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

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Code for manipulating bucket marks for garbage collection.
   4  *
   5  * Copyright 2014 Datera, Inc.
   6  */
   7
   8 #include "bcachefs.h"
   9 #include "alloc_background.h"
  10 #include "bset.h"
  11 #include "btree_gc.h"
  12 #include "btree_update.h"
  13 #include "buckets.h"
  14 #include "buckets_waiting_for_journal.h"
  15 #include "ec.h"
  16 #include "error.h"
  17 #include "inode.h"
  18 #include "movinggc.h"
  19 #include "recovery.h"
  20 #include "reflink.h"
  21 #include "replicas.h"
  22 #include "subvolume.h"
  23
  24 #include <linux/preempt.h>
  25 #include <trace/events/bcachefs.h>
  26
  27 static inline void fs_usage_data_type_to_base(struct bch_fs_usage *fs_usage,
  28                                               enum bch_data_type data_type,
  29                                               s64 sectors)
  30 {
  31         switch (data_type) {
  32         case BCH_DATA_btree:
  33                 fs_usage->btree         += sectors;
  34                 break;
  35         case BCH_DATA_user:
  36         case BCH_DATA_parity:
  37                 fs_usage->data          += sectors;
  38                 break;
  39         case BCH_DATA_cached:
  40                 fs_usage->cached        += sectors;
  41                 break;
  42         default:
  43                 break;
  44         }
  45 }
  46
  47 void bch2_fs_usage_initialize(struct bch_fs *c)
  48 {
  49         struct bch_fs_usage *usage;
  50         struct bch_dev *ca;
  51         unsigned i;
  52
  53         percpu_down_write(&c->mark_lock);
  54         usage = c->usage_base;
  55
  56         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
  57                 bch2_fs_usage_acc_to_base(c, i);
  58
  59         for (i = 0; i < BCH_REPLICAS_MAX; i++)
  60                 usage->reserved += usage->persistent_reserved[i];
  61
  62         for (i = 0; i < c->replicas.nr; i++) {
  63                 struct bch_replicas_entry *e =
  64                         cpu_replicas_entry(&c->replicas, i);
  65
  66                 fs_usage_data_type_to_base(usage, e->data_type, usage->replicas[i]);
  67         }
  68
  69         for_each_member_device(ca, c, i) {
  70                 struct bch_dev_usage dev = bch2_dev_usage_read(ca);
  71
  72                 usage->hidden += (dev.d[BCH_DATA_sb].buckets +
  73                                   dev.d[BCH_DATA_journal].buckets) *
  74                         ca->mi.bucket_size;
  75         }
  76
  77         percpu_up_write(&c->mark_lock);
  78 }
  79
  80 static inline struct bch_dev_usage *dev_usage_ptr(struct bch_dev *ca,
  81                                                   unsigned journal_seq,
  82                                                   bool gc)
  83 {
  84         BUG_ON(!gc && !journal_seq);
  85
  86         return this_cpu_ptr(gc
  87                             ? ca->usage_gc
  88                             : ca->usage[journal_seq & JOURNAL_BUF_MASK]);
  89 }
  90
  91 struct bch_dev_usage bch2_dev_usage_read(struct bch_dev *ca)
  92 {
  93         struct bch_fs *c = ca->fs;
  94         struct bch_dev_usage ret;
  95         unsigned seq, i, u64s = dev_usage_u64s();
  96
  97         do {
  98                 seq = read_seqcount_begin(&c->usage_lock);
  99                 memcpy(&ret, ca->usage_base, u64s * sizeof(u64));
 100                 for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
 101                         acc_u64s_percpu((u64 *) &ret, (u64 __percpu *) ca->usage[i], u64s);
 102         } while (read_seqcount_retry(&c->usage_lock, seq));
 103
 104         return ret;
 105 }
 106
 107 static inline struct bch_fs_usage *fs_usage_ptr(struct bch_fs *c,
 108                                                 unsigned journal_seq,
 109                                                 bool gc)
 110 {
 111         percpu_rwsem_assert_held(&c->mark_lock);
 112         BUG_ON(!gc && !journal_seq);
 113
 114         return this_cpu_ptr(gc
 115                             ? c->usage_gc
 116                             : c->usage[journal_seq & JOURNAL_BUF_MASK]);
 117 }
 118
 119 u64 bch2_fs_usage_read_one(struct bch_fs *c, u64 *v)
 120 {
 121         ssize_t offset = v - (u64 *) c->usage_base;
 122         unsigned i, seq;
 123         u64 ret;
 124
 125         BUG_ON(offset < 0 || offset >= fs_usage_u64s(c));
 126         percpu_rwsem_assert_held(&c->mark_lock);
 127
 128         do {
 129                 seq = read_seqcount_begin(&c->usage_lock);
 130                 ret = *v;
 131
 132                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 133                         ret += percpu_u64_get((u64 __percpu *) c->usage[i] + offset);
 134         } while (read_seqcount_retry(&c->usage_lock, seq));
 135
 136         return ret;
 137 }
 138
 139 struct bch_fs_usage_online *bch2_fs_usage_read(struct bch_fs *c)
 140 {
 141         struct bch_fs_usage_online *ret;
 142         unsigned seq, i, u64s;
 143
 144         percpu_down_read(&c->mark_lock);
 145
 146         ret = kmalloc(sizeof(struct bch_fs_usage_online) +
 147                       sizeof(u64) * c->replicas.nr, GFP_NOFS);
 148         if (unlikely(!ret)) {
 149                 percpu_up_read(&c->mark_lock);
 150                 return NULL;
 151         }
 152
 153         ret->online_reserved = percpu_u64_get(c->online_reserved);
 154
 155         u64s = fs_usage_u64s(c);
 156         do {
 157                 seq = read_seqcount_begin(&c->usage_lock);
 158                 memcpy(&ret->u, c->usage_base, u64s * sizeof(u64));
 159                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 160                         acc_u64s_percpu((u64 *) &ret->u, (u64 __percpu *) c->usage[i], u64s);
 161         } while (read_seqcount_retry(&c->usage_lock, seq));
 162
 163         return ret;
 164 }
 165
 166 void bch2_fs_usage_acc_to_base(struct bch_fs *c, unsigned idx)
 167 {
 168         struct bch_dev *ca;
 169         unsigned i, u64s = fs_usage_u64s(c);
 170
 171         BUG_ON(idx >= ARRAY_SIZE(c->usage));
 172
 173         preempt_disable();
 174         write_seqcount_begin(&c->usage_lock);
 175
 176         acc_u64s_percpu((u64 *) c->usage_base,
 177                         (u64 __percpu *) c->usage[idx], u64s);
 178         percpu_memset(c->usage[idx], 0, u64s * sizeof(u64));
 179
 180         rcu_read_lock();
 181         for_each_member_device_rcu(ca, c, i, NULL) {
 182                 u64s = dev_usage_u64s();
 183
 184                 acc_u64s_percpu((u64 *) ca->usage_base,
 185                                 (u64 __percpu *) ca->usage[idx], u64s);
 186                 percpu_memset(ca->usage[idx], 0, u64s * sizeof(u64));
 187         }
 188         rcu_read_unlock();
 189
 190         write_seqcount_end(&c->usage_lock);
 191         preempt_enable();
 192 }
 193
 194 void bch2_fs_usage_to_text(struct printbuf *out,
 195                            struct bch_fs *c,
 196                            struct bch_fs_usage_online *fs_usage)
 197 {
 198         unsigned i;
 199
 200         pr_buf(out, "capacity:\t\t\t%llu\n", c->capacity);
 201
 202         pr_buf(out, "hidden:\t\t\t\t%llu\n",
 203                fs_usage->u.hidden);
 204         pr_buf(out, "data:\t\t\t\t%llu\n",
 205                fs_usage->u.data);
 206         pr_buf(out, "cached:\t\t\t\t%llu\n",
 207                fs_usage->u.cached);
 208         pr_buf(out, "reserved:\t\t\t%llu\n",
 209                fs_usage->u.reserved);
 210         pr_buf(out, "nr_inodes:\t\t\t%llu\n",
 211                fs_usage->u.nr_inodes);
 212         pr_buf(out, "online reserved:\t\t%llu\n",
 213                fs_usage->online_reserved);
 214
 215         for (i = 0;
 216              i < ARRAY_SIZE(fs_usage->u.persistent_reserved);
 217              i++) {
 218                 pr_buf(out, "%u replicas:\n", i + 1);
 219                 pr_buf(out, "\treserved:\t\t%llu\n",
 220                        fs_usage->u.persistent_reserved[i]);
 221         }
 222
 223         for (i = 0; i < c->replicas.nr; i++) {
 224                 struct bch_replicas_entry *e =
 225                         cpu_replicas_entry(&c->replicas, i);
 226
 227                 pr_buf(out, "\t");
 228                 bch2_replicas_entry_to_text(out, e);
 229                 pr_buf(out, ":\t%llu\n", fs_usage->u.replicas[i]);
 230         }
 231 }
 232
 233 static u64 reserve_factor(u64 r)
 234 {
 235         return r + (round_up(r, (1 << RESERVE_FACTOR)) >> RESERVE_FACTOR);
 236 }
 237
 238 u64 bch2_fs_sectors_used(struct bch_fs *c, struct bch_fs_usage_online *fs_usage)
 239 {
 240         return min(fs_usage->u.hidden +
 241                    fs_usage->u.btree +
 242                    fs_usage->u.data +
 243                    reserve_factor(fs_usage->u.reserved +
 244                                   fs_usage->online_reserved),
 245                    c->capacity);
 246 }
 247
 248 static struct bch_fs_usage_short
 249 __bch2_fs_usage_read_short(struct bch_fs *c)
 250 {
 251         struct bch_fs_usage_short ret;
 252         u64 data, reserved;
 253
 254         ret.capacity = c->capacity -
 255                 bch2_fs_usage_read_one(c, &c->usage_base->hidden);
 256
 257         data            = bch2_fs_usage_read_one(c, &c->usage_base->data) +
 258                 bch2_fs_usage_read_one(c, &c->usage_base->btree);
 259         reserved        = bch2_fs_usage_read_one(c, &c->usage_base->reserved) +
 260                 percpu_u64_get(c->online_reserved);
 261
 262         ret.used        = min(ret.capacity, data + reserve_factor(reserved));
 263         ret.free        = ret.capacity - ret.used;
 264
 265         ret.nr_inodes   = bch2_fs_usage_read_one(c, &c->usage_base->nr_inodes);
 266
 267         return ret;
 268 }
 269
 270 struct bch_fs_usage_short
