553         if (gc) {
 554                 struct bucket *g = gc_bucket(ca, new.k->p.offset);
 555
 556                 bucket_lock(g);
 557
 558                 g->gen_valid            = 1;
 559                 g->gen                  = new_a.gen;
 560                 g->data_type            = new_a.data_type;
 561                 g->stripe               = new_a.stripe;
 562                 g->stripe_redundancy    = new_a.stripe_redundancy;
 563                 g->dirty_sectors        = new_a.dirty_sectors;
 564                 g->cached_sectors       = new_a.cached_sectors;
 565
 566                 bucket_unlock(g);
 567         }
 568         percpu_up_read(&c->mark_lock);
 569
 570         /*
 571          * need to know if we're getting called from the invalidate path or
 572          * not:
 573          */
 574
 575         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 576             old_a.cached_sectors) {
 577                 ret = update_cached_sectors(c, new, ca->dev_idx,
 578                                             -((s64) old_a.cached_sectors),
 579                                             journal_seq, gc);
 580                 if (ret) {
 581                         bch2_fs_fatal_error(c, "bch2_mark_alloc(): no replicas entry while updating cached sectors");
 582                         return ret;
 583                 }
 584         }
 585
 586         if (new_a.data_type == BCH_DATA_free &&
 587             (!new_a.journal_seq || new_a.journal_seq < c->journal.flushed_seq_ondisk))
 588                 closure_wake_up(&c->freelist_wait);
 589
 590         if (new_a.data_type == BCH_DATA_need_discard &&
 591             (!new_a.journal_seq || new_a.journal_seq < c->journal.flushed_seq_ondisk))
 592                 bch2_do_discards(c);
 593
 594         if (old_a.data_type != BCH_DATA_cached &&
 595             new_a.data_type == BCH_DATA_cached &&
 596             should_invalidate_buckets(ca, bch2_dev_usage_read(ca)))
 597                 bch2_do_invalidates(c);
 598
 599         if (new_a.data_type == BCH_DATA_need_gc_gens)
 600                 bch2_do_gc_gens(c);
 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 ((bucket_data_type == BCH_DATA_stripe && ptr_data_type == BCH_DATA_user) ||
 678             (bucket_data_type == BCH_DATA_user   && ptr_data_type == BCH_DATA_stripe))
 679                 bucket_data_type = ptr_data_type = BCH_DATA_stripe;
 680
 681         if (gen_after(ptr->gen, b_gen)) {
 682                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 683                         "bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\n"
 684                         "while marking %s",
 685                         ptr->dev, bucket_nr, b_gen,
 686                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 687                         ptr->gen,
 688                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 689                 ret = -EIO;
 690                 goto err;
 691         }
 692
 693         if (gen_cmp(b_gen, ptr->gen) > BUCKET_GC_GEN_MAX) {
 694                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 695                         "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 696                         "while marking %s",
 697                         ptr->dev, bucket_nr, b_gen,
 698                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 699                         ptr->gen,
 700                         (printbuf_reset(&buf),
 701                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 702                 ret = -EIO;
 703                 goto err;
 704         }
 705
 706         if (b_gen != ptr->gen && !ptr->cached) {
 707                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 708                         "bucket %u:%zu gen %u (mem gen %u) data type %s: stale dirty ptr (gen %u)\n"
 709                         "while marking %s",
 710                         ptr->dev, bucket_nr, b_gen,
 711                         *bucket_gen(ca, bucket_nr),
 712                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 713                         ptr->gen,
 714                         (printbuf_reset(&buf),
 715                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 716                 ret = -EIO;
 717                 goto err;
 718         }
 719
 720         if (b_gen != ptr->gen) {
 721                 ret = 1;
 722                 goto err;
 723         }
 724
 725         if (!data_type_is_empty(bucket_data_type) &&
 726             ptr_data_type &&
 727             bucket_data_type != ptr_data_type) {
 728                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 729                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 730                         "while marking %s",
 731                         ptr->dev, bucket_nr, b_gen,
 732                         bch2_data_types[bucket_data_type],
 733                         bch2_data_types[ptr_data_type],
 734                         (printbuf_reset(&buf),
 735                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 736                 ret = -EIO;
 737                 goto err;
 738         }
 739
 740         if ((unsigned) (bucket_sectors + sectors) > U32_MAX) {
 741                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 742                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U16_MAX\n"
 743                         "while marking %s",
 744                         ptr->dev, bucket_nr, b_gen,
 745                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 746                         bucket_sectors, sectors,
 747                         (printbuf_reset(&buf),
 748                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 749                 ret = -EIO;
 750                 goto err;
 751         }
 752 err:
 753         printbuf_exit(&buf);
 754         return ret;
 755 }
 756
 757 static int mark_stripe_bucket(struct btree_trans *trans,
 758                               struct bkey_s_c k,
 759                               unsigned ptr_idx,
 760                               unsigned flags)
 761 {
 762         struct bch_fs *c = trans->c;
 763         u64 journal_seq = trans->journal_res.seq;
 764         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 765         unsigned nr_data = s->nr_blocks - s->nr_redundant;
 766         bool parity = ptr_idx >= nr_data;
 767         enum bch_data_type data_type = parity ? BCH_DATA_parity : 0;
 768         s64 sectors = parity ? le16_to_cpu(s->sectors) : 0;
 769         const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
 770         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 771         struct bucket old, new, *g;
 772         struct printbuf buf = PRINTBUF;
 773         int ret = 0;
 774
 775         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 776
 777         /* * XXX doesn't handle deletion */
 778
 779         percpu_down_read(&c->mark_lock);
 780         buf.atomic++;
 781         g = PTR_GC_BUCKET(ca, ptr);
 782
 783         if (g->dirty_sectors ||
 784             (g->stripe && g->stripe != k.k->p.offset)) {
 785                 bch2_fs_inconsistent(c,
 786                               "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
 787                               ptr->dev, PTR_BUCKET_NR(ca, ptr), g->gen,
 788                               (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 789                 ret = -EINVAL;
