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 u64 bch2_fs_usage_read_one(struct bch_fs *c, u64 *v)
 106 {
 107         ssize_t offset = v - (u64 *) c->usage_base;
 108         unsigned i, seq;
 109         u64 ret;
 110
 111         BUG_ON(offset < 0 || offset >= fs_usage_u64s(c));
 112         percpu_rwsem_assert_held(&c->mark_lock);
 113
 114         do {
 115                 seq = read_seqcount_begin(&c->usage_lock);
 116                 ret = *v;
 117
 118                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 119                         ret += percpu_u64_get((u64 __percpu *) c->usage[i] + offset);
 120         } while (read_seqcount_retry(&c->usage_lock, seq));
 121
 122         return ret;
 123 }
 124
 125 struct bch_fs_usage_online *bch2_fs_usage_read(struct bch_fs *c)
 126 {
 127         struct bch_fs_usage_online *ret;
 128         unsigned nr_replicas = READ_ONCE(c->replicas.nr);
 129         unsigned seq, i;
 130 retry:
 131         ret = kmalloc(__fs_usage_online_u64s(nr_replicas) * sizeof(u64), GFP_KERNEL);
 132         if (unlikely(!ret))
 133                 return NULL;
 134
 135         percpu_down_read(&c->mark_lock);
 136
 137         if (nr_replicas != c->replicas.nr) {
 138                 nr_replicas = c->replicas.nr;
 139                 percpu_up_read(&c->mark_lock);
 140                 kfree(ret);
 141                 goto retry;
 142         }
 143
 144         ret->online_reserved = percpu_u64_get(c->online_reserved);
 145
 146         do {
 147                 seq = read_seqcount_begin(&c->usage_lock);
 148                 unsafe_memcpy(&ret->u, c->usage_base,
 149                               __fs_usage_u64s(nr_replicas) * sizeof(u64),
 150                               "embedded variable length struct");
 151                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 152                         acc_u64s_percpu((u64 *) &ret->u, (u64 __percpu *) c->usage[i],
 153                                         __fs_usage_u64s(nr_replicas));
 154         } while (read_seqcount_retry(&c->usage_lock, seq));
 155
 156         return ret;
 157 }
 158
 159 void bch2_fs_usage_acc_to_base(struct bch_fs *c, unsigned idx)
 160 {
 161         struct bch_dev *ca;
 162         unsigned i, u64s = fs_usage_u64s(c);
 163
 164         BUG_ON(idx >= ARRAY_SIZE(c->usage));
 165
 166         preempt_disable();
 167         write_seqcount_begin(&c->usage_lock);
 168
 169         acc_u64s_percpu((u64 *) c->usage_base,
 170                         (u64 __percpu *) c->usage[idx], u64s);
 171         percpu_memset(c->usage[idx], 0, u64s * sizeof(u64));
 172
 173         rcu_read_lock();
 174         for_each_member_device_rcu(ca, c, i, NULL) {
 175                 u64s = dev_usage_u64s();
 176
 177                 acc_u64s_percpu((u64 *) ca->usage_base,
 178                                 (u64 __percpu *) ca->usage[idx], u64s);
 179                 percpu_memset(ca->usage[idx], 0, u64s * sizeof(u64));
 180         }
 181         rcu_read_unlock();
 182
 183         write_seqcount_end(&c->usage_lock);
 184         preempt_enable();
 185 }
 186
 187 void bch2_fs_usage_to_text(struct printbuf *out,
 188                            struct bch_fs *c,
 189                            struct bch_fs_usage_online *fs_usage)
 190 {
 191         unsigned i;
 192
 193         prt_printf(out, "capacity:\t\t\t%llu\n", c->capacity);
 194
 195         prt_printf(out, "hidden:\t\t\t\t%llu\n",
 196                fs_usage->u.hidden);
 197         prt_printf(out, "data:\t\t\t\t%llu\n",
 198                fs_usage->u.data);
 199         prt_printf(out, "cached:\t\t\t\t%llu\n",
 200                fs_usage->u.cached);
 201         prt_printf(out, "reserved:\t\t\t%llu\n",
 202                fs_usage->u.reserved);
 203         prt_printf(out, "nr_inodes:\t\t\t%llu\n",
 204                fs_usage->u.nr_inodes);
 205         prt_printf(out, "online reserved:\t\t%llu\n",
 206                fs_usage->online_reserved);
 207
 208         for (i = 0;
 209              i < ARRAY_SIZE(fs_usage->u.persistent_reserved);
 210              i++) {
 211                 prt_printf(out, "%u replicas:\n", i + 1);
 212                 prt_printf(out, "\treserved:\t\t%llu\n",
 213                        fs_usage->u.persistent_reserved[i]);
 214         }
 215
 216         for (i = 0; i < c->replicas.nr; i++) {
 217                 struct bch_replicas_entry *e =
 218                         cpu_replicas_entry(&c->replicas, i);
 219
 220                 prt_printf(out, "\t");
 221                 bch2_replicas_entry_to_text(out, e);
 222                 prt_printf(out, ":\t%llu\n", fs_usage->u.replicas[i]);
 223         }
 224 }
 225
 226 static u64 reserve_factor(u64 r)
 227 {
 228         return r + (round_up(r, (1 << RESERVE_FACTOR)) >> RESERVE_FACTOR);
 229 }
 230
 231 u64 bch2_fs_sectors_used(struct bch_fs *c, struct bch_fs_usage_online *fs_usage)
 232 {
 233         return min(fs_usage->u.hidden +
 234                    fs_usage->u.btree +
 235                    fs_usage->u.data +
 236                    reserve_factor(fs_usage->u.reserved +
 237                                   fs_usage->online_reserved),
 238                    c->capacity);
 239 }
 240
 241 static struct bch_fs_usage_short
 242 __bch2_fs_usage_read_short(struct bch_fs *c)
 243 {
 244         struct bch_fs_usage_short ret;
 245         u64 data, reserved;
 246
 247         ret.capacity = c->capacity -
 248                 bch2_fs_usage_read_one(c, &c->usage_base->hidden);
 249
 250         data            = bch2_fs_usage_read_one(c, &c->usage_base->data) +
 251                 bch2_fs_usage_read_one(c, &c->usage_base->btree);
 252         reserved        = bch2_fs_usage_read_one(c, &c->usage_base->reserved) +
 253                 percpu_u64_get(c->online_reserved);
 254
 255         ret.used        = min(ret.capacity, data + reserve_factor(reserved));
 256         ret.free        = ret.capacity - ret.used;
 257
 258         ret.nr_inodes   = bch2_fs_usage_read_one(c, &c->usage_base->nr_inodes);
 259
 260         return ret;
 261 }
 262
 263 struct bch_fs_usage_short
 264 bch2_fs_usage_read_short(struct bch_fs *c)
 265 {
 266         struct bch_fs_usage_short ret;
 267
 268         percpu_down_read(&c->mark_lock);
 269         ret = __bch2_fs_usage_read_short(c);
 270         percpu_up_read(&c->mark_lock);
 271
 272         return ret;
 273 }
 274
 275 void bch2_dev_usage_init(struct bch_dev *ca)
 276 {
 277         ca->usage_base->d[BCH_DATA_free].buckets = ca->mi.nbuckets - ca->mi.first_bucket;
 278 }
 279
 280 static inline int bucket_sectors_fragmented(struct bch_dev *ca,
 281                                             struct bch_alloc_v4 a)
 282 {
 283         return a.dirty_sectors
 284                 ? max(0, (int) ca->mi.bucket_size - (int) a.dirty_sectors)
 285                 : 0;
 286 }
 287
 288 static void bch2_dev_usage_update(struct bch_fs *c, struct bch_dev *ca,
 289                                   struct bch_alloc_v4 old,
 290                                   struct bch_alloc_v4 new,
 291                                   u64 journal_seq, bool gc)
 292 {
 293         struct bch_fs_usage *fs_usage;
 294         struct bch_dev_usage *u;
 295
 296         preempt_disable();
 297         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 298
 299         if (data_type_is_hidden(old.data_type))
 300                 fs_usage->hidden -= ca->mi.bucket_size;
 301         if (data_type_is_hidden(new.data_type))
 302                 fs_usage->hidden += ca->mi.bucket_size;
 303
 304         u = dev_usage_ptr(ca, journal_seq, gc);
 305
 306         u->d[old.data_type].buckets--;
 307         u->d[new.data_type].buckets++;
 308
 309         u->buckets_ec -= (int) !!old.stripe;
 310         u->buckets_ec += (int) !!new.stripe;
 311
 312         u->d[old.data_type].sectors -= old.dirty_sectors;
 313         u->d[new.data_type].sectors += new.dirty_sectors;
 314
 315         u->d[BCH_DATA_cached].sectors += new.cached_sectors;
 316         u->d[BCH_DATA_cached].sectors -= old.cached_sectors;
 317
 318         u->d[old.data_type].fragmented -= bucket_sectors_fragmented(ca, old);
 319         u->d[new.data_type].fragmented += bucket_sectors_fragmented(ca, new);
 320
 321         preempt_enable();
 322 }
 323
 324 static void bch2_dev_usage_update_m(struct bch_fs *c, struct bch_dev *ca,