 271 bch2_fs_usage_read_short(struct bch_fs *c)
 272 {
 273         struct bch_fs_usage_short ret;
 274
 275         percpu_down_read(&c->mark_lock);
 276         ret = __bch2_fs_usage_read_short(c);
 277         percpu_up_read(&c->mark_lock);
 278
 279         return ret;
 280 }
 281
 282 void bch2_dev_usage_init(struct bch_dev *ca)
 283 {
 284         ca->usage_base->d[BCH_DATA_free].buckets = ca->mi.nbuckets - ca->mi.first_bucket;
 285 }
 286
 287 static inline int bucket_sectors_fragmented(struct bch_dev *ca,
 288                                             struct bch_alloc_v4 a)
 289 {
 290         return a.dirty_sectors
 291                 ? max(0, (int) ca->mi.bucket_size - (int) a.dirty_sectors)
 292                 : 0;
 293 }
 294
 295 static void bch2_dev_usage_update(struct bch_fs *c, struct bch_dev *ca,
 296                                   struct bch_alloc_v4 old,
 297                                   struct bch_alloc_v4 new,
 298                                   u64 journal_seq, bool gc)
 299 {
 300         struct bch_fs_usage *fs_usage;
 301         struct bch_dev_usage *u;
 302
 303         preempt_disable();
 304         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 305
 306         if (data_type_is_hidden(old.data_type))
 307                 fs_usage->hidden -= ca->mi.bucket_size;
 308         if (data_type_is_hidden(new.data_type))
 309                 fs_usage->hidden += ca->mi.bucket_size;
 310
 311         u = dev_usage_ptr(ca, journal_seq, gc);
 312
 313         u->d[old.data_type].buckets--;
 314         u->d[new.data_type].buckets++;
 315
 316         u->buckets_ec -= (int) !!old.stripe;
 317         u->buckets_ec += (int) !!new.stripe;
 318
 319         u->d[old.data_type].sectors -= old.dirty_sectors;
 320         u->d[new.data_type].sectors += new.dirty_sectors;
 321
 322         u->d[BCH_DATA_cached].sectors += new.cached_sectors;
 323         u->d[BCH_DATA_cached].sectors -= old.cached_sectors;
 324
 325         u->d[old.data_type].fragmented -= bucket_sectors_fragmented(ca, old);
 326         u->d[new.data_type].fragmented += bucket_sectors_fragmented(ca, new);
 327
 328         preempt_enable();
 329 }
 330
 331 static void bch2_dev_usage_update_m(struct bch_fs *c, struct bch_dev *ca,
 332                                     struct bucket old, struct bucket new,
 333                                     u64 journal_seq, bool gc)
 334 {
 335         struct bch_alloc_v4 old_a = {
 336                 .gen            = old.gen,
 337                 .data_type      = old.data_type,
 338                 .dirty_sectors  = old.dirty_sectors,
 339                 .cached_sectors = old.cached_sectors,
 340                 .stripe         = old.stripe,
 341         };
 342         struct bch_alloc_v4 new_a = {
 343                 .gen            = new.gen,
 344                 .data_type      = new.data_type,
 345                 .dirty_sectors  = new.dirty_sectors,
 346                 .cached_sectors = new.cached_sectors,
 347                 .stripe         = new.stripe,
 348         };
 349
 350         bch2_dev_usage_update(c, ca, old_a, new_a, journal_seq, gc);
 351 }
 352
 353 static inline int __update_replicas(struct bch_fs *c,
 354                                     struct bch_fs_usage *fs_usage,
 355                                     struct bch_replicas_entry *r,
 356                                     s64 sectors)
 357 {
 358         int idx = bch2_replicas_entry_idx(c, r);
 359
 360         if (idx < 0)
 361                 return -1;
 362
 363         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 364         fs_usage->replicas[idx]         += sectors;
 365         return 0;
 366 }
 367
 368 static inline int update_replicas(struct bch_fs *c, struct bkey_s_c k,
 369                         struct bch_replicas_entry *r, s64 sectors,
 370                         unsigned journal_seq, bool gc)
 371 {
 372         struct bch_fs_usage __percpu *fs_usage;
 373         int idx, ret = 0;
 374         struct printbuf buf = PRINTBUF;
 375
 376         percpu_down_read(&c->mark_lock);
 377         buf.atomic++;
 378
 379         idx = bch2_replicas_entry_idx(c, r);
 380         if (idx < 0 &&
 381             (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
 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(&n->r, r, replicas_entry_bytes(r));
 471         bch2_replicas_entry_sort(&n->r);
 472         d->used += b;
 473 }
 474
 475 static inline void update_cached_sectors_list(struct btree_trans *trans,
 476                                               unsigned dev, s64 sectors)
 477 {
 478         struct bch_replicas_padded r;
 479
 480         bch2_replicas_entry_cached(&r.e, dev);
 481
 482         update_replicas_list(trans, &r.e, sectors);
 483 }
 484
 485 int bch2_mark_alloc(struct btree_trans *trans,
 486                     struct bkey_s_c old, struct bkey_s_c new,
 487                     unsigned flags)
 488 {
 489         bool gc = flags & BTREE_TRIGGER_GC;
 490         u64 journal_seq = trans->journal_res.seq;
 491         struct bch_fs *c = trans->c;
 492         struct bch_alloc_v4 old_a, new_a;
 493         struct bch_dev *ca;
 494         int ret = 0;
 495
 496         /*
 497          * alloc btree is read in by bch2_alloc_read, not gc:
 498          */
 499         if ((flags & BTREE_TRIGGER_GC) &&
 500             !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
 501                 return 0;
 502
 503         if (bch2_trans_inconsistent_on(!bch2_dev_bucket_exists(c, new.k->p), trans,
 504                                        "alloc key for invalid device or bucket"))
 505                 return -EIO;
 506
 507         ca = bch_dev_bkey_exists(c, new.k->p.inode);
 508
 509         bch2_alloc_to_v4(old, &old_a);
 510         bch2_alloc_to_v4(new, &new_a);
 511
 512         if ((flags & BTREE_TRIGGER_INSERT) &&
 513             data_type_is_empty(old_a.data_type) !=
 514             data_type_is_empty(new_a.data_type) &&
 515             new.k->type == KEY_TYPE_alloc_v4) {
 516                 struct bch_alloc_v4 *v = (struct bch_alloc_v4 *) new.v;
 517
 518                 BUG_ON(!journal_seq);
 519
 520                 /*
 521                  * If the btree updates referring to a bucket weren't flushed
 522                  * before the bucket became empty again, then the we don't have
 523                  * to wait on a journal flush before we can reuse the bucket:
 524                  */
 525                 new_a.journal_seq = data_type_is_empty(new_a.data_type) &&
 526                         (journal_seq == v->journal_seq ||
 527                          bch2_journal_noflush_seq(&c->journal, v->journal_seq))
 528                         ? 0 : journal_seq;
 529                 v->journal_seq = new_a.journal_seq;
 530         }
 531
 532         if (!data_type_is_empty(old_a.data_type) &&
 533             data_type_is_empty(new_a.data_type) &&
 534             new_a.journal_seq) {
 535                 ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 536                                 c->journal.flushed_seq_ondisk,
 537                                 new.k->p.inode, new.k->p.offset,
 538                                 new_a.journal_seq);
 539                 if (ret) {
 540                         bch2_fs_fatal_error(c,
 541                                 "error setting bucket_needs_journal_commit: %i", ret);
 542                         return ret;
 543                 }
 544         }
 545
 546         if (new_a.data_type == BCH_DATA_free &&
 547             (!new_a.journal_seq || new_a.journal_seq < c->journal.flushed_seq_ondisk))
 548                 closure_wake_up(&c->freelist_wait);
 549
 550         if (new_a.data_type == BCH_DATA_need_discard &&
 551             (!new_a.journal_seq || new_a.journal_seq < c->journal.flushed_seq_ondisk))
 552                 bch2_do_discards(c);
 553
 554         if (old_a.data_type != BCH_DATA_cached &&
 555             new_a.data_type == BCH_DATA_cached &&
 556             should_invalidate_buckets(ca, bch2_dev_usage_read(ca)))
 557                 bch2_do_invalidates(c);
 558
 559         if (new_a.data_type == BCH_DATA_need_gc_gens)
 560                 bch2_do_gc_gens(c);
 561
 562         percpu_down_read(&c->mark_lock);
 563         if (!gc && new_a.gen != old_a.gen)
 564                 *bucket_gen(ca, new.k->p.offset) = new_a.gen;
 565
 566         bch2_dev_usage_update(c, ca, old_a, new_a, journal_seq, gc);
 567
 568         if (gc) {
 569                 struct bucket *g = gc_bucket(ca, new.k->p.offset);
 570
 571                 bucket_lock(g);
 572
 573                 g->gen_valid            = 1;
 574                 g->gen                  = new_a.gen;
 575                 g->data_type            = new_a.data_type;
 576                 g->stripe               = new_a.stripe;
 577                 g->stripe_redundancy    = new_a.stripe_redundancy;
 578                 g->dirty_sectors        = new_a.dirty_sectors;
 579                 g->cached_sectors       = new_a.cached_sectors;
 580
 581                 bucket_unlock(g);
 582         }
 583         percpu_up_read(&c->mark_lock);
 584
 585         /*
 586          * need to know if we're getting called from the invalidate path or
 587          * not:
 588          */
 589
 590         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 591             old_a.cached_sectors) {
 592                 ret = update_cached_sectors(c, new, ca->dev_idx,
 593                                             -old_a.cached_sectors,
 594                                             journal_seq, gc);
 595                 if (ret) {
 596                         bch2_fs_fatal_error(c, "bch2_mark_alloc(): no replicas entry while updating cached sectors");
 597                         return ret;
 598                 }
 599
 600                 trace_invalidate(ca, bucket_to_sector(ca, new.k->p.offset),
