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

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Code for manipulating bucket marks for garbage collection.
   4  *
   5  * Copyright 2014 Datera, Inc.
   6  */
   7
   8 #include "bcachefs.h"
   9 #include "alloc_background.h"
  10 #include "backpointers.h"
  11 #include "bset.h"
  12 #include "btree_gc.h"
  13 #include "btree_update.h"
  14 #include "buckets.h"
  15 #include "buckets_waiting_for_journal.h"
  16 #include "ec.h"
  17 #include "error.h"
  18 #include "inode.h"
  19 #include "movinggc.h"
  20 #include "recovery.h"
  21 #include "reflink.h"
  22 #include "replicas.h"
  23 #include "subvolume.h"
  24 #include "trace.h"
  25
  26 #include <linux/preempt.h>
  27
  28 static inline void fs_usage_data_type_to_base(struct bch_fs_usage *fs_usage,
  29                                               enum bch_data_type data_type,
  30                                               s64 sectors)
  31 {
  32         switch (data_type) {
  33         case BCH_DATA_btree:
  34                 fs_usage->btree         += sectors;
  35                 break;
  36         case BCH_DATA_user:
  37         case BCH_DATA_parity:
  38                 fs_usage->data          += sectors;
  39                 break;
  40         case BCH_DATA_cached:
  41                 fs_usage->cached        += sectors;
  42                 break;
  43         default:
  44                 break;
  45         }
  46 }
  47
  48 void bch2_fs_usage_initialize(struct bch_fs *c)
  49 {
  50         struct bch_fs_usage *usage;
  51         struct bch_dev *ca;
  52         unsigned i;
  53
  54         percpu_down_write(&c->mark_lock);
  55         usage = c->usage_base;
  56
  57         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
  58                 bch2_fs_usage_acc_to_base(c, i);
  59
  60         for (i = 0; i < BCH_REPLICAS_MAX; i++)
  61                 usage->reserved += usage->persistent_reserved[i];
  62
  63         for (i = 0; i < c->replicas.nr; i++) {
  64                 struct bch_replicas_entry *e =
  65                         cpu_replicas_entry(&c->replicas, i);
  66
  67                 fs_usage_data_type_to_base(usage, e->data_type, usage->replicas[i]);
  68         }
  69
  70         for_each_member_device(ca, c, i) {
  71                 struct bch_dev_usage dev = bch2_dev_usage_read(ca);
  72
  73                 usage->hidden += (dev.d[BCH_DATA_sb].buckets +
  74                                   dev.d[BCH_DATA_journal].buckets) *
  75                         ca->mi.bucket_size;
  76         }
  77
  78         percpu_up_write(&c->mark_lock);
  79 }
  80
  81 static inline struct bch_dev_usage *dev_usage_ptr(struct bch_dev *ca,
  82                                                   unsigned journal_seq,
  83                                                   bool gc)
  84 {
  85         BUG_ON(!gc && !journal_seq);
  86
  87         return this_cpu_ptr(gc
  88                             ? ca->usage_gc
  89                             : ca->usage[journal_seq & JOURNAL_BUF_MASK]);
  90 }
  91
  92 void bch2_dev_usage_read_fast(struct bch_dev *ca, struct bch_dev_usage *usage)
  93 {
  94         struct bch_fs *c = ca->fs;
  95         unsigned seq, i, u64s = dev_usage_u64s();
  96
  97         do {
  98                 seq = read_seqcount_begin(&c->usage_lock);
  99                 memcpy(usage, ca->usage_base, u64s * sizeof(u64));
 100                 for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
 101                         acc_u64s_percpu((u64 *) usage, (u64 __percpu *) ca->usage[i], u64s);
 102         } while (read_seqcount_retry(&c->usage_lock, seq));
 103 }
 104
 105 static inline struct bch_fs_usage *fs_usage_ptr(struct bch_fs *c,
 106                                                 unsigned journal_seq,
 107                                                 bool gc)
 108 {
 109         percpu_rwsem_assert_held(&c->mark_lock);
 110         BUG_ON(!gc && !journal_seq);
 111
 112         return this_cpu_ptr(gc
 113                             ? c->usage_gc
 114                             : c->usage[journal_seq & JOURNAL_BUF_MASK]);
 115 }
 116
 117 u64 bch2_fs_usage_read_one(struct bch_fs *c, u64 *v)
 118 {
 119         ssize_t offset = v - (u64 *) c->usage_base;
 120         unsigned i, seq;
 121         u64 ret;
 122
 123         BUG_ON(offset < 0 || offset >= fs_usage_u64s(c));
 124         percpu_rwsem_assert_held(&c->mark_lock);
 125
 126         do {
 127                 seq = read_seqcount_begin(&c->usage_lock);
 128                 ret = *v;
 129
 130                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 131                         ret += percpu_u64_get((u64 __percpu *) c->usage[i] + offset);
 132         } while (read_seqcount_retry(&c->usage_lock, seq));
 133
 134         return ret;
 135 }
 136
 137 struct bch_fs_usage_online *bch2_fs_usage_read(struct bch_fs *c)
 138 {
 139         struct bch_fs_usage_online *ret;
 140         unsigned nr_replicas = READ_ONCE(c->replicas.nr);
 141         unsigned seq, i;
 142 retry:
 143         ret = kmalloc(__fs_usage_online_u64s(nr_replicas) * sizeof(u64), GFP_KERNEL);
 144         if (unlikely(!ret))
 145                 return NULL;
 146
 147         percpu_down_read(&c->mark_lock);
 148
 149         if (nr_replicas != c->replicas.nr) {
 150                 nr_replicas = c->replicas.nr;
 151                 percpu_up_read(&c->mark_lock);
 152                 kfree(ret);
 153                 goto retry;
 154         }
 155
 156         ret->online_reserved = percpu_u64_get(c->online_reserved);
 157
 158         do {
 159                 seq = read_seqcount_begin(&c->usage_lock);
 160                 unsafe_memcpy(&ret->u, c->usage_base,
 161                               __fs_usage_u64s(nr_replicas) * sizeof(u64),
 162                               "embedded variable length struct");
 163                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 164                         acc_u64s_percpu((u64 *) &ret->u, (u64 __percpu *) c->usage[i],
 165                                         __fs_usage_u64s(nr_replicas));
 166         } while (read_seqcount_retry(&c->usage_lock, seq));
 167
 168         return ret;
 169 }
 170
 171 void bch2_fs_usage_acc_to_base(struct bch_fs *c, unsigned idx)
 172 {
 173         struct bch_dev *ca;
 174         unsigned i, u64s = fs_usage_u64s(c);
 175
 176         BUG_ON(idx >= ARRAY_SIZE(c->usage));
 177
 178         preempt_disable();
 179         write_seqcount_begin(&c->usage_lock);
 180
 181         acc_u64s_percpu((u64 *) c->usage_base,
 182                         (u64 __percpu *) c->usage[idx], u64s);
 183         percpu_memset(c->usage[idx], 0, u64s * sizeof(u64));
 184
 185         rcu_read_lock();
 186         for_each_member_device_rcu(ca, c, i, NULL) {
 187                 u64s = dev_usage_u64s();
 188
 189                 acc_u64s_percpu((u64 *) ca->usage_base,
 190                                 (u64 __percpu *) ca->usage[idx], u64s);
 191                 percpu_memset(ca->usage[idx], 0, u64s * sizeof(u64));
 192         }
 193         rcu_read_unlock();
 194
 195         write_seqcount_end(&c->usage_lock);
 196         preempt_enable();
 197 }
 198
 199 void bch2_fs_usage_to_text(struct printbuf *out,
 200                            struct bch_fs *c,
 201                            struct bch_fs_usage_online *fs_usage)
 202 {
 203         unsigned i;
 204
 205         prt_printf(out, "capacity:\t\t\t%llu\n", c->capacity);
 206
 207         prt_printf(out, "hidden:\t\t\t\t%llu\n",
 208                fs_usage->u.hidden);
 209         prt_printf(out, "data:\t\t\t\t%llu\n",
 210                fs_usage->u.data);
 211         prt_printf(out, "cached:\t\t\t\t%llu\n",
 212                fs_usage->u.cached);
 213         prt_printf(out, "reserved:\t\t\t%llu\n",
 214                fs_usage->u.reserved);
 215         prt_printf(out, "nr_inodes:\t\t\t%llu\n",
 216                fs_usage->u.nr_inodes);
 217         prt_printf(out, "online reserved:\t\t%llu\n",
 218                fs_usage->online_reserved);
 219
 220         for (i = 0;
 221              i < ARRAY_SIZE(fs_usage->u.persistent_reserved);
 222              i++) {
 223                 prt_printf(out, "%u replicas:\n", i + 1);
 224                 prt_printf(out, "\treserved:\t\t%llu\n",
 225                        fs_usage->u.persistent_reserved[i]);
 226         }
 227
 228         for (i = 0; i < c->replicas.nr; i++) {
 229                 struct bch_replicas_entry *e =
 230                         cpu_replicas_entry(&c->replicas, i);
 231
 232                 prt_printf(out, "\t");
 233                 bch2_replicas_entry_to_text(out, e);
 234                 prt_printf(out, ":\t%llu\n", fs_usage->u.replicas[i]);
 235         }
 236 }
 237
 238 static u64 reserve_factor(u64 r)
 239 {
 240         return r + (round_up(r, (1 << RESERVE_FACTOR)) >> RESERVE_FACTOR);
 241 }
 242
 243 u64 bch2_fs_sectors_used(struct bch_fs *c, struct bch_fs_usage_online *fs_usage)
 244 {
 245         return min(fs_usage->u.hidden +
 246                    fs_usage->u.btree +
 247                    fs_usage->u.data +
 248                    reserve_factor(fs_usage->u.reserved +
 249                                   fs_usage->online_reserved),
 250                    c->capacity);
 251 }
 252
 253 static struct bch_fs_usage_short
 254 __bch2_fs_usage_read_short(struct bch_fs *c)
 255 {
 256         struct bch_fs_usage_short ret;
 257         u64 data, reserved;
 258
 259         ret.capacity = c->capacity -
 260                 bch2_fs_usage_read_one(c, &c->usage_base->hidden);
 261
 262         data            = bch2_fs_usage_read_one(c, &c->usage_base->data) +
 263                 bch2_fs_usage_read_one(c, &c->usage_base->btree);
 264         reserved        = bch2_fs_usage_read_one(c, &c->usage_base->reserved) +
 265                 percpu_u64_get(c->online_reserved);
 266
 267         ret.used        = min(ret.capacity, data + reserve_factor(reserved));
 268         ret.free        = ret.capacity - ret.used;
 269
 270         ret.nr_inodes   = bch2_fs_usage_read_one(c, &c->usage_base->nr_inodes);
 271
 272         return ret;
 273 }
 274
 275 struct bch_fs_usage_short
 276 bch2_fs_usage_read_short(struct bch_fs *c)
 277 {
 278         struct bch_fs_usage_short ret;
 279
 280         percpu_down_read(&c->mark_lock);
 281         ret = __bch2_fs_usage_read_short(c);
 282         percpu_up_read(&c->mark_lock);
 283
 284         return ret;
 285 }
 286
 287 void bch2_dev_usage_init(struct bch_dev *ca)
 288 {
 289         ca->usage_base->d[BCH_DATA_free].buckets = ca->mi.nbuckets - ca->mi.first_bucket;
 290 }
 291
 292 static inline int bucket_sectors_fragmented(struct bch_dev *ca,
