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             fsck_err(c, "no replicas entry\n"
 382                      "  while marking %s",
 383                      (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 384                 percpu_up_read(&c->mark_lock);
 385                 ret = bch2_mark_replicas(c, r);
 386                 percpu_down_read(&c->mark_lock);
 387
 388                 if (ret)
 389                         goto err;
 390                 idx = bch2_replicas_entry_idx(c, r);
 391         }
 392         if (idx < 0) {
 393                 ret = -1;
 394                 goto err;
 395         }
 396
 397         preempt_disable();
 398         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 399         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 400         fs_usage->replicas[idx]         += sectors;
 401         preempt_enable();
 402 err:
 403 fsck_err:
 404         percpu_up_read(&c->mark_lock);
 405         printbuf_exit(&buf);
 406         return ret;
 407 }
 408
 409 static inline int update_cached_sectors(struct bch_fs *c,
 410                         struct bkey_s_c k,
 411                         unsigned dev, s64 sectors,
 412                         unsigned journal_seq, bool gc)
 413 {
 414         struct bch_replicas_padded r;
 415
 416         bch2_replicas_entry_cached(&r.e, dev);
 417
 418         return update_replicas(c, k, &r.e, sectors, journal_seq, gc);
 419 }
 420
 421 static struct replicas_delta_list *
 422 replicas_deltas_realloc(struct btree_trans *trans, unsigned more)
 423 {
 424         struct replicas_delta_list *d = trans->fs_usage_deltas;
 425         unsigned new_size = d ? (d->size + more) * 2 : 128;
 426         unsigned alloc_size = sizeof(*d) + new_size;
 427
 428         WARN_ON_ONCE(alloc_size > REPLICAS_DELTA_LIST_MAX);
 429
 430         if (!d || d->used + more > d->size) {
 431                 d = krealloc(d, alloc_size, GFP_NOIO|__GFP_ZERO);
 432
 433                 BUG_ON(!d && alloc_size > REPLICAS_DELTA_LIST_MAX);
 434
 435                 if (!d) {
 436                         d = mempool_alloc(&trans->c->replicas_delta_pool, GFP_NOIO);
 437                         memset(d, 0, REPLICAS_DELTA_LIST_MAX);
 438
 439                         if (trans->fs_usage_deltas)
 440                                 memcpy(d, trans->fs_usage_deltas,
 441                                        trans->fs_usage_deltas->size + sizeof(*d));
 442
 443                         new_size = REPLICAS_DELTA_LIST_MAX - sizeof(*d);
 444                         kfree(trans->fs_usage_deltas);
 445                 }
 446
 447                 d->size = new_size;
 448                 trans->fs_usage_deltas = d;
 449         }
 450         return d;
 451 }
 452
 453 static inline void update_replicas_list(struct btree_trans *trans,
 454                                         struct bch_replicas_entry *r,
 455                                         s64 sectors)
 456 {
 457         struct replicas_delta_list *d;
 458         struct replicas_delta *n;
 459         unsigned b;
 460
 461         if (!sectors)
 462                 return;
 463
 464         b = replicas_entry_bytes(r) + 8;
 465         d = replicas_deltas_realloc(trans, b);
 466
 467         n = (void *) d->d + d->used;
 468         n->delta = sectors;
 469         memcpy(&n->r, r, replicas_entry_bytes(r));
 470         bch2_replicas_entry_sort(&n->r);
 471         d->used += b;
 472 }
 473
 474 static inline void update_cached_sectors_list(struct btree_trans *trans,
 475                                               unsigned dev, s64 sectors)
 476 {
 477         struct bch_replicas_padded r;
 478
 479         bch2_replicas_entry_cached(&r.e, dev);
 480
 481         update_replicas_list(trans, &r.e, sectors);
 482 }
 483
 484 int bch2_mark_alloc(struct btree_trans *trans,
 485                     struct bkey_s_c old, struct bkey_s_c new,
 486                     unsigned flags)
 487 {
 488         bool gc = flags & BTREE_TRIGGER_GC;
 489         u64 journal_seq = trans->journal_res.seq;
 490         struct bch_fs *c = trans->c;
 491         struct bch_alloc_v4 old_a, new_a;
 492         struct bch_dev *ca;
 493         int ret = 0;
 494
 495         /*
 496          * alloc btree is read in by bch2_alloc_read, not gc:
 497          */
 498         if ((flags & BTREE_TRIGGER_GC) &&
 499             !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
 500                 return 0;
 501
 502         if (bch2_trans_inconsistent_on(!bch2_dev_bucket_exists(c, new.k->p), trans,
 503                                        "alloc key for invalid device or bucket"))
 504                 return -EIO;
 505
 506         ca = bch_dev_bkey_exists(c, new.k->p.inode);
 507
 508         bch2_alloc_to_v4(old, &old_a);
 509         bch2_alloc_to_v4(new, &new_a);
 510
 511         if ((flags & BTREE_TRIGGER_INSERT) &&
 512             data_type_is_empty(old_a.data_type) !=
 513             data_type_is_empty(new_a.data_type) &&
 514             new.k->type == KEY_TYPE_alloc_v4) {
 515                 struct bch_alloc_v4 *v = (struct bch_alloc_v4 *) new.v;
 516
 517                 BUG_ON(!journal_seq);
 518
 519                 /*
 520                  * If the btree updates referring to a bucket weren't flushed
 521                  * before the bucket became empty again, then the we don't have
 522                  * to wait on a journal flush before we can reuse the bucket:
 523                  */
 524                 new_a.journal_seq = data_type_is_empty(new_a.data_type) &&
 525                         (journal_seq == v->journal_seq ||
 526                          bch2_journal_noflush_seq(&c->journal, v->journal_seq))
 527                         ? 0 : journal_seq;
 528                 v->journal_seq = new_a.journal_seq;
 529         }
 530
 531         if (!data_type_is_empty(old_a.data_type) &&
 532             data_type_is_empty(new_a.data_type) &&
 533             new_a.journal_seq) {
 534                 ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 535                                 c->journal.flushed_seq_ondisk,
 536                                 new.k->p.inode, new.k->p.offset,
 537                                 new_a.journal_seq);
 538                 if (ret) {
 539                         bch2_fs_fatal_error(c,
 540                                 "error setting bucket_needs_journal_commit: %i", ret);
 541                         return ret;
 542                 }
 543         }
 544
 545         if (new_a.data_type == BCH_DATA_free &&
 546             (!new_a.journal_seq || new_a.journal_seq < c->journal.flushed_seq_ondisk))
 547                 closure_wake_up(&c->freelist_wait);
 548
 549         if (new_a.data_type == BCH_DATA_need_discard &&
 550             (!new_a.journal_seq || new_a.journal_seq < c->journal.flushed_seq_ondisk))
 551                 bch2_do_discards(c);
 552
 553         if (old_a.data_type != BCH_DATA_cached &&
 554             new_a.data_type == BCH_DATA_cached &&
 555             should_invalidate_buckets(ca, bch2_dev_usage_read(ca)))
 556                 bch2_do_invalidates(c);
 557
 558         if (new_a.data_type == BCH_DATA_need_gc_gens)
 559                 bch2_do_gc_gens(c);
 560
 561         percpu_down_read(&c->mark_lock);
 562         if (!gc && new_a.gen != old_a.gen)
 563                 *bucket_gen(ca, new.k->p.offset) = new_a.gen;
 564
 565         bch2_dev_usage_update(c, ca, old_a, new_a, journal_seq, gc);
 566
 567         if (gc) {
 568                 struct bucket *g = gc_bucket(ca, new.k->p.offset);
 569
 570                 bucket_lock(g);
 571
 572                 g->gen_valid            = 1;
 573                 g->gen                  = new_a.gen;
 574                 g->data_type            = new_a.data_type;
 575                 g->stripe               = new_a.stripe;
 576                 g->stripe_redundancy    = new_a.stripe_redundancy;
 577                 g->dirty_sectors        = new_a.dirty_sectors;
 578                 g->cached_sectors       = new_a.cached_sectors;
 579
 580                 bucket_unlock(g);