 601                                  old_a.cached_sectors);
 602         }
 603
 604         return 0;
 605 }
 606
 607 int bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 608                               size_t b, enum bch_data_type data_type,
 609                               unsigned sectors, struct gc_pos pos,
 610                               unsigned flags)
 611 {
 612         struct bucket old, new, *g;
 613         int ret = 0;
 614
 615         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 616         BUG_ON(data_type != BCH_DATA_sb &&
 617                data_type != BCH_DATA_journal);
 618
 619         /*
 620          * Backup superblock might be past the end of our normal usable space:
 621          */
 622         if (b >= ca->mi.nbuckets)
 623                 return 0;
 624
 625         percpu_down_read(&c->mark_lock);
 626         g = gc_bucket(ca, b);
 627
 628         bucket_lock(g);
 629         old = *g;
 630
 631         if (bch2_fs_inconsistent_on(g->data_type &&
 632                         g->data_type != data_type, c,
 633                         "different types of data in same bucket: %s, %s",
 634                         bch2_data_types[g->data_type],
 635                         bch2_data_types[data_type])) {
 636                 ret = -EIO;
 637                 goto err;
 638         }
 639
 640         if (bch2_fs_inconsistent_on((u64) g->dirty_sectors + sectors > ca->mi.bucket_size, c,
 641                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > bucket size",
 642                         ca->dev_idx, b, g->gen,
 643                         bch2_data_types[g->data_type ?: data_type],
 644                         g->dirty_sectors, sectors)) {
 645                 ret = -EIO;
 646                 goto err;
 647         }
 648
 649
 650         g->data_type = data_type;
 651         g->dirty_sectors += sectors;
 652         new = *g;
 653 err:
 654         bucket_unlock(g);
 655         if (!ret)
 656                 bch2_dev_usage_update_m(c, ca, old, new, 0, true);
 657         percpu_up_read(&c->mark_lock);
 658         return ret;
 659 }
 660
 661 static s64 ptr_disk_sectors(s64 sectors, struct extent_ptr_decoded p)
 662 {
 663         EBUG_ON(sectors < 0);
 664
 665         return crc_is_compressed(p.crc)
 666                 ? DIV_ROUND_UP_ULL(sectors * p.crc.compressed_size,
 667                                    p.crc.uncompressed_size)
 668                 : sectors;
 669 }
 670
 671 static int check_bucket_ref(struct bch_fs *c,
 672                             struct bkey_s_c k,
 673                             const struct bch_extent_ptr *ptr,
 674                             s64 sectors, enum bch_data_type ptr_data_type,
 675                             u8 b_gen, u8 bucket_data_type,
 676                             u32 dirty_sectors, u32 cached_sectors)
 677 {
 678         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 679         size_t bucket_nr = PTR_BUCKET_NR(ca, ptr);
 680         u16 bucket_sectors = !ptr->cached
 681                 ? dirty_sectors
 682                 : cached_sectors;
 683         struct printbuf buf = PRINTBUF;
 684         int ret = 0;
 685
 686         if (bucket_data_type == BCH_DATA_cached)
 687                 bucket_data_type = BCH_DATA_user;
 688
 689         if (gen_after(ptr->gen, b_gen)) {
 690                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 691                         "bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\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                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 697                 ret = -EIO;
 698                 goto err;
 699         }
 700
 701         if (gen_cmp(b_gen, ptr->gen) > BUCKET_GC_GEN_MAX) {
 702                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 703                         "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 704                         "while marking %s",
 705                         ptr->dev, bucket_nr, b_gen,
 706                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 707                         ptr->gen,
 708                         (printbuf_reset(&buf),
 709                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 710                 ret = -EIO;
 711                 goto err;
 712         }
 713
 714         if (b_gen != ptr->gen && !ptr->cached) {
 715                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 716                         "bucket %u:%zu gen %u (mem gen %u) data type %s: stale dirty ptr (gen %u)\n"
 717                         "while marking %s",
 718                         ptr->dev, bucket_nr, b_gen,
 719                         *bucket_gen(ca, bucket_nr),
 720                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 721                         ptr->gen,
 722                         (printbuf_reset(&buf),
 723                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 724                 ret = -EIO;
 725                 goto err;
 726         }
 727
 728         if (b_gen != ptr->gen) {
 729                 ret = 1;
 730                 goto err;
 731         }
 732
 733         if (!data_type_is_empty(bucket_data_type) &&
 734             ptr_data_type &&
 735             bucket_data_type != ptr_data_type) {
 736                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 737                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 738                         "while marking %s",
 739                         ptr->dev, bucket_nr, b_gen,
 740                         bch2_data_types[bucket_data_type],
 741                         bch2_data_types[ptr_data_type],
 742                         (printbuf_reset(&buf),
 743                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 744                 ret = -EIO;
 745                 goto err;
 746         }
 747
 748         if ((unsigned) (bucket_sectors + sectors) > U32_MAX) {
 749                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 750                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U16_MAX\n"
 751                         "while marking %s",
 752                         ptr->dev, bucket_nr, b_gen,
 753                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 754                         bucket_sectors, sectors,
 755                         (printbuf_reset(&buf),
 756                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 757                 ret = -EIO;
 758                 goto err;
 759         }
 760 err:
 761         printbuf_exit(&buf);
 762         return ret;
 763 }
 764
 765 static int mark_stripe_bucket(struct btree_trans *trans,
 766                               struct bkey_s_c k,
 767                               unsigned ptr_idx,
 768                               unsigned flags)
 769 {
 770         struct bch_fs *c = trans->c;
 771         u64 journal_seq = trans->journal_res.seq;
 772         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 773         unsigned nr_data = s->nr_blocks - s->nr_redundant;
 774         bool parity = ptr_idx >= nr_data;
 775         enum bch_data_type data_type = parity ? BCH_DATA_parity : 0;
 776         s64 sectors = parity ? le16_to_cpu(s->sectors) : 0;
 777         const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
 778         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 779         struct bucket old, new, *g;
 780         struct printbuf buf = PRINTBUF;
 781         int ret = 0;
 782
 783         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 784
 785         /* * XXX doesn't handle deletion */
 786
 787         percpu_down_read(&c->mark_lock);
 788         buf.atomic++;
 789         g = PTR_GC_BUCKET(ca, ptr);
 790
 791         if (g->dirty_sectors ||
 792             (g->stripe && g->stripe != k.k->p.offset)) {
 793                 bch2_fs_inconsistent(c,
 794                               "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
 795                               ptr->dev, PTR_BUCKET_NR(ca, ptr), g->gen,
 796                               (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 797                 ret = -EINVAL;
 798                 goto err;
 799         }
 800
 801         bucket_lock(g);
 802         old = *g;
 803
 804         ret = check_bucket_ref(c, k, ptr, sectors, data_type,
 805                                g->gen, g->data_type,
 806                                g->dirty_sectors, g->cached_sectors);
 807         if (ret)
 808                 goto err;
 809
 810         if (data_type)
 811                 g->data_type = data_type;
 812         g->dirty_sectors += sectors;
 813
 814         g->stripe               = k.k->p.offset;
 815         g->stripe_redundancy    = s->nr_redundant;
 816         new = *g;
 817 err:
 818         bucket_unlock(g);
 819         if (!ret)
 820                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 821         percpu_up_read(&c->mark_lock);
 822         printbuf_exit(&buf);
 823         return ret;
 824 }
 825
 826 static int __mark_pointer(struct btree_trans *trans,
 827                           struct bkey_s_c k,
 828                           const struct bch_extent_ptr *ptr,
 829                           s64 sectors, enum bch_data_type ptr_data_type,
 830                           u8 bucket_gen, u8 *bucket_data_type,
 831                           u32 *dirty_sectors, u32 *cached_sectors)
 832 {
 833         u32 *dst_sectors = !ptr->cached
 834                 ? dirty_sectors
 835                 : cached_sectors;
 836         int ret = check_bucket_ref(trans->c, k, ptr, sectors, ptr_data_type,
 837                                    bucket_gen, *bucket_data_type,
 838                                    *dirty_sectors, *cached_sectors);
 839
 840         if (ret)
 841                 return ret;