 293                                             struct bch_alloc_v4 a)
 294 {
 295         return a.dirty_sectors
 296                 ? max(0, (int) ca->mi.bucket_size - (int) a.dirty_sectors)
 297                 : 0;
 298 }
 299
 300 static void bch2_dev_usage_update(struct bch_fs *c, struct bch_dev *ca,
 301                                   struct bch_alloc_v4 old,
 302                                   struct bch_alloc_v4 new,
 303                                   u64 journal_seq, bool gc)
 304 {
 305         struct bch_fs_usage *fs_usage;
 306         struct bch_dev_usage *u;
 307
 308         preempt_disable();
 309         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 310
 311         if (data_type_is_hidden(old.data_type))
 312                 fs_usage->hidden -= ca->mi.bucket_size;
 313         if (data_type_is_hidden(new.data_type))
 314                 fs_usage->hidden += ca->mi.bucket_size;
 315
 316         u = dev_usage_ptr(ca, journal_seq, gc);
 317
 318         u->d[old.data_type].buckets--;
 319         u->d[new.data_type].buckets++;
 320
 321         u->buckets_ec -= (int) !!old.stripe;
 322         u->buckets_ec += (int) !!new.stripe;
 323
 324         u->d[old.data_type].sectors -= old.dirty_sectors;
 325         u->d[new.data_type].sectors += new.dirty_sectors;
 326
 327         u->d[BCH_DATA_cached].sectors += new.cached_sectors;
 328         u->d[BCH_DATA_cached].sectors -= old.cached_sectors;
 329
 330         u->d[old.data_type].fragmented -= bucket_sectors_fragmented(ca, old);
 331         u->d[new.data_type].fragmented += bucket_sectors_fragmented(ca, new);
 332
 333         preempt_enable();
 334 }
 335
 336 static void bch2_dev_usage_update_m(struct bch_fs *c, struct bch_dev *ca,
 337                                     struct bucket old, struct bucket new,
 338                                     u64 journal_seq, bool gc)
 339 {
 340         struct bch_alloc_v4 old_a = {
 341                 .gen            = old.gen,
 342                 .data_type      = old.data_type,
 343                 .dirty_sectors  = old.dirty_sectors,
 344                 .cached_sectors = old.cached_sectors,
 345                 .stripe         = old.stripe,
 346         };
 347         struct bch_alloc_v4 new_a = {
 348                 .gen            = new.gen,
 349                 .data_type      = new.data_type,
 350                 .dirty_sectors  = new.dirty_sectors,
 351                 .cached_sectors = new.cached_sectors,
 352                 .stripe         = new.stripe,
 353         };
 354
 355         bch2_dev_usage_update(c, ca, old_a, new_a, journal_seq, gc);
 356 }
 357
 358 static inline int __update_replicas(struct bch_fs *c,
 359                                     struct bch_fs_usage *fs_usage,
 360                                     struct bch_replicas_entry *r,
 361                                     s64 sectors)
 362 {
 363         int idx = bch2_replicas_entry_idx(c, r);
 364
 365         if (idx < 0)
 366                 return -1;
 367
 368         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 369         fs_usage->replicas[idx]         += sectors;
 370         return 0;
 371 }
 372
 373 static inline int update_replicas(struct bch_fs *c, struct bkey_s_c k,
 374                         struct bch_replicas_entry *r, s64 sectors,
 375                         unsigned journal_seq, bool gc)
 376 {
 377         struct bch_fs_usage __percpu *fs_usage;
 378         int idx, ret = 0;
 379         struct printbuf buf = PRINTBUF;
 380
 381         percpu_down_read(&c->mark_lock);
 382         buf.atomic++;
 383
 384         idx = bch2_replicas_entry_idx(c, r);
 385         if (idx < 0 &&
 386             fsck_err(c, "no replicas entry\n"
 387                      "  while marking %s",
 388                      (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 389                 percpu_up_read(&c->mark_lock);
 390                 ret = bch2_mark_replicas(c, r);
 391                 percpu_down_read(&c->mark_lock);
 392
 393                 if (ret)
 394                         goto err;
 395                 idx = bch2_replicas_entry_idx(c, r);
 396         }
 397         if (idx < 0) {
 398                 ret = -1;
 399                 goto err;
 400         }
 401
 402         preempt_disable();
 403         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 404         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 405         fs_usage->replicas[idx]         += sectors;
 406         preempt_enable();
 407 err:
 408 fsck_err:
 409         percpu_up_read(&c->mark_lock);
 410         printbuf_exit(&buf);
 411         return ret;
 412 }
 413
 414 static inline int update_cached_sectors(struct bch_fs *c,
 415                         struct bkey_s_c k,
 416                         unsigned dev, s64 sectors,
 417                         unsigned journal_seq, bool gc)
 418 {
 419         struct bch_replicas_padded r;
 420
 421         bch2_replicas_entry_cached(&r.e, dev);
 422
 423         return update_replicas(c, k, &r.e, sectors, journal_seq, gc);
 424 }
 425
 426 static int __replicas_deltas_realloc(struct btree_trans *trans, unsigned more,
 427                                      gfp_t gfp)
 428 {
 429         struct replicas_delta_list *d = trans->fs_usage_deltas;
 430         unsigned new_size = d ? (d->size + more) * 2 : 128;
 431         unsigned alloc_size = sizeof(*d) + new_size;
 432
 433         WARN_ON_ONCE(alloc_size > REPLICAS_DELTA_LIST_MAX);
 434
 435         if (!d || d->used + more > d->size) {
 436                 d = krealloc(d, alloc_size, gfp|__GFP_ZERO);
 437
 438                 if (unlikely(!d)) {
 439                         if (alloc_size > REPLICAS_DELTA_LIST_MAX)
 440                                 return -ENOMEM;
 441
 442                         d = mempool_alloc(&trans->c->replicas_delta_pool, gfp);
 443                         if (!d)
 444                                 return -ENOMEM;
 445
 446                         memset(d, 0, REPLICAS_DELTA_LIST_MAX);
 447
 448                         if (trans->fs_usage_deltas)
 449                                 memcpy(d, trans->fs_usage_deltas,
 450                                        trans->fs_usage_deltas->size + sizeof(*d));
 451
 452                         new_size = REPLICAS_DELTA_LIST_MAX - sizeof(*d);
 453                         kfree(trans->fs_usage_deltas);
 454                 }
 455
 456                 d->size = new_size;
 457                 trans->fs_usage_deltas = d;
 458         }
 459
 460         return 0;
 461 }
 462
 463 static int replicas_deltas_realloc(struct btree_trans *trans, unsigned more)
 464 {
 465         return allocate_dropping_locks_errcode(trans,
 466                                 __replicas_deltas_realloc(trans, more, _gfp));
 467 }
 468
 469 static inline int update_replicas_list(struct btree_trans *trans,
 470                                         struct bch_replicas_entry *r,
 471                                         s64 sectors)
 472 {
 473         struct replicas_delta_list *d;
 474         struct replicas_delta *n;
 475         unsigned b;
 476         int ret;
 477
 478         if (!sectors)
 479                 return 0;
 480
 481         b = replicas_entry_bytes(r) + 8;
 482         ret = replicas_deltas_realloc(trans, b);
 483         if (ret)
 484                 return ret;
 485
 486         d = trans->fs_usage_deltas;
 487         n = (void *) d->d + d->used;
 488         n->delta = sectors;
 489         memcpy((void *) n + offsetof(struct replicas_delta, r),
 490                r, replicas_entry_bytes(r));
 491         bch2_replicas_entry_sort(&n->r);
 492         d->used += b;
 493         return 0;
 494 }
 495
 496 static inline int update_cached_sectors_list(struct btree_trans *trans,
 497                                               unsigned dev, s64 sectors)
 498 {
 499         struct bch_replicas_padded r;
 500
 501         bch2_replicas_entry_cached(&r.e, dev);
 502
 503         return update_replicas_list(trans, &r.e, sectors);
 504 }
 505
 506 int bch2_mark_alloc(struct btree_trans *trans,
 507                     enum btree_id btree, unsigned level,
 508                     struct bkey_s_c old, struct bkey_s_c new,
 509                     unsigned flags)
 510 {
 511         bool gc = flags & BTREE_TRIGGER_GC;
 512         u64 journal_seq = trans->journal_res.seq;
 513         u64 bucket_journal_seq;
 514         struct bch_fs *c = trans->c;
 515         struct bch_alloc_v4 old_a_convert, new_a_convert;
 516         const struct bch_alloc_v4 *old_a, *new_a;
 517         struct bch_dev *ca;
 518         int ret = 0;
 519
 520         /*
 521          * alloc btree is read in by bch2_alloc_read, not gc:
 522          */
 523         if ((flags & BTREE_TRIGGER_GC) &&
 524             !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
 525                 return 0;
 526
 527         if (bch2_trans_inconsistent_on(!bch2_dev_bucket_exists(c, new.k->p), trans,
 528                                        "alloc key for invalid device or bucket"))
 529                 return -EIO;
 530
 531         ca = bch_dev_bkey_exists(c, new.k->p.inode);
 532
 533         old_a = bch2_alloc_to_v4(old, &old_a_convert);
 534         new_a = bch2_alloc_to_v4(new, &new_a_convert);
 535
 536         bucket_journal_seq = new_a->journal_seq;
 537
 538         if ((flags & BTREE_TRIGGER_INSERT) &&
 539             data_type_is_empty(old_a->data_type) !=
 540             data_type_is_empty(new_a->data_type) &&
 541             new.k->type == KEY_TYPE_alloc_v4) {
 542                 struct bch_alloc_v4 *v = (struct bch_alloc_v4 *) new.v;
 543
 544                 EBUG_ON(!journal_seq);
 545
 546                 /*
 547                  * If the btree updates referring to a bucket weren't flushed
 548                  * before the bucket became empty again, then the we don't have
 549                  * to wait on a journal flush before we can reuse the bucket:
 550                  */
 551                 v->journal_seq = bucket_journal_seq =
 552                         data_type_is_empty(new_a->data_type) &&
 553                         (journal_seq == v->journal_seq ||
 554                          bch2_journal_noflush_seq(&c->journal, v->journal_seq))
 555                         ? 0 : journal_seq;
 556         }
 557
 558         if (!data_type_is_empty(old_a->data_type) &&
 559             data_type_is_empty(new_a->data_type) &&
 560             bucket_journal_seq) {
 561                 ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 562                                 c->journal.flushed_seq_ondisk,
 563                                 new.k->p.inode, new.k->p.offset,
 564                                 bucket_journal_seq);