 581         }
 582         percpu_up_read(&c->mark_lock);
 583
 584         /*
 585          * need to know if we're getting called from the invalidate path or
 586          * not:
 587          */
 588
 589         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 590             old_a.cached_sectors) {
 591                 ret = update_cached_sectors(c, new, ca->dev_idx,
 592                                             -old_a.cached_sectors,
 593                                             journal_seq, gc);
 594                 if (ret) {
 595                         bch2_fs_fatal_error(c, "bch2_mark_alloc(): no replicas entry while updating cached sectors");
 596                         return ret;
 597                 }
 598         }
 599
 600         return 0;
 601 }
 602
 603 int bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 604                               size_t b, enum bch_data_type data_type,
 605                               unsigned sectors, struct gc_pos pos,
 606                               unsigned flags)
 607 {
 608         struct bucket old, new, *g;
 609         int ret = 0;
 610
 611         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 612         BUG_ON(data_type != BCH_DATA_sb &&
 613                data_type != BCH_DATA_journal);
 614
 615         /*
 616          * Backup superblock might be past the end of our normal usable space:
 617          */
 618         if (b >= ca->mi.nbuckets)
 619                 return 0;
 620
 621         percpu_down_read(&c->mark_lock);
 622         g = gc_bucket(ca, b);
 623
 624         bucket_lock(g);
 625         old = *g;
 626
 627         if (bch2_fs_inconsistent_on(g->data_type &&
 628                         g->data_type != data_type, c,
 629                         "different types of data in same bucket: %s, %s",
 630                         bch2_data_types[g->data_type],
 631                         bch2_data_types[data_type])) {
 632                 ret = -EIO;
 633                 goto err;
 634         }
 635
 636         if (bch2_fs_inconsistent_on((u64) g->dirty_sectors + sectors > ca->mi.bucket_size, c,
 637                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > bucket size",
 638                         ca->dev_idx, b, g->gen,
 639                         bch2_data_types[g->data_type ?: data_type],
 640                         g->dirty_sectors, sectors)) {
 641                 ret = -EIO;
 642                 goto err;
 643         }
 644
 645
 646         g->data_type = data_type;
 647         g->dirty_sectors += sectors;
 648         new = *g;
 649 err:
 650         bucket_unlock(g);
 651         if (!ret)
 652                 bch2_dev_usage_update_m(c, ca, old, new, 0, true);
 653         percpu_up_read(&c->mark_lock);
 654         return ret;
 655 }
 656
 657 static s64 ptr_disk_sectors(s64 sectors, struct extent_ptr_decoded p)
 658 {
 659         EBUG_ON(sectors < 0);
 660
 661         return crc_is_compressed(p.crc)
 662                 ? DIV_ROUND_UP_ULL(sectors * p.crc.compressed_size,
 663                                    p.crc.uncompressed_size)
 664                 : sectors;
 665 }
 666
 667 static int check_bucket_ref(struct bch_fs *c,
 668                             struct bkey_s_c k,
 669                             const struct bch_extent_ptr *ptr,
 670                             s64 sectors, enum bch_data_type ptr_data_type,
 671                             u8 b_gen, u8 bucket_data_type,
 672                             u32 dirty_sectors, u32 cached_sectors)
 673 {
 674         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 675         size_t bucket_nr = PTR_BUCKET_NR(ca, ptr);
 676         u16 bucket_sectors = !ptr->cached
 677                 ? dirty_sectors
 678                 : cached_sectors;
 679         struct printbuf buf = PRINTBUF;
 680         int ret = 0;
 681
 682         if (bucket_data_type == BCH_DATA_cached)
 683                 bucket_data_type = BCH_DATA_user;
 684
 685         if (gen_after(ptr->gen, b_gen)) {
 686                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 687                         "bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\n"
 688                         "while marking %s",
 689                         ptr->dev, bucket_nr, b_gen,
 690                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 691                         ptr->gen,
 692                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 693                 ret = -EIO;
 694                 goto err;
 695         }
 696
 697         if (gen_cmp(b_gen, ptr->gen) > BUCKET_GC_GEN_MAX) {
 698                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 699                         "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 700                         "while marking %s",
 701                         ptr->dev, bucket_nr, b_gen,
 702                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 703                         ptr->gen,
 704                         (printbuf_reset(&buf),
 705                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 706                 ret = -EIO;
 707                 goto err;
 708         }
 709
 710         if (b_gen != ptr->gen && !ptr->cached) {
 711                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 712                         "bucket %u:%zu gen %u (mem gen %u) data type %s: stale dirty ptr (gen %u)\n"
 713                         "while marking %s",
 714                         ptr->dev, bucket_nr, b_gen,
 715                         *bucket_gen(ca, bucket_nr),
 716                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 717                         ptr->gen,
 718                         (printbuf_reset(&buf),
 719                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 720                 ret = -EIO;
 721                 goto err;
 722         }
 723
 724         if (b_gen != ptr->gen) {
 725                 ret = 1;
 726                 goto err;
 727         }
 728
 729         if (!data_type_is_empty(bucket_data_type) &&
 730             ptr_data_type &&
 731             bucket_data_type != ptr_data_type) {
 732                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 733                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 734                         "while marking %s",
 735                         ptr->dev, bucket_nr, b_gen,
 736                         bch2_data_types[bucket_data_type],
 737                         bch2_data_types[ptr_data_type],
 738                         (printbuf_reset(&buf),
 739                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 740                 ret = -EIO;
 741                 goto err;
 742         }
 743
 744         if ((unsigned) (bucket_sectors + sectors) > U32_MAX) {
 745                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 746                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U16_MAX\n"
 747                         "while marking %s",
 748                         ptr->dev, bucket_nr, b_gen,
 749                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 750                         bucket_sectors, sectors,
 751                         (printbuf_reset(&buf),
 752                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 753                 ret = -EIO;
 754                 goto err;
 755         }
 756 err:
 757         printbuf_exit(&buf);
 758         return ret;
 759 }
 760
 761 static int mark_stripe_bucket(struct btree_trans *trans,
 762                               struct bkey_s_c k,
 763                               unsigned ptr_idx,
 764                               unsigned flags)
 765 {
 766         struct bch_fs *c = trans->c;
 767         u64 journal_seq = trans->journal_res.seq;
 768         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 769         unsigned nr_data = s->nr_blocks - s->nr_redundant;
 770         bool parity = ptr_idx >= nr_data;
 771         enum bch_data_type data_type = parity ? BCH_DATA_parity : 0;
 772         s64 sectors = parity ? le16_to_cpu(s->sectors) : 0;