 790                 goto err;
 791         }
 792
 793         bucket_lock(g);
 794         old = *g;
 795
 796         ret = check_bucket_ref(c, k, ptr, sectors, data_type,
 797                                g->gen, g->data_type,
 798                                g->dirty_sectors, g->cached_sectors);
 799         if (ret)
 800                 goto err;
 801
 802         if (data_type)
 803                 g->data_type = data_type;
 804         g->dirty_sectors += sectors;
 805
 806         g->stripe               = k.k->p.offset;
 807         g->stripe_redundancy    = s->nr_redundant;
 808         new = *g;
 809 err:
 810         bucket_unlock(g);
 811         if (!ret)
 812                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 813         percpu_up_read(&c->mark_lock);
 814         printbuf_exit(&buf);
 815         return ret;
 816 }
 817
 818 static int __mark_pointer(struct btree_trans *trans,
 819                           struct bkey_s_c k,
 820                           const struct bch_extent_ptr *ptr,
 821                           s64 sectors, enum bch_data_type ptr_data_type,
 822                           u8 bucket_gen, u8 *bucket_data_type,
 823                           u32 *dirty_sectors, u32 *cached_sectors)
 824 {
 825         u32 *dst_sectors = !ptr->cached
 826                 ? dirty_sectors
 827                 : cached_sectors;
 828         int ret = check_bucket_ref(trans->c, k, ptr, sectors, ptr_data_type,
 829                                    bucket_gen, *bucket_data_type,
 830                                    *dirty_sectors, *cached_sectors);
 831
 832         if (ret)
 833                 return ret;
 834
 835         *dst_sectors += sectors;
 836         *bucket_data_type = *dirty_sectors || *cached_sectors
 837                 ? ptr_data_type : 0;
 838         return 0;
 839 }
 840
 841 static int bch2_mark_pointer(struct btree_trans *trans,
 842                              struct bkey_s_c k,
 843                              struct extent_ptr_decoded p,
 844                              s64 sectors, enum bch_data_type data_type,
 845                              unsigned flags)
 846 {
 847         u64 journal_seq = trans->journal_res.seq;
 848         struct bch_fs *c = trans->c;
 849         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 850         struct bucket old, new, *g;
 851         u8 bucket_data_type;
 852         int ret = 0;
 853
 854         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 855
 856         percpu_down_read(&c->mark_lock);
 857         g = PTR_GC_BUCKET(ca, &p.ptr);
 858         bucket_lock(g);
 859         old = *g;
 860
 861         bucket_data_type = g->data_type;
 862         ret = __mark_pointer(trans, k, &p.ptr, sectors,
 863                              data_type, g->gen,
 864                              &bucket_data_type,
 865                              &g->dirty_sectors,
 866                              &g->cached_sectors);
 867         if (!ret)
 868                 g->data_type = bucket_data_type;
 869
 870         new = *g;
 871         bucket_unlock(g);
 872         if (!ret)
 873                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 874         percpu_up_read(&c->mark_lock);
 875
 876         return ret;
 877 }
 878
 879 static int bch2_mark_stripe_ptr(struct btree_trans *trans,
 880                                 struct bkey_s_c k,
 881                                 struct bch_extent_stripe_ptr p,
 882                                 enum bch_data_type data_type,
 883                                 s64 sectors,
 884                                 unsigned flags)
 885 {
 886         struct bch_fs *c = trans->c;
 887         struct bch_replicas_padded r;
 888         struct gc_stripe *m;
 889
 890         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 891
 892         m = genradix_ptr_alloc(&c->gc_stripes, p.idx, GFP_KERNEL);
 893         if (!m) {
 894                 bch_err(c, "error allocating memory for gc_stripes, idx %llu",
 895                         (u64) p.idx);
 896                 return -ENOMEM;
 897         }
 898
 899         spin_lock(&c->ec_stripes_heap_lock);
 900
 901         if (!m || !m->alive) {
 902                 spin_unlock(&c->ec_stripes_heap_lock);
 903                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
 904                                     (u64) p.idx);
 905                 bch2_inconsistent_error(c);
 906                 return -EIO;
 907         }
 908
 909         m->block_sectors[p.block] += sectors;
 910
 911         r = m->r;
 912         spin_unlock(&c->ec_stripes_heap_lock);
 913
 914         r.e.data_type = data_type;
 915         update_replicas(c, k, &r.e, sectors, trans->journal_res.seq, true);
 916
 917         return 0;
 918 }
 919
 920 int bch2_mark_extent(struct btree_trans *trans,
 921                      struct bkey_s_c old, struct bkey_s_c new,
 922                      unsigned flags)
 923 {
 924         u64 journal_seq = trans->journal_res.seq;
 925         struct bch_fs *c = trans->c;
 926         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
 927         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 928         const union bch_extent_entry *entry;
 929         struct extent_ptr_decoded p;
 930         struct bch_replicas_padded r;
 931         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
 932                 ? BCH_DATA_btree
 933                 : BCH_DATA_user;
 934         s64 sectors = bkey_is_btree_ptr(k.k)
 935                 ? btree_sectors(c)
 936                 : k.k->size;
 937         s64 dirty_sectors = 0;
 938         bool stale;
 939         int ret;
 940
 941         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 942
 943         r.e.data_type   = data_type;
 944         r.e.nr_devs     = 0;
 945         r.e.nr_required = 1;
 946
 947         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 948                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
 949
 950                 if (flags & BTREE_TRIGGER_OVERWRITE)
 951                         disk_sectors = -disk_sectors;
 952
 953                 ret = bch2_mark_pointer(trans, k, p, disk_sectors,
 954                                         data_type, flags);
 955                 if (ret < 0)
 956                         return ret;
 957
 958                 stale = ret > 0;
 959
 960                 if (p.ptr.cached) {
 961                         if (!stale) {
 962                                 ret = update_cached_sectors(c, k, p.ptr.dev,
 963                                                 disk_sectors, journal_seq, true);
 964                                 if (ret) {
 965                                         bch2_fs_fatal_error(c, "bch2_mark_extent(): no replicas entry while updating cached sectors");
 966                                         return ret;
 967                                 }
 968                         }
 969                 } else if (!p.has_ec) {
 970                         dirty_sectors          += disk_sectors;
 971                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
 972                 } else {
 973                         ret = bch2_mark_stripe_ptr(trans, k, p.ec, data_type,
 974                                         disk_sectors, flags);
 975                         if (ret)
 976                                 return ret;
 977
 978                         /*