 565                 if (ret) {
 566                         bch2_fs_fatal_error(c,
 567                                 "error setting bucket_needs_journal_commit: %i", ret);
 568                         return ret;
 569                 }
 570         }
 571
 572         percpu_down_read(&c->mark_lock);
 573         if (!gc && new_a->gen != old_a->gen)
 574                 *bucket_gen(ca, new.k->p.offset) = new_a->gen;
 575
 576         bch2_dev_usage_update(c, ca, *old_a, *new_a, journal_seq, gc);
 577
 578         if (gc) {
 579                 struct bucket *g = gc_bucket(ca, new.k->p.offset);
 580
 581                 bucket_lock(g);
 582
 583                 g->gen_valid            = 1;
 584                 g->gen                  = new_a->gen;
 585                 g->data_type            = new_a->data_type;
 586                 g->stripe               = new_a->stripe;
 587                 g->stripe_redundancy    = new_a->stripe_redundancy;
 588                 g->dirty_sectors        = new_a->dirty_sectors;
 589                 g->cached_sectors       = new_a->cached_sectors;
 590
 591                 bucket_unlock(g);
 592         }
 593         percpu_up_read(&c->mark_lock);
 594
 595         /*
 596          * need to know if we're getting called from the invalidate path or
 597          * not:
 598          */
 599
 600         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 601             old_a->cached_sectors) {
 602                 ret = update_cached_sectors(c, new, ca->dev_idx,
 603                                             -((s64) old_a->cached_sectors),
 604                                             journal_seq, gc);
 605                 if (ret) {
 606                         bch2_fs_fatal_error(c, "%s(): no replicas entry while updating cached sectors",
 607                                             __func__);
 608                         return ret;
 609                 }
 610         }
 611
 612         if (new_a->data_type == BCH_DATA_free &&
 613             (!new_a->journal_seq || new_a->journal_seq < c->journal.flushed_seq_ondisk))
 614                 closure_wake_up(&c->freelist_wait);
 615
 616         if (new_a->data_type == BCH_DATA_need_discard &&
 617             (!bucket_journal_seq || bucket_journal_seq < c->journal.flushed_seq_ondisk))
 618                 bch2_do_discards(c);
 619
 620         if (old_a->data_type != BCH_DATA_cached &&
 621             new_a->data_type == BCH_DATA_cached &&
 622             should_invalidate_buckets(ca, bch2_dev_usage_read(ca)))
 623                 bch2_do_invalidates(c);
 624
 625         if (new_a->data_type == BCH_DATA_need_gc_gens)
 626                 bch2_do_gc_gens(c);
 627
 628         return 0;
 629 }
 630
 631 int bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 632                               size_t b, enum bch_data_type data_type,
 633                               unsigned sectors, struct gc_pos pos,
 634                               unsigned flags)
 635 {
 636         struct bucket old, new, *g;
 637         int ret = 0;
 638
 639         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 640         BUG_ON(data_type != BCH_DATA_sb &&
 641                data_type != BCH_DATA_journal);
 642
 643         /*
 644          * Backup superblock might be past the end of our normal usable space:
 645          */
 646         if (b >= ca->mi.nbuckets)
 647                 return 0;
 648
 649         percpu_down_read(&c->mark_lock);
 650         g = gc_bucket(ca, b);
 651
 652         bucket_lock(g);
 653         old = *g;
 654
 655         if (bch2_fs_inconsistent_on(g->data_type &&
 656                         g->data_type != data_type, c,
 657                         "different types of data in same bucket: %s, %s",
 658                         bch2_data_types[g->data_type],
 659                         bch2_data_types[data_type])) {
 660                 ret = -EIO;
 661                 goto err;
 662         }
 663
 664         if (bch2_fs_inconsistent_on((u64) g->dirty_sectors + sectors > ca->mi.bucket_size, c,
 665                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > bucket size",
 666                         ca->dev_idx, b, g->gen,
 667                         bch2_data_types[g->data_type ?: data_type],
 668                         g->dirty_sectors, sectors)) {
 669                 ret = -EIO;
 670                 goto err;
 671         }
 672
 673
 674         g->data_type = data_type;
 675         g->dirty_sectors += sectors;
 676         new = *g;
 677 err:
 678         bucket_unlock(g);
 679         if (!ret)
 680                 bch2_dev_usage_update_m(c, ca, old, new, 0, true);
 681         percpu_up_read(&c->mark_lock);
 682         return ret;
 683 }
 684
 685 static int check_bucket_ref(struct btree_trans *trans,
 686                             struct bkey_s_c k,
 687                             const struct bch_extent_ptr *ptr,
 688                             s64 sectors, enum bch_data_type ptr_data_type,
 689                             u8 b_gen, u8 bucket_data_type,
 690                             u32 dirty_sectors, u32 cached_sectors)
 691 {
 692         struct bch_fs *c = trans->c;
 693         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 694         size_t bucket_nr = PTR_BUCKET_NR(ca, ptr);
 695         u16 bucket_sectors = !ptr->cached
 696                 ? dirty_sectors
 697                 : cached_sectors;
 698         struct printbuf buf = PRINTBUF;
 699         int ret = 0;
 700
 701         if (bucket_data_type == BCH_DATA_cached)
 702                 bucket_data_type = BCH_DATA_user;
 703
 704         if ((bucket_data_type == BCH_DATA_stripe && ptr_data_type == BCH_DATA_user) ||
 705             (bucket_data_type == BCH_DATA_user   && ptr_data_type == BCH_DATA_stripe))
 706                 bucket_data_type = ptr_data_type = BCH_DATA_stripe;
 707
 708         if (gen_after(ptr->gen, b_gen)) {
 709                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 710                         "bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\n"
 711                         "while marking %s",
 712                         ptr->dev, bucket_nr, b_gen,
 713                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 714                         ptr->gen,
 715                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 716                 ret = -EIO;
 717                 goto err;
 718         }
 719
 720         if (gen_cmp(b_gen, ptr->gen) > BUCKET_GC_GEN_MAX) {
 721                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 722                         "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 723                         "while marking %s",
 724                         ptr->dev, bucket_nr, b_gen,
 725                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 726                         ptr->gen,
 727                         (printbuf_reset(&buf),
 728                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 729                 ret = -EIO;
 730                 goto err;
 731         }
 732
 733         if (b_gen != ptr->gen && !ptr->cached) {
 734                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 735                         "bucket %u:%zu gen %u (mem gen %u) data type %s: stale dirty ptr (gen %u)\n"
 736                         "while marking %s",
 737                         ptr->dev, bucket_nr, b_gen,
 738                         *bucket_gen(ca, bucket_nr),
 739                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 740                         ptr->gen,
 741                         (printbuf_reset(&buf),
 742                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 743                 ret = -EIO;
 744                 goto err;
 745         }
 746
 747         if (b_gen != ptr->gen) {
 748                 ret = 1;
 749                 goto out;
 750         }
 751
 752         if (!data_type_is_empty(bucket_data_type) &&
 753             ptr_data_type &&
 754             bucket_data_type != ptr_data_type) {
 755                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 756                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 757                         "while marking %s",
 758                         ptr->dev, bucket_nr, b_gen,
 759                         bch2_data_types[bucket_data_type],
 760                         bch2_data_types[ptr_data_type],
 761                         (printbuf_reset(&buf),
 762                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 763                 ret = -EIO;
 764                 goto err;
 765         }
 766
 767         if ((unsigned) (bucket_sectors + sectors) > U32_MAX) {
 768                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 769                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U16_MAX\n"
 770                         "while marking %s",
 771                         ptr->dev, bucket_nr, b_gen,
 772                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 773                         bucket_sectors, sectors,
 774                         (printbuf_reset(&buf),
 775                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 776                 ret = -EIO;
 777                 goto err;
 778         }
 779 out:
 780         printbuf_exit(&buf);
 781         return ret;
 782 err:
 783         bch2_dump_trans_updates(trans);
 784         goto out;
 785 }
 786
 787 static int mark_stripe_bucket(struct btree_trans *trans,
 788                               struct bkey_s_c k,
 789                               unsigned ptr_idx,
 790                               unsigned flags)
 791 {
 792         struct bch_fs *c = trans->c;
 793         u64 journal_seq = trans->journal_res.seq;
 794         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 795         unsigned nr_data = s->nr_blocks - s->nr_redundant;
 796         bool parity = ptr_idx >= nr_data;
 797         enum bch_data_type data_type = parity ? BCH_DATA_parity : BCH_DATA_stripe;
 798         s64 sectors = parity ? le16_to_cpu(s->sectors) : 0;
 799         const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
 800         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 801         struct bucket old, new, *g;
 802         struct printbuf buf = PRINTBUF;
 803         int ret = 0;
 804
 805         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 806
 807         /* * XXX doesn't handle deletion */
 808
 809         percpu_down_read(&c->mark_lock);
 810         buf.atomic++;
 811         g = PTR_GC_BUCKET(ca, ptr);
 812
 813         if (g->dirty_sectors ||
 814             (g->stripe && g->stripe != k.k->p.offset)) {
 815                 bch2_fs_inconsistent(c,