 842
 843         *dst_sectors += sectors;
 844         *bucket_data_type = *dirty_sectors || *cached_sectors
 845                 ? ptr_data_type : 0;
 846         return 0;
 847 }
 848
 849 static int bch2_mark_pointer(struct btree_trans *trans,
 850                              struct bkey_s_c k,
 851                              struct extent_ptr_decoded p,
 852                              s64 sectors, enum bch_data_type data_type,
 853                              unsigned flags)
 854 {
 855         u64 journal_seq = trans->journal_res.seq;
 856         struct bch_fs *c = trans->c;
 857         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 858         struct bucket old, new, *g;
 859         u8 bucket_data_type;
 860         int ret = 0;
 861
 862         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 863
 864         percpu_down_read(&c->mark_lock);
 865         g = PTR_GC_BUCKET(ca, &p.ptr);
 866         bucket_lock(g);
 867         old = *g;
 868
 869         bucket_data_type = g->data_type;
 870         ret = __mark_pointer(trans, k, &p.ptr, sectors,
 871                              data_type, g->gen,
 872                              &bucket_data_type,
 873                              &g->dirty_sectors,
 874                              &g->cached_sectors);
 875         if (!ret)
 876                 g->data_type = bucket_data_type;
 877
 878         new = *g;
 879         bucket_unlock(g);
 880         if (!ret)
 881                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 882         percpu_up_read(&c->mark_lock);
 883
 884         return ret;
 885 }
 886
 887 static int bch2_mark_stripe_ptr(struct btree_trans *trans,
 888                                 struct bkey_s_c k,
 889                                 struct bch_extent_stripe_ptr p,
 890                                 enum bch_data_type data_type,
 891                                 s64 sectors,
 892                                 unsigned flags)
 893 {
 894         struct bch_fs *c = trans->c;
 895         struct bch_replicas_padded r;
 896         struct gc_stripe *m;
 897
 898         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 899
 900         m = genradix_ptr_alloc(&c->gc_stripes, p.idx, GFP_KERNEL);
 901         if (!m) {
 902                 bch_err(c, "error allocating memory for gc_stripes, idx %llu",
 903                         (u64) p.idx);
 904                 return -ENOMEM;
 905         }
 906
 907         spin_lock(&c->ec_stripes_heap_lock);
 908
 909         if (!m || !m->alive) {
 910                 spin_unlock(&c->ec_stripes_heap_lock);
 911                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
 912                                     (u64) p.idx);
 913                 bch2_inconsistent_error(c);
 914                 return -EIO;
 915         }
 916
 917         m->block_sectors[p.block] += sectors;
 918
 919         r = m->r;
 920         spin_unlock(&c->ec_stripes_heap_lock);
 921
 922         r.e.data_type = data_type;
 923         update_replicas(c, k, &r.e, sectors, trans->journal_res.seq, true);
 924
 925         return 0;
 926 }
 927
 928 int bch2_mark_extent(struct btree_trans *trans,
 929                      struct bkey_s_c old, struct bkey_s_c new,
 930                      unsigned flags)
 931 {
 932         u64 journal_seq = trans->journal_res.seq;
 933         struct bch_fs *c = trans->c;
 934         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
 935         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 936         const union bch_extent_entry *entry;
 937         struct extent_ptr_decoded p;
 938         struct bch_replicas_padded r;
 939         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
 940                 ? BCH_DATA_btree
 941                 : BCH_DATA_user;
 942         s64 sectors = bkey_is_btree_ptr(k.k)
 943                 ? btree_sectors(c)
 944                 : k.k->size;
 945         s64 dirty_sectors = 0;
 946         bool stale;
 947         int ret;
 948
 949         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 950
 951         r.e.data_type   = data_type;
 952         r.e.nr_devs     = 0;
 953         r.e.nr_required = 1;
 954
 955         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 956                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
 957
 958                 if (flags & BTREE_TRIGGER_OVERWRITE)
 959                         disk_sectors = -disk_sectors;
 960
 961                 ret = bch2_mark_pointer(trans, k, p, disk_sectors,
 962                                         data_type, flags);
 963                 if (ret < 0)
 964                         return ret;
 965
 966                 stale = ret > 0;
 967
 968                 if (p.ptr.cached) {
 969                         if (!stale) {
 970                                 ret = update_cached_sectors(c, k, p.ptr.dev,
 971                                                 disk_sectors, journal_seq, true);
 972                                 if (ret) {
 973                                         bch2_fs_fatal_error(c, "bch2_mark_extent(): no replicas entry while updating cached sectors");
 974                                         return ret;
 975                                 }
 976                         }
 977                 } else if (!p.has_ec) {
 978                         dirty_sectors          += disk_sectors;
 979                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
 980                 } else {
 981                         ret = bch2_mark_stripe_ptr(trans, k, p.ec, data_type,
 982                                         disk_sectors, flags);
 983                         if (ret)
 984                                 return ret;
 985
 986                         /*
 987                          * There may be other dirty pointers in this extent, but
 988                          * if so they're not required for mounting if we have an
 989                          * erasure coded pointer in this extent:
 990                          */
 991                         r.e.nr_required = 0;
 992                 }
 993         }
 994
 995         if (r.e.nr_devs) {
 996                 ret = update_replicas(c, k, &r.e, dirty_sectors, journal_seq, true);
 997                 if (ret) {
 998                         struct printbuf buf = PRINTBUF;
 999
1000                         bch2_bkey_val_to_text(&buf, c, k);
1001                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
1002                         printbuf_exit(&buf);
1003                         return ret;
1004                 }
1005         }
1006
1007         return 0;
1008 }
1009
1010 int bch2_mark_stripe(struct btree_trans *trans,
1011                      struct bkey_s_c old, struct bkey_s_c new,
1012                      unsigned flags)
1013 {
1014         bool gc = flags & BTREE_TRIGGER_GC;
1015         u64 journal_seq = trans->journal_res.seq;
1016         struct bch_fs *c = trans->c;
1017         u64 idx = new.k->p.offset;
1018         const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
1019                 ? bkey_s_c_to_stripe(old).v : NULL;
1020         const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
1021                 ? bkey_s_c_to_stripe(new).v : NULL;
1022         unsigned i;
1023         int ret;
1024
1025         BUG_ON(gc && old_s);
1026
1027         if (!gc) {
1028                 struct stripe *m = genradix_ptr(&c->stripes, idx);
1029
1030                 if (!m || (old_s && !m->alive)) {
1031                         struct printbuf buf1 = PRINTBUF;
1032                         struct printbuf buf2 = PRINTBUF;
1033
1034                         bch2_bkey_val_to_text(&buf1, c, old);
1035                         bch2_bkey_val_to_text(&buf2, c, new);
1036                         bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
1037                                             "old %s\n"
1038                                             "new %s", idx, buf1.buf, buf2.buf);
1039                         printbuf_exit(&buf2);
1040                         printbuf_exit(&buf1);
1041                         bch2_inconsistent_error(c);
1042                         return -1;
1043                 }
1044
1045                 if (!new_s) {
1046                         spin_lock(&c->ec_stripes_heap_lock);
1047                         bch2_stripes_heap_del(c, m, idx);
1048                         spin_unlock(&c->ec_stripes_heap_lock);
1049
1050                         memset(m, 0, sizeof(*m));
1051                 } else {
1052                         m->alive        = true;
1053                         m->sectors      = le16_to_cpu(new_s->sectors);
1054                         m->algorithm    = new_s->algorithm;
1055                         m->nr_blocks    = new_s->nr_blocks;
1056                         m->nr_redundant = new_s->nr_redundant;
1057                         m->blocks_nonempty = 0;
1058
1059                         for (i = 0; i < new_s->nr_blocks; i++)
1060                                 m->blocks_nonempty += !!stripe_blockcount_get(new_s, i);
1061
1062                         spin_lock(&c->ec_stripes_heap_lock);
1063                         bch2_stripes_heap_update(c, m, idx);
1064                         spin_unlock(&c->ec_stripes_heap_lock);
1065                 }
1066         } else {
1067                 struct gc_stripe *m =
1068                         genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
1069
1070                 if (!m) {
1071                         bch_err(c, "error allocating memory for gc_stripes, idx %llu",
1072                                 idx);
1073                         return -ENOMEM;
1074                 }
1075                 /*
1076                  * This will be wrong when we bring back runtime gc: we should
1077                  * be unmarking the old key and then marking the new key