 773         const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
 774         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 775         struct bucket old, new, *g;
 776         struct printbuf buf = PRINTBUF;
 777         int ret = 0;
 778
 779         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 780
 781         /* * XXX doesn't handle deletion */
 782
 783         percpu_down_read(&c->mark_lock);
 784         buf.atomic++;
 785         g = PTR_GC_BUCKET(ca, ptr);
 786
 787         if (g->dirty_sectors ||
 788             (g->stripe && g->stripe != k.k->p.offset)) {
 789                 bch2_fs_inconsistent(c,
 790                               "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
 791                               ptr->dev, PTR_BUCKET_NR(ca, ptr), g->gen,
 792                               (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 793                 ret = -EINVAL;
 794                 goto err;
 795         }
 796
 797         bucket_lock(g);
 798         old = *g;
 799
 800         ret = check_bucket_ref(c, k, ptr, sectors, data_type,
 801                                g->gen, g->data_type,
 802                                g->dirty_sectors, g->cached_sectors);
 803         if (ret)
 804                 goto err;
 805
 806         if (data_type)
 807                 g->data_type = data_type;
 808         g->dirty_sectors += sectors;
 809
 810         g->stripe               = k.k->p.offset;
 811         g->stripe_redundancy    = s->nr_redundant;
 812         new = *g;
 813 err:
 814         bucket_unlock(g);
 815         if (!ret)
 816                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 817         percpu_up_read(&c->mark_lock);
 818         printbuf_exit(&buf);
 819         return ret;
 820 }
 821
 822 static int __mark_pointer(struct btree_trans *trans,
 823                           struct bkey_s_c k,
 824                           const struct bch_extent_ptr *ptr,
 825                           s64 sectors, enum bch_data_type ptr_data_type,
 826                           u8 bucket_gen, u8 *bucket_data_type,
 827                           u32 *dirty_sectors, u32 *cached_sectors)
 828 {
 829         u32 *dst_sectors = !ptr->cached
 830                 ? dirty_sectors
 831                 : cached_sectors;
 832         int ret = check_bucket_ref(trans->c, k, ptr, sectors, ptr_data_type,
 833                                    bucket_gen, *bucket_data_type,
 834                                    *dirty_sectors, *cached_sectors);
 835
 836         if (ret)
 837                 return ret;
 838
 839         *dst_sectors += sectors;
 840         *bucket_data_type = *dirty_sectors || *cached_sectors
 841                 ? ptr_data_type : 0;
 842         return 0;
 843 }
 844
 845 static int bch2_mark_pointer(struct btree_trans *trans,
 846                              struct bkey_s_c k,
 847                              struct extent_ptr_decoded p,
 848                              s64 sectors, enum bch_data_type data_type,
 849                              unsigned flags)
 850 {
 851         u64 journal_seq = trans->journal_res.seq;
 852         struct bch_fs *c = trans->c;
 853         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 854         struct bucket old, new, *g;
 855         u8 bucket_data_type;
 856         int ret = 0;
 857
 858         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 859
 860         percpu_down_read(&c->mark_lock);
 861         g = PTR_GC_BUCKET(ca, &p.ptr);
 862         bucket_lock(g);
 863         old = *g;
 864
 865         bucket_data_type = g->data_type;
 866         ret = __mark_pointer(trans, k, &p.ptr, sectors,
 867                              data_type, g->gen,
 868                              &bucket_data_type,
 869                              &g->dirty_sectors,
 870                              &g->cached_sectors);
 871         if (!ret)
 872                 g->data_type = bucket_data_type;
 873
 874         new = *g;
 875         bucket_unlock(g);
 876         if (!ret)
 877                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 878         percpu_up_read(&c->mark_lock);
 879
 880         return ret;
 881 }
 882
 883 static int bch2_mark_stripe_ptr(struct btree_trans *trans,
 884                                 struct bkey_s_c k,
 885                                 struct bch_extent_stripe_ptr p,
 886                                 enum bch_data_type data_type,
 887                                 s64 sectors,
 888                                 unsigned flags)
 889 {
 890         struct bch_fs *c = trans->c;
 891         struct bch_replicas_padded r;
 892         struct gc_stripe *m;
 893
 894         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 895
 896         m = genradix_ptr_alloc(&c->gc_stripes, p.idx, GFP_KERNEL);
 897         if (!m) {
 898                 bch_err(c, "error allocating memory for gc_stripes, idx %llu",
 899                         (u64) p.idx);
 900                 return -ENOMEM;
 901         }
 902
 903         spin_lock(&c->ec_stripes_heap_lock);
 904
 905         if (!m || !m->alive) {
 906                 spin_unlock(&c->ec_stripes_heap_lock);
 907                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
 908                                     (u64) p.idx);
 909                 bch2_inconsistent_error(c);
 910                 return -EIO;
 911         }
 912
 913         m->block_sectors[p.block] += sectors;
 914
 915         r = m->r;
 916         spin_unlock(&c->ec_stripes_heap_lock);
 917
 918         r.e.data_type = data_type;
 919         update_replicas(c, k, &r.e, sectors, trans->journal_res.seq, true);
 920
 921         return 0;
 922 }
 923
 924 int bch2_mark_extent(struct btree_trans *trans,
 925                      struct bkey_s_c old, struct bkey_s_c new,
 926                      unsigned flags)
 927 {
 928         u64 journal_seq = trans->journal_res.seq;
 929         struct bch_fs *c = trans->c;
 930         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
 931         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 932         const union bch_extent_entry *entry;
 933         struct extent_ptr_decoded p;
 934         struct bch_replicas_padded r;
 935         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
 936                 ? BCH_DATA_btree
 937                 : BCH_DATA_user;
 938         s64 sectors = bkey_is_btree_ptr(k.k)
 939                 ? btree_sectors(c)
 940                 : k.k->size;
 941         s64 dirty_sectors = 0;
 942         bool stale;
 943         int ret;
 944
 945         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 946
 947         r.e.data_type   = data_type;
 948         r.e.nr_devs     = 0;
 949         r.e.nr_required = 1;
 950
 951         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 952                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
 953
 954                 if (flags & BTREE_TRIGGER_OVERWRITE)
 955                         disk_sectors = -disk_sectors;
 956
 957                 ret = bch2_mark_pointer(trans, k, p, disk_sectors,
 958                                         data_type, flags);
 959                 if (ret < 0)
 960                         return ret;
 961
 962                 stale = ret > 0;
 963
 964                 if (p.ptr.cached) {
 965                         if (!stale) {
 966                                 ret = update_cached_sectors(c, k, p.ptr.dev,
 967                                                 disk_sectors, journal_seq, true);
 968                                 if (ret) {
 969                                         bch2_fs_fatal_error(c, "bch2_mark_extent(): no replicas entry while updating cached sectors");
 970                                         return ret;
 971                                 }
 972                         }
 973                 } else if (!p.has_ec) {
 974                         dirty_sectors          += disk_sectors;
 975                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
 976                 } else {