 816                               "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
 817                               ptr->dev, PTR_BUCKET_NR(ca, ptr), g->gen,
 818                               (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 819                 ret = -EINVAL;
 820                 goto err;
 821         }
 822
 823         bucket_lock(g);
 824         old = *g;
 825
 826         ret = check_bucket_ref(trans, k, ptr, sectors, data_type,
 827                                g->gen, g->data_type,
 828                                g->dirty_sectors, g->cached_sectors);
 829         if (ret)
 830                 goto err;
 831
 832         g->data_type = data_type;
 833         g->dirty_sectors += sectors;
 834
 835         g->stripe               = k.k->p.offset;
 836         g->stripe_redundancy    = s->nr_redundant;
 837         new = *g;
 838 err:
 839         bucket_unlock(g);
 840         if (!ret)
 841                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 842         percpu_up_read(&c->mark_lock);
 843         printbuf_exit(&buf);
 844         return ret;
 845 }
 846
 847 static int __mark_pointer(struct btree_trans *trans,
 848                           struct bkey_s_c k,
 849                           const struct bch_extent_ptr *ptr,
 850                           s64 sectors, enum bch_data_type ptr_data_type,
 851                           u8 bucket_gen, u8 *bucket_data_type,
 852                           u32 *dirty_sectors, u32 *cached_sectors)
 853 {
 854         u32 *dst_sectors = !ptr->cached
 855                 ? dirty_sectors
 856                 : cached_sectors;
 857         int ret = check_bucket_ref(trans, k, ptr, sectors, ptr_data_type,
 858                                    bucket_gen, *bucket_data_type,
 859                                    *dirty_sectors, *cached_sectors);
 860
 861         if (ret)
 862                 return ret;
 863
 864         *dst_sectors += sectors;
 865         *bucket_data_type = *dirty_sectors || *cached_sectors
 866                 ? ptr_data_type : 0;
 867         return 0;
 868 }
 869
 870 static int bch2_mark_pointer(struct btree_trans *trans,
 871                              enum btree_id btree_id, unsigned level,
 872                              struct bkey_s_c k,
 873                              struct extent_ptr_decoded p,
 874                              s64 sectors,
 875                              unsigned flags)
 876 {
 877         u64 journal_seq = trans->journal_res.seq;
 878         struct bch_fs *c = trans->c;
 879         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 880         struct bucket old, new, *g;
 881         enum bch_data_type data_type = bkey_ptr_data_type(btree_id, level, k, p);
 882         u8 bucket_data_type;
 883         int ret = 0;
 884
 885         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 886
 887         percpu_down_read(&c->mark_lock);
 888         g = PTR_GC_BUCKET(ca, &p.ptr);
 889         bucket_lock(g);
 890         old = *g;
 891
 892         bucket_data_type = g->data_type;
 893         ret = __mark_pointer(trans, k, &p.ptr, sectors,
 894                              data_type, g->gen,
 895                              &bucket_data_type,
 896                              &g->dirty_sectors,
 897                              &g->cached_sectors);
 898         if (!ret)
 899                 g->data_type = bucket_data_type;
 900
 901         new = *g;
 902         bucket_unlock(g);
 903         if (!ret)
 904                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 905         percpu_up_read(&c->mark_lock);
 906
 907         return ret;
 908 }
 909
 910 static int bch2_mark_stripe_ptr(struct btree_trans *trans,
 911                                 struct bkey_s_c k,
 912                                 struct bch_extent_stripe_ptr p,
 913                                 enum bch_data_type data_type,
 914                                 s64 sectors,
 915                                 unsigned flags)
 916 {
 917         struct bch_fs *c = trans->c;
 918         struct bch_replicas_padded r;
 919         struct gc_stripe *m;
 920
 921         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 922
 923         m = genradix_ptr_alloc(&c->gc_stripes, p.idx, GFP_KERNEL);
 924         if (!m) {
 925                 bch_err(c, "error allocating memory for gc_stripes, idx %llu",
 926                         (u64) p.idx);
 927                 return -BCH_ERR_ENOMEM_mark_stripe_ptr;
 928         }
 929
 930         mutex_lock(&c->ec_stripes_heap_lock);
 931
 932         if (!m || !m->alive) {
 933                 mutex_unlock(&c->ec_stripes_heap_lock);
 934                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
 935                                     (u64) p.idx);
 936                 bch2_inconsistent_error(c);
 937                 return -EIO;
 938         }
 939
 940         m->block_sectors[p.block] += sectors;
 941
 942         r = m->r;
 943         mutex_unlock(&c->ec_stripes_heap_lock);
 944
 945         r.e.data_type = data_type;
 946         update_replicas(c, k, &r.e, sectors, trans->journal_res.seq, true);
 947
 948         return 0;
 949 }
 950
 951 int bch2_mark_extent(struct btree_trans *trans,
 952                      enum btree_id btree_id, unsigned level,
 953                      struct bkey_s_c old, struct bkey_s_c new,
 954                      unsigned flags)
 955 {
 956         u64 journal_seq = trans->journal_res.seq;
 957         struct bch_fs *c = trans->c;
 958         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old : new;
 959         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 960         const union bch_extent_entry *entry;
 961         struct extent_ptr_decoded p;
 962         struct bch_replicas_padded r;
 963         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
 964                 ? BCH_DATA_btree
 965                 : BCH_DATA_user;
 966         s64 sectors = bkey_is_btree_ptr(k.k)
 967                 ? btree_sectors(c)
 968                 : k.k->size;
 969         s64 dirty_sectors = 0;
 970         bool stale;
 971         int ret;
 972
 973         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 974
 975         r.e.data_type   = data_type;
 976         r.e.nr_devs     = 0;
 977         r.e.nr_required = 1;
 978
 979         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 980                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
 981
 982                 if (flags & BTREE_TRIGGER_OVERWRITE)
 983                         disk_sectors = -disk_sectors;
 984
 985                 ret = bch2_mark_pointer(trans, btree_id, level, k, p, disk_sectors, flags);
 986                 if (ret < 0)
 987                         return ret;
 988
 989                 stale = ret > 0;
 990
 991                 if (p.ptr.cached) {
 992                         if (!stale) {
 993                                 ret = update_cached_sectors(c, k, p.ptr.dev,
 994                                                 disk_sectors, journal_seq, true);
 995                                 if (ret) {
 996                                         bch2_fs_fatal_error(c, "%s(): no replicas entry while updating cached sectors",
 997                                                             __func__);
 998                                         return ret;
 999                                 }
1000                         }
1001                 } else if (!p.has_ec) {
1002                         dirty_sectors          += disk_sectors;
1003                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1004                 } else {
1005                         ret = bch2_mark_stripe_ptr(trans, k, p.ec, data_type,
1006                                         disk_sectors, flags);
1007                         if (ret)
1008                                 return ret;
1009
1010                         /*
1011                          * There may be other dirty pointers in this extent, but
1012                          * if so they're not required for mounting if we have an
1013                          * erasure coded pointer in this extent:
1014                          */
1015                         r.e.nr_required = 0;
1016                 }
1017         }
1018
1019         if (r.e.nr_devs) {
1020                 ret = update_replicas(c, k, &r.e, dirty_sectors, journal_seq, true);
1021                 if (ret) {
1022                         struct printbuf buf = PRINTBUF;
1023
1024                         bch2_bkey_val_to_text(&buf, c, k);
1025                         bch2_fs_fatal_error(c, "%s(): no replicas entry for %s", __func__, buf.buf);
1026                         printbuf_exit(&buf);
1027                         return ret;
1028                 }
1029         }
1030
1031         return 0;
1032 }
1033
1034 int bch2_mark_stripe(struct btree_trans *trans,
1035                      enum btree_id btree_id, unsigned level,
1036                      struct bkey_s_c old, struct bkey_s_c new,
1037                      unsigned flags)
1038 {
1039         bool gc = flags & BTREE_TRIGGER_GC;
1040         u64 journal_seq = trans->journal_res.seq;
1041         struct bch_fs *c = trans->c;
1042         u64 idx = new.k->p.offset;
1043         const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
1044                 ? bkey_s_c_to_stripe(old).v : NULL;
1045         const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
1046                 ? bkey_s_c_to_stripe(new).v : NULL;
1047         unsigned i;
1048         int ret;
1049
1050         BUG_ON(gc && old_s);
1051
1052         if (!gc) {
1053                 struct stripe *m = genradix_ptr(&c->stripes, idx);
1054
1055                 if (!m) {
1056                         struct printbuf buf1 = PRINTBUF;
1057                         struct printbuf buf2 = PRINTBUF;
1058
1059                         bch2_bkey_val_to_text(&buf1, c, old);
1060                         bch2_bkey_val_to_text(&buf2, c, new);
1061                         bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
1062                                             "old %s\n"
1063                                             "new %s", idx, buf1.buf, buf2.buf);
1064                         printbuf_exit(&buf2);
1065                         printbuf_exit(&buf1);
1066                         bch2_inconsistent_error(c);
1067                         return -1;
1068                 }
1069
1070                 if (!new_s) {
1071                         bch2_stripes_heap_del(c, m, idx);
1072
1073                         memset(m, 0, sizeof(*m));
1074                 } else {
1075                         m->sectors      = le16_to_cpu(new_s->sectors);