 325                                     struct bucket old, struct bucket new,
 326                                     u64 journal_seq, bool gc)
 327 {
 328         struct bch_alloc_v4 old_a = {
 329                 .gen            = old.gen,
 330                 .data_type      = old.data_type,
 331                 .dirty_sectors  = old.dirty_sectors,
 332                 .cached_sectors = old.cached_sectors,
 333                 .stripe         = old.stripe,
 334         };
 335         struct bch_alloc_v4 new_a = {
 336                 .gen            = new.gen,
 337                 .data_type      = new.data_type,
 338                 .dirty_sectors  = new.dirty_sectors,
 339                 .cached_sectors = new.cached_sectors,
 340                 .stripe         = new.stripe,
 341         };
 342
 343         bch2_dev_usage_update(c, ca, old_a, new_a, journal_seq, gc);
 344 }
 345
 346 static inline int __update_replicas(struct bch_fs *c,
 347                                     struct bch_fs_usage *fs_usage,
 348                                     struct bch_replicas_entry *r,
 349                                     s64 sectors)
 350 {
 351         int idx = bch2_replicas_entry_idx(c, r);
 352
 353         if (idx < 0)
 354                 return -1;
 355
 356         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 357         fs_usage->replicas[idx]         += sectors;
 358         return 0;
 359 }
 360
 361 static inline int update_replicas(struct bch_fs *c, struct bkey_s_c k,
 362                         struct bch_replicas_entry *r, s64 sectors,
 363                         unsigned journal_seq, bool gc)
 364 {
 365         struct bch_fs_usage *fs_usage;
 366         int idx, ret = 0;
 367         struct printbuf buf = PRINTBUF;
 368
 369         percpu_down_read(&c->mark_lock);
 370
 371         idx = bch2_replicas_entry_idx(c, r);
 372         if (idx < 0 &&
 373             fsck_err(c, "no replicas entry\n"
 374                      "  while marking %s",
 375                      (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 376                 percpu_up_read(&c->mark_lock);
 377                 ret = bch2_mark_replicas(c, r);
 378                 percpu_down_read(&c->mark_lock);
 379
 380                 if (ret)
 381                         goto err;
 382                 idx = bch2_replicas_entry_idx(c, r);
 383         }
 384         if (idx < 0) {
 385                 ret = -1;
 386                 goto err;
 387         }
 388
 389         preempt_disable();
 390         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 391         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 392         fs_usage->replicas[idx]         += sectors;
 393         preempt_enable();
 394 err:
 395 fsck_err:
 396         percpu_up_read(&c->mark_lock);
 397         printbuf_exit(&buf);
 398         return ret;
 399 }
 400
 401 static inline int update_cached_sectors(struct bch_fs *c,
 402                         struct bkey_s_c k,
 403                         unsigned dev, s64 sectors,
 404                         unsigned journal_seq, bool gc)
 405 {
 406         struct bch_replicas_padded r;
 407
 408         bch2_replicas_entry_cached(&r.e, dev);
 409
 410         return update_replicas(c, k, &r.e, sectors, journal_seq, gc);
 411 }
 412
 413 static int __replicas_deltas_realloc(struct btree_trans *trans, unsigned more,
 414                                      gfp_t gfp)
 415 {
 416         struct replicas_delta_list *d = trans->fs_usage_deltas;
 417         unsigned new_size = d ? (d->size + more) * 2 : 128;
 418         unsigned alloc_size = sizeof(*d) + new_size;
 419
 420         WARN_ON_ONCE(alloc_size > REPLICAS_DELTA_LIST_MAX);
 421
 422         if (!d || d->used + more > d->size) {
 423                 d = krealloc(d, alloc_size, gfp|__GFP_ZERO);
 424
 425                 if (unlikely(!d)) {
 426                         if (alloc_size > REPLICAS_DELTA_LIST_MAX)
 427                                 return -ENOMEM;
 428
 429                         d = mempool_alloc(&trans->c->replicas_delta_pool, gfp);
 430                         if (!d)
 431                                 return -ENOMEM;
 432
 433                         memset(d, 0, REPLICAS_DELTA_LIST_MAX);
 434
 435                         if (trans->fs_usage_deltas)
 436                                 memcpy(d, trans->fs_usage_deltas,
 437                                        trans->fs_usage_deltas->size + sizeof(*d));
 438
 439                         new_size = REPLICAS_DELTA_LIST_MAX - sizeof(*d);
 440                         kfree(trans->fs_usage_deltas);
 441                 }
 442
 443                 d->size = new_size;
 444                 trans->fs_usage_deltas = d;
 445         }
 446
 447         return 0;
 448 }
 449
 450 int bch2_replicas_deltas_realloc(struct btree_trans *trans, unsigned more)
 451 {
 452         return allocate_dropping_locks_errcode(trans,
 453                                 __replicas_deltas_realloc(trans, more, _gfp));
 454 }
 455
 456 static inline int update_replicas_list(struct btree_trans *trans,
 457                                         struct bch_replicas_entry *r,
 458                                         s64 sectors)
 459 {
 460         struct replicas_delta_list *d;
 461         struct replicas_delta *n;
 462         unsigned b;
 463         int ret;
 464
 465         if (!sectors)
 466                 return 0;
 467
 468         b = replicas_entry_bytes(r) + 8;
 469         ret = bch2_replicas_deltas_realloc(trans, b);
 470         if (ret)
 471                 return ret;
 472
 473         d = trans->fs_usage_deltas;
 474         n = (void *) d->d + d->used;
 475         n->delta = sectors;
 476         unsafe_memcpy((void *) n + offsetof(struct replicas_delta, r),
 477                       r, replicas_entry_bytes(r),
 478                       "flexible array member embedded in strcuct with padding");
 479         bch2_replicas_entry_sort(&n->r);
 480         d->used += b;
 481         return 0;
 482 }
 483
 484 static inline int update_cached_sectors_list(struct btree_trans *trans,
 485                                               unsigned dev, s64 sectors)
 486 {
 487         struct bch_replicas_padded r;
 488
 489         bch2_replicas_entry_cached(&r.e, dev);
 490
 491         return update_replicas_list(trans, &r.e, sectors);
 492 }
 493
 494 int bch2_mark_alloc(struct btree_trans *trans,
 495                     enum btree_id btree, unsigned level,
 496                     struct bkey_s_c old, struct bkey_s_c new,
 497                     unsigned flags)
 498 {
 499         bool gc = flags & BTREE_TRIGGER_GC;
 500         u64 journal_seq = trans->journal_res.seq;
 501         u64 bucket_journal_seq;
 502         struct bch_fs *c = trans->c;
 503         struct bch_alloc_v4 old_a_convert, new_a_convert;
 504         const struct bch_alloc_v4 *old_a, *new_a;
 505         struct bch_dev *ca;
 506         int ret = 0;
 507
 508         /*
 509          * alloc btree is read in by bch2_alloc_read, not gc:
 510          */
 511         if ((flags & BTREE_TRIGGER_GC) &&
 512             !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
 513                 return 0;
 514
 515         if (bch2_trans_inconsistent_on(!bch2_dev_bucket_exists(c, new.k->p), trans,
 516                                        "alloc key for invalid device or bucket"))
 517                 return -EIO;
 518
 519         ca = bch_dev_bkey_exists(c, new.k->p.inode);
 520
 521         old_a = bch2_alloc_to_v4(old, &old_a_convert);
 522         new_a = bch2_alloc_to_v4(new, &new_a_convert);
 523
 524         bucket_journal_seq = new_a->journal_seq;
 525
 526         if ((flags & BTREE_TRIGGER_INSERT) &&
 527             data_type_is_empty(old_a->data_type) !=
 528             data_type_is_empty(new_a->data_type) &&
 529             new.k->type == KEY_TYPE_alloc_v4) {
 530                 struct bch_alloc_v4 *v = (struct bch_alloc_v4 *) new.v;
 531
 532                 EBUG_ON(!journal_seq);
 533
 534                 /*
 535                  * If the btree updates referring to a bucket weren't flushed
 536                  * before the bucket became empty again, then the we don't have
 537                  * to wait on a journal flush before we can reuse the bucket:
 538                  */
 539                 v->journal_seq = bucket_journal_seq =