 977                         ret = bch2_mark_stripe_ptr(trans, k, p.ec, data_type,
 978                                         disk_sectors, flags);
 979                         if (ret)
 980                                 return ret;
 981
 982                         /*
 983                          * There may be other dirty pointers in this extent, but
 984                          * if so they're not required for mounting if we have an
 985                          * erasure coded pointer in this extent:
 986                          */
 987                         r.e.nr_required = 0;
 988                 }
 989         }
 990
 991         if (r.e.nr_devs) {
 992                 ret = update_replicas(c, k, &r.e, dirty_sectors, journal_seq, true);
 993                 if (ret) {
 994                         struct printbuf buf = PRINTBUF;
 995
 996                         bch2_bkey_val_to_text(&buf, c, k);
 997                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
 998                         printbuf_exit(&buf);
 999                         return ret;
1000                 }
1001         }
1002
1003         return 0;
1004 }
1005
1006 int bch2_mark_stripe(struct btree_trans *trans,
1007                      struct bkey_s_c old, struct bkey_s_c new,
1008                      unsigned flags)
1009 {
1010         bool gc = flags & BTREE_TRIGGER_GC;
1011         u64 journal_seq = trans->journal_res.seq;
1012         struct bch_fs *c = trans->c;
1013         u64 idx = new.k->p.offset;
1014         const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
1015                 ? bkey_s_c_to_stripe(old).v : NULL;
1016         const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
1017                 ? bkey_s_c_to_stripe(new).v : NULL;
1018         unsigned i;
1019         int ret;
1020
1021         BUG_ON(gc && old_s);
1022
1023         if (!gc) {
1024                 struct stripe *m = genradix_ptr(&c->stripes, idx);
1025
1026                 if (!m || (old_s && !m->alive)) {
1027                         struct printbuf buf1 = PRINTBUF;
1028                         struct printbuf buf2 = PRINTBUF;
1029
1030                         bch2_bkey_val_to_text(&buf1, c, old);
1031                         bch2_bkey_val_to_text(&buf2, c, new);
1032                         bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
1033                                             "old %s\n"
1034                                             "new %s", idx, buf1.buf, buf2.buf);
1035                         printbuf_exit(&buf2);
1036                         printbuf_exit(&buf1);
1037                         bch2_inconsistent_error(c);
1038                         return -1;
1039                 }
1040
1041                 if (!new_s) {
1042                         spin_lock(&c->ec_stripes_heap_lock);
1043                         bch2_stripes_heap_del(c, m, idx);
1044                         spin_unlock(&c->ec_stripes_heap_lock);
1045
1046                         memset(m, 0, sizeof(*m));
1047                 } else {
1048                         m->alive        = true;
1049                         m->sectors      = le16_to_cpu(new_s->sectors);
1050                         m->algorithm    = new_s->algorithm;
1051                         m->nr_blocks    = new_s->nr_blocks;
1052                         m->nr_redundant = new_s->nr_redundant;
1053                         m->blocks_nonempty = 0;
1054
1055                         for (i = 0; i < new_s->nr_blocks; i++)
1056                                 m->blocks_nonempty += !!stripe_blockcount_get(new_s, i);
1057
1058                         spin_lock(&c->ec_stripes_heap_lock);
1059                         bch2_stripes_heap_update(c, m, idx);
1060                         spin_unlock(&c->ec_stripes_heap_lock);
1061                 }
1062         } else {
1063                 struct gc_stripe *m =
1064                         genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
1065
1066                 if (!m) {
1067                         bch_err(c, "error allocating memory for gc_stripes, idx %llu",
1068                                 idx);
1069                         return -ENOMEM;
1070                 }
1071                 /*
1072                  * This will be wrong when we bring back runtime gc: we should
1073                  * be unmarking the old key and then marking the new key
1074                  */
1075                 m->alive        = true;
1076                 m->sectors      = le16_to_cpu(new_s->sectors);
1077                 m->nr_blocks    = new_s->nr_blocks;
1078                 m->nr_redundant = new_s->nr_redundant;
1079
1080                 for (i = 0; i < new_s->nr_blocks; i++)
1081                         m->ptrs[i] = new_s->ptrs[i];
1082
1083                 bch2_bkey_to_replicas(&m->r.e, new);
1084
1085                 /*
1086                  * gc recalculates this field from stripe ptr
1087                  * references:
1088                  */
1089                 memset(m->block_sectors, 0, sizeof(m->block_sectors));
1090
1091                 for (i = 0; i < new_s->nr_blocks; i++) {
1092                         ret = mark_stripe_bucket(trans, new, i, flags);
1093                         if (ret)
1094                                 return ret;
1095                 }
1096
1097                 ret = update_replicas(c, new, &m->r.e,
1098                                       ((s64) m->sectors * m->nr_redundant),
1099                                       journal_seq, gc);
1100                 if (ret) {
1101                         struct printbuf buf = PRINTBUF;
1102
1103                         bch2_bkey_val_to_text(&buf, c, new);
1104                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
1105                         printbuf_exit(&buf);
1106                         return ret;
1107                 }
1108         }
1109
1110         return 0;
1111 }
1112
1113 int bch2_mark_inode(struct btree_trans *trans,
1114                     struct bkey_s_c old, struct bkey_s_c new,
1115                     unsigned flags)
1116 {
1117         struct bch_fs *c = trans->c;
1118         struct bch_fs_usage __percpu *fs_usage;
1119         u64 journal_seq = trans->journal_res.seq;
1120
1121         if (flags & BTREE_TRIGGER_INSERT) {
1122                 struct bch_inode_v2 *v = (struct bch_inode_v2 *) new.v;
1123
1124                 BUG_ON(!journal_seq);
1125                 BUG_ON(new.k->type != KEY_TYPE_inode_v2);
1126
1127                 v->bi_journal_seq = cpu_to_le64(journal_seq);
1128         }
1129
1130         if (flags & BTREE_TRIGGER_GC) {
1131                 percpu_down_read(&c->mark_lock);
1132                 preempt_disable();
1133
1134                 fs_usage = fs_usage_ptr(c, journal_seq, flags & BTREE_TRIGGER_GC);
1135                 fs_usage->nr_inodes += bkey_is_inode(new.k);
1136                 fs_usage->nr_inodes -= bkey_is_inode(old.k);
1137
1138                 preempt_enable();
1139                 percpu_up_read(&c->mark_lock);
1140         }
1141         return 0;
1142 }
1143
1144 int bch2_mark_reservation(struct btree_trans *trans,
1145                           struct bkey_s_c old, struct bkey_s_c new,
1146                           unsigned flags)
1147 {
1148         struct bch_fs *c = trans->c;
1149         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1150         struct bch_fs_usage __percpu *fs_usage;
1151         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1152         s64 sectors = (s64) k.k->size;
1153
1154         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1155
1156         if (flags & BTREE_TRIGGER_OVERWRITE)
1157                 sectors = -sectors;
1158         sectors *= replicas;
1159
1160         percpu_down_read(&c->mark_lock);
1161         preempt_disable();
1162
1163         fs_usage = fs_usage_ptr(c, trans->journal_res.seq, flags & BTREE_TRIGGER_GC);
1164         replicas = clamp_t(unsigned, replicas, 1,
1165                            ARRAY_SIZE(fs_usage->persistent_reserved));
1166
1167         fs_usage->reserved                              += sectors;