 979                          * There may be other dirty pointers in this extent, but
 980                          * if so they're not required for mounting if we have an
 981                          * erasure coded pointer in this extent:
 982                          */
 983                         r.e.nr_required = 0;
 984                 }
 985         }
 986
 987         if (r.e.nr_devs) {
 988                 ret = update_replicas(c, k, &r.e, dirty_sectors, journal_seq, true);
 989                 if (ret) {
 990                         struct printbuf buf = PRINTBUF;
 991
 992                         bch2_bkey_val_to_text(&buf, c, k);
 993                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
 994                         printbuf_exit(&buf);
 995                         return ret;
 996                 }
 997         }
 998
 999         return 0;
1000 }
1001
1002 int bch2_mark_stripe(struct btree_trans *trans,
1003                      struct bkey_s_c old, struct bkey_s_c new,
1004                      unsigned flags)
1005 {
1006         bool gc = flags & BTREE_TRIGGER_GC;
1007         u64 journal_seq = trans->journal_res.seq;
1008         struct bch_fs *c = trans->c;
1009         u64 idx = new.k->p.offset;
1010         const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
1011                 ? bkey_s_c_to_stripe(old).v : NULL;
1012         const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
1013                 ? bkey_s_c_to_stripe(new).v : NULL;
1014         unsigned i;
1015         int ret;
1016
1017         BUG_ON(gc && old_s);
1018
1019         if (!gc) {
1020                 struct stripe *m = genradix_ptr(&c->stripes, idx);
1021
1022                 if (!m || (old_s && !m->alive)) {
1023                         struct printbuf buf1 = PRINTBUF;
1024                         struct printbuf buf2 = PRINTBUF;
1025
1026                         bch2_bkey_val_to_text(&buf1, c, old);
1027                         bch2_bkey_val_to_text(&buf2, c, new);
1028                         bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
1029                                             "old %s\n"
1030                                             "new %s", idx, buf1.buf, buf2.buf);
1031                         printbuf_exit(&buf2);
1032                         printbuf_exit(&buf1);
1033                         bch2_inconsistent_error(c);
1034                         return -1;
1035                 }
1036
1037                 if (!new_s) {
1038                         spin_lock(&c->ec_stripes_heap_lock);
1039                         bch2_stripes_heap_del(c, m, idx);
1040                         spin_unlock(&c->ec_stripes_heap_lock);
1041
1042                         memset(m, 0, sizeof(*m));
1043                 } else {
1044                         m->alive        = true;
1045                         m->sectors      = le16_to_cpu(new_s->sectors);
1046                         m->algorithm    = new_s->algorithm;
1047                         m->nr_blocks    = new_s->nr_blocks;
1048                         m->nr_redundant = new_s->nr_redundant;
1049                         m->blocks_nonempty = 0;
1050
1051                         for (i = 0; i < new_s->nr_blocks; i++)
1052                                 m->blocks_nonempty += !!stripe_blockcount_get(new_s, i);
1053
1054                         spin_lock(&c->ec_stripes_heap_lock);
1055                         bch2_stripes_heap_update(c, m, idx);
1056                         spin_unlock(&c->ec_stripes_heap_lock);
1057                 }
1058         } else {
1059                 struct gc_stripe *m =
1060                         genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
1061
1062                 if (!m) {
1063                         bch_err(c, "error allocating memory for gc_stripes, idx %llu",
1064                                 idx);
1065                         return -ENOMEM;
1066                 }
1067                 /*
1068                  * This will be wrong when we bring back runtime gc: we should
1069                  * be unmarking the old key and then marking the new key
1070                  */
1071                 m->alive        = true;
1072                 m->sectors      = le16_to_cpu(new_s->sectors);
1073                 m->nr_blocks    = new_s->nr_blocks;
1074                 m->nr_redundant = new_s->nr_redundant;
1075
1076                 for (i = 0; i < new_s->nr_blocks; i++)
1077                         m->ptrs[i] = new_s->ptrs[i];
1078
1079                 bch2_bkey_to_replicas(&m->r.e, new);
1080
1081                 /*
1082                  * gc recalculates this field from stripe ptr
1083                  * references:
1084                  */
1085                 memset(m->block_sectors, 0, sizeof(m->block_sectors));
1086
1087                 for (i = 0; i < new_s->nr_blocks; i++) {
1088                         ret = mark_stripe_bucket(trans, new, i, flags);
1089                         if (ret)
1090                                 return ret;
1091                 }
1092
1093                 ret = update_replicas(c, new, &m->r.e,
1094                                       ((s64) m->sectors * m->nr_redundant),
1095                                       journal_seq, gc);
1096                 if (ret) {
1097                         struct printbuf buf = PRINTBUF;
1098
1099                         bch2_bkey_val_to_text(&buf, c, new);
1100                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
1101                         printbuf_exit(&buf);
1102                         return ret;
1103                 }
1104         }
1105
1106         return 0;
1107 }
1108
1109 int bch2_mark_inode(struct btree_trans *trans,
1110                     struct bkey_s_c old, struct bkey_s_c new,
1111                     unsigned flags)
1112 {
1113         struct bch_fs *c = trans->c;
1114         struct bch_fs_usage __percpu *fs_usage;
1115         u64 journal_seq = trans->journal_res.seq;
1116
1117         if (flags & BTREE_TRIGGER_INSERT) {
1118                 struct bch_inode_v2 *v = (struct bch_inode_v2 *) new.v;
1119
1120                 BUG_ON(!journal_seq);
1121                 BUG_ON(new.k->type != KEY_TYPE_inode_v2);
1122
1123                 v->bi_journal_seq = cpu_to_le64(journal_seq);
1124         }
1125
1126         if (flags & BTREE_TRIGGER_GC) {
1127                 percpu_down_read(&c->mark_lock);
1128                 preempt_disable();
1129
1130                 fs_usage = fs_usage_ptr(c, journal_seq, flags & BTREE_TRIGGER_GC);
1131                 fs_usage->nr_inodes += bkey_is_inode(new.k);
1132                 fs_usage->nr_inodes -= bkey_is_inode(old.k);
1133
1134                 preempt_enable();
1135                 percpu_up_read(&c->mark_lock);
1136         }
1137         return 0;
1138 }
1139
1140 int bch2_mark_reservation(struct btree_trans *trans,
1141                           struct bkey_s_c old, struct bkey_s_c new,
1142                           unsigned flags)
1143 {
1144         struct bch_fs *c = trans->c;
1145         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1146         struct bch_fs_usage __percpu *fs_usage;
1147         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1148         s64 sectors = (s64) k.k->size;
1149
1150         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1151
1152         if (flags & BTREE_TRIGGER_OVERWRITE)
1153                 sectors = -sectors;
1154         sectors *= replicas;
1155