1078                  */
1079                 m->alive        = true;
1080                 m->sectors      = le16_to_cpu(new_s->sectors);
1081                 m->nr_blocks    = new_s->nr_blocks;
1082                 m->nr_redundant = new_s->nr_redundant;
1083
1084                 for (i = 0; i < new_s->nr_blocks; i++)
1085                         m->ptrs[i] = new_s->ptrs[i];
1086
1087                 bch2_bkey_to_replicas(&m->r.e, new);
1088
1089                 /*
1090                  * gc recalculates this field from stripe ptr
1091                  * references:
1092                  */
1093                 memset(m->block_sectors, 0, sizeof(m->block_sectors));
1094
1095                 for (i = 0; i < new_s->nr_blocks; i++) {
1096                         ret = mark_stripe_bucket(trans, new, i, flags);
1097                         if (ret)
1098                                 return ret;
1099                 }
1100
1101                 ret = update_replicas(c, new, &m->r.e,
1102                                       ((s64) m->sectors * m->nr_redundant),
1103                                       journal_seq, gc);
1104                 if (ret) {
1105                         struct printbuf buf = PRINTBUF;
1106
1107                         bch2_bkey_val_to_text(&buf, c, new);
1108                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
1109                         printbuf_exit(&buf);
1110                         return ret;
1111                 }
1112         }
1113
1114         return 0;
1115 }
1116
1117 int bch2_mark_inode(struct btree_trans *trans,
1118                     struct bkey_s_c old, struct bkey_s_c new,
1119                     unsigned flags)
1120 {
1121         struct bch_fs *c = trans->c;
1122         struct bch_fs_usage __percpu *fs_usage;
1123         u64 journal_seq = trans->journal_res.seq;
1124
1125         if (flags & BTREE_TRIGGER_INSERT) {
1126                 struct bch_inode_v2 *v = (struct bch_inode_v2 *) new.v;
1127
1128                 BUG_ON(!journal_seq);
1129                 BUG_ON(new.k->type != KEY_TYPE_inode_v2);
1130
1131                 v->bi_journal_seq = cpu_to_le64(journal_seq);
1132         }
1133
1134         if (flags & BTREE_TRIGGER_GC) {
1135                 percpu_down_read(&c->mark_lock);
1136                 preempt_disable();
1137
1138                 fs_usage = fs_usage_ptr(c, journal_seq, flags & BTREE_TRIGGER_GC);
1139                 fs_usage->nr_inodes += bkey_is_inode(new.k);
1140                 fs_usage->nr_inodes -= bkey_is_inode(old.k);
1141
1142                 preempt_enable();
1143                 percpu_up_read(&c->mark_lock);
1144         }
1145         return 0;
1146 }
1147
1148 int bch2_mark_reservation(struct btree_trans *trans,
1149                           struct bkey_s_c old, struct bkey_s_c new,
1150                           unsigned flags)
1151 {
1152         struct bch_fs *c = trans->c;
1153         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1154         struct bch_fs_usage __percpu *fs_usage;
1155         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1156         s64 sectors = (s64) k.k->size;
1157
1158         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1159
1160         if (flags & BTREE_TRIGGER_OVERWRITE)
1161                 sectors = -sectors;
1162         sectors *= replicas;
1163
1164         percpu_down_read(&c->mark_lock);
1165         preempt_disable();
1166
1167         fs_usage = fs_usage_ptr(c, trans->journal_res.seq, flags & BTREE_TRIGGER_GC);
1168         replicas = clamp_t(unsigned, replicas, 1,
1169                            ARRAY_SIZE(fs_usage->persistent_reserved));
1170
1171         fs_usage->reserved                              += sectors;
1172         fs_usage->persistent_reserved[replicas - 1]     += sectors;
1173
1174         preempt_enable();
1175         percpu_up_read(&c->mark_lock);
1176
1177         return 0;
1178 }
1179
1180 static s64 __bch2_mark_reflink_p(struct btree_trans *trans,
1181                                  struct bkey_s_c_reflink_p p,
1182                                  u64 start, u64 end,
1183                                  u64 *idx, unsigned flags, size_t r_idx)
1184 {
1185         struct bch_fs *c = trans->c;
1186         struct reflink_gc *r;
1187         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1188         u64 next_idx = end;
1189         s64 ret = 0;
1190         struct printbuf buf = PRINTBUF;
1191
1192         if (r_idx >= c->reflink_gc_nr)
1193                 goto not_found;
1194
1195         r = genradix_ptr(&c->reflink_gc_table, r_idx);
1196         next_idx = min(next_idx, r->offset - r->size);
1197         if (*idx < next_idx)
1198                 goto not_found;
1199
1200         BUG_ON((s64) r->refcount + add < 0);
1201
1202         r->refcount += add;
1203         *idx = r->offset;
1204         return 0;
1205 not_found:
1206         if (fsck_err(c, "pointer to missing indirect extent\n"
1207                      "  %s\n"
1208                      "  missing range %llu-%llu",
1209                      (bch2_bkey_val_to_text(&buf, c, p.s_c), buf.buf),
1210                      *idx, next_idx)) {
1211                 struct bkey_i_error new;
1212
1213                 bkey_init(&new.k);
1214                 new.k.type      = KEY_TYPE_error;
1215                 new.k.p         = bkey_start_pos(p.k);
1216                 new.k.p.offset += *idx - start;
1217                 bch2_key_resize(&new.k, next_idx - *idx);
1218                 ret = __bch2_btree_insert(trans, BTREE_ID_extents, &new.k_i);
1219         }
1220
1221         *idx = next_idx;
1222 fsck_err:
1223         printbuf_exit(&buf);
1224         return ret;
1225 }
1226
1227 int bch2_mark_reflink_p(struct btree_trans *trans,
1228                         struct bkey_s_c old, struct bkey_s_c new,
1229                         unsigned flags)
1230 {
1231         struct bch_fs *c = trans->c;
1232         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1233         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1234         struct reflink_gc *ref;
1235         size_t l, r, m;
1236         u64 idx = le64_to_cpu(p.v->idx), start = idx;
1237         u64 end = le64_to_cpu(p.v->idx) + p.k->size;
1238         int ret = 0;
1239
1240         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1241
1242         if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix) {
1243                 idx -= le32_to_cpu(p.v->front_pad);
1244                 end += le32_to_cpu(p.v->back_pad);
1245         }
1246
1247         l = 0;
1248         r = c->reflink_gc_nr;
1249         while (l < r) {
1250                 m = l + (r - l) / 2;
1251
1252                 ref = genradix_ptr(&c->reflink_gc_table, m);
1253                 if (ref->offset <= idx)
1254                         l = m + 1;
1255                 else
1256                         r = m;
1257         }
1258
1259         while (idx < end && !ret)
1260                 ret = __bch2_mark_reflink_p(trans, p, start, end,
1261                                             &idx, flags, l++);
1262
1263         return ret;
1264 }
1265
1266 static noinline __cold
1267 void fs_usage_apply_warn(struct btree_trans *trans,
1268                          unsigned disk_res_sectors,
1269                          s64 should_not_have_added)
1270 {
1271         struct bch_fs *c = trans->c;
1272         struct btree_insert_entry *i;
1273         struct printbuf buf = PRINTBUF;
1274
1275         bch_err(c, "disk usage increased %lli more than %u sectors reserved",
1276                 should_not_have_added, disk_res_sectors);
1277
1278         trans_for_each_update(trans, i) {
1279                 struct bkey_s_c old = { &i->old_k, i->old_v };
1280
1281                 pr_err("while inserting");
1282                 printbuf_reset(&buf);
1283                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(i->k));
1284                 pr_err("  %s", buf.buf);
1285                 pr_err("overlapping with");
1286                 printbuf_reset(&buf);
1287                 bch2_bkey_val_to_text(&buf, c, old);
1288                 pr_err("  %s", buf.buf);
1289         }
1290
1291         __WARN();
1292         printbuf_exit(&buf);
1293 }
1294
1295 int bch2_trans_fs_usage_apply(struct btree_trans *trans,
1296                               struct replicas_delta_list *deltas)
1297 {
1298         struct bch_fs *c = trans->c;
1299         static int warned_disk_usage = 0;
1300         bool warn = false;
1301         unsigned disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1302         struct replicas_delta *d = deltas->d, *d2;
1303         struct replicas_delta *top = (void *) deltas->d + deltas->used;
1304         struct bch_fs_usage *dst;
1305         s64 added = 0, should_not_have_added;
1306         unsigned i;
1307
1308         percpu_down_read(&c->mark_lock);
1309         preempt_disable();
1310         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1311
1312         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1313                 switch (d->r.data_type) {
1314                 case BCH_DATA_btree:
1315                 case BCH_DATA_user:
1316                 case BCH_DATA_parity:
1317                         added += d->delta;
1318                 }
1319
1320                 if (__update_replicas(c, dst, &d->r, d->delta))
1321                         goto need_mark;
1322         }
1323
1324         dst->nr_inodes += deltas->nr_inodes;
1325
1326         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1327                 added                           += deltas->persistent_reserved[i];
1328                 dst->reserved                   += deltas->persistent_reserved[i];
1329                 dst->persistent_reserved[i]     += deltas->persistent_reserved[i];
1330         }
1331
1332         /*
1333          * Not allowed to reduce sectors_available except by getting a