1076                         m->algorithm    = new_s->algorithm;
1077                         m->nr_blocks    = new_s->nr_blocks;
1078                         m->nr_redundant = new_s->nr_redundant;
1079                         m->blocks_nonempty = 0;
1080
1081                         for (i = 0; i < new_s->nr_blocks; i++)
1082                                 m->blocks_nonempty += !!stripe_blockcount_get(new_s, i);
1083
1084                         if (!old_s)
1085                                 bch2_stripes_heap_insert(c, m, idx);
1086                         else
1087                                 bch2_stripes_heap_update(c, m, idx);
1088                 }
1089         } else {
1090                 struct gc_stripe *m =
1091                         genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
1092
1093                 if (!m) {
1094                         bch_err(c, "error allocating memory for gc_stripes, idx %llu",
1095                                 idx);
1096                         return -BCH_ERR_ENOMEM_mark_stripe;
1097                 }
1098                 /*
1099                  * This will be wrong when we bring back runtime gc: we should
1100                  * be unmarking the old key and then marking the new key
1101                  */
1102                 m->alive        = true;
1103                 m->sectors      = le16_to_cpu(new_s->sectors);
1104                 m->nr_blocks    = new_s->nr_blocks;
1105                 m->nr_redundant = new_s->nr_redundant;
1106
1107                 for (i = 0; i < new_s->nr_blocks; i++)
1108                         m->ptrs[i] = new_s->ptrs[i];
1109
1110                 bch2_bkey_to_replicas(&m->r.e, new);
1111
1112                 /*
1113                  * gc recalculates this field from stripe ptr
1114                  * references:
1115                  */
1116                 memset(m->block_sectors, 0, sizeof(m->block_sectors));
1117
1118                 for (i = 0; i < new_s->nr_blocks; i++) {
1119                         ret = mark_stripe_bucket(trans, new, i, flags);
1120                         if (ret)
1121                                 return ret;
1122                 }
1123
1124                 ret = update_replicas(c, new, &m->r.e,
1125                                       ((s64) m->sectors * m->nr_redundant),
1126                                       journal_seq, gc);
1127                 if (ret) {
1128                         struct printbuf buf = PRINTBUF;
1129
1130                         bch2_bkey_val_to_text(&buf, c, new);
1131                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
1132                         printbuf_exit(&buf);
1133                         return ret;
1134                 }
1135         }
1136
1137         return 0;
1138 }
1139
1140 int bch2_mark_inode(struct btree_trans *trans,
1141                     enum btree_id btree_id, unsigned level,
1142                     struct bkey_s_c old, struct bkey_s_c new,
1143                     unsigned flags)
1144 {
1145         struct bch_fs *c = trans->c;
1146         struct bch_fs_usage __percpu *fs_usage;
1147         u64 journal_seq = trans->journal_res.seq;
1148
1149         if (flags & BTREE_TRIGGER_INSERT) {
1150                 struct bch_inode_v3 *v = (struct bch_inode_v3 *) new.v;
1151
1152                 BUG_ON(!journal_seq);
1153                 BUG_ON(new.k->type != KEY_TYPE_inode_v3);
1154
1155                 v->bi_journal_seq = cpu_to_le64(journal_seq);
1156         }
1157
1158         if (flags & BTREE_TRIGGER_GC) {
1159                 percpu_down_read(&c->mark_lock);
1160                 preempt_disable();
1161
1162                 fs_usage = fs_usage_ptr(c, journal_seq, flags & BTREE_TRIGGER_GC);
1163                 fs_usage->nr_inodes += bkey_is_inode(new.k);
1164                 fs_usage->nr_inodes -= bkey_is_inode(old.k);
1165
1166                 preempt_enable();
1167                 percpu_up_read(&c->mark_lock);
1168         }
1169         return 0;
1170 }
1171
1172 int bch2_mark_reservation(struct btree_trans *trans,
1173                           enum btree_id btree_id, unsigned level,
1174                           struct bkey_s_c old, struct bkey_s_c new,
1175                           unsigned flags)
1176 {
1177         struct bch_fs *c = trans->c;
1178         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old : new;
1179         struct bch_fs_usage __percpu *fs_usage;
1180         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1181         s64 sectors = (s64) k.k->size;
1182
1183         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1184
1185         if (flags & BTREE_TRIGGER_OVERWRITE)
1186                 sectors = -sectors;
1187         sectors *= replicas;
1188
1189         percpu_down_read(&c->mark_lock);
1190         preempt_disable();
1191
1192         fs_usage = fs_usage_ptr(c, trans->journal_res.seq, flags & BTREE_TRIGGER_GC);
1193         replicas = clamp_t(unsigned, replicas, 1,
1194                            ARRAY_SIZE(fs_usage->persistent_reserved));
1195
1196         fs_usage->reserved                              += sectors;
1197         fs_usage->persistent_reserved[replicas - 1]     += sectors;
1198
1199         preempt_enable();
1200         percpu_up_read(&c->mark_lock);
1201
1202         return 0;
1203 }
1204
1205 static s64 __bch2_mark_reflink_p(struct btree_trans *trans,
1206                                  struct bkey_s_c_reflink_p p,
1207                                  u64 start, u64 end,
1208                                  u64 *idx, unsigned flags, size_t r_idx)
1209 {
1210         struct bch_fs *c = trans->c;
1211         struct reflink_gc *r;
1212         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1213         u64 next_idx = end;
1214         s64 ret = 0;
1215         struct printbuf buf = PRINTBUF;
1216
1217         if (r_idx >= c->reflink_gc_nr)
1218                 goto not_found;
1219
1220         r = genradix_ptr(&c->reflink_gc_table, r_idx);
1221         next_idx = min(next_idx, r->offset - r->size);
1222         if (*idx < next_idx)
1223                 goto not_found;
1224
1225         BUG_ON((s64) r->refcount + add < 0);
1226
1227         r->refcount += add;
1228         *idx = r->offset;
1229         return 0;
1230 not_found:
1231         if (fsck_err(c, "pointer to missing indirect extent\n"
1232                      "  %s\n"
1233                      "  missing range %llu-%llu",
1234                      (bch2_bkey_val_to_text(&buf, c, p.s_c), buf.buf),
1235                      *idx, next_idx)) {
1236                 struct bkey_i_error *new;
1237
1238                 new = bch2_trans_kmalloc(trans, sizeof(*new));
1239                 ret = PTR_ERR_OR_ZERO(new);
1240                 if (ret)
1241                         goto err;
1242
1243                 bkey_init(&new->k);
1244                 new->k.type     = KEY_TYPE_error;
1245                 new->k.p                = bkey_start_pos(p.k);
1246                 new->k.p.offset += *idx - start;
1247                 bch2_key_resize(&new->k, next_idx - *idx);
1248                 ret = __bch2_btree_insert(trans, BTREE_ID_extents, &new->k_i,
1249                                           BTREE_TRIGGER_NORUN);
1250         }
1251
1252         *idx = next_idx;
1253 err:
1254 fsck_err:
1255         printbuf_exit(&buf);
1256         return ret;
1257 }
1258
1259 int bch2_mark_reflink_p(struct btree_trans *trans,
1260                         enum btree_id btree_id, unsigned level,
1261                         struct bkey_s_c old, struct bkey_s_c new,
1262                         unsigned flags)
1263 {
1264         struct bch_fs *c = trans->c;
1265         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old : new;
1266         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1267         struct reflink_gc *ref;
1268         size_t l, r, m;
1269         u64 idx = le64_to_cpu(p.v->idx), start = idx;
1270         u64 end = le64_to_cpu(p.v->idx) + p.k->size;
1271         int ret = 0;
1272
1273         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1274
1275         if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix) {
1276                 idx -= le32_to_cpu(p.v->front_pad);
1277                 end += le32_to_cpu(p.v->back_pad);
1278         }
1279
1280         l = 0;
1281         r = c->reflink_gc_nr;
1282         while (l < r) {
1283                 m = l + (r - l) / 2;
1284
1285                 ref = genradix_ptr(&c->reflink_gc_table, m);
1286                 if (ref->offset <= idx)
1287                         l = m + 1;
1288                 else
1289                         r = m;
1290         }
1291
1292         while (idx < end && !ret)
1293                 ret = __bch2_mark_reflink_p(trans, p, start, end,
1294                                             &idx, flags, l++);
1295
1296         return ret;
1297 }
1298
1299 void bch2_trans_fs_usage_revert(struct btree_trans *trans,
1300                                 struct replicas_delta_list *deltas)
1301 {
1302         struct bch_fs *c = trans->c;
1303         struct bch_fs_usage *dst;
1304         struct replicas_delta *d, *top = (void *) deltas->d + deltas->used;
1305         s64 added = 0;
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         /* revert changes: */
1313         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1314                 switch (d->r.data_type) {
1315                 case BCH_DATA_btree:
1316                 case BCH_DATA_user:
1317                 case BCH_DATA_parity:
1318                         added += d->delta;
1319                 }
1320                 BUG_ON(__update_replicas(c, dst, &d->r, -d->delta));
1321         }
1322
1323         dst->nr_inodes -= deltas->nr_inodes;
1324
1325         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1326                 added                           -= deltas->persistent_reserved[i];
1327                 dst->reserved                   -= deltas->persistent_reserved[i];
1328                 dst->persistent_reserved[i]     -= deltas->persistent_reserved[i];
1329         }
1330
1331         if (added > 0) {
1332                 trans->disk_res->sectors += added;
1333                 this_cpu_add(*c->online_reserved, added);
1334         }
1335
1336         preempt_enable();
1337         percpu_up_read(&c->mark_lock);
1338 }
1339
1340 int bch2_trans_fs_usage_apply(struct btree_trans *trans,
1341                               struct replicas_delta_list *deltas)
1342 {
1343         struct bch_fs *c = trans->c;
1344         static int warned_disk_usage = 0;
1345         bool warn = false;
1346         unsigned disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1347         struct replicas_delta *d = deltas->d, *d2;