1156         percpu_down_read(&c->mark_lock);
1157         preempt_disable();
1158
1159         fs_usage = fs_usage_ptr(c, trans->journal_res.seq, flags & BTREE_TRIGGER_GC);
1160         replicas = clamp_t(unsigned, replicas, 1,
1161                            ARRAY_SIZE(fs_usage->persistent_reserved));
1162
1163         fs_usage->reserved                              += sectors;
1164         fs_usage->persistent_reserved[replicas - 1]     += sectors;
1165
1166         preempt_enable();
1167         percpu_up_read(&c->mark_lock);
1168
1169         return 0;
1170 }
1171
1172 static s64 __bch2_mark_reflink_p(struct btree_trans *trans,
1173                                  struct bkey_s_c_reflink_p p,
1174                                  u64 start, u64 end,
1175                                  u64 *idx, unsigned flags, size_t r_idx)
1176 {
1177         struct bch_fs *c = trans->c;
1178         struct reflink_gc *r;
1179         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1180         u64 next_idx = end;
1181         s64 ret = 0;
1182         struct printbuf buf = PRINTBUF;
1183
1184         if (r_idx >= c->reflink_gc_nr)
1185                 goto not_found;
1186
1187         r = genradix_ptr(&c->reflink_gc_table, r_idx);
1188         next_idx = min(next_idx, r->offset - r->size);
1189         if (*idx < next_idx)
1190                 goto not_found;
1191
1192         BUG_ON((s64) r->refcount + add < 0);
1193
1194         r->refcount += add;
1195         *idx = r->offset;
1196         return 0;
1197 not_found:
1198         if (fsck_err(c, "pointer to missing indirect extent\n"
1199                      "  %s\n"
1200                      "  missing range %llu-%llu",
1201                      (bch2_bkey_val_to_text(&buf, c, p.s_c), buf.buf),
1202                      *idx, next_idx)) {
1203                 struct bkey_i_error new;
1204
1205                 bkey_init(&new.k);
1206                 new.k.type      = KEY_TYPE_error;
1207                 new.k.p         = bkey_start_pos(p.k);
1208                 new.k.p.offset += *idx - start;
1209                 bch2_key_resize(&new.k, next_idx - *idx);
1210                 ret = __bch2_btree_insert(trans, BTREE_ID_extents, &new.k_i);
1211         }
1212
1213         *idx = next_idx;
1214 fsck_err:
1215         printbuf_exit(&buf);
1216         return ret;
1217 }
1218
1219 int bch2_mark_reflink_p(struct btree_trans *trans,
1220                         struct bkey_s_c old, struct bkey_s_c new,
1221                         unsigned flags)
1222 {
1223         struct bch_fs *c = trans->c;
1224         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1225         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1226         struct reflink_gc *ref;
1227         size_t l, r, m;
1228         u64 idx = le64_to_cpu(p.v->idx), start = idx;
1229         u64 end = le64_to_cpu(p.v->idx) + p.k->size;
1230         int ret = 0;
1231
1232         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1233
1234         if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix) {
1235                 idx -= le32_to_cpu(p.v->front_pad);
1236                 end += le32_to_cpu(p.v->back_pad);
1237         }
1238
1239         l = 0;
1240         r = c->reflink_gc_nr;
1241         while (l < r) {
1242                 m = l + (r - l) / 2;
1243
1244                 ref = genradix_ptr(&c->reflink_gc_table, m);
1245                 if (ref->offset <= idx)
1246                         l = m + 1;
1247                 else
1248                         r = m;
1249         }
1250
1251         while (idx < end && !ret)
1252                 ret = __bch2_mark_reflink_p(trans, p, start, end,
1253                                             &idx, flags, l++);
1254
1255         return ret;
1256 }
1257
1258 static noinline __cold
1259 void fs_usage_apply_warn(struct btree_trans *trans,
1260                          unsigned disk_res_sectors,
1261                          s64 should_not_have_added)
1262 {
1263         struct bch_fs *c = trans->c;
1264         struct btree_insert_entry *i;
1265         struct printbuf buf = PRINTBUF;
1266
1267         bch_err(c, "disk usage increased %lli more than %u sectors reserved",
1268                 should_not_have_added, disk_res_sectors);
1269
1270         trans_for_each_update(trans, i) {
1271                 struct bkey_s_c old = { &i->old_k, i->old_v };
1272
1273                 pr_err("while inserting");
1274                 printbuf_reset(&buf);
1275                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(i->k));
1276                 pr_err("  %s", buf.buf);
1277                 pr_err("overlapping with");
1278                 printbuf_reset(&buf);
1279                 bch2_bkey_val_to_text(&buf, c, old);
1280                 pr_err("  %s", buf.buf);
1281         }
1282
1283         __WARN();
1284         printbuf_exit(&buf);
1285 }
1286
1287 int bch2_trans_fs_usage_apply(struct btree_trans *trans,
1288                               struct replicas_delta_list *deltas)
1289 {
1290         struct bch_fs *c = trans->c;
1291         static int warned_disk_usage = 0;
1292         bool warn = false;
1293         unsigned disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1294         struct replicas_delta *d = deltas->d, *d2;
1295         struct replicas_delta *top = (void *) deltas->d + deltas->used;
1296         struct bch_fs_usage *dst;
1297         s64 added = 0, should_not_have_added;
1298         unsigned i;
1299
1300         percpu_down_read(&c->mark_lock);
1301         preempt_disable();
1302         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1303
1304         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1305                 switch (d->r.data_type) {
1306                 case BCH_DATA_btree:
1307                 case BCH_DATA_user:
1308                 case BCH_DATA_parity:
1309                         added += d->delta;
1310                 }
1311
1312                 if (__update_replicas(c, dst, &d->r, d->delta))
1313                         goto need_mark;
1314         }
1315
1316         dst->nr_inodes += deltas->nr_inodes;
1317
1318         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1319                 added                           += deltas->persistent_reserved[i];
1320                 dst->reserved                   += deltas->persistent_reserved[i];
1321                 dst->persistent_reserved[i]     += deltas->persistent_reserved[i];
1322         }
1323
1324         /*
1325          * Not allowed to reduce sectors_available except by getting a
1326          * reservation:
1327          */
1328         should_not_have_added = added - (s64) disk_res_sectors;
1329         if (unlikely(should_not_have_added > 0)) {
1330                 u64 old, new, v = atomic64_read(&c->sectors_available);
1331
1332                 do {
1333                         old = v;
1334                         new = max_t(s64, 0, old - should_not_have_added);
1335                 } while ((v = atomic64_cmpxchg(&c->sectors_available,
1336                                                old, new)) != old);
1337
1338                 added -= should_not_have_added;
1339                 warn = true;
1340         }
1341
1342         if (added > 0) {
1343                 trans->disk_res->sectors -= added;
1344                 this_cpu_sub(*c->online_reserved, added);
1345         }
1346
1347         preempt_enable();
1348         percpu_up_read(&c->mark_lock);
1349
1350         if (unlikely(warn) && !xchg(&warned_disk_usage, 1))