1334          * reservation:
1335          */
1336         should_not_have_added = added - (s64) disk_res_sectors;
1337         if (unlikely(should_not_have_added > 0)) {
1338                 u64 old, new, v = atomic64_read(&c->sectors_available);
1339
1340                 do {
1341                         old = v;
1342                         new = max_t(s64, 0, old - should_not_have_added);
1343                 } while ((v = atomic64_cmpxchg(&c->sectors_available,
1344                                                old, new)) != old);
1345
1346                 added -= should_not_have_added;
1347                 warn = true;
1348         }
1349
1350         if (added > 0) {
1351                 trans->disk_res->sectors -= added;
1352                 this_cpu_sub(*c->online_reserved, added);
1353         }
1354
1355         preempt_enable();
1356         percpu_up_read(&c->mark_lock);
1357
1358         if (unlikely(warn) && !xchg(&warned_disk_usage, 1))
1359                 fs_usage_apply_warn(trans, disk_res_sectors, should_not_have_added);
1360         return 0;
1361 need_mark:
1362         /* revert changes: */
1363         for (d2 = deltas->d; d2 != d; d2 = replicas_delta_next(d2))
1364                 BUG_ON(__update_replicas(c, dst, &d2->r, -d2->delta));
1365
1366         preempt_enable();
1367         percpu_up_read(&c->mark_lock);
1368         return -1;
1369 }
1370
1371 /* trans_mark: */
1372
1373 static int bch2_trans_mark_pointer(struct btree_trans *trans,
1374                         struct bkey_s_c k, struct extent_ptr_decoded p,
1375                         s64 sectors, enum bch_data_type data_type)
1376 {
1377         struct btree_iter iter;
1378         struct bkey_i_alloc_v4 *a;
1379         int ret;
1380
1381         a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(trans->c, &p.ptr));
1382         if (IS_ERR(a))
1383                 return PTR_ERR(a);
1384
1385         ret = __mark_pointer(trans, k, &p.ptr, sectors, data_type,
1386                              a->v.gen, &a->v.data_type,
1387                              &a->v.dirty_sectors, &a->v.cached_sectors) ?:
1388                 bch2_trans_update(trans, &iter, &a->k_i, 0);
1389         bch2_trans_iter_exit(trans, &iter);
1390         return ret;
1391 }
1392
1393 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1394                         struct extent_ptr_decoded p,
1395                         s64 sectors, enum bch_data_type data_type)
1396 {
1397         struct btree_iter iter;
1398         struct bkey_s_c k;
1399         struct bkey_i_stripe *s;
1400         struct bch_replicas_padded r;
1401         int ret = 0;
1402
1403         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes, POS(0, p.ec.idx),
1404                              BTREE_ITER_INTENT|
1405                              BTREE_ITER_WITH_UPDATES);
1406         k = bch2_btree_iter_peek_slot(&iter);
1407         ret = bkey_err(k);
1408         if (ret)
1409                 goto err;
1410
1411         if (k.k->type != KEY_TYPE_stripe) {
1412                 bch2_trans_inconsistent(trans,
1413                         "pointer to nonexistent stripe %llu",
1414                         (u64) p.ec.idx);
1415                 ret = -EIO;
1416                 goto err;
1417         }
1418
1419         if (!bch2_ptr_matches_stripe(bkey_s_c_to_stripe(k).v, p)) {
1420                 bch2_trans_inconsistent(trans,
1421                         "stripe pointer doesn't match stripe %llu",
1422                         (u64) p.ec.idx);
1423                 ret = -EIO;
1424                 goto err;
1425         }
1426
1427         s = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1428         ret = PTR_ERR_OR_ZERO(s);
1429         if (ret)
1430                 goto err;
1431
1432         bkey_reassemble(&s->k_i, k);
1433         stripe_blockcount_set(&s->v, p.ec.block,
1434                 stripe_blockcount_get(&s->v, p.ec.block) +
1435                 sectors);
1436
1437         ret = bch2_trans_update(trans, &iter, &s->k_i, 0);
1438         if (ret)
1439                 goto err;
1440
1441         bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1442         r.e.data_type = data_type;
1443         update_replicas_list(trans, &r.e, sectors);
1444 err:
1445         bch2_trans_iter_exit(trans, &iter);
1446         return ret;
1447 }
1448
1449 int bch2_trans_mark_extent(struct btree_trans *trans,
1450                            struct bkey_s_c old, struct bkey_i *new,
1451                            unsigned flags)
1452 {
1453         struct bch_fs *c = trans->c;
1454         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1455                 ? old
1456                 : bkey_i_to_s_c(new);
1457         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1458         const union bch_extent_entry *entry;
1459         struct extent_ptr_decoded p;
1460         struct bch_replicas_padded r;
1461         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
1462                 ? BCH_DATA_btree
1463                 : BCH_DATA_user;
1464         s64 sectors = bkey_is_btree_ptr(k.k)
1465                 ? btree_sectors(c)
1466                 : k.k->size;
1467         s64 dirty_sectors = 0;
1468         bool stale;
1469         int ret;
1470
1471         r.e.data_type   = data_type;
1472         r.e.nr_devs     = 0;
1473         r.e.nr_required = 1;
1474
1475         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1476                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
1477
1478                 if (flags & BTREE_TRIGGER_OVERWRITE)
1479                         disk_sectors = -disk_sectors;
1480
1481                 ret = bch2_trans_mark_pointer(trans, k, p,
1482                                         disk_sectors, data_type);
1483                 if (ret < 0)
1484                         return ret;
1485
1486                 stale = ret > 0;
1487
1488                 if (p.ptr.cached) {
1489                         if (!stale)
1490                                 update_cached_sectors_list(trans, p.ptr.dev,
1491                                                            disk_sectors);
1492                 } else if (!p.has_ec) {
1493                         dirty_sectors          += disk_sectors;
1494                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1495                 } else {
1496                         ret = bch2_trans_mark_stripe_ptr(trans, p,
1497                                         disk_sectors, data_type);
1498                         if (ret)
1499                                 return ret;
1500
1501                         r.e.nr_required = 0;
1502                 }
1503         }
1504
1505         if (r.e.nr_devs)
1506                 update_replicas_list(trans, &r.e, dirty_sectors);
1507
1508         return 0;
1509 }
1510
1511 static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
1512                                          struct bkey_s_c_stripe s,
1513                                          unsigned idx, bool deleting)
1514 {
1515         struct bch_fs *c = trans->c;
1516         const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
1517         struct btree_iter iter;
1518         struct bkey_i_alloc_v4 *a;
1519         enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
1520                 ? BCH_DATA_parity : 0;
1521         s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
1522         int ret = 0;
1523
1524         if (deleting)
1525                 sectors = -sectors;
1526
1527         a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(c, ptr));
1528         if (IS_ERR(a))
1529                 return PTR_ERR(a);
1530
1531         ret = check_bucket_ref(c, s.s_c, ptr, sectors, data_type,
1532                                a->v.gen, a->v.data_type,
1533                                a->v.dirty_sectors, a->v.cached_sectors);
1534         if (ret)
1535                 goto err;
1536
1537         if (!deleting) {
1538                 if (bch2_trans_inconsistent_on(a->v.stripe ||
1539                                                a->v.stripe_redundancy, trans,
1540                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
1541                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1542                                 bch2_data_types[a->v.data_type],
1543                                 a->v.dirty_sectors,
1544                                 a->v.stripe, s.k->p.offset)) {
1545                         ret = -EIO;
1546                         goto err;
1547                 }
1548
1549                 if (bch2_trans_inconsistent_on(data_type && a->v.dirty_sectors, trans,
1550                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %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                                 s.k->p.offset)) {
1555                         ret = -EIO;
1556                         goto err;
1557                 }
1558
1559                 a->v.stripe             = s.k->p.offset;
1560                 a->v.stripe_redundancy  = s.v->nr_redundant;
1561         } else {
1562                 if (bch2_trans_inconsistent_on(a->v.stripe != s.k->p.offset ||
1563                                                a->v.stripe_redundancy != s.v->nr_redundant, trans,
1564                                 "bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
1565                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1566                                 s.k->p.offset, a->v.stripe)) {
1567                         ret = -EIO;
1568                         goto err;
1569                 }
1570
1571                 a->v.stripe             = 0;
1572                 a->v.stripe_redundancy  = 0;
1573         }
1574
1575         a->v.dirty_sectors += sectors;
1576         if (data_type)
1577                 a->v.data_type = !deleting ? data_type : 0;