1348         struct replicas_delta *top = (void *) deltas->d + deltas->used;
1349         struct bch_fs_usage *dst;
1350         s64 added = 0, should_not_have_added;
1351         unsigned i;
1352
1353         percpu_down_read(&c->mark_lock);
1354         preempt_disable();
1355         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1356
1357         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1358                 switch (d->r.data_type) {
1359                 case BCH_DATA_btree:
1360                 case BCH_DATA_user:
1361                 case BCH_DATA_parity:
1362                         added += d->delta;
1363                 }
1364
1365                 if (__update_replicas(c, dst, &d->r, d->delta))
1366                         goto need_mark;
1367         }
1368
1369         dst->nr_inodes += deltas->nr_inodes;
1370
1371         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1372                 added                           += deltas->persistent_reserved[i];
1373                 dst->reserved                   += deltas->persistent_reserved[i];
1374                 dst->persistent_reserved[i]     += deltas->persistent_reserved[i];
1375         }
1376
1377         /*
1378          * Not allowed to reduce sectors_available except by getting a
1379          * reservation:
1380          */
1381         should_not_have_added = added - (s64) disk_res_sectors;
1382         if (unlikely(should_not_have_added > 0)) {
1383                 u64 old, new, v = atomic64_read(&c->sectors_available);
1384
1385                 do {
1386                         old = v;
1387                         new = max_t(s64, 0, old - should_not_have_added);
1388                 } while ((v = atomic64_cmpxchg(&c->sectors_available,
1389                                                old, new)) != old);
1390
1391                 added -= should_not_have_added;
1392                 warn = true;
1393         }
1394
1395         if (added > 0) {
1396                 trans->disk_res->sectors -= added;
1397                 this_cpu_sub(*c->online_reserved, added);
1398         }
1399
1400         preempt_enable();
1401         percpu_up_read(&c->mark_lock);
1402
1403         if (unlikely(warn) && !xchg(&warned_disk_usage, 1))
1404                 bch2_trans_inconsistent(trans,
1405                                         "disk usage increased %lli more than %u sectors reserved)",
1406                                         should_not_have_added, disk_res_sectors);
1407         return 0;
1408 need_mark:
1409         /* revert changes: */
1410         for (d2 = deltas->d; d2 != d; d2 = replicas_delta_next(d2))
1411                 BUG_ON(__update_replicas(c, dst, &d2->r, -d2->delta));
1412
1413         preempt_enable();
1414         percpu_up_read(&c->mark_lock);
1415         return -1;
1416 }
1417
1418 /* trans_mark: */
1419
1420 static inline int bch2_trans_mark_pointer(struct btree_trans *trans,
1421                                    enum btree_id btree_id, unsigned level,
1422                                    struct bkey_s_c k, struct extent_ptr_decoded p,
1423                                    unsigned flags)
1424 {
1425         bool insert = !(flags & BTREE_TRIGGER_OVERWRITE);
1426         struct btree_iter iter;
1427         struct bkey_i_alloc_v4 *a;
1428         struct bpos bucket;
1429         struct bch_backpointer bp;
1430         s64 sectors;
1431         int ret;
1432
1433         bch2_extent_ptr_to_bp(trans->c, btree_id, level, k, p, &bucket, &bp);
1434         sectors = bp.bucket_len;
1435         if (!insert)
1436                 sectors = -sectors;
1437
1438         a = bch2_trans_start_alloc_update(trans, &iter, bucket);
1439         if (IS_ERR(a))
1440                 return PTR_ERR(a);
1441
1442         ret = __mark_pointer(trans, k, &p.ptr, sectors, bp.data_type,
1443                              a->v.gen, &a->v.data_type,
1444                              &a->v.dirty_sectors, &a->v.cached_sectors);
1445         if (ret)
1446                 goto err;
1447
1448         if (!p.ptr.cached) {
1449                 ret = bch2_bucket_backpointer_mod(trans, bucket, bp, k, insert);
1450                 if (ret)
1451                         goto err;
1452         }
1453
1454         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1455 err:
1456         bch2_trans_iter_exit(trans, &iter);
1457         return ret;
1458 }
1459
1460 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1461                         struct extent_ptr_decoded p,
1462                         s64 sectors, enum bch_data_type data_type)
1463 {
1464         struct btree_iter iter;
1465         struct bkey_i_stripe *s;
1466         struct bch_replicas_padded r;
1467         int ret = 0;
1468
1469         s = bch2_bkey_get_mut_typed(trans, &iter,
1470                         BTREE_ID_stripes, POS(0, p.ec.idx),
1471                         BTREE_ITER_WITH_UPDATES, stripe);
1472         ret = PTR_ERR_OR_ZERO(s);
1473         if (unlikely(ret)) {
1474                 bch2_trans_inconsistent_on(bch2_err_matches(ret, ENOENT), trans,
1475                         "pointer to nonexistent stripe %llu",
1476                         (u64) p.ec.idx);
1477                 goto err;
1478         }
1479
1480         if (!bch2_ptr_matches_stripe(&s->v, p)) {
1481                 bch2_trans_inconsistent(trans,
1482                         "stripe pointer doesn't match stripe %llu",
1483                         (u64) p.ec.idx);
1484                 ret = -EIO;
1485                 goto err;
1486         }
1487
1488         stripe_blockcount_set(&s->v, p.ec.block,
1489                 stripe_blockcount_get(&s->v, p.ec.block) +
1490                 sectors);
1491
1492         bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1493         r.e.data_type = data_type;
1494         ret = update_replicas_list(trans, &r.e, sectors);
1495 err:
1496         bch2_trans_iter_exit(trans, &iter);
1497         return ret;
1498 }
1499
1500 int bch2_trans_mark_extent(struct btree_trans *trans,
1501                            enum btree_id btree_id, unsigned level,
1502                            struct bkey_s_c old, struct bkey_i *new,
1503                            unsigned flags)
1504 {
1505         struct bch_fs *c = trans->c;
1506         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1507                 ? old
1508                 : bkey_i_to_s_c(new);
1509         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1510         const union bch_extent_entry *entry;
1511         struct extent_ptr_decoded p;
1512         struct bch_replicas_padded r;
1513         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
1514                 ? BCH_DATA_btree
1515                 : BCH_DATA_user;
1516         s64 sectors = bkey_is_btree_ptr(k.k)
1517                 ? btree_sectors(c)
1518                 : k.k->size;
1519         s64 dirty_sectors = 0;
1520         bool stale;
1521         int ret = 0;
1522
1523         r.e.data_type   = data_type;
1524         r.e.nr_devs     = 0;
1525         r.e.nr_required = 1;
1526
1527         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1528                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
1529
1530                 if (flags & BTREE_TRIGGER_OVERWRITE)
1531                         disk_sectors = -disk_sectors;
1532
1533                 ret = bch2_trans_mark_pointer(trans, btree_id, level, k, p, flags);
1534                 if (ret < 0)
1535                         return ret;
1536
1537                 stale = ret > 0;
1538
1539                 if (p.ptr.cached) {
1540                         if (!stale) {
1541                                 ret = update_cached_sectors_list(trans, p.ptr.dev,
1542                                                                  disk_sectors);
1543                                 if (ret)
1544                                         return ret;
1545                         }
1546                 } else if (!p.has_ec) {
1547                         dirty_sectors          += disk_sectors;
1548                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1549                 } else {
1550                         ret = bch2_trans_mark_stripe_ptr(trans, p,
1551                                         disk_sectors, data_type);
1552                         if (ret)
1553                                 return ret;
1554
1555                         r.e.nr_required = 0;
1556                 }
1557         }
1558
1559         if (r.e.nr_devs)
1560                 ret = update_replicas_list(trans, &r.e, dirty_sectors);
1561
1562         return ret;
1563 }
1564
1565 static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
1566                                          struct bkey_s_c_stripe s,
1567                                          unsigned idx, bool deleting)
1568 {
1569         struct bch_fs *c = trans->c;
1570         const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
1571         struct btree_iter iter;
1572         struct bkey_i_alloc_v4 *a;
1573         enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
1574                 ? BCH_DATA_parity : 0;
1575         s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
1576         int ret = 0;
1577
1578         if (deleting)
1579                 sectors = -sectors;
1580
1581         a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(c, ptr));
1582         if (IS_ERR(a))
1583                 return PTR_ERR(a);
1584
1585         ret = check_bucket_ref(trans, s.s_c, ptr, sectors, data_type,
1586                                a->v.gen, a->v.data_type,
1587                                a->v.dirty_sectors, a->v.cached_sectors);
1588         if (ret)
1589                 goto err;
1590
1591         if (!deleting) {
1592                 if (bch2_trans_inconsistent_on(a->v.stripe ||
1593                                                a->v.stripe_redundancy, trans,
1594                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
1595                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1596                                 bch2_data_types[a->v.data_type],
1597                                 a->v.dirty_sectors,
1598                                 a->v.stripe, s.k->p.offset)) {
1599                         ret = -EIO;
1600                         goto err;
1601                 }
1602
1603                 if (bch2_trans_inconsistent_on(data_type && a->v.dirty_sectors, trans,
1604                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
1605                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1606                                 bch2_data_types[a->v.data_type],
1607                                 a->v.dirty_sectors,
1608                                 s.k->p.offset)) {