 540                         data_type_is_empty(new_a->data_type) &&
 541                         (journal_seq == v->journal_seq ||
 542                          bch2_journal_noflush_seq(&c->journal, v->journal_seq))
 543                         ? 0 : journal_seq;
 544         }
 545
 546         if (!data_type_is_empty(old_a->data_type) &&
 547             data_type_is_empty(new_a->data_type) &&
 548             bucket_journal_seq) {
 549                 ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 550                                 c->journal.flushed_seq_ondisk,
 551                                 new.k->p.inode, new.k->p.offset,
 552                                 bucket_journal_seq);
 553                 if (ret) {
 554                         bch2_fs_fatal_error(c,
 555                                 "error setting bucket_needs_journal_commit: %i", ret);
 556                         return ret;
 557                 }
 558         }
 559
 560         percpu_down_read(&c->mark_lock);
 561         if (!gc && new_a->gen != old_a->gen)
 562                 *bucket_gen(ca, new.k->p.offset) = new_a->gen;
 563
 564         bch2_dev_usage_update(c, ca, *old_a, *new_a, journal_seq, gc);
 565
 566         if (gc) {
 567                 struct bucket *g = gc_bucket(ca, new.k->p.offset);
 568
 569                 bucket_lock(g);
 570
 571                 g->gen_valid            = 1;
 572                 g->gen                  = new_a->gen;
 573                 g->data_type            = new_a->data_type;
 574                 g->stripe               = new_a->stripe;
 575                 g->stripe_redundancy    = new_a->stripe_redundancy;
 576                 g->dirty_sectors        = new_a->dirty_sectors;
 577                 g->cached_sectors       = new_a->cached_sectors;
 578
 579                 bucket_unlock(g);
 580         }
 581         percpu_up_read(&c->mark_lock);
 582
 583         /*
 584          * need to know if we're getting called from the invalidate path or
 585          * not:
 586          */
 587
 588         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 589             old_a->cached_sectors) {
 590                 ret = update_cached_sectors(c, new, ca->dev_idx,
 591                                             -((s64) old_a->cached_sectors),
 592                                             journal_seq, gc);
 593                 if (ret) {
 594                         bch2_fs_fatal_error(c, "%s(): no replicas entry while updating cached sectors",
 595                                             __func__);
 596                         return ret;
 597                 }
 598         }
 599
 600         if (new_a->data_type == BCH_DATA_free &&
 601             (!new_a->journal_seq || new_a->journal_seq < c->journal.flushed_seq_ondisk))
 602                 closure_wake_up(&c->freelist_wait);
 603
 604         if (new_a->data_type == BCH_DATA_need_discard &&
 605             (!bucket_journal_seq || bucket_journal_seq < c->journal.flushed_seq_ondisk))
 606                 bch2_do_discards(c);
 607
 608         if (old_a->data_type != BCH_DATA_cached &&
 609             new_a->data_type == BCH_DATA_cached &&
 610             should_invalidate_buckets(ca, bch2_dev_usage_read(ca)))
 611                 bch2_do_invalidates(c);
 612
 613         if (new_a->data_type == BCH_DATA_need_gc_gens)
 614                 bch2_do_gc_gens(c);
 615
 616         return 0;
 617 }
 618
 619 int bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 620                               size_t b, enum bch_data_type data_type,
 621                               unsigned sectors, struct gc_pos pos,
 622                               unsigned flags)
 623 {
 624         struct bucket old, new, *g;
 625         int ret = 0;
 626
 627         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 628         BUG_ON(data_type != BCH_DATA_sb &&
 629                data_type != BCH_DATA_journal);
 630
 631         /*
 632          * Backup superblock might be past the end of our normal usable space:
 633          */
 634         if (b >= ca->mi.nbuckets)
 635                 return 0;
 636
 637         percpu_down_read(&c->mark_lock);
 638         g = gc_bucket(ca, b);
 639
 640         bucket_lock(g);
 641         old = *g;
 642
 643         if (bch2_fs_inconsistent_on(g->data_type &&
 644                         g->data_type != data_type, c,
 645                         "different types of data in same bucket: %s, %s",
 646                         bch2_data_types[g->data_type],
 647                         bch2_data_types[data_type])) {
 648                 ret = -EIO;
 649                 goto err;
 650         }
 651
 652         if (bch2_fs_inconsistent_on((u64) g->dirty_sectors + sectors > ca->mi.bucket_size, c,
 653                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > bucket size",
 654                         ca->dev_idx, b, g->gen,
 655                         bch2_data_types[g->data_type ?: data_type],
 656                         g->dirty_sectors, sectors)) {
 657                 ret = -EIO;
 658                 goto err;
 659         }
 660
 661
 662         g->data_type = data_type;
 663         g->dirty_sectors += sectors;
 664         new = *g;
 665 err:
 666         bucket_unlock(g);
 667         if (!ret)
 668                 bch2_dev_usage_update_m(c, ca, old, new, 0, true);
 669         percpu_up_read(&c->mark_lock);
 670         return ret;
 671 }
 672
 673 static int check_bucket_ref(struct btree_trans *trans,
 674                             struct bkey_s_c k,
 675                             const struct bch_extent_ptr *ptr,
 676                             s64 sectors, enum bch_data_type ptr_data_type,
 677                             u8 b_gen, u8 bucket_data_type,
 678                             u32 dirty_sectors, u32 cached_sectors)
 679 {
 680         struct bch_fs *c = trans->c;
 681         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 682         size_t bucket_nr = PTR_BUCKET_NR(ca, ptr);
 683         u32 bucket_sectors = !ptr->cached
 684                 ? dirty_sectors
 685                 : cached_sectors;
 686         struct printbuf buf = PRINTBUF;
 687         int ret = 0;
 688
 689         if (bucket_data_type == BCH_DATA_cached)
 690                 bucket_data_type = BCH_DATA_user;
 691
 692         if ((bucket_data_type == BCH_DATA_stripe && ptr_data_type == BCH_DATA_user) ||
 693             (bucket_data_type == BCH_DATA_user   && ptr_data_type == BCH_DATA_stripe))
 694                 bucket_data_type = ptr_data_type = BCH_DATA_stripe;
 695
 696         if (gen_after(ptr->gen, b_gen)) {
 697                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 698                         "bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\n"
 699                         "while marking %s",
 700                         ptr->dev, bucket_nr, b_gen,
 701                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 702                         ptr->gen,
 703                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 704                 ret = -EIO;
 705                 goto err;
 706         }
 707
 708         if (gen_cmp(b_gen, ptr->gen) > BUCKET_GC_GEN_MAX) {
 709                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 710                         "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\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                         (printbuf_reset(&buf),
 716                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 717                 ret = -EIO;
 718                 goto err;
 719         }
 720
 721         if (b_gen != ptr->gen && !ptr->cached) {
 722                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 723                         "bucket %u:%zu gen %u (mem gen %u) data type %s: stale dirty ptr (gen %u)\n"
 724                         "while marking %s",
 725                         ptr->dev, bucket_nr, b_gen,
 726                         *bucket_gen(ca, bucket_nr),
 727                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 728                         ptr->gen,
 729                         (printbuf_reset(&buf),
 730                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 731                 ret = -EIO;
 732                 goto err;
 733         }
 734
 735         if (b_gen != ptr->gen) {
 736                 ret = 1;