1351                 fs_usage_apply_warn(trans, disk_res_sectors, should_not_have_added);
1352         return 0;
1353 need_mark:
1354         /* revert changes: */
1355         for (d2 = deltas->d; d2 != d; d2 = replicas_delta_next(d2))
1356                 BUG_ON(__update_replicas(c, dst, &d2->r, -d2->delta));
1357
1358         preempt_enable();
1359         percpu_up_read(&c->mark_lock);
1360         return -1;
1361 }
1362
1363 /* trans_mark: */
1364
1365 static int bch2_trans_mark_pointer(struct btree_trans *trans,
1366                                    enum btree_id btree_id, unsigned level,
1367                                    struct bkey_s_c k, struct extent_ptr_decoded p,
1368                                    unsigned flags)
1369 {
1370         bool insert = !(flags & BTREE_TRIGGER_OVERWRITE);
1371         struct btree_iter iter;
1372         struct bkey_i_alloc_v4 *a;
1373         struct bpos bucket_pos;
1374         struct bch_backpointer bp;
1375         s64 sectors;
1376         int ret;
1377
1378         bch2_extent_ptr_to_bp(trans->c, btree_id, level, k, p, &bucket_pos, &bp);
1379         sectors = bp.bucket_len;
1380         if (!insert)
1381                 sectors = -sectors;
1382
1383         a = bch2_trans_start_alloc_update(trans, &iter, bucket_pos);
1384         if (IS_ERR(a))
1385                 return PTR_ERR(a);
1386
1387         ret = __mark_pointer(trans, k, &p.ptr, sectors, bp.data_type,
1388                              a->v.gen, &a->v.data_type,
1389                              &a->v.dirty_sectors, &a->v.cached_sectors);
1390         if (ret)
1391                 goto err;
1392
1393         if (!p.ptr.cached) {
1394                 ret = insert
1395                         ? bch2_bucket_backpointer_add(trans, a, bp, k)
1396                         : bch2_bucket_backpointer_del(trans, a, bp, k);
1397                 if (ret)
1398                         goto err;
1399         }
1400
1401         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1402 err:
1403         bch2_trans_iter_exit(trans, &iter);
1404         return ret;
1405 }
1406
1407 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1408                         struct extent_ptr_decoded p,
1409                         s64 sectors, enum bch_data_type data_type)
1410 {
1411         struct btree_iter iter;
1412         struct bkey_s_c k;
1413         struct bkey_i_stripe *s;
1414         struct bch_replicas_padded r;
1415         int ret = 0;
1416
1417         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes, POS(0, p.ec.idx),
1418                              BTREE_ITER_INTENT|
1419                              BTREE_ITER_WITH_UPDATES);
1420         k = bch2_btree_iter_peek_slot(&iter);
1421         ret = bkey_err(k);
1422         if (ret)
1423                 goto err;
1424
1425         if (k.k->type != KEY_TYPE_stripe) {
1426                 bch2_trans_inconsistent(trans,
1427                         "pointer to nonexistent stripe %llu",
1428                         (u64) p.ec.idx);
1429                 ret = -EIO;
1430                 goto err;
1431         }
1432
1433         if (!bch2_ptr_matches_stripe(bkey_s_c_to_stripe(k).v, p)) {
1434                 bch2_trans_inconsistent(trans,
1435                         "stripe pointer doesn't match stripe %llu",
1436                         (u64) p.ec.idx);
1437                 ret = -EIO;
1438                 goto err;
1439         }
1440
1441         s = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1442         ret = PTR_ERR_OR_ZERO(s);
1443         if (ret)
1444                 goto err;
1445
1446         bkey_reassemble(&s->k_i, k);
1447         stripe_blockcount_set(&s->v, p.ec.block,
1448                 stripe_blockcount_get(&s->v, p.ec.block) +
1449                 sectors);
1450
1451         ret = bch2_trans_update(trans, &iter, &s->k_i, 0);
1452         if (ret)
1453                 goto err;
1454
1455         bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1456         r.e.data_type = data_type;
1457         update_replicas_list(trans, &r.e, sectors);
1458 err:
1459         bch2_trans_iter_exit(trans, &iter);
1460         return ret;
1461 }
1462
1463 int bch2_trans_mark_extent(struct btree_trans *trans,
1464                            enum btree_id btree_id, unsigned level,
1465                            struct bkey_s_c old, struct bkey_i *new,
1466                            unsigned flags)
1467 {
1468         struct bch_fs *c = trans->c;
1469         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1470                 ? old
1471                 : bkey_i_to_s_c(new);
1472         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1473         const union bch_extent_entry *entry;
1474         struct extent_ptr_decoded p;
1475         struct bch_replicas_padded r;
1476         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
1477                 ? BCH_DATA_btree
1478                 : BCH_DATA_user;
1479         s64 sectors = bkey_is_btree_ptr(k.k)
1480                 ? btree_sectors(c)
1481                 : k.k->size;
1482         s64 dirty_sectors = 0;
1483         bool stale;
1484         int ret;
1485
1486         r.e.data_type   = data_type;
1487         r.e.nr_devs     = 0;
1488         r.e.nr_required = 1;
1489
1490         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1491                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
1492
1493                 if (flags & BTREE_TRIGGER_OVERWRITE)
1494                         disk_sectors = -disk_sectors;
1495
1496                 ret = bch2_trans_mark_pointer(trans, btree_id, level, k, p, flags);
1497                 if (ret < 0)
1498                         return ret;
1499
1500                 stale = ret > 0;
1501
1502                 if (p.ptr.cached) {
1503                         if (!stale)
1504                                 update_cached_sectors_list(trans, p.ptr.dev,
1505                                                            disk_sectors);
1506                 } else if (!p.has_ec) {
1507                         dirty_sectors          += disk_sectors;
1508                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1509                 } else {
1510                         ret = bch2_trans_mark_stripe_ptr(trans, p,
1511                                         disk_sectors, data_type);
1512                         if (ret)
1513                                 return ret;
1514
1515                         r.e.nr_required = 0;
1516                 }
1517         }
1518
1519         if (r.e.nr_devs)
1520                 update_replicas_list(trans, &r.e, dirty_sectors);
1521
1522         return 0;
1523 }
1524
1525 static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
1526                                          struct bkey_s_c_stripe s,
1527                                          unsigned idx, bool deleting)
1528 {
1529         struct bch_fs *c = trans->c;
1530         const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
1531         struct btree_iter iter;
1532         struct bkey_i_alloc_v4 *a;
1533         enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
1534                 ? BCH_DATA_parity : 0;
1535         s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
1536         int ret = 0;
1537
1538         if (deleting)
1539                 sectors = -sectors;
1540
1541         a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(c, ptr));