1578
1579         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1580         if (ret)
1581                 goto err;
1582 err:
1583         bch2_trans_iter_exit(trans, &iter);
1584         return ret;
1585 }
1586
1587 int bch2_trans_mark_stripe(struct btree_trans *trans,
1588                            struct bkey_s_c old, struct bkey_i *new,
1589                            unsigned flags)
1590 {
1591         const struct bch_stripe *old_s = NULL;
1592         struct bch_stripe *new_s = NULL;
1593         struct bch_replicas_padded r;
1594         unsigned i, nr_blocks;
1595         int ret = 0;
1596
1597         if (old.k->type == KEY_TYPE_stripe)
1598                 old_s = bkey_s_c_to_stripe(old).v;
1599         if (new->k.type == KEY_TYPE_stripe)
1600                 new_s = &bkey_i_to_stripe(new)->v;
1601
1602         /*
1603          * If the pointers aren't changing, we don't need to do anything:
1604          */
1605         if (new_s && old_s &&
1606             new_s->nr_blocks    == old_s->nr_blocks &&
1607             new_s->nr_redundant == old_s->nr_redundant &&
1608             !memcmp(old_s->ptrs, new_s->ptrs,
1609                     new_s->nr_blocks * sizeof(struct bch_extent_ptr)))
1610                 return 0;
1611
1612         BUG_ON(new_s && old_s &&
1613                (new_s->nr_blocks        != old_s->nr_blocks ||
1614                 new_s->nr_redundant     != old_s->nr_redundant));
1615
1616         nr_blocks = new_s ? new_s->nr_blocks : old_s->nr_blocks;
1617
1618         if (new_s) {
1619                 s64 sectors = le16_to_cpu(new_s->sectors);
1620
1621                 bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(new));
1622                 update_replicas_list(trans, &r.e, sectors * new_s->nr_redundant);
1623         }
1624
1625         if (old_s) {
1626                 s64 sectors = -((s64) le16_to_cpu(old_s->sectors));
1627
1628                 bch2_bkey_to_replicas(&r.e, old);
1629                 update_replicas_list(trans, &r.e, sectors * old_s->nr_redundant);
1630         }
1631
1632         for (i = 0; i < nr_blocks; i++) {
1633                 if (new_s && old_s &&
1634                     !memcmp(&new_s->ptrs[i],
1635                             &old_s->ptrs[i],
1636                             sizeof(new_s->ptrs[i])))
1637                         continue;
1638
1639                 if (new_s) {
1640                         ret = bch2_trans_mark_stripe_bucket(trans,
1641                                         bkey_i_to_s_c_stripe(new), i, false);
1642                         if (ret)
1643                                 break;
1644                 }
1645
1646                 if (old_s) {
1647                         ret = bch2_trans_mark_stripe_bucket(trans,
1648                                         bkey_s_c_to_stripe(old), i, true);
1649                         if (ret)
1650                                 break;
1651                 }
1652         }
1653
1654         return ret;
1655 }
1656
1657 int bch2_trans_mark_inode(struct btree_trans *trans,
1658                           struct bkey_s_c old,
1659                           struct bkey_i *new,
1660                           unsigned flags)
1661 {
1662         int nr = bkey_is_inode(&new->k) - bkey_is_inode(old.k);
1663
1664         if (nr) {
1665                 struct replicas_delta_list *d =
1666                         replicas_deltas_realloc(trans, 0);
1667                 d->nr_inodes += nr;
1668         }
1669
1670         return 0;
1671 }
1672
1673 int bch2_trans_mark_reservation(struct btree_trans *trans,
1674                                 struct bkey_s_c old,
1675                                 struct bkey_i *new,
1676                                 unsigned flags)
1677 {
1678         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1679                 ? old
1680                 : bkey_i_to_s_c(new);
1681         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1682         s64 sectors = (s64) k.k->size;
1683         struct replicas_delta_list *d;
1684
1685         if (flags & BTREE_TRIGGER_OVERWRITE)
1686                 sectors = -sectors;
1687         sectors *= replicas;
1688
1689         d = replicas_deltas_realloc(trans, 0);
1690
1691         replicas = clamp_t(unsigned, replicas, 1,
1692                            ARRAY_SIZE(d->persistent_reserved));
1693
1694         d->persistent_reserved[replicas - 1] += sectors;
1695         return 0;
1696 }
1697
1698 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1699                         struct bkey_s_c_reflink_p p,
1700                         u64 *idx, unsigned flags)
1701 {
1702         struct bch_fs *c = trans->c;
1703         struct btree_iter iter;
1704         struct bkey_s_c k;
1705         struct bkey_i *n;
1706         __le64 *refcount;
1707         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1708         struct printbuf buf = PRINTBUF;
1709         int ret;
1710
1711         bch2_trans_iter_init(trans, &iter, BTREE_ID_reflink, POS(0, *idx),
1712                              BTREE_ITER_INTENT|
1713                              BTREE_ITER_WITH_UPDATES);
1714         k = bch2_btree_iter_peek_slot(&iter);
1715         ret = bkey_err(k);
1716         if (ret)
1717                 goto err;
1718
1719         n = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1720         ret = PTR_ERR_OR_ZERO(n);
1721         if (ret)
1722                 goto err;
1723
1724         bkey_reassemble(n, k);
1725
1726         refcount = bkey_refcount(n);
1727         if (!refcount) {
1728                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1729                 bch2_trans_inconsistent(trans,
1730                         "nonexistent indirect extent at %llu while marking\n  %s",
1731                         *idx, buf.buf);
1732                 ret = -EIO;
1733                 goto err;
1734         }
1735
1736         if (!*refcount && (flags & BTREE_TRIGGER_OVERWRITE)) {
1737                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1738                 bch2_trans_inconsistent(trans,
1739                         "indirect extent refcount underflow at %llu while marking\n  %s",
1740                         *idx, buf.buf);
1741                 ret = -EIO;
1742                 goto err;
1743         }
1744
1745         if (flags & BTREE_TRIGGER_INSERT) {
1746                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1747                 u64 pad;
1748
1749                 pad = max_t(s64, le32_to_cpu(v->front_pad),
1750                             le64_to_cpu(v->idx) - bkey_start_offset(k.k));
1751                 BUG_ON(pad > U32_MAX);
1752                 v->front_pad = cpu_to_le32(pad);
1753
1754                 pad = max_t(s64, le32_to_cpu(v->back_pad),
1755                             k.k->p.offset - p.k->size - le64_to_cpu(v->idx));
1756                 BUG_ON(pad > U32_MAX);
1757                 v->back_pad = cpu_to_le32(pad);
1758         }
1759
1760         le64_add_cpu(refcount, add);
1761
1762         bch2_btree_iter_set_pos_to_extent_start(&iter);
1763         ret = bch2_trans_update(trans, &iter, n, 0);
1764         if (ret)
1765                 goto err;
1766
1767         *idx = k.k->p.offset;
1768 err:
1769         bch2_trans_iter_exit(trans, &iter);
1770         printbuf_exit(&buf);
1771         return ret;
1772 }
1773
1774 int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1775                               struct bkey_s_c old,
1776                               struct bkey_i *new,
1777                               unsigned flags)
1778 {
1779         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1780                 ? old
1781                 : bkey_i_to_s_c(new);
1782         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1783         u64 idx, end_idx;
1784         int ret = 0;
1785
1786         if (flags & BTREE_TRIGGER_INSERT) {
1787                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1788
1789                 v->front_pad = v->back_pad = 0;
1790         }
1791
1792         idx     = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
1793         end_idx = le64_to_cpu(p.v->idx) + p.k->size +
1794                 le32_to_cpu(p.v->back_pad);
1795
1796         while (idx < end_idx && !ret)
1797                 ret = __bch2_trans_mark_reflink_p(trans, p, &idx, flags);
1798
1799         return ret;
1800 }
1801
1802 static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1803                                     struct bch_dev *ca, size_t b,
1804                                     enum bch_data_type type,
1805                                     unsigned sectors)
1806 {
1807         struct bch_fs *c = trans->c;
1808         struct btree_iter iter;
1809         struct bkey_i_alloc_v4 *a;
1810         int ret = 0;
1811
1812         /*
1813          * Backup superblock might be past the end of our normal usable space:
1814          */
1815         if (b >= ca->mi.nbuckets)
1816                 return 0;
1817
1818         a = bch2_trans_start_alloc_update(trans, &iter, POS(ca->dev_idx, b));
1819         if (IS_ERR(a))
1820                 return PTR_ERR(a);
1821
1822         if (a->v.data_type && a->v.data_type != type) {
1823                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1824                         "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
1825                         "while marking %s",
1826                         iter.pos.inode, iter.pos.offset, a->v.gen,
1827                         bch2_data_types[a->v.data_type],
1828                         bch2_data_types[type],
1829                         bch2_data_types[type]);
1830                 ret = -EIO;
1831                 goto out;
1832         }
1833
1834         a->v.data_type          = type;
1835         a->v.dirty_sectors      = sectors;
1836