 737                 goto out;
 738         }
 739
 740         if (!data_type_is_empty(bucket_data_type) &&
 741             ptr_data_type &&
 742             bucket_data_type != ptr_data_type) {
 743                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 744                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 745                         "while marking %s",
 746                         ptr->dev, bucket_nr, b_gen,
 747                         bch2_data_types[bucket_data_type],
 748                         bch2_data_types[ptr_data_type],
 749                         (printbuf_reset(&buf),
 750                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 751                 ret = -EIO;
 752                 goto err;
 753         }
 754
 755         if ((u64) bucket_sectors + sectors > U32_MAX) {
 756                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 757                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U32_MAX\n"
 758                         "while marking %s",
 759                         ptr->dev, bucket_nr, b_gen,
 760                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 761                         bucket_sectors, sectors,
 762                         (printbuf_reset(&buf),
 763                          bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 764                 ret = -EIO;
 765                 goto err;
 766         }
 767 out:
 768         printbuf_exit(&buf);
 769         return ret;
 770 err:
 771         bch2_dump_trans_updates(trans);
 772         goto out;
 773 }
 774
 775 static int mark_stripe_bucket(struct btree_trans *trans,
 776                               struct bkey_s_c k,
 777                               unsigned ptr_idx,
 778                               unsigned flags)
 779 {
 780         struct bch_fs *c = trans->c;
 781         u64 journal_seq = trans->journal_res.seq;
 782         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 783         unsigned nr_data = s->nr_blocks - s->nr_redundant;
 784         bool parity = ptr_idx >= nr_data;
 785         enum bch_data_type data_type = parity ? BCH_DATA_parity : BCH_DATA_stripe;
 786         s64 sectors = parity ? le16_to_cpu(s->sectors) : 0;
 787         const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
 788         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 789         struct bucket old, new, *g;
 790         struct printbuf buf = PRINTBUF;
 791         int ret = 0;
 792
 793         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 794
 795         /* * XXX doesn't handle deletion */
 796
 797         percpu_down_read(&c->mark_lock);
 798         g = PTR_GC_BUCKET(ca, ptr);
 799
 800         if (g->dirty_sectors ||
 801             (g->stripe && g->stripe != k.k->p.offset)) {
 802                 bch2_fs_inconsistent(c,
 803                               "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
 804                               ptr->dev, PTR_BUCKET_NR(ca, ptr), g->gen,
 805                               (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
 806                 ret = -EINVAL;
 807                 goto err;
 808         }
 809
 810         bucket_lock(g);
 811         old = *g;
 812
 813         ret = check_bucket_ref(trans, k, ptr, sectors, data_type,
 814                                g->gen, g->data_type,
 815                                g->dirty_sectors, g->cached_sectors);
 816         if (ret)
 817                 goto err;
 818
 819         g->data_type = data_type;
 820         g->dirty_sectors += sectors;
 821
 822         g->stripe               = k.k->p.offset;
 823         g->stripe_redundancy    = s->nr_redundant;
 824         new = *g;
 825 err:
 826         bucket_unlock(g);
 827         if (!ret)
 828                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 829         percpu_up_read(&c->mark_lock);
 830         printbuf_exit(&buf);
 831         return ret;
 832 }
 833
 834 static int __mark_pointer(struct btree_trans *trans,
 835                           struct bkey_s_c k,
 836                           const struct bch_extent_ptr *ptr,
 837                           s64 sectors, enum bch_data_type ptr_data_type,
 838                           u8 bucket_gen, u8 *bucket_data_type,
 839                           u32 *dirty_sectors, u32 *cached_sectors)
 840 {
 841         u32 *dst_sectors = !ptr->cached
 842                 ? dirty_sectors
 843                 : cached_sectors;
 844         int ret = check_bucket_ref(trans, k, ptr, sectors, ptr_data_type,
 845                                    bucket_gen, *bucket_data_type,
 846                                    *dirty_sectors, *cached_sectors);
 847
 848         if (ret)
 849                 return ret;
 850
 851         *dst_sectors += sectors;
 852         *bucket_data_type = *dirty_sectors || *cached_sectors
 853                 ? ptr_data_type : 0;
 854         return 0;
 855 }
 856
 857 static int bch2_mark_pointer(struct btree_trans *trans,
 858                              enum btree_id btree_id, unsigned level,
 859                              struct bkey_s_c k,
 860                              struct extent_ptr_decoded p,
 861                              s64 sectors,
 862                              unsigned flags)
 863 {
 864         u64 journal_seq = trans->journal_res.seq;
 865         struct bch_fs *c = trans->c;
 866         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 867         struct bucket old, new, *g;
 868         enum bch_data_type data_type = bkey_ptr_data_type(btree_id, level, k, p);
 869         u8 bucket_data_type;
 870         int ret = 0;
 871
 872         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 873
 874         percpu_down_read(&c->mark_lock);
 875         g = PTR_GC_BUCKET(ca, &p.ptr);
 876         bucket_lock(g);
 877         old = *g;
 878
 879         bucket_data_type = g->data_type;
 880         ret = __mark_pointer(trans, k, &p.ptr, sectors,
 881                              data_type, g->gen,
 882                              &bucket_data_type,
 883                              &g->dirty_sectors,
 884                              &g->cached_sectors);
 885         if (!ret)
 886                 g->data_type = bucket_data_type;
 887
 888         new = *g;
 889         bucket_unlock(g);
 890         if (!ret)
 891                 bch2_dev_usage_update_m(c, ca, old, new, journal_seq, true);
 892         percpu_up_read(&c->mark_lock);
 893
 894         return ret;
 895 }
 896
 897 static int bch2_mark_stripe_ptr(struct btree_trans *trans,
 898                                 struct bkey_s_c k,
 899                                 struct bch_extent_stripe_ptr p,
 900                                 enum bch_data_type data_type,
 901                                 s64 sectors,
 902                                 unsigned flags)
 903 {
 904         struct bch_fs *c = trans->c;
 905         struct bch_replicas_padded r;
 906         struct gc_stripe *m;
 907
 908         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 909
 910         m = genradix_ptr_alloc(&c->gc_stripes, p.idx, GFP_KERNEL);
 911         if (!m) {
 912                 bch_err(c, "error allocating memory for gc_stripes, idx %llu",
 913                         (u64) p.idx);
 914                 return -BCH_ERR_ENOMEM_mark_stripe_ptr;
 915         }
 916
 917         mutex_lock(&c->ec_stripes_heap_lock);
 918
 919         if (!m || !m->alive) {
 920                 mutex_unlock(&c->ec_stripes_heap_lock);
 921                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
 922                                     (u64) p.idx);
 923                 bch2_inconsistent_error(c);
 924                 return -EIO;
 925         }
 926
 927         m->block_sectors[p.block] += sectors;
 928
 929         r = m->r;
 930         mutex_unlock(&c->ec_stripes_heap_lock);
 931
 932         r.e.data_type = data_type;
 933         update_replicas(c, k, &r.e, sectors, trans->journal_res.seq, true);
 934
 935         return 0;
 936 }
 937
 938 static int __mark_extent(struct btree_trans *trans,
 939                          enum btree_id btree_id, unsigned level,
 940                          struct bkey_s_c k, unsigned flags)
 941 {
 942         u64 journal_seq = trans->journal_res.seq;
 943         struct bch_fs *c = trans->c;