1542         if (IS_ERR(a))
1543                 return PTR_ERR(a);
1544
1545         ret = check_bucket_ref(c, s.s_c, ptr, sectors, data_type,
1546                                a->v.gen, a->v.data_type,
1547                                a->v.dirty_sectors, a->v.cached_sectors);
1548         if (ret)
1549                 goto err;
1550
1551         if (!deleting) {
1552                 if (bch2_trans_inconsistent_on(a->v.stripe ||
1553                                                a->v.stripe_redundancy, trans,
1554                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
1555                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1556                                 bch2_data_types[a->v.data_type],
1557                                 a->v.dirty_sectors,
1558                                 a->v.stripe, s.k->p.offset)) {
1559                         ret = -EIO;
1560                         goto err;
1561                 }
1562
1563                 if (bch2_trans_inconsistent_on(data_type && a->v.dirty_sectors, trans,
1564                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
1565                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1566                                 bch2_data_types[a->v.data_type],
1567                                 a->v.dirty_sectors,
1568                                 s.k->p.offset)) {
1569                         ret = -EIO;
1570                         goto err;
1571                 }
1572
1573                 a->v.stripe             = s.k->p.offset;
1574                 a->v.stripe_redundancy  = s.v->nr_redundant;
1575         } else {
1576                 if (bch2_trans_inconsistent_on(a->v.stripe != s.k->p.offset ||
1577                                                a->v.stripe_redundancy != s.v->nr_redundant, trans,
1578                                 "bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
1579                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1580                                 s.k->p.offset, a->v.stripe)) {
1581                         ret = -EIO;
1582                         goto err;
1583                 }
1584
1585                 a->v.stripe             = 0;
1586                 a->v.stripe_redundancy  = 0;
1587         }
1588
1589         a->v.dirty_sectors += sectors;
1590         if (data_type)
1591                 a->v.data_type = !deleting ? data_type : 0;
1592
1593         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1594         if (ret)
1595                 goto err;
1596 err:
1597         bch2_trans_iter_exit(trans, &iter);
1598         return ret;
1599 }
1600
1601 int bch2_trans_mark_stripe(struct btree_trans *trans,
1602                            enum btree_id btree_id, unsigned level,
1603                            struct bkey_s_c old, struct bkey_i *new,
1604                            unsigned flags)
1605 {
1606         const struct bch_stripe *old_s = NULL;
1607         struct bch_stripe *new_s = NULL;
1608         struct bch_replicas_padded r;
1609         unsigned i, nr_blocks;
1610         int ret = 0;
1611
1612         if (old.k->type == KEY_TYPE_stripe)
1613                 old_s = bkey_s_c_to_stripe(old).v;
1614         if (new->k.type == KEY_TYPE_stripe)
1615                 new_s = &bkey_i_to_stripe(new)->v;
1616
1617         /*
1618          * If the pointers aren't changing, we don't need to do anything:
1619          */
1620         if (new_s && old_s &&
1621             new_s->nr_blocks    == old_s->nr_blocks &&
1622             new_s->nr_redundant == old_s->nr_redundant &&
1623             !memcmp(old_s->ptrs, new_s->ptrs,
1624                     new_s->nr_blocks * sizeof(struct bch_extent_ptr)))
1625                 return 0;
1626
1627         BUG_ON(new_s && old_s &&
1628                (new_s->nr_blocks        != old_s->nr_blocks ||
1629                 new_s->nr_redundant     != old_s->nr_redundant));
1630
1631         nr_blocks = new_s ? new_s->nr_blocks : old_s->nr_blocks;
1632
1633         if (new_s) {
1634                 s64 sectors = le16_to_cpu(new_s->sectors);
1635
1636                 bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(new));
1637                 update_replicas_list(trans, &r.e, sectors * new_s->nr_redundant);
1638         }
1639
1640         if (old_s) {
1641                 s64 sectors = -((s64) le16_to_cpu(old_s->sectors));
1642
1643                 bch2_bkey_to_replicas(&r.e, old);
1644                 update_replicas_list(trans, &r.e, sectors * old_s->nr_redundant);
1645         }
1646
1647         for (i = 0; i < nr_blocks; i++) {
1648                 if (new_s && old_s &&
1649                     !memcmp(&new_s->ptrs[i],
1650                             &old_s->ptrs[i],
1651                             sizeof(new_s->ptrs[i])))
1652                         continue;
1653
1654                 if (new_s) {
1655                         ret = bch2_trans_mark_stripe_bucket(trans,
1656                                         bkey_i_to_s_c_stripe(new), i, false);
1657                         if (ret)
1658                                 break;
1659                 }
1660
1661                 if (old_s) {
1662                         ret = bch2_trans_mark_stripe_bucket(trans,
1663                                         bkey_s_c_to_stripe(old), i, true);
1664                         if (ret)
1665                                 break;
1666                 }
1667         }
1668
1669         return ret;
1670 }
1671
1672 int bch2_trans_mark_inode(struct btree_trans *trans,
1673                           enum btree_id btree_id, unsigned level,
1674                           struct bkey_s_c old,
1675                           struct bkey_i *new,
1676                           unsigned flags)
1677 {
1678         int nr = bkey_is_inode(&new->k) - bkey_is_inode(old.k);
1679
1680         if (nr) {
1681                 struct replicas_delta_list *d =
1682                         replicas_deltas_realloc(trans, 0);
1683                 d->nr_inodes += nr;
1684         }
1685
1686         return 0;
1687 }
1688
1689 int bch2_trans_mark_reservation(struct btree_trans *trans,
1690                                 enum btree_id btree_id, unsigned level,
1691                                 struct bkey_s_c old,
1692                                 struct bkey_i *new,
1693                                 unsigned flags)
1694 {
1695         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1696                 ? old
1697                 : bkey_i_to_s_c(new);
1698         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1699         s64 sectors = (s64) k.k->size;
1700         struct replicas_delta_list *d;
1701
1702         if (flags & BTREE_TRIGGER_OVERWRITE)
1703                 sectors = -sectors;
1704         sectors *= replicas;
1705
1706         d = replicas_deltas_realloc(trans, 0);
1707
1708         replicas = clamp_t(unsigned, replicas, 1,
1709                            ARRAY_SIZE(d->persistent_reserved));
1710
1711         d->persistent_reserved[replicas - 1] += sectors;
1712         return 0;
1713 }
1714
1715 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1716                         struct bkey_s_c_reflink_p p,
1717                         u64 *idx, unsigned flags)
1718 {
1719         struct bch_fs *c = trans->c;
1720         struct btree_iter iter;
1721         struct bkey_s_c k;
1722         struct bkey_i *n;
1723         __le64 *refcount;