1837         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1838         if (ret)
1839                 goto out;
1840 out:
1841         bch2_trans_iter_exit(trans, &iter);
1842         return ret;
1843 }
1844
1845 int bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1846                                     struct bch_dev *ca, size_t b,
1847                                     enum bch_data_type type,
1848                                     unsigned sectors)
1849 {
1850         return __bch2_trans_do(trans, NULL, NULL, 0,
1851                         __bch2_trans_mark_metadata_bucket(trans, ca, b, type, sectors));
1852 }
1853
1854 static int bch2_trans_mark_metadata_sectors(struct btree_trans *trans,
1855                                             struct bch_dev *ca,
1856                                             u64 start, u64 end,
1857                                             enum bch_data_type type,
1858                                             u64 *bucket, unsigned *bucket_sectors)
1859 {
1860         do {
1861                 u64 b = sector_to_bucket(ca, start);
1862                 unsigned sectors =
1863                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1864
1865                 if (b != *bucket && *bucket_sectors) {
1866                         int ret = bch2_trans_mark_metadata_bucket(trans, ca, *bucket,
1867                                                                   type, *bucket_sectors);
1868                         if (ret)
1869                                 return ret;
1870
1871                         *bucket_sectors = 0;
1872                 }
1873
1874                 *bucket         = b;
1875                 *bucket_sectors += sectors;
1876                 start += sectors;
1877         } while (start < end);
1878
1879         return 0;
1880 }
1881
1882 static int __bch2_trans_mark_dev_sb(struct btree_trans *trans,
1883                                     struct bch_dev *ca)
1884 {
1885         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1886         u64 bucket = 0;
1887         unsigned i, bucket_sectors = 0;
1888         int ret;
1889
1890         for (i = 0; i < layout->nr_superblocks; i++) {
1891                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1892
1893                 if (offset == BCH_SB_SECTOR) {
1894                         ret = bch2_trans_mark_metadata_sectors(trans, ca,
1895                                                 0, BCH_SB_SECTOR,
1896                                                 BCH_DATA_sb, &bucket, &bucket_sectors);
1897                         if (ret)
1898                                 return ret;
1899                 }
1900
1901                 ret = bch2_trans_mark_metadata_sectors(trans, ca, offset,
1902                                       offset + (1 << layout->sb_max_size_bits),
1903                                       BCH_DATA_sb, &bucket, &bucket_sectors);
1904                 if (ret)
1905                         return ret;
1906         }
1907
1908         if (bucket_sectors) {
1909                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1910                                 bucket, BCH_DATA_sb, bucket_sectors);
1911                 if (ret)
1912                         return ret;
1913         }
1914
1915         for (i = 0; i < ca->journal.nr; i++) {
1916                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1917                                 ca->journal.buckets[i],
1918                                 BCH_DATA_journal, ca->mi.bucket_size);
1919                 if (ret)
1920                         return ret;
1921         }
1922
1923         return 0;
1924 }
1925
1926 int bch2_trans_mark_dev_sb(struct bch_fs *c, struct bch_dev *ca)
1927 {
1928         return bch2_trans_do(c, NULL, NULL, BTREE_INSERT_LAZY_RW,
1929                         __bch2_trans_mark_dev_sb(&trans, ca));
1930 }
1931
1932 /* Disk reservations: */
1933
1934 #define SECTORS_CACHE   1024
1935
1936 int bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
1937                               u64 sectors, int flags)
1938 {
1939         struct bch_fs_pcpu *pcpu;
1940         u64 old, v, get;
1941         s64 sectors_available;
1942         int ret;
1943
1944         percpu_down_read(&c->mark_lock);
1945         preempt_disable();
1946         pcpu = this_cpu_ptr(c->pcpu);
1947
1948         if (sectors <= pcpu->sectors_available)
1949                 goto out;
1950
1951         v = atomic64_read(&c->sectors_available);
1952         do {
1953                 old = v;
1954                 get = min((u64) sectors + SECTORS_CACHE, old);
1955
1956                 if (get < sectors) {
1957                         preempt_enable();
1958                         goto recalculate;
1959                 }
1960         } while ((v = atomic64_cmpxchg(&c->sectors_available,
1961                                        old, old - get)) != old);
1962
1963         pcpu->sectors_available         += get;
1964
1965 out:
1966         pcpu->sectors_available         -= sectors;
1967         this_cpu_add(*c->online_reserved, sectors);
1968         res->sectors                    += sectors;
1969
1970         preempt_enable();
1971         percpu_up_read(&c->mark_lock);
1972         return 0;
1973
1974 recalculate:
1975         mutex_lock(&c->sectors_available_lock);
1976
1977         percpu_u64_set(&c->pcpu->sectors_available, 0);
1978         sectors_available = avail_factor(__bch2_fs_usage_read_short(c).free);
1979
1980         if (sectors <= sectors_available ||
1981             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
1982                 atomic64_set(&c->sectors_available,
1983                              max_t(s64, 0, sectors_available - sectors));
1984                 this_cpu_add(*c->online_reserved, sectors);
1985                 res->sectors                    += sectors;
1986                 ret = 0;
1987         } else {
1988                 atomic64_set(&c->sectors_available, sectors_available);
1989                 ret = -ENOSPC;
1990         }
1991
1992         mutex_unlock(&c->sectors_available_lock);
1993         percpu_up_read(&c->mark_lock);
1994
1995         return ret;
1996 }
1997
1998 /* Startup/shutdown: */
1999
2000 static void bucket_gens_free_rcu(struct rcu_head *rcu)
2001 {
2002         struct bucket_gens *buckets =
2003                 container_of(rcu, struct bucket_gens, rcu);
2004
2005         kvpfree(buckets, sizeof(*buckets) + buckets->nbuckets);
2006 }
2007
2008 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
2009 {
2010         struct bucket_gens *bucket_gens = NULL, *old_bucket_gens = NULL;
2011         unsigned long *buckets_nouse = NULL;
2012         bool resize = ca->bucket_gens != NULL;
2013         int ret = -ENOMEM;
2014
2015         if (!(bucket_gens       = kvpmalloc(sizeof(struct bucket_gens) + nbuckets,
2016                                             GFP_KERNEL|__GFP_ZERO)) ||
2017             (c->opts.buckets_nouse &&
2018              !(buckets_nouse    = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2019                                             sizeof(unsigned long),
2020                                             GFP_KERNEL|__GFP_ZERO))))
2021                 goto err;
2022
2023         bucket_gens->first_bucket = ca->mi.first_bucket;
2024         bucket_gens->nbuckets   = nbuckets;
2025
2026         bch2_copygc_stop(c);
2027
2028         if (resize) {
2029                 down_write(&c->gc_lock);
2030                 down_write(&ca->bucket_lock);
2031                 percpu_down_write(&c->mark_lock);
2032         }
2033
2034         old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
2035
2036         if (resize) {
2037                 size_t n = min(bucket_gens->nbuckets, old_bucket_gens->nbuckets);
2038
2039                 memcpy(bucket_gens->b,
2040                        old_bucket_gens->b,
2041                        n);
2042                 if (buckets_nouse)
2043                         memcpy(buckets_nouse,
2044                                ca->buckets_nouse,
2045                                BITS_TO_LONGS(n) * sizeof(unsigned long));
2046         }
2047
2048         rcu_assign_pointer(ca->bucket_gens, bucket_gens);
2049         bucket_gens     = old_bucket_gens;
2050
2051         swap(ca->buckets_nouse, buckets_nouse);
2052
2053         nbuckets = ca->mi.nbuckets;
2054
2055         if (resize) {
2056                 percpu_up_write(&c->mark_lock);
2057                 up_write(&ca->bucket_lock);
2058                 up_write(&c->gc_lock);
2059         }
2060
2061         ret = 0;
2062 err:
2063         kvpfree(buckets_nouse,
2064                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2065         if (bucket_gens)
2066                 call_rcu(&bucket_gens->rcu, bucket_gens_free_rcu);
2067
2068         return ret;
2069 }
2070
2071 void bch2_dev_buckets_free(struct bch_dev *ca)
2072 {
2073         unsigned i;
2074
2075         kvpfree(ca->buckets_nouse,
2076                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2077         kvpfree(rcu_dereference_protected(ca->bucket_gens, 1),
2078                 sizeof(struct bucket_gens) + ca->mi.nbuckets);
2079
2080         for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
2081                 free_percpu(ca->usage[i]);
2082         kfree(ca->usage_base);
2083 }
2084
2085 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2086 {
2087         unsigned i;
2088
2089         ca->usage_base = kzalloc(sizeof(struct bch_dev_usage), GFP_KERNEL);
2090         if (!ca->usage_base)
2091                 return -ENOMEM;
2092
2093         for (i = 0; i < ARRAY_SIZE(ca->usage); i++) {
2094                 ca->usage[i] = alloc_percpu(struct bch_dev_usage);
2095                 if (!ca->usage[i])
2096                         return -ENOMEM;
2097         }
2098
2099         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);;
2100 }