 944         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 945         const union bch_extent_entry *entry;
 946         struct extent_ptr_decoded p;
 947         struct bch_replicas_padded r;
 948         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
 949                 ? BCH_DATA_btree
 950                 : BCH_DATA_user;
 951         s64 sectors = bkey_is_btree_ptr(k.k)
 952                 ? btree_sectors(c)
 953                 : k.k->size;
 954         s64 dirty_sectors = 0;
 955         bool stale;
 956         int ret;
 957
 958         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 959
 960         r.e.data_type   = data_type;
 961         r.e.nr_devs     = 0;
 962         r.e.nr_required = 1;
 963
 964         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 965                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
 966
 967                 if (flags & BTREE_TRIGGER_OVERWRITE)
 968                         disk_sectors = -disk_sectors;
 969
 970                 ret = bch2_mark_pointer(trans, btree_id, level, k, p, disk_sectors, flags);
 971                 if (ret < 0)
 972                         return ret;
 973
 974                 stale = ret > 0;
 975
 976                 if (p.ptr.cached) {
 977                         if (!stale) {
 978                                 ret = update_cached_sectors(c, k, p.ptr.dev,
 979                                                 disk_sectors, journal_seq, true);
 980                                 if (ret) {
 981                                         bch2_fs_fatal_error(c, "%s(): no replicas entry while updating cached sectors",
 982                                                             __func__);
 983                                         return ret;
 984                                 }
 985                         }
 986                 } else if (!p.has_ec) {
 987                         dirty_sectors          += disk_sectors;
 988                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
 989                 } else {
 990                         ret = bch2_mark_stripe_ptr(trans, k, p.ec, data_type,
 991                                         disk_sectors, flags);
 992                         if (ret)
 993                                 return ret;
 994
 995                         /*
 996                          * There may be other dirty pointers in this extent, but
 997                          * if so they're not required for mounting if we have an
 998                          * erasure coded pointer in this extent:
 999                          */
1000                         r.e.nr_required = 0;
1001                 }
1002         }
1003
1004         if (r.e.nr_devs) {
1005                 ret = update_replicas(c, k, &r.e, dirty_sectors, journal_seq, true);
1006                 if (ret) {
1007                         struct printbuf buf = PRINTBUF;
1008
1009                         bch2_bkey_val_to_text(&buf, c, k);
1010                         bch2_fs_fatal_error(c, "%s(): no replicas entry for %s", __func__, buf.buf);
1011                         printbuf_exit(&buf);
1012                         return ret;
1013                 }
1014         }
1015
1016         return 0;
1017 }
1018
1019 int bch2_mark_extent(struct btree_trans *trans,
1020                      enum btree_id btree_id, unsigned level,
1021                      struct bkey_s_c old, struct bkey_s_c new,
1022                      unsigned flags)
1023 {
1024         return mem_trigger_run_overwrite_then_insert(__mark_extent, trans, btree_id, level, old, new, flags);
1025 }
1026
1027 int bch2_mark_stripe(struct btree_trans *trans,
1028                      enum btree_id btree_id, unsigned level,
1029                      struct bkey_s_c old, struct bkey_s_c new,
1030                      unsigned flags)
1031 {
1032         bool gc = flags & BTREE_TRIGGER_GC;
1033         u64 journal_seq = trans->journal_res.seq;
1034         struct bch_fs *c = trans->c;
1035         u64 idx = new.k->p.offset;
1036         const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
1037                 ? bkey_s_c_to_stripe(old).v : NULL;
1038         const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
1039                 ? bkey_s_c_to_stripe(new).v : NULL;
1040         unsigned i;
1041         int ret;
1042
1043         BUG_ON(gc && old_s);
1044
1045         if (!gc) {
1046                 struct stripe *m = genradix_ptr(&c->stripes, idx);
1047
1048                 if (!m) {
1049                         struct printbuf buf1 = PRINTBUF;
1050                         struct printbuf buf2 = PRINTBUF;
1051
1052                         bch2_bkey_val_to_text(&buf1, c, old);
1053                         bch2_bkey_val_to_text(&buf2, c, new);
1054                         bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
1055                                             "old %s\n"
1056                                             "new %s", idx, buf1.buf, buf2.buf);
1057                         printbuf_exit(&buf2);
1058                         printbuf_exit(&buf1);
1059                         bch2_inconsistent_error(c);
1060                         return -1;
1061                 }
1062
1063                 if (!new_s) {
1064                         bch2_stripes_heap_del(c, m, idx);
1065
1066                         memset(m, 0, sizeof(*m));
1067                 } else {
1068                         m->sectors      = le16_to_cpu(new_s->sectors);
1069                         m->algorithm    = new_s->algorithm;
1070                         m->nr_blocks    = new_s->nr_blocks;
1071                         m->nr_redundant = new_s->nr_redundant;
1072                         m->blocks_nonempty = 0;
1073
1074                         for (i = 0; i < new_s->nr_blocks; i++)
1075                                 m->blocks_nonempty += !!stripe_blockcount_get(new_s, i);
1076
1077                         if (!old_s)
1078                                 bch2_stripes_heap_insert(c, m, idx);
1079                         else
1080                                 bch2_stripes_heap_update(c, m, idx);
1081                 }
1082         } else {
1083                 struct gc_stripe *m =
1084                         genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
1085
1086                 if (!m) {
1087                         bch_err(c, "error allocating memory for gc_stripes, idx %llu",
1088                                 idx);
1089                         return -BCH_ERR_ENOMEM_mark_stripe;
1090                 }
1091                 /*
1092                  * This will be wrong when we bring back runtime gc: we should
1093                  * be unmarking the old key and then marking the new key
1094                  */
1095                 m->alive        = true;
1096                 m->sectors      = le16_to_cpu(new_s->sectors);
1097                 m->nr_blocks    = new_s->nr_blocks;
1098                 m->nr_redundant = new_s->nr_redundant;
1099
1100                 for (i = 0; i < new_s->nr_blocks; i++)
1101                         m->ptrs[i] = new_s->ptrs[i];
1102
1103                 bch2_bkey_to_replicas(&m->r.e, new);
1104
1105                 /*
1106                  * gc recalculates this field from stripe ptr
1107                  * references:
1108                  */
1109                 memset(m->block_sectors, 0, sizeof(m->block_sectors));
1110
1111                 for (i = 0; i < new_s->nr_blocks; i++) {
1112                         ret = mark_stripe_bucket(trans, new, i, flags);
1113                         if (ret)
1114                                 return ret;
1115                 }
1116
1117                 ret = update_replicas(c, new, &m->r.e,
1118                                       ((s64) m->sectors * m->nr_redundant),
1119                                       journal_seq, gc);
1120                 if (ret) {
1121                         struct printbuf buf = PRINTBUF;
1122
1123                         bch2_bkey_val_to_text(&buf, c, new);
1124                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
1125                         printbuf_exit(&buf);
1126                         return ret;
1127                 }
1128         }
1129
1130         return 0;
1131 }
1132
1133 static int __mark_reservation(struct btree_trans *trans,
1134                               enum btree_id btree_id, unsigned level,