1609                         ret = -EIO;
1610                         goto err;
1611                 }
1612
1613                 a->v.stripe             = s.k->p.offset;
1614                 a->v.stripe_redundancy  = s.v->nr_redundant;
1615                 a->v.data_type          = BCH_DATA_stripe;
1616         } else {
1617                 if (bch2_trans_inconsistent_on(a->v.stripe != s.k->p.offset ||
1618                                                a->v.stripe_redundancy != s.v->nr_redundant, trans,
1619                                 "bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
1620                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1621                                 s.k->p.offset, a->v.stripe)) {
1622                         ret = -EIO;
1623                         goto err;
1624                 }
1625
1626                 a->v.stripe             = 0;
1627                 a->v.stripe_redundancy  = 0;
1628                 a->v.data_type          = alloc_data_type(a->v, BCH_DATA_user);
1629         }
1630
1631         a->v.dirty_sectors += sectors;
1632         if (data_type)
1633                 a->v.data_type = !deleting ? data_type : 0;
1634
1635         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1636         if (ret)
1637                 goto err;
1638 err:
1639         bch2_trans_iter_exit(trans, &iter);
1640         return ret;
1641 }
1642
1643 int bch2_trans_mark_stripe(struct btree_trans *trans,
1644                            enum btree_id btree_id, unsigned level,
1645                            struct bkey_s_c old, struct bkey_i *new,
1646                            unsigned flags)
1647 {
1648         const struct bch_stripe *old_s = NULL;
1649         struct bch_stripe *new_s = NULL;
1650         struct bch_replicas_padded r;
1651         unsigned i, nr_blocks;
1652         int ret = 0;
1653
1654         if (old.k->type == KEY_TYPE_stripe)
1655                 old_s = bkey_s_c_to_stripe(old).v;
1656         if (new->k.type == KEY_TYPE_stripe)
1657                 new_s = &bkey_i_to_stripe(new)->v;
1658
1659         /*
1660          * If the pointers aren't changing, we don't need to do anything:
1661          */
1662         if (new_s && old_s &&
1663             new_s->nr_blocks    == old_s->nr_blocks &&
1664             new_s->nr_redundant == old_s->nr_redundant &&
1665             !memcmp(old_s->ptrs, new_s->ptrs,
1666                     new_s->nr_blocks * sizeof(struct bch_extent_ptr)))
1667                 return 0;
1668
1669         BUG_ON(new_s && old_s &&
1670                (new_s->nr_blocks        != old_s->nr_blocks ||
1671                 new_s->nr_redundant     != old_s->nr_redundant));
1672
1673         nr_blocks = new_s ? new_s->nr_blocks : old_s->nr_blocks;
1674
1675         if (new_s) {
1676                 s64 sectors = le16_to_cpu(new_s->sectors);
1677
1678                 bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(new));
1679                 ret = update_replicas_list(trans, &r.e, sectors * new_s->nr_redundant);
1680                 if (ret)
1681                         return ret;
1682         }
1683
1684         if (old_s) {
1685                 s64 sectors = -((s64) le16_to_cpu(old_s->sectors));
1686
1687                 bch2_bkey_to_replicas(&r.e, old);
1688                 ret = update_replicas_list(trans, &r.e, sectors * old_s->nr_redundant);
1689                 if (ret)
1690                         return ret;
1691         }
1692
1693         for (i = 0; i < nr_blocks; i++) {
1694                 if (new_s && old_s &&
1695                     !memcmp(&new_s->ptrs[i],
1696                             &old_s->ptrs[i],
1697                             sizeof(new_s->ptrs[i])))
1698                         continue;
1699
1700                 if (new_s) {
1701                         ret = bch2_trans_mark_stripe_bucket(trans,
1702                                         bkey_i_to_s_c_stripe(new), i, false);
1703                         if (ret)
1704                                 break;
1705                 }
1706
1707                 if (old_s) {
1708                         ret = bch2_trans_mark_stripe_bucket(trans,
1709                                         bkey_s_c_to_stripe(old), i, true);
1710                         if (ret)
1711                                 break;
1712                 }
1713         }
1714
1715         return ret;
1716 }
1717
1718 int bch2_trans_mark_inode(struct btree_trans *trans,
1719                           enum btree_id btree_id, unsigned level,
1720                           struct bkey_s_c old,
1721                           struct bkey_i *new,
1722                           unsigned flags)
1723 {
1724         int nr = bkey_is_inode(&new->k) - bkey_is_inode(old.k);
1725
1726         if (nr) {
1727                 int ret = replicas_deltas_realloc(trans, 0);
1728                 struct replicas_delta_list *d = trans->fs_usage_deltas;
1729
1730                 if (ret)
1731                         return ret;
1732
1733                 d->nr_inodes += nr;
1734         }
1735
1736         return 0;
1737 }
1738
1739 int bch2_trans_mark_reservation(struct btree_trans *trans,
1740                                 enum btree_id btree_id, unsigned level,
1741                                 struct bkey_s_c old,
1742                                 struct bkey_i *new,
1743                                 unsigned flags)
1744 {
1745         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1746                 ? old
1747                 : bkey_i_to_s_c(new);
1748         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1749         s64 sectors = (s64) k.k->size;
1750         struct replicas_delta_list *d;
1751         int ret;
1752
1753         if (flags & BTREE_TRIGGER_OVERWRITE)
1754                 sectors = -sectors;
1755         sectors *= replicas;
1756
1757         ret = replicas_deltas_realloc(trans, 0);
1758         if (ret)
1759                 return ret;
1760
1761         d = trans->fs_usage_deltas;
1762         replicas = clamp_t(unsigned, replicas, 1,
1763                            ARRAY_SIZE(d->persistent_reserved));
1764
1765         d->persistent_reserved[replicas - 1] += sectors;
1766         return 0;
1767 }
1768
1769 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1770                         struct bkey_s_c_reflink_p p,
1771                         u64 *idx, unsigned flags)
1772 {
1773         struct bch_fs *c = trans->c;
1774         struct btree_iter iter;
1775         struct bkey_i *k;
1776         __le64 *refcount;
1777         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1778         struct printbuf buf = PRINTBUF;
1779         int ret;
1780
1781         k = bch2_bkey_get_mut_noupdate(trans, &iter,
1782                         BTREE_ID_reflink, POS(0, *idx),
1783                         BTREE_ITER_WITH_UPDATES);
1784         ret = PTR_ERR_OR_ZERO(k);
1785         if (ret)
1786                 goto err;
1787
1788         refcount = bkey_refcount(k);
1789         if (!refcount) {
1790                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1791                 bch2_trans_inconsistent(trans,
1792                         "nonexistent indirect extent at %llu while marking\n  %s",
1793                         *idx, buf.buf);
1794                 ret = -EIO;
1795                 goto err;
1796         }
1797
1798         if (!*refcount && (flags & BTREE_TRIGGER_OVERWRITE)) {
1799                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1800                 bch2_trans_inconsistent(trans,
1801                         "indirect extent refcount underflow at %llu while marking\n  %s",
1802                         *idx, buf.buf);
1803                 ret = -EIO;
1804                 goto err;
1805         }
1806
1807         if (flags & BTREE_TRIGGER_INSERT) {
1808                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1809                 u64 pad;
1810
1811                 pad = max_t(s64, le32_to_cpu(v->front_pad),
1812                             le64_to_cpu(v->idx) - bkey_start_offset(&k->k));
1813                 BUG_ON(pad > U32_MAX);
1814                 v->front_pad = cpu_to_le32(pad);
1815
1816                 pad = max_t(s64, le32_to_cpu(v->back_pad),
1817                             k->k.p.offset - p.k->size - le64_to_cpu(v->idx));
1818                 BUG_ON(pad > U32_MAX);
1819                 v->back_pad = cpu_to_le32(pad);
1820         }
1821
1822         le64_add_cpu(refcount, add);
1823
1824         bch2_btree_iter_set_pos_to_extent_start(&iter);
1825         ret = bch2_trans_update(trans, &iter, k, 0);
1826         if (ret)
1827                 goto err;
1828
1829         *idx = k->k.p.offset;
1830 err:
1831         bch2_trans_iter_exit(trans, &iter);
1832         printbuf_exit(&buf);
1833         return ret;
1834 }
1835
1836 int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1837                               enum btree_id btree_id, unsigned level,
1838                               struct bkey_s_c old,
1839                               struct bkey_i *new,
1840                               unsigned flags)
1841 {
1842         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE
1843                 ? old
1844                 : bkey_i_to_s_c(new);
1845         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1846         u64 idx, end_idx;
1847         int ret = 0;
1848
1849         if (flags & BTREE_TRIGGER_INSERT) {
1850                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1851
1852                 v->front_pad = v->back_pad = 0;
1853         }
1854
1855         idx     = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
1856         end_idx = le64_to_cpu(p.v->idx) + p.k->size +
1857                 le32_to_cpu(p.v->back_pad);
1858
1859         while (idx < end_idx && !ret)
1860                 ret = __bch2_trans_mark_reflink_p(trans, p, &idx, flags);
1861
1862         return ret;
1863 }
1864
1865 static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1866                                     struct bch_dev *ca, size_t b,
1867                                     enum bch_data_type type,
1868                                     unsigned sectors)
1869 {
1870         struct bch_fs *c = trans->c;
1871         struct btree_iter iter;
1872         struct bkey_i_alloc_v4 *a;
1873         int ret = 0;
1874
1875         /*
1876          * Backup superblock might be past the end of our normal usable space:
1877          */
1878         if (b >= ca->mi.nbuckets)
1879                 return 0;
1880
1881         a = bch2_trans_start_alloc_update(trans, &iter, POS(ca->dev_idx, b));
1882         if (IS_ERR(a))
1883                 return PTR_ERR(a);
1884