1168         fs_usage->persistent_reserved[replicas - 1]     += sectors;
1169
1170         preempt_enable();
1171         percpu_up_read(&c->mark_lock);
1172
1173         return 0;
1174 }
1175
1176 static s64 __bch2_mark_reflink_p(struct btree_trans *trans,
1177                                  struct bkey_s_c_reflink_p p,
1178                                  u64 start, u64 end,
1179                                  u64 *idx, unsigned flags, size_t r_idx)
1180 {
1181         struct bch_fs *c = trans->c;
1182         struct reflink_gc *r;
1183         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1184         u64 next_idx = end;
1185         s64 ret = 0;
1186         struct printbuf buf = PRINTBUF;
1187
1188         if (r_idx >= c->reflink_gc_nr)
1189                 goto not_found;
1190
1191         r = genradix_ptr(&c->reflink_gc_table, r_idx);
1192         next_idx = min(next_idx, r->offset - r->size);
1193         if (*idx < next_idx)
1194                 goto not_found;
1195
1196         BUG_ON((s64) r->refcount + add < 0);
1197
1198         r->refcount += add;
1199         *idx = r->offset;
1200         return 0;
1201 not_found:
1202         if (fsck_err(c, "pointer to missing indirect extent\n"
1203                      "  %s\n"
1204                      "  missing range %llu-%llu",
1205                      (bch2_bkey_val_to_text(&buf, c, p.s_c), buf.buf),
1206                      *idx, next_idx)) {
1207                 struct bkey_i_error new;
1208
1209                 bkey_init(&new.k);
1210                 new.k.type      = KEY_TYPE_error;
1211                 new.k.p         = bkey_start_pos(p.k);
1212                 new.k.p.offset += *idx - start;
1213                 bch2_key_resize(&new.k, next_idx - *idx);
1214                 ret = __bch2_btree_insert(trans, BTREE_ID_extents, &new.k_i);
1215         }
1216
1217         *idx = next_idx;
1218 fsck_err:
1219         printbuf_exit(&buf);
1220         return ret;
1221 }
1222
1223 int bch2_mark_reflink_p(struct btree_trans *trans,
1224                         struct bkey_s_c old, struct bkey_s_c new,
1225                         unsigned flags)
1226 {
1227         struct bch_fs *c = trans->c;
1228         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1229         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1230         struct reflink_gc *ref;
1231         size_t l, r, m;
1232         u64 idx = le64_to_cpu(p.v->idx), start = idx;
1233         u64 end = le64_to_cpu(p.v->idx) + p.k->size;
1234         int ret = 0;
1235
1236         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1237
1238         if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix) {
1239                 idx -= le32_to_cpu(p.v->front_pad);
1240                 end += le32_to_cpu(p.v->back_pad);
1241         }
1242
1243         l = 0;
1244         r = c->reflink_gc_nr;
1245         while (l < r) {
1246                 m = l + (r - l) / 2;
1247
1248                 ref = genradix_ptr(&c->reflink_gc_table, m);
1249                 if (ref->offset <= idx)
1250                         l = m + 1;
1251                 else
1252                         r = m;
1253         }
1254
1255         while (idx < end && !ret)
1256                 ret = __bch2_mark_reflink_p(trans, p, start, end,
1257                                             &idx, flags, l++);
1258
1259         return ret;
1260 }
1261
1262 static noinline __cold
1263 void fs_usage_apply_warn(struct btree_trans *trans,
1264                          unsigned disk_res_sectors,
1265                          s64 should_not_have_added)
1266 {
1267         struct bch_fs *c = trans->c;
1268         struct btree_insert_entry *i;
1269         struct printbuf buf = PRINTBUF;
1270
1271         bch_err(c, "disk usage increased %lli more than %u sectors reserved",
1272                 should_not_have_added, disk_res_sectors);
1273
1274         trans_for_each_update(trans, i) {
1275                 struct bkey_s_c old = { &i->old_k, i->old_v };
1276
1277                 pr_err("while inserting");
1278                 printbuf_reset(&buf);
1279                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(i->k));
1280                 pr_err("  %s", buf.buf);
1281                 pr_err("overlapping with");
1282                 printbuf_reset(&buf);
1283                 bch2_bkey_val_to_text(&buf, c, old);
1284                 pr_err("  %s", buf.buf);
1285         }
1286
1287         __WARN();
1288         printbuf_exit(&buf);
1289 }
1290
1291 int bch2_trans_fs_usage_apply(struct btree_trans *trans,
1292                               struct replicas_delta_list *deltas)
1293 {
1294         struct bch_fs *c = trans->c;
1295         static int warned_disk_usage = 0;
1296         bool warn = false;
1297         unsigned disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1298         struct replicas_delta *d = deltas->d, *d2;
1299         struct replicas_delta *top = (void *) deltas->d + deltas->used;
1300         struct bch_fs_usage *dst;
1301         s64 added = 0, should_not_have_added;
1302         unsigned i;
1303
1304         percpu_down_read(&c->mark_lock);
1305         preempt_disable();
1306         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1307
1308         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1309                 switch (d->r.data_type) {
1310                 case BCH_DATA_btree:
1311                 case BCH_DATA_user:
1312                 case BCH_DATA_parity:
1313                         added += d->delta;
1314                 }
1315
1316                 if (__update_replicas(c, dst, &d->r, d->delta))
1317                         goto need_mark;
1318         }
1319
1320         dst->nr_inodes += deltas->nr_inodes;
1321
1322         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1323                 added                           += deltas->persistent_reserved[i];
1324                 dst->reserved                   += deltas->persistent_reserved[i];
1325                 dst->persistent_reserved[i]     += deltas->persistent_reserved[i];
1326         }
1327
1328         /*
1329          * Not allowed to reduce sectors_available except by getting a
1330          * reservation:
1331          */
1332         should_not_have_added = added - (s64) disk_res_sectors;
1333         if (unlikely(should_not_have_added > 0)) {
1334                 u64 old, new, v = atomic64_read(&c->sectors_available);
1335
1336                 do {
1337                         old = v;
1338                         new = max_t(s64, 0, old - should_not_have_added);
1339                 } while ((v = atomic64_cmpxchg(&c->sectors_available,
1340                                                old, new)) != old);
1341
1342                 added -= should_not_have_added;
1343                 warn = true;
1344         }
1345
1346         if (added > 0) {
1347                 trans->disk_res->sectors -= added;
1348                 this_cpu_sub(*c->online_reserved, added);
1349         }
1350
1351         preempt_enable();
1352         percpu_up_read(&c->mark_lock);
1353
1354         if (unlikely(warn) && !xchg(&warned_disk_usage, 1))
1355                 fs_usage_apply_warn(trans, disk_res_sectors, should_not_have_added);
1356         return 0;
1357 need_mark:
1358         /* revert changes: */
1359         for (d2 = deltas->d; d2 != d; d2 = replicas_delta_next(d2))
1360                 BUG_ON(__update_replicas(c, dst, &d2->r, -d2->delta));
1361
1362         preempt_enable();
1363         percpu_up_read(&c->mark_lock);
1364         return -1;
1365 }
1366
1367 /* trans_mark: */
1368
1369 static int bch2_trans_mark_pointer(struct btree_trans *trans,
1370                         struct bkey_s_c k, struct extent_ptr_decoded p,
1371                         s64 sectors, enum bch_data_type data_type)
1372 {
1373         struct btree_iter iter;
1374         struct bkey_i_alloc_v4 *a;
1375         int ret;
1376
1377         a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(trans->c, &p.ptr));
1378         if (IS_ERR(a))
1379                 return PTR_ERR(a);
1380
1381         ret = __mark_pointer(trans, k, &p.ptr, sectors, data_type,