1135                               struct bkey_s_c k, unsigned flags)
1136 {
1137         struct bch_fs *c = trans->c;
1138         struct bch_fs_usage *fs_usage;
1139         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1140         s64 sectors = (s64) k.k->size;
1141
1142         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1143
1144         if (flags & BTREE_TRIGGER_OVERWRITE)
1145                 sectors = -sectors;
1146         sectors *= replicas;
1147
1148         percpu_down_read(&c->mark_lock);
1149         preempt_disable();
1150
1151         fs_usage = fs_usage_ptr(c, trans->journal_res.seq, flags & BTREE_TRIGGER_GC);
1152         replicas = clamp_t(unsigned, replicas, 1,
1153                            ARRAY_SIZE(fs_usage->persistent_reserved));
1154
1155         fs_usage->reserved                              += sectors;
1156         fs_usage->persistent_reserved[replicas - 1]     += sectors;
1157
1158         preempt_enable();
1159         percpu_up_read(&c->mark_lock);
1160
1161         return 0;
1162 }
1163
1164 int bch2_mark_reservation(struct btree_trans *trans,
1165                           enum btree_id btree_id, unsigned level,
1166                           struct bkey_s_c old, struct bkey_s_c new,
1167                           unsigned flags)
1168 {
1169         return mem_trigger_run_overwrite_then_insert(__mark_reservation, trans, btree_id, level, old, new, flags);
1170 }
1171
1172 static s64 __bch2_mark_reflink_p(struct btree_trans *trans,
1173                                  struct bkey_s_c_reflink_p p,
1174                                  u64 start, u64 end,
1175                                  u64 *idx, unsigned flags, size_t r_idx)
1176 {
1177         struct bch_fs *c = trans->c;
1178         struct reflink_gc *r;
1179         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1180         u64 next_idx = end;
1181         s64 ret = 0;
1182         struct printbuf buf = PRINTBUF;
1183
1184         if (r_idx >= c->reflink_gc_nr)
1185                 goto not_found;
1186
1187         r = genradix_ptr(&c->reflink_gc_table, r_idx);
1188         next_idx = min(next_idx, r->offset - r->size);
1189         if (*idx < next_idx)
1190                 goto not_found;
1191
1192         BUG_ON((s64) r->refcount + add < 0);
1193
1194         r->refcount += add;
1195         *idx = r->offset;
1196         return 0;
1197 not_found:
1198         if (fsck_err(c, "pointer to missing indirect extent\n"
1199                      "  %s\n"
1200                      "  missing range %llu-%llu",
1201                      (bch2_bkey_val_to_text(&buf, c, p.s_c), buf.buf),
1202                      *idx, next_idx)) {
1203                 struct bkey_i_error *new;
1204
1205                 new = bch2_trans_kmalloc(trans, sizeof(*new));
1206                 ret = PTR_ERR_OR_ZERO(new);
1207                 if (ret)
1208                         goto err;
1209
1210                 bkey_init(&new->k);
1211                 new->k.type     = KEY_TYPE_error;
1212                 new->k.p                = bkey_start_pos(p.k);
1213                 new->k.p.offset += *idx - start;
1214                 bch2_key_resize(&new->k, next_idx - *idx);
1215                 ret = bch2_btree_insert_trans(trans, BTREE_ID_extents, &new->k_i,
1216                                           BTREE_TRIGGER_NORUN);
1217         }
1218
1219         *idx = next_idx;
1220 err:
1221 fsck_err:
1222         printbuf_exit(&buf);
1223         return ret;
1224 }
1225
1226 static int __mark_reflink_p(struct btree_trans *trans,
1227                             enum btree_id btree_id, unsigned level,
1228                             struct bkey_s_c k, unsigned flags)
1229 {
1230         struct bch_fs *c = trans->c;
1231         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1232         struct reflink_gc *ref;
1233         size_t l, r, m;
1234         u64 idx = le64_to_cpu(p.v->idx), start = idx;
1235         u64 end = le64_to_cpu(p.v->idx) + p.k->size;
1236         int ret = 0;
1237
1238         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1239
1240         if (c->sb.version_upgrade_complete >= bcachefs_metadata_version_reflink_p_fix) {
1241                 idx -= le32_to_cpu(p.v->front_pad);
1242                 end += le32_to_cpu(p.v->back_pad);
1243         }
1244
1245         l = 0;
1246         r = c->reflink_gc_nr;
1247         while (l < r) {
1248                 m = l + (r - l) / 2;
1249
1250                 ref = genradix_ptr(&c->reflink_gc_table, m);
1251                 if (ref->offset <= idx)
1252                         l = m + 1;
1253                 else
1254                         r = m;
1255         }
1256
1257         while (idx < end && !ret)
1258                 ret = __bch2_mark_reflink_p(trans, p, start, end,
1259                                             &idx, flags, l++);
1260
1261         return ret;
1262 }
1263
1264 int bch2_mark_reflink_p(struct btree_trans *trans,
1265                         enum btree_id btree_id, unsigned level,
1266                         struct bkey_s_c old, struct bkey_s_c new,
1267                         unsigned flags)
1268 {
1269         return mem_trigger_run_overwrite_then_insert(__mark_reflink_p, trans, btree_id, level, old, new, flags);
1270 }
1271
1272 void bch2_trans_fs_usage_revert(struct btree_trans *trans,
1273                                 struct replicas_delta_list *deltas)
1274 {
1275         struct bch_fs *c = trans->c;
1276         struct bch_fs_usage *dst;
1277         struct replicas_delta *d, *top = (void *) deltas->d + deltas->used;
1278         s64 added = 0;
1279         unsigned i;
1280
1281         percpu_down_read(&c->mark_lock);
1282         preempt_disable();
1283         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1284
1285         /* revert changes: */
1286         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1287                 switch (d->r.data_type) {
1288                 case BCH_DATA_btree:
1289                 case BCH_DATA_user:
1290                 case BCH_DATA_parity:
1291                         added += d->delta;
1292                 }
1293                 BUG_ON(__update_replicas(c, dst, &d->r, -d->delta));
1294         }
1295
1296         dst->nr_inodes -= deltas->nr_inodes;
1297
1298         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1299                 added                           -= deltas->persistent_reserved[i];
1300                 dst->reserved                   -= deltas->persistent_reserved[i];
1301                 dst->persistent_reserved[i]     -= deltas->persistent_reserved[i];
1302         }
1303
1304         if (added > 0) {
1305                 trans->disk_res->sectors += added;
1306                 this_cpu_add(*c->online_reserved, added);
1307         }
1308
1309         preempt_enable();
1310         percpu_up_read(&c->mark_lock);
1311 }
1312
1313 int bch2_trans_fs_usage_apply(struct btree_trans *trans,
1314                               struct replicas_delta_list *deltas)
1315 {
1316         struct bch_fs *c = trans->c;
1317         static int warned_disk_usage = 0;
1318         bool warn = false;
1319         unsigned disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1320         struct replicas_delta *d, *d2;
1321         struct replicas_delta *top = (void *) deltas->d + deltas->used;
1322         struct bch_fs_usage *dst;
1323         s64 added = 0, should_not_have_added;
1324         unsigned i;
1325
1326         percpu_down_read(&c->mark_lock);
1327         preempt_disable();
1328         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1329
1330         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1331                 switch (d->r.data_type) {
1332                 case BCH_DATA_btree:
1333                 case BCH_DATA_user:
1334                 case BCH_DATA_parity:
1335                         added += d->delta;
1336                 }
1337
1338                 if (__update_replicas(c, dst, &d->r, d->delta))
1339                         goto need_mark;
1340         }
1341
1342         dst->nr_inodes += deltas->nr_inodes;
1343
1344         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1345                 added                           += deltas->persistent_reserved[i];
1346                 dst->reserved                   += deltas->persistent_reserved[i];