1724         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1725         struct printbuf buf = PRINTBUF;
1726         int ret;
1727
1728         bch2_trans_iter_init(trans, &iter, BTREE_ID_reflink, POS(0, *idx),
1729                              BTREE_ITER_INTENT|
1730                              BTREE_ITER_WITH_UPDATES);
1731         k = bch2_btree_iter_peek_slot(&iter);
1732         ret = bkey_err(k);
1733         if (ret)
1734                 goto err;
1735
1736         n = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1737         ret = PTR_ERR_OR_ZERO(n);
1738         if (ret)
1739                 goto err;
1740
1741         bkey_reassemble(n, k);
1742
1743         refcount = bkey_refcount(n);
1744         if (!refcount) {
1745                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1746                 bch2_trans_inconsistent(trans,
1747                         "nonexistent indirect extent at %llu while marking\n  %s",
1748                         *idx, buf.buf);
1749                 ret = -EIO;
1750                 goto err;
1751         }
1752
1753         if (!*refcount && (flags & BTREE_TRIGGER_OVERWRITE)) {
1754                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1755                 bch2_trans_inconsistent(trans,
1756                         "indirect extent refcount underflow at %llu while marking\n  %s",
1757                         *idx, buf.buf);
1758                 ret = -EIO;
1759                 goto err;
1760         }
1761
1762         if (flags & BTREE_TRIGGER_INSERT) {
1763                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1764                 u64 pad;
1765
1766                 pad = max_t(s64, le32_to_cpu(v->front_pad),
1767                             le64_to_cpu(v->idx) - bkey_start_offset(k.k));
1768                 BUG_ON(pad > U32_MAX);
1769                 v->front_pad = cpu_to_le32(pad);
1770
1771                 pad = max_t(s64, le32_to_cpu(v->back_pad),
1772                             k.k->p.offset - p.k->size - le64_to_cpu(v->idx));
1773                 BUG_ON(pad > U32_MAX);
1774                 v->back_pad = cpu_to_le32(pad);
1775         }
1776
1777         le64_add_cpu(refcount, add);
1778
1779         bch2_btree_iter_set_pos_to_extent_start(&iter);
1780         ret = bch2_trans_update(trans, &iter, n, 0);
1781         if (ret)
1782                 goto err;
1783
1784         *idx = k.k->p.offset;
1785 err:
1786         bch2_trans_iter_exit(trans, &iter);
1787         printbuf_exit(&buf);
1788         return ret;
1789 }
1790
1791 int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1792                               enum btree_id btree_id, unsigned level,
1793                               struct bkey_s_c old,
1794                               struct bkey_i *new,
1795                               unsigned flags)
1796 {
1797         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1798                 ? old
1799                 : bkey_i_to_s_c(new);
1800         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1801         u64 idx, end_idx;
1802         int ret = 0;
1803
1804         if (flags & BTREE_TRIGGER_INSERT) {
1805                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1806
1807                 v->front_pad = v->back_pad = 0;
1808         }
1809
1810         idx     = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
1811         end_idx = le64_to_cpu(p.v->idx) + p.k->size +
1812                 le32_to_cpu(p.v->back_pad);
1813
1814         while (idx < end_idx && !ret)
1815                 ret = __bch2_trans_mark_reflink_p(trans, p, &idx, flags);
1816
1817         return ret;
1818 }
1819
1820 static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1821                                     struct bch_dev *ca, size_t b,
1822                                     enum bch_data_type type,
1823                                     unsigned sectors)
1824 {
1825         struct bch_fs *c = trans->c;
1826         struct btree_iter iter;
1827         struct bkey_i_alloc_v4 *a;
1828         int ret = 0;
1829
1830         /*
1831          * Backup superblock might be past the end of our normal usable space:
1832          */
1833         if (b >= ca->mi.nbuckets)
1834                 return 0;
1835
1836         a = bch2_trans_start_alloc_update(trans, &iter, POS(ca->dev_idx, b));
1837         if (IS_ERR(a))
1838                 return PTR_ERR(a);
1839
1840         if (a->v.data_type && a->v.data_type != type) {
1841                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1842                         "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
1843                         "while marking %s",
1844                         iter.pos.inode, iter.pos.offset, a->v.gen,
1845                         bch2_data_types[a->v.data_type],
1846                         bch2_data_types[type],
1847                         bch2_data_types[type]);
1848                 ret = -EIO;
1849                 goto out;
1850         }
1851
1852         a->v.data_type          = type;
1853         a->v.dirty_sectors      = sectors;
1854
1855         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1856         if (ret)
1857                 goto out;
1858 out:
1859         bch2_trans_iter_exit(trans, &iter);
1860         return ret;
1861 }
1862
1863 int bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1864                                     struct bch_dev *ca, size_t b,
1865                                     enum bch_data_type type,
1866                                     unsigned sectors)
1867 {
1868         return commit_do(trans, NULL, NULL, 0,
1869                         __bch2_trans_mark_metadata_bucket(trans, ca, b, type, sectors));
1870 }
1871
1872 static int bch2_trans_mark_metadata_sectors(struct btree_trans *trans,
1873                                             struct bch_dev *ca,
1874                                             u64 start, u64 end,
1875                                             enum bch_data_type type,
1876                                             u64 *bucket, unsigned *bucket_sectors)
1877 {
1878         do {
1879                 u64 b = sector_to_bucket(ca, start);
1880                 unsigned sectors =
1881                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1882
1883                 if (b != *bucket && *bucket_sectors) {
1884                         int ret = bch2_trans_mark_metadata_bucket(trans, ca, *bucket,
1885                                                                   type, *bucket_sectors);
1886                         if (ret)
1887                                 return ret;
1888
1889                         *bucket_sectors = 0;
1890                 }
1891
1892                 *bucket         = b;
1893                 *bucket_sectors += sectors;
1894                 start += sectors;
1895         } while (start < end);
1896
1897         return 0;
1898 }
1899
1900 static int __bch2_trans_mark_dev_sb(struct btree_trans *trans,
1901                                     struct bch_dev *ca)
1902 {
1903         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1904         u64 bucket = 0;
1905         unsigned i, bucket_sectors = 0;
1906         int ret;
1907
1908         for (i = 0; i < layout->nr_superblocks; i++) {
1909                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1910
1911                 if (offset == BCH_SB_SECTOR) {
1912                         ret = bch2_trans_mark_metadata_sectors(trans, ca,