1382                              a->v.gen, &a->v.data_type,
1383                              &a->v.dirty_sectors, &a->v.cached_sectors) ?:
1384                 bch2_trans_update(trans, &iter, &a->k_i, 0);
1385         bch2_trans_iter_exit(trans, &iter);
1386         return ret;
1387 }
1388
1389 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1390                         struct extent_ptr_decoded p,
1391                         s64 sectors, enum bch_data_type data_type)
1392 {
1393         struct btree_iter iter;
1394         struct bkey_s_c k;
1395         struct bkey_i_stripe *s;
1396         struct bch_replicas_padded r;
1397         int ret = 0;
1398
1399         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes, POS(0, p.ec.idx),
1400                              BTREE_ITER_INTENT|
1401                              BTREE_ITER_WITH_UPDATES);
1402         k = bch2_btree_iter_peek_slot(&iter);
1403         ret = bkey_err(k);
1404         if (ret)
1405                 goto err;
1406
1407         if (k.k->type != KEY_TYPE_stripe) {
1408                 bch2_trans_inconsistent(trans,
1409                         "pointer to nonexistent stripe %llu",
1410                         (u64) p.ec.idx);
1411                 ret = -EIO;
1412                 goto err;
1413         }
1414
1415         if (!bch2_ptr_matches_stripe(bkey_s_c_to_stripe(k).v, p)) {
1416                 bch2_trans_inconsistent(trans,
1417                         "stripe pointer doesn't match stripe %llu",
1418                         (u64) p.ec.idx);
1419                 ret = -EIO;
1420                 goto err;
1421         }
1422
1423         s = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1424         ret = PTR_ERR_OR_ZERO(s);
1425         if (ret)
1426                 goto err;
1427
1428         bkey_reassemble(&s->k_i, k);
1429         stripe_blockcount_set(&s->v, p.ec.block,
1430                 stripe_blockcount_get(&s->v, p.ec.block) +
1431                 sectors);
1432
1433         ret = bch2_trans_update(trans, &iter, &s->k_i, 0);
1434         if (ret)
1435                 goto err;
1436
1437         bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1438         r.e.data_type = data_type;
1439         update_replicas_list(trans, &r.e, sectors);
1440 err:
1441         bch2_trans_iter_exit(trans, &iter);
1442         return ret;
1443 }
1444
1445 int bch2_trans_mark_extent(struct btree_trans *trans,
1446                            enum btree_id btree_id, unsigned level,
1447                            struct bkey_s_c old, struct bkey_i *new,
1448                            unsigned flags)
1449 {
1450         struct bch_fs *c = trans->c;
1451         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1452                 ? old
1453                 : bkey_i_to_s_c(new);
1454         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1455         const union bch_extent_entry *entry;
1456         struct extent_ptr_decoded p;
1457         struct bch_replicas_padded r;
1458         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
1459                 ? BCH_DATA_btree
1460                 : BCH_DATA_user;
1461         s64 sectors = bkey_is_btree_ptr(k.k)
1462                 ? btree_sectors(c)
1463                 : k.k->size;
1464         s64 dirty_sectors = 0;
1465         bool stale;
1466         int ret;
1467
1468         r.e.data_type   = data_type;
1469         r.e.nr_devs     = 0;
1470         r.e.nr_required = 1;
1471
1472         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1473                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
1474
1475                 if (flags & BTREE_TRIGGER_OVERWRITE)
1476                         disk_sectors = -disk_sectors;
1477
1478                 ret = bch2_trans_mark_pointer(trans, k, p,
1479                                         disk_sectors, data_type);
1480                 if (ret < 0)
1481                         return ret;
1482
1483                 stale = ret > 0;
1484
1485                 if (p.ptr.cached) {
1486                         if (!stale)
1487                                 update_cached_sectors_list(trans, p.ptr.dev,
1488                                                            disk_sectors);
1489                 } else if (!p.has_ec) {
1490                         dirty_sectors          += disk_sectors;
1491                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1492                 } else {
1493                         ret = bch2_trans_mark_stripe_ptr(trans, p,
1494                                         disk_sectors, data_type);
1495                         if (ret)
1496                                 return ret;
1497
1498                         r.e.nr_required = 0;
1499                 }
1500         }
1501
1502         if (r.e.nr_devs)
1503                 update_replicas_list(trans, &r.e, dirty_sectors);
1504
1505         return 0;
1506 }
1507
1508 static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
1509                                          struct bkey_s_c_stripe s,
1510                                          unsigned idx, bool deleting)
1511 {
1512         struct bch_fs *c = trans->c;
1513         const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
1514         struct btree_iter iter;
1515         struct bkey_i_alloc_v4 *a;
1516         enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
1517                 ? BCH_DATA_parity : 0;
1518         s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
1519         int ret = 0;
1520
1521         if (deleting)
1522                 sectors = -sectors;
1523
1524         a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(c, ptr));
1525         if (IS_ERR(a))
1526                 return PTR_ERR(a);
1527
1528         ret = check_bucket_ref(c, s.s_c, ptr, sectors, data_type,
1529                                a->v.gen, a->v.data_type,
1530                                a->v.dirty_sectors, a->v.cached_sectors);
1531         if (ret)
1532                 goto err;
1533
1534         if (!deleting) {
1535                 if (bch2_trans_inconsistent_on(a->v.stripe ||
1536                                                a->v.stripe_redundancy, trans,
1537                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
1538                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1539                                 bch2_data_types[a->v.data_type],
1540                                 a->v.dirty_sectors,
1541                                 a->v.stripe, s.k->p.offset)) {
1542                         ret = -EIO;
1543                         goto err;
1544                 }
1545
1546                 if (bch2_trans_inconsistent_on(data_type && a->v.dirty_sectors, trans,
1547                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
1548                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1549                                 bch2_data_types[a->v.data_type],
1550                                 a->v.dirty_sectors,
1551                                 s.k->p.offset)) {
1552                         ret = -EIO;
1553                         goto err;
1554                 }
1555
1556                 a->v.stripe             = s.k->p.offset;
1557                 a->v.stripe_redundancy  = s.v->nr_redundant;
1558         } else {
1559                 if (bch2_trans_inconsistent_on(a->v.stripe != s.k->p.offset ||
1560                                                a->v.stripe_redundancy != s.v->nr_redundant, trans,
1561                                 "bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
1562                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1563                                 s.k->p.offset, a->v.stripe)) {
1564                         ret = -EIO;
1565                         goto err;
1566                 }
1567
1568                 a->v.stripe             = 0;
1569                 a->v.stripe_redundancy  = 0;
1570         }
1571
1572         a->v.dirty_sectors += sectors;
1573         if (data_type)
1574                 a->v.data_type = !deleting ? data_type : 0;