1347                 dst->persistent_reserved[i]     += deltas->persistent_reserved[i];
1348         }
1349
1350         /*
1351          * Not allowed to reduce sectors_available except by getting a
1352          * reservation:
1353          */
1354         should_not_have_added = added - (s64) disk_res_sectors;
1355         if (unlikely(should_not_have_added > 0)) {
1356                 u64 old, new, v = atomic64_read(&c->sectors_available);
1357
1358                 do {
1359                         old = v;
1360                         new = max_t(s64, 0, old - should_not_have_added);
1361                 } while ((v = atomic64_cmpxchg(&c->sectors_available,
1362                                                old, new)) != old);
1363
1364                 added -= should_not_have_added;
1365                 warn = true;
1366         }
1367
1368         if (added > 0) {
1369                 trans->disk_res->sectors -= added;
1370                 this_cpu_sub(*c->online_reserved, added);
1371         }
1372
1373         preempt_enable();
1374         percpu_up_read(&c->mark_lock);
1375
1376         if (unlikely(warn) && !xchg(&warned_disk_usage, 1))
1377                 bch2_trans_inconsistent(trans,
1378                                         "disk usage increased %lli more than %u sectors reserved)",
1379                                         should_not_have_added, disk_res_sectors);
1380         return 0;
1381 need_mark:
1382         /* revert changes: */
1383         for (d2 = deltas->d; d2 != d; d2 = replicas_delta_next(d2))
1384                 BUG_ON(__update_replicas(c, dst, &d2->r, -d2->delta));
1385
1386         preempt_enable();
1387         percpu_up_read(&c->mark_lock);
1388         return -1;
1389 }
1390
1391 /* trans_mark: */
1392
1393 static inline int bch2_trans_mark_pointer(struct btree_trans *trans,
1394                                    enum btree_id btree_id, unsigned level,
1395                                    struct bkey_s_c k, struct extent_ptr_decoded p,
1396                                    unsigned flags)
1397 {
1398         bool insert = !(flags & BTREE_TRIGGER_OVERWRITE);
1399         struct btree_iter iter;
1400         struct bkey_i_alloc_v4 *a;
1401         struct bpos bucket;
1402         struct bch_backpointer bp;
1403         s64 sectors;
1404         int ret;
1405
1406         bch2_extent_ptr_to_bp(trans->c, btree_id, level, k, p, &bucket, &bp);
1407         sectors = bp.bucket_len;
1408         if (!insert)
1409                 sectors = -sectors;
1410
1411         a = bch2_trans_start_alloc_update(trans, &iter, bucket);
1412         if (IS_ERR(a))
1413                 return PTR_ERR(a);
1414
1415         ret = __mark_pointer(trans, k, &p.ptr, sectors, bp.data_type,
1416                              a->v.gen, &a->v.data_type,
1417                              &a->v.dirty_sectors, &a->v.cached_sectors) ?:
1418                 bch2_trans_update(trans, &iter, &a->k_i, 0);
1419         bch2_trans_iter_exit(trans, &iter);
1420
1421         if (ret)
1422                 return ret;
1423
1424         if (!p.ptr.cached) {
1425                 ret = bch2_bucket_backpointer_mod(trans, bucket, bp, k, insert);
1426                 if (ret)
1427                         return ret;
1428         }
1429
1430         return 0;
1431 }
1432
1433 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1434                         struct extent_ptr_decoded p,
1435                         s64 sectors, enum bch_data_type data_type)
1436 {
1437         struct btree_iter iter;
1438         struct bkey_i_stripe *s;
1439         struct bch_replicas_padded r;
1440         int ret = 0;
1441
1442         s = bch2_bkey_get_mut_typed(trans, &iter,
1443                         BTREE_ID_stripes, POS(0, p.ec.idx),
1444                         BTREE_ITER_WITH_UPDATES, stripe);
1445         ret = PTR_ERR_OR_ZERO(s);
1446         if (unlikely(ret)) {
1447                 bch2_trans_inconsistent_on(bch2_err_matches(ret, ENOENT), trans,
1448                         "pointer to nonexistent stripe %llu",
1449                         (u64) p.ec.idx);
1450                 goto err;
1451         }
1452
1453         if (!bch2_ptr_matches_stripe(&s->v, p)) {
1454                 bch2_trans_inconsistent(trans,
1455                         "stripe pointer doesn't match stripe %llu",
1456                         (u64) p.ec.idx);
1457                 ret = -EIO;
1458                 goto err;
1459         }
1460
1461         stripe_blockcount_set(&s->v, p.ec.block,
1462                 stripe_blockcount_get(&s->v, p.ec.block) +
1463                 sectors);
1464
1465         bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1466         r.e.data_type = data_type;
1467         ret = update_replicas_list(trans, &r.e, sectors);
1468 err:
1469         bch2_trans_iter_exit(trans, &iter);
1470         return ret;
1471 }
1472
1473 static int __trans_mark_extent(struct btree_trans *trans,
1474                                enum btree_id btree_id, unsigned level,
1475                                struct bkey_s_c k, unsigned flags)
1476 {
1477         struct bch_fs *c = trans->c;
1478         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1479         const union bch_extent_entry *entry;
1480         struct extent_ptr_decoded p;
1481         struct bch_replicas_padded r;
1482         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
1483                 ? BCH_DATA_btree
1484                 : BCH_DATA_user;
1485         s64 sectors = bkey_is_btree_ptr(k.k)
1486                 ? btree_sectors(c)
1487                 : k.k->size;
1488         s64 dirty_sectors = 0;
1489         bool stale;
1490         int ret = 0;
1491
1492         r.e.data_type   = data_type;
1493         r.e.nr_devs     = 0;
1494         r.e.nr_required = 1;
1495
1496         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1497                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
1498
1499                 if (flags & BTREE_TRIGGER_OVERWRITE)
1500                         disk_sectors = -disk_sectors;
1501
1502                 ret = bch2_trans_mark_pointer(trans, btree_id, level, k, p, flags);
1503                 if (ret < 0)
1504                         return ret;
1505
1506                 stale = ret > 0;
1507
1508                 if (p.ptr.cached) {
1509                         if (!stale) {
1510                                 ret = update_cached_sectors_list(trans, p.ptr.dev,
1511                                                                  disk_sectors);
1512                                 if (ret)
1513                                         return ret;
1514                         }
1515                 } else if (!p.has_ec) {
1516                         dirty_sectors          += disk_sectors;
1517                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1518                 } else {
1519                         ret = bch2_trans_mark_stripe_ptr(trans, p,
1520                                         disk_sectors, data_type);
1521                         if (ret)
1522                                 return ret;
1523
1524                         r.e.nr_required = 0;
1525                 }
1526         }
1527
1528         if (r.e.nr_devs)
1529                 ret = update_replicas_list(trans, &r.e, dirty_sectors);
1530
1531         return ret;
1532 }
1533
1534 int bch2_trans_mark_extent(struct btree_trans *trans,
1535                            enum btree_id btree_id, unsigned level,
1536                            struct bkey_s_c old, struct bkey_i *new,
1537                            unsigned flags)
1538 {
1539         struct bch_fs *c = trans->c;
1540         int mod = (int) bch2_bkey_needs_rebalance(c, bkey_i_to_s_c(new)) -
1541                   (int) bch2_bkey_needs_rebalance(c, old);
1542
1543         if (mod) {
1544                 int ret = bch2_btree_bit_mod(trans, BTREE_ID_rebalance_work, new->k.p, mod > 0);
1545                 if (ret)
1546                         return ret;
1547         }
1548
1549         return trigger_run_overwrite_then_insert(__trans_mark_extent, trans, btree_id, level, old, new, flags);
1550 }
1551
1552 static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
1553                                          struct bkey_s_c_stripe s,
1554                                          unsigned idx, bool deleting)
1555 {
1556         struct bch_fs *c = trans->c;