1885         if (a->v.data_type && type && a->v.data_type != type) {
1886                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1887                         "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
1888                         "while marking %s",
1889                         iter.pos.inode, iter.pos.offset, a->v.gen,
1890                         bch2_data_types[a->v.data_type],
1891                         bch2_data_types[type],
1892                         bch2_data_types[type]);
1893                 ret = -EIO;
1894                 goto out;
1895         }
1896
1897         a->v.data_type          = type;
1898         a->v.dirty_sectors      = sectors;
1899
1900         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1901         if (ret)
1902                 goto out;
1903 out:
1904         bch2_trans_iter_exit(trans, &iter);
1905         return ret;
1906 }
1907
1908 int bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1909                                     struct bch_dev *ca, size_t b,
1910                                     enum bch_data_type type,
1911                                     unsigned sectors)
1912 {
1913         return commit_do(trans, NULL, NULL, 0,
1914                         __bch2_trans_mark_metadata_bucket(trans, ca, b, type, sectors));
1915 }
1916
1917 static int bch2_trans_mark_metadata_sectors(struct btree_trans *trans,
1918                                             struct bch_dev *ca,
1919                                             u64 start, u64 end,
1920                                             enum bch_data_type type,
1921                                             u64 *bucket, unsigned *bucket_sectors)
1922 {
1923         do {
1924                 u64 b = sector_to_bucket(ca, start);
1925                 unsigned sectors =
1926                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1927
1928                 if (b != *bucket && *bucket_sectors) {
1929                         int ret = bch2_trans_mark_metadata_bucket(trans, ca, *bucket,
1930                                                                   type, *bucket_sectors);
1931                         if (ret)
1932                                 return ret;
1933
1934                         *bucket_sectors = 0;
1935                 }
1936
1937                 *bucket         = b;
1938                 *bucket_sectors += sectors;
1939                 start += sectors;
1940         } while (start < end);
1941
1942         return 0;
1943 }
1944
1945 static int __bch2_trans_mark_dev_sb(struct btree_trans *trans,
1946                                     struct bch_dev *ca)
1947 {
1948         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1949         u64 bucket = 0;
1950         unsigned i, bucket_sectors = 0;
1951         int ret;
1952
1953         for (i = 0; i < layout->nr_superblocks; i++) {
1954                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1955
1956                 if (offset == BCH_SB_SECTOR) {
1957                         ret = bch2_trans_mark_metadata_sectors(trans, ca,
1958                                                 0, BCH_SB_SECTOR,
1959                                                 BCH_DATA_sb, &bucket, &bucket_sectors);
1960                         if (ret)
1961                                 return ret;
1962                 }
1963
1964                 ret = bch2_trans_mark_metadata_sectors(trans, ca, offset,
1965                                       offset + (1 << layout->sb_max_size_bits),
1966                                       BCH_DATA_sb, &bucket, &bucket_sectors);
1967                 if (ret)
1968                         return ret;
1969         }
1970
1971         if (bucket_sectors) {
1972                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1973                                 bucket, BCH_DATA_sb, bucket_sectors);
1974                 if (ret)
1975                         return ret;
1976         }
1977
1978         for (i = 0; i < ca->journal.nr; i++) {
1979                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1980                                 ca->journal.buckets[i],
1981                                 BCH_DATA_journal, ca->mi.bucket_size);
1982                 if (ret)
1983                         return ret;
1984         }
1985
1986         return 0;
1987 }
1988
1989 int bch2_trans_mark_dev_sb(struct bch_fs *c, struct bch_dev *ca)
1990 {
1991         return bch2_trans_run(c, __bch2_trans_mark_dev_sb(&trans, ca));
1992 }
1993
1994 /* Disk reservations: */
1995
1996 #define SECTORS_CACHE   1024
1997
1998 int __bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
1999                               u64 sectors, int flags)
2000 {
2001         struct bch_fs_pcpu *pcpu;
2002         u64 old, v, get;
2003         s64 sectors_available;
2004         int ret;
2005
2006         percpu_down_read(&c->mark_lock);
2007         preempt_disable();
2008         pcpu = this_cpu_ptr(c->pcpu);
2009
2010         if (sectors <= pcpu->sectors_available)
2011                 goto out;
2012
2013         v = atomic64_read(&c->sectors_available);
2014         do {
2015                 old = v;
2016                 get = min((u64) sectors + SECTORS_CACHE, old);
2017
2018                 if (get < sectors) {
2019                         preempt_enable();
2020                         goto recalculate;
2021                 }
2022         } while ((v = atomic64_cmpxchg(&c->sectors_available,
2023                                        old, old - get)) != old);
2024
2025         pcpu->sectors_available         += get;
2026
2027 out:
2028         pcpu->sectors_available         -= sectors;
2029         this_cpu_add(*c->online_reserved, sectors);
2030         res->sectors                    += sectors;
2031
2032         preempt_enable();
2033         percpu_up_read(&c->mark_lock);
2034         return 0;
2035
2036 recalculate:
2037         mutex_lock(&c->sectors_available_lock);
2038
2039         percpu_u64_set(&c->pcpu->sectors_available, 0);
2040         sectors_available = avail_factor(__bch2_fs_usage_read_short(c).free);
2041
2042         if (sectors <= sectors_available ||
2043             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
2044                 atomic64_set(&c->sectors_available,
2045                              max_t(s64, 0, sectors_available - sectors));
2046                 this_cpu_add(*c->online_reserved, sectors);
2047                 res->sectors                    += sectors;
2048                 ret = 0;
2049         } else {
2050                 atomic64_set(&c->sectors_available, sectors_available);
2051                 ret = -BCH_ERR_ENOSPC_disk_reservation;
2052         }
2053
2054         mutex_unlock(&c->sectors_available_lock);
2055         percpu_up_read(&c->mark_lock);
2056
2057         return ret;
2058 }
2059
2060 /* Startup/shutdown: */
2061
2062 static void bucket_gens_free_rcu(struct rcu_head *rcu)
2063 {
2064         struct bucket_gens *buckets =
2065                 container_of(rcu, struct bucket_gens, rcu);
2066
2067         kvpfree(buckets, sizeof(*buckets) + buckets->nbuckets);
2068 }
2069
2070 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
2071 {
2072         struct bucket_gens *bucket_gens = NULL, *old_bucket_gens = NULL;
2073         unsigned long *buckets_nouse = NULL;
2074         bool resize = ca->bucket_gens != NULL;
2075         int ret;
2076
2077         if (!(bucket_gens       = kvpmalloc(sizeof(struct bucket_gens) + nbuckets,
2078                                             GFP_KERNEL|__GFP_ZERO))) {
2079                 ret = -BCH_ERR_ENOMEM_bucket_gens;
2080                 goto err;
2081         }
2082
2083         if ((c->opts.buckets_nouse &&
2084              !(buckets_nouse    = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2085                                             sizeof(unsigned long),
2086                                             GFP_KERNEL|__GFP_ZERO)))) {
2087                 ret = -BCH_ERR_ENOMEM_buckets_nouse;
2088                 goto err;
2089         }
2090
2091         bucket_gens->first_bucket = ca->mi.first_bucket;
2092         bucket_gens->nbuckets   = nbuckets;
2093
2094         bch2_copygc_stop(c);
2095
2096         if (resize) {
2097                 down_write(&c->gc_lock);
2098                 down_write(&ca->bucket_lock);
2099                 percpu_down_write(&c->mark_lock);
2100         }
2101
2102         old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
2103
2104         if (resize) {
2105                 size_t n = min(bucket_gens->nbuckets, old_bucket_gens->nbuckets);
2106
2107                 memcpy(bucket_gens->b,
2108                        old_bucket_gens->b,
2109                        n);
2110                 if (buckets_nouse)
2111                         memcpy(buckets_nouse,
2112                                ca->buckets_nouse,
2113                                BITS_TO_LONGS(n) * sizeof(unsigned long));
2114         }
2115
2116         rcu_assign_pointer(ca->bucket_gens, bucket_gens);
2117         bucket_gens     = old_bucket_gens;
2118
2119         swap(ca->buckets_nouse, buckets_nouse);
2120
2121         nbuckets = ca->mi.nbuckets;
2122
2123         if (resize) {
2124                 percpu_up_write(&c->mark_lock);
2125                 up_write(&ca->bucket_lock);
2126                 up_write(&c->gc_lock);
2127         }
2128
2129         ret = 0;
2130 err:
2131         kvpfree(buckets_nouse,
2132                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2133         if (bucket_gens)
2134                 call_rcu(&bucket_gens->rcu, bucket_gens_free_rcu);
2135
2136         return ret;
2137 }
2138
2139 void bch2_dev_buckets_free(struct bch_dev *ca)
2140 {
2141         unsigned i;
2142
2143         kvpfree(ca->buckets_nouse,
2144                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2145         kvpfree(rcu_dereference_protected(ca->bucket_gens, 1),
2146                 sizeof(struct bucket_gens) + ca->mi.nbuckets);
2147
2148         for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
2149                 free_percpu(ca->usage[i]);
2150         kfree(ca->usage_base);
2151 }
2152
2153 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2154 {
2155         unsigned i;
2156
2157         ca->usage_base = kzalloc(sizeof(struct bch_dev_usage), GFP_KERNEL);
2158         if (!ca->usage_base)
2159                 return -BCH_ERR_ENOMEM_usage_init;
2160
2161         for (i = 0; i < ARRAY_SIZE(ca->usage); i++) {
2162                 ca->usage[i] = alloc_percpu(struct bch_dev_usage);
2163                 if (!ca->usage[i])
2164                         return -BCH_ERR_ENOMEM_usage_init;
2165         }
2166
2167         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);
2168 }