1575
1576         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1577         if (ret)
1578                 goto err;
1579 err:
1580         bch2_trans_iter_exit(trans, &iter);
1581         return ret;
1582 }
1583
1584 int bch2_trans_mark_stripe(struct btree_trans *trans,
1585                            enum btree_id btree_id, unsigned level,
1586                            struct bkey_s_c old, struct bkey_i *new,
1587                            unsigned flags)
1588 {
1589         const struct bch_stripe *old_s = NULL;
1590         struct bch_stripe *new_s = NULL;
1591         struct bch_replicas_padded r;
1592         unsigned i, nr_blocks;
1593         int ret = 0;
1594
1595         if (old.k->type == KEY_TYPE_stripe)
1596                 old_s = bkey_s_c_to_stripe(old).v;
1597         if (new->k.type == KEY_TYPE_stripe)
1598                 new_s = &bkey_i_to_stripe(new)->v;
1599
1600         /*
1601          * If the pointers aren't changing, we don't need to do anything:
1602          */
1603         if (new_s && old_s &&
1604             new_s->nr_blocks    == old_s->nr_blocks &&
1605             new_s->nr_redundant == old_s->nr_redundant &&
1606             !memcmp(old_s->ptrs, new_s->ptrs,
1607                     new_s->nr_blocks * sizeof(struct bch_extent_ptr)))
1608                 return 0;
1609
1610         BUG_ON(new_s && old_s &&
1611                (new_s->nr_blocks        != old_s->nr_blocks ||
1612                 new_s->nr_redundant     != old_s->nr_redundant));
1613
1614         nr_blocks = new_s ? new_s->nr_blocks : old_s->nr_blocks;
1615
1616         if (new_s) {
1617                 s64 sectors = le16_to_cpu(new_s->sectors);
1618
1619                 bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(new));
1620                 update_replicas_list(trans, &r.e, sectors * new_s->nr_redundant);
1621         }
1622
1623         if (old_s) {
1624                 s64 sectors = -((s64) le16_to_cpu(old_s->sectors));
1625
1626                 bch2_bkey_to_replicas(&r.e, old);
1627                 update_replicas_list(trans, &r.e, sectors * old_s->nr_redundant);
1628         }
1629
1630         for (i = 0; i < nr_blocks; i++) {
1631                 if (new_s && old_s &&
1632                     !memcmp(&new_s->ptrs[i],
1633                             &old_s->ptrs[i],
1634                             sizeof(new_s->ptrs[i])))
1635                         continue;
1636
1637                 if (new_s) {
1638                         ret = bch2_trans_mark_stripe_bucket(trans,
1639                                         bkey_i_to_s_c_stripe(new), i, false);
1640                         if (ret)
1641                                 break;
1642                 }
1643
1644                 if (old_s) {
1645                         ret = bch2_trans_mark_stripe_bucket(trans,
1646                                         bkey_s_c_to_stripe(old), i, true);
1647                         if (ret)
1648                                 break;
1649                 }
1650         }
1651
1652         return ret;
1653 }
1654
1655 int bch2_trans_mark_inode(struct btree_trans *trans,
1656                           enum btree_id btree_id, unsigned level,
1657                           struct bkey_s_c old,
1658                           struct bkey_i *new,
1659                           unsigned flags)
1660 {
1661         int nr = bkey_is_inode(&new->k) - bkey_is_inode(old.k);
1662
1663         if (nr) {
1664                 struct replicas_delta_list *d =
1665                         replicas_deltas_realloc(trans, 0);
1666                 d->nr_inodes += nr;
1667         }
1668
1669         return 0;
1670 }
1671
1672 int bch2_trans_mark_reservation(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         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                               enum btree_id btree_id, unsigned level,
1776                               struct bkey_s_c old,
1777                               struct bkey_i *new,
1778                               unsigned flags)
1779 {
1780         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1781                 ? old
1782                 : bkey_i_to_s_c(new);
1783         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1784         u64 idx, end_idx;
1785         int ret = 0;
1786
1787         if (flags & BTREE_TRIGGER_INSERT) {
1788                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1789
1790                 v->front_pad = v->back_pad = 0;
1791         }
1792
1793         idx     = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
1794         end_idx = le64_to_cpu(p.v->idx) + p.k->size +
1795                 le32_to_cpu(p.v->back_pad);
1796
1797         while (idx < end_idx && !ret)
1798                 ret = __bch2_trans_mark_reflink_p(trans, p, &idx, flags);
1799
1800         return ret;
1801 }
1802
1803 static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1804                                     struct bch_dev *ca, size_t b,
1805                                     enum bch_data_type type,
1806                                     unsigned sectors)
1807 {
1808         struct bch_fs *c = trans->c;
1809         struct btree_iter iter;
1810         struct bkey_i_alloc_v4 *a;
1811         int ret = 0;
1812
1813         /*
1814          * Backup superblock might be past the end of our normal usable space:
1815          */
1816         if (b >= ca->mi.nbuckets)
1817                 return 0;
1818
1819         a = bch2_trans_start_alloc_update(trans, &iter, POS(ca->dev_idx, b));
1820         if (IS_ERR(a))
1821                 return PTR_ERR(a);
1822
1823         if (a->v.data_type && a->v.data_type != type) {
1824                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1825                         "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
1826                         "while marking %s",
1827                         iter.pos.inode, iter.pos.offset, a->v.gen,
1828                         bch2_data_types[a->v.data_type],
1829                         bch2_data_types[type],
1830                         bch2_data_types[type]);
1831                 ret = -EIO;
1832                 goto out;
1833         }
1834
1835         a->v.data_type          = type;
1836         a->v.dirty_sectors      = sectors;
1837
1838         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1839         if (ret)
1840                 goto out;
1841 out:
1842         bch2_trans_iter_exit(trans, &iter);
1843         return ret;
1844 }
1845
1846 int bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1847                                     struct bch_dev *ca, size_t b,
1848                                     enum bch_data_type type,
1849                                     unsigned sectors)
1850 {
1851         return __bch2_trans_do(trans, NULL, NULL, 0,
1852                         __bch2_trans_mark_metadata_bucket(trans, ca, b, type, sectors));
1853 }
1854
1855 static int bch2_trans_mark_metadata_sectors(struct btree_trans *trans,
1856                                             struct bch_dev *ca,
1857                                             u64 start, u64 end,
1858                                             enum bch_data_type type,
1859                                             u64 *bucket, unsigned *bucket_sectors)
1860 {
1861         do {
1862                 u64 b = sector_to_bucket(ca, start);
1863                 unsigned sectors =
1864                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1865
1866                 if (b != *bucket && *bucket_sectors) {
1867                         int ret = bch2_trans_mark_metadata_bucket(trans, ca, *bucket,
1868                                                                   type, *bucket_sectors);
1869                         if (ret)
1870                                 return ret;
1871
1872                         *bucket_sectors = 0;
1873                 }
1874
1875                 *bucket         = b;
1876                 *bucket_sectors += sectors;
1877                 start += sectors;
1878         } while (start < end);
1879