1913                                                 0, BCH_SB_SECTOR,
1914                                                 BCH_DATA_sb, &bucket, &bucket_sectors);
1915                         if (ret)
1916                                 return ret;
1917                 }
1918
1919                 ret = bch2_trans_mark_metadata_sectors(trans, ca, offset,
1920                                       offset + (1 << layout->sb_max_size_bits),
1921                                       BCH_DATA_sb, &bucket, &bucket_sectors);
1922                 if (ret)
1923                         return ret;
1924         }
1925
1926         if (bucket_sectors) {
1927                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1928                                 bucket, BCH_DATA_sb, bucket_sectors);
1929                 if (ret)
1930                         return ret;
1931         }
1932
1933         for (i = 0; i < ca->journal.nr; i++) {
1934                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1935                                 ca->journal.buckets[i],
1936                                 BCH_DATA_journal, ca->mi.bucket_size);
1937                 if (ret)
1938                         return ret;
1939         }
1940
1941         return 0;
1942 }
1943
1944 int bch2_trans_mark_dev_sb(struct bch_fs *c, struct bch_dev *ca)
1945 {
1946         return bch2_trans_run(c, __bch2_trans_mark_dev_sb(&trans, ca));
1947 }
1948
1949 /* Disk reservations: */
1950
1951 #define SECTORS_CACHE   1024
1952
1953 int bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
1954                               u64 sectors, int flags)
1955 {
1956         struct bch_fs_pcpu *pcpu;
1957         u64 old, v, get;
1958         s64 sectors_available;
1959         int ret;
1960
1961         percpu_down_read(&c->mark_lock);
1962         preempt_disable();
1963         pcpu = this_cpu_ptr(c->pcpu);
1964
1965         if (sectors <= pcpu->sectors_available)
1966                 goto out;
1967
1968         v = atomic64_read(&c->sectors_available);
1969         do {
1970                 old = v;
1971                 get = min((u64) sectors + SECTORS_CACHE, old);
1972
1973                 if (get < sectors) {
1974                         preempt_enable();
1975                         goto recalculate;
1976                 }
1977         } while ((v = atomic64_cmpxchg(&c->sectors_available,
1978                                        old, old - get)) != old);
1979
1980         pcpu->sectors_available         += get;
1981
1982 out:
1983         pcpu->sectors_available         -= sectors;
1984         this_cpu_add(*c->online_reserved, sectors);
1985         res->sectors                    += sectors;
1986
1987         preempt_enable();
1988         percpu_up_read(&c->mark_lock);
1989         return 0;
1990
1991 recalculate:
1992         mutex_lock(&c->sectors_available_lock);
1993
1994         percpu_u64_set(&c->pcpu->sectors_available, 0);
1995         sectors_available = avail_factor(__bch2_fs_usage_read_short(c).free);
1996
1997         if (sectors <= sectors_available ||
1998             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
1999                 atomic64_set(&c->sectors_available,
2000                              max_t(s64, 0, sectors_available - sectors));
2001                 this_cpu_add(*c->online_reserved, sectors);
2002                 res->sectors                    += sectors;
2003                 ret = 0;
2004         } else {
2005                 atomic64_set(&c->sectors_available, sectors_available);
2006                 ret = -BCH_ERR_ENOSPC_disk_reservation;
2007         }
2008
2009         mutex_unlock(&c->sectors_available_lock);
2010         percpu_up_read(&c->mark_lock);
2011
2012         return ret;
2013 }
2014
2015 /* Startup/shutdown: */
2016
2017 static void bucket_gens_free_rcu(struct rcu_head *rcu)
2018 {
2019         struct bucket_gens *buckets =
2020                 container_of(rcu, struct bucket_gens, rcu);
2021
2022         kvpfree(buckets, sizeof(*buckets) + buckets->nbuckets);
2023 }
2024
2025 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
2026 {
2027         struct bucket_gens *bucket_gens = NULL, *old_bucket_gens = NULL;
2028         unsigned long *buckets_nouse = NULL;
2029         bool resize = ca->bucket_gens != NULL;
2030         int ret = -ENOMEM;
2031
2032         if (!(bucket_gens       = kvpmalloc(sizeof(struct bucket_gens) + nbuckets,
2033                                             GFP_KERNEL|__GFP_ZERO)) ||
2034             (c->opts.buckets_nouse &&
2035              !(buckets_nouse    = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2036                                             sizeof(unsigned long),
2037                                             GFP_KERNEL|__GFP_ZERO))))
2038                 goto err;
2039
2040         bucket_gens->first_bucket = ca->mi.first_bucket;
2041         bucket_gens->nbuckets   = nbuckets;
2042
2043         bch2_copygc_stop(c);
2044
2045         if (resize) {
2046                 down_write(&c->gc_lock);
2047                 down_write(&ca->bucket_lock);
2048                 percpu_down_write(&c->mark_lock);
2049         }
2050
2051         old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
2052
2053         if (resize) {
2054                 size_t n = min(bucket_gens->nbuckets, old_bucket_gens->nbuckets);
2055
2056                 memcpy(bucket_gens->b,
2057                        old_bucket_gens->b,
2058                        n);
2059                 if (buckets_nouse)
2060                         memcpy(buckets_nouse,
2061                                ca->buckets_nouse,
2062                                BITS_TO_LONGS(n) * sizeof(unsigned long));
2063         }
2064
2065         rcu_assign_pointer(ca->bucket_gens, bucket_gens);
2066         bucket_gens     = old_bucket_gens;
2067
2068         swap(ca->buckets_nouse, buckets_nouse);
2069
2070         nbuckets = ca->mi.nbuckets;
2071
2072         if (resize) {
2073                 percpu_up_write(&c->mark_lock);
2074                 up_write(&ca->bucket_lock);
2075                 up_write(&c->gc_lock);
2076         }
2077
2078         ret = 0;
2079 err:
2080         kvpfree(buckets_nouse,
2081                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2082         if (bucket_gens)
2083                 call_rcu(&bucket_gens->rcu, bucket_gens_free_rcu);
2084
2085         return ret;
2086 }
2087
2088 void bch2_dev_buckets_free(struct bch_dev *ca)
2089 {
2090         unsigned i;
2091
2092         kvpfree(ca->buckets_nouse,
2093                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2094         kvpfree(rcu_dereference_protected(ca->bucket_gens, 1),
2095                 sizeof(struct bucket_gens) + ca->mi.nbuckets);
2096
2097         for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
2098                 free_percpu(ca->usage[i]);
2099         kfree(ca->usage_base);
2100 }
2101
2102 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2103 {
2104         unsigned i;
2105
2106         ca->usage_base = kzalloc(sizeof(struct bch_dev_usage), GFP_KERNEL);
2107         if (!ca->usage_base)
2108                 return -ENOMEM;
2109
2110         for (i = 0; i < ARRAY_SIZE(ca->usage); i++) {
2111                 ca->usage[i] = alloc_percpu(struct bch_dev_usage);
2112                 if (!ca->usage[i])
2113                         return -ENOMEM;
2114         }
2115
2116         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);;
2117 }