1880         return 0;
1881 }
1882
1883 static int __bch2_trans_mark_dev_sb(struct btree_trans *trans,
1884                                     struct bch_dev *ca)
1885 {
1886         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1887         u64 bucket = 0;
1888         unsigned i, bucket_sectors = 0;
1889         int ret;
1890
1891         for (i = 0; i < layout->nr_superblocks; i++) {
1892                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1893
1894                 if (offset == BCH_SB_SECTOR) {
1895                         ret = bch2_trans_mark_metadata_sectors(trans, ca,
1896                                                 0, BCH_SB_SECTOR,
1897                                                 BCH_DATA_sb, &bucket, &bucket_sectors);
1898                         if (ret)
1899                                 return ret;
1900                 }
1901
1902                 ret = bch2_trans_mark_metadata_sectors(trans, ca, offset,
1903                                       offset + (1 << layout->sb_max_size_bits),
1904                                       BCH_DATA_sb, &bucket, &bucket_sectors);
1905                 if (ret)
1906                         return ret;
1907         }
1908
1909         if (bucket_sectors) {
1910                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1911                                 bucket, BCH_DATA_sb, bucket_sectors);
1912                 if (ret)
1913                         return ret;
1914         }
1915
1916         for (i = 0; i < ca->journal.nr; i++) {
1917                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1918                                 ca->journal.buckets[i],
1919                                 BCH_DATA_journal, ca->mi.bucket_size);
1920                 if (ret)
1921                         return ret;
1922         }
1923
1924         return 0;
1925 }
1926
1927 int bch2_trans_mark_dev_sb(struct bch_fs *c, struct bch_dev *ca)
1928 {
1929         return bch2_trans_do(c, NULL, NULL, BTREE_INSERT_LAZY_RW,
1930                         __bch2_trans_mark_dev_sb(&trans, ca));
1931 }
1932
1933 /* Disk reservations: */
1934
1935 #define SECTORS_CACHE   1024
1936
1937 int bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
1938                               u64 sectors, int flags)
1939 {
1940         struct bch_fs_pcpu *pcpu;
1941         u64 old, v, get;
1942         s64 sectors_available;
1943         int ret;
1944
1945         percpu_down_read(&c->mark_lock);
1946         preempt_disable();
1947         pcpu = this_cpu_ptr(c->pcpu);
1948
1949         if (sectors <= pcpu->sectors_available)
1950                 goto out;
1951
1952         v = atomic64_read(&c->sectors_available);
1953         do {
1954                 old = v;
1955                 get = min((u64) sectors + SECTORS_CACHE, old);
1956
1957                 if (get < sectors) {
1958                         preempt_enable();
1959                         goto recalculate;
1960                 }
1961         } while ((v = atomic64_cmpxchg(&c->sectors_available,
1962                                        old, old - get)) != old);
1963
1964         pcpu->sectors_available         += get;
1965
1966 out:
1967         pcpu->sectors_available         -= sectors;
1968         this_cpu_add(*c->online_reserved, sectors);
1969         res->sectors                    += sectors;
1970
1971         preempt_enable();
1972         percpu_up_read(&c->mark_lock);
1973         return 0;
1974
1975 recalculate:
1976         mutex_lock(&c->sectors_available_lock);
1977
1978         percpu_u64_set(&c->pcpu->sectors_available, 0);
1979         sectors_available = avail_factor(__bch2_fs_usage_read_short(c).free);
1980
1981         if (sectors <= sectors_available ||
1982             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
1983                 atomic64_set(&c->sectors_available,
1984                              max_t(s64, 0, sectors_available - sectors));
1985                 this_cpu_add(*c->online_reserved, sectors);
1986                 res->sectors                    += sectors;
1987                 ret = 0;
1988         } else {
1989                 atomic64_set(&c->sectors_available, sectors_available);
1990                 ret = -ENOSPC;
1991         }
1992
1993         mutex_unlock(&c->sectors_available_lock);
1994         percpu_up_read(&c->mark_lock);
1995
1996         return ret;
1997 }
1998
1999 /* Startup/shutdown: */
2000
2001 static void bucket_gens_free_rcu(struct rcu_head *rcu)
2002 {
2003         struct bucket_gens *buckets =
2004                 container_of(rcu, struct bucket_gens, rcu);
2005
2006         kvpfree(buckets, sizeof(*buckets) + buckets->nbuckets);
2007 }
2008
2009 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
2010 {
2011         struct bucket_gens *bucket_gens = NULL, *old_bucket_gens = NULL;
2012         unsigned long *buckets_nouse = NULL;
2013         bool resize = ca->bucket_gens != NULL;
2014         int ret = -ENOMEM;
2015
2016         if (!(bucket_gens       = kvpmalloc(sizeof(struct bucket_gens) + nbuckets,
2017                                             GFP_KERNEL|__GFP_ZERO)) ||
2018             (c->opts.buckets_nouse &&
2019              !(buckets_nouse    = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2020                                             sizeof(unsigned long),
2021                                             GFP_KERNEL|__GFP_ZERO))))
2022                 goto err;
2023
2024         bucket_gens->first_bucket = ca->mi.first_bucket;
2025         bucket_gens->nbuckets   = nbuckets;
2026
2027         bch2_copygc_stop(c);
2028
2029         if (resize) {
2030                 down_write(&c->gc_lock);
2031                 down_write(&ca->bucket_lock);
2032                 percpu_down_write(&c->mark_lock);
2033         }
2034
2035         old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
2036
2037         if (resize) {
2038                 size_t n = min(bucket_gens->nbuckets, old_bucket_gens->nbuckets);
2039
2040                 memcpy(bucket_gens->b,
2041                        old_bucket_gens->b,
2042                        n);
2043                 if (buckets_nouse)
2044                         memcpy(buckets_nouse,
2045                                ca->buckets_nouse,
2046                                BITS_TO_LONGS(n) * sizeof(unsigned long));
2047         }
2048
2049         rcu_assign_pointer(ca->bucket_gens, bucket_gens);
2050         bucket_gens     = old_bucket_gens;
2051
2052         swap(ca->buckets_nouse, buckets_nouse);
2053
2054         nbuckets = ca->mi.nbuckets;
2055
2056         if (resize) {
2057                 percpu_up_write(&c->mark_lock);
2058                 up_write(&ca->bucket_lock);
2059                 up_write(&c->gc_lock);
2060         }
2061
2062         ret = 0;
2063 err:
2064         kvpfree(buckets_nouse,
2065                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2066         if (bucket_gens)
2067                 call_rcu(&bucket_gens->rcu, bucket_gens_free_rcu);
2068
2069         return ret;
2070 }
2071
2072 void bch2_dev_buckets_free(struct bch_dev *ca)
2073 {
2074         unsigned i;
2075
2076         kvpfree(ca->buckets_nouse,
2077                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2078         kvpfree(rcu_dereference_protected(ca->bucket_gens, 1),
2079                 sizeof(struct bucket_gens) + ca->mi.nbuckets);
2080
2081         for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
2082                 free_percpu(ca->usage[i]);
2083         kfree(ca->usage_base);
2084 }
2085
2086 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2087 {
2088         unsigned i;
2089
2090         ca->usage_base = kzalloc(sizeof(struct bch_dev_usage), GFP_KERNEL);
2091         if (!ca->usage_base)
2092                 return -ENOMEM;
2093
2094         for (i = 0; i < ARRAY_SIZE(ca->usage); i++) {
2095                 ca->usage[i] = alloc_percpu(struct bch_dev_usage);
2096                 if (!ca->usage[i])
2097                         return -ENOMEM;
2098         }
2099
2100         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);;
2101 }