1557         const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
1558         struct btree_iter iter;
1559         struct bkey_i_alloc_v4 *a;
1560         enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
1561                 ? BCH_DATA_parity : 0;
1562         s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
1563         int ret = 0;
1564
1565         if (deleting)
1566                 sectors = -sectors;
1567
1568         a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(c, ptr));
1569         if (IS_ERR(a))
1570                 return PTR_ERR(a);
1571
1572         ret = check_bucket_ref(trans, s.s_c, ptr, sectors, data_type,
1573                                a->v.gen, a->v.data_type,
1574                                a->v.dirty_sectors, a->v.cached_sectors);
1575         if (ret)
1576                 goto err;
1577
1578         if (!deleting) {
1579                 if (bch2_trans_inconsistent_on(a->v.stripe ||
1580                                                a->v.stripe_redundancy, trans,
1581                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
1582                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1583                                 bch2_data_types[a->v.data_type],
1584                                 a->v.dirty_sectors,
1585                                 a->v.stripe, s.k->p.offset)) {
1586                         ret = -EIO;
1587                         goto err;
1588                 }
1589
1590                 if (bch2_trans_inconsistent_on(data_type && a->v.dirty_sectors, trans,
1591                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
1592                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1593                                 bch2_data_types[a->v.data_type],
1594                                 a->v.dirty_sectors,
1595                                 s.k->p.offset)) {
1596                         ret = -EIO;
1597                         goto err;
1598                 }
1599
1600                 a->v.stripe             = s.k->p.offset;
1601                 a->v.stripe_redundancy  = s.v->nr_redundant;
1602                 a->v.data_type          = BCH_DATA_stripe;
1603         } else {
1604                 if (bch2_trans_inconsistent_on(a->v.stripe != s.k->p.offset ||
1605                                                a->v.stripe_redundancy != s.v->nr_redundant, trans,
1606                                 "bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
1607                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1608                                 s.k->p.offset, a->v.stripe)) {
1609                         ret = -EIO;
1610                         goto err;
1611                 }
1612
1613                 a->v.stripe             = 0;
1614                 a->v.stripe_redundancy  = 0;
1615                 a->v.data_type          = alloc_data_type(a->v, BCH_DATA_user);
1616         }
1617
1618         a->v.dirty_sectors += sectors;
1619         if (data_type)
1620                 a->v.data_type = !deleting ? data_type : 0;
1621
1622         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1623         if (ret)
1624                 goto err;
1625 err:
1626         bch2_trans_iter_exit(trans, &iter);
1627         return ret;
1628 }
1629
1630 int bch2_trans_mark_stripe(struct btree_trans *trans,
1631                            enum btree_id btree_id, unsigned level,
1632                            struct bkey_s_c old, struct bkey_i *new,
1633                            unsigned flags)
1634 {
1635         const struct bch_stripe *old_s = NULL;
1636         struct bch_stripe *new_s = NULL;
1637         struct bch_replicas_padded r;
1638         unsigned i, nr_blocks;
1639         int ret = 0;
1640
1641         if (old.k->type == KEY_TYPE_stripe)
1642                 old_s = bkey_s_c_to_stripe(old).v;
1643         if (new->k.type == KEY_TYPE_stripe)
1644                 new_s = &bkey_i_to_stripe(new)->v;
1645
1646         /*
1647          * If the pointers aren't changing, we don't need to do anything:
1648          */
1649         if (new_s && old_s &&
1650             new_s->nr_blocks    == old_s->nr_blocks &&
1651             new_s->nr_redundant == old_s->nr_redundant &&
1652             !memcmp(old_s->ptrs, new_s->ptrs,
1653                     new_s->nr_blocks * sizeof(struct bch_extent_ptr)))
1654                 return 0;
1655
1656         BUG_ON(new_s && old_s &&
1657                (new_s->nr_blocks        != old_s->nr_blocks ||
1658                 new_s->nr_redundant     != old_s->nr_redundant));
1659
1660         nr_blocks = new_s ? new_s->nr_blocks : old_s->nr_blocks;
1661
1662         if (new_s) {
1663                 s64 sectors = le16_to_cpu(new_s->sectors);
1664
1665                 bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(new));
1666                 ret = update_replicas_list(trans, &r.e, sectors * new_s->nr_redundant);
1667                 if (ret)
1668                         return ret;
1669         }
1670
1671         if (old_s) {
1672                 s64 sectors = -((s64) le16_to_cpu(old_s->sectors));
1673
1674                 bch2_bkey_to_replicas(&r.e, old);
1675                 ret = update_replicas_list(trans, &r.e, sectors * old_s->nr_redundant);
1676                 if (ret)
1677                         return ret;
1678         }
1679
1680         for (i = 0; i < nr_blocks; i++) {
1681                 if (new_s && old_s &&
1682                     !memcmp(&new_s->ptrs[i],
1683                             &old_s->ptrs[i],
1684                             sizeof(new_s->ptrs[i])))
1685                         continue;
1686
1687                 if (new_s) {
1688                         ret = bch2_trans_mark_stripe_bucket(trans,
1689                                         bkey_i_to_s_c_stripe(new), i, false);
1690                         if (ret)
1691                                 break;
1692                 }
1693
1694                 if (old_s) {
1695                         ret = bch2_trans_mark_stripe_bucket(trans,
1696                                         bkey_s_c_to_stripe(old), i, true);
1697                         if (ret)
1698                                 break;
1699                 }
1700         }
1701
1702         return ret;
1703 }
1704
1705 static int __trans_mark_reservation(struct btree_trans *trans,
1706                                     enum btree_id btree_id, unsigned level,
1707                                     struct bkey_s_c k, unsigned flags)
1708 {
1709         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1710         s64 sectors = (s64) k.k->size;
1711         struct replicas_delta_list *d;
1712         int ret;
1713
1714         if (flags & BTREE_TRIGGER_OVERWRITE)
1715                 sectors = -sectors;
1716         sectors *= replicas;
1717
1718         ret = bch2_replicas_deltas_realloc(trans, 0);
1719         if (ret)
1720                 return ret;
1721
1722         d = trans->fs_usage_deltas;
1723         replicas = clamp_t(unsigned, replicas, 1,
1724                            ARRAY_SIZE(d->persistent_reserved));
1725
1726         d->persistent_reserved[replicas - 1] += sectors;
1727         return 0;
1728 }
1729
1730 int bch2_trans_mark_reservation(struct btree_trans *trans,
1731                                 enum btree_id btree_id, unsigned level,
1732                                 struct bkey_s_c old,
1733                                 struct bkey_i *new,
1734                                 unsigned flags)
1735 {
1736         return trigger_run_overwrite_then_insert(__trans_mark_reservation, trans, btree_id, level, old, new, flags);
1737 }
1738
1739 static int trans_mark_reflink_p_segment(struct btree_trans *trans,
1740                         struct bkey_s_c_reflink_p p,
1741                         u64 *idx, unsigned flags)
1742 {
1743         struct bch_fs *c = trans->c;
1744         struct btree_iter iter;
1745         struct bkey_i *k;
1746         __le64 *refcount;
1747         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1748         struct printbuf buf = PRINTBUF;
1749         int ret;
1750
1751         k = bch2_bkey_get_mut_noupdate(trans, &iter,
1752                         BTREE_ID_reflink, POS(0, *idx),
1753                         BTREE_ITER_WITH_UPDATES);
1754         ret = PTR_ERR_OR_ZERO(k);
1755         if (ret)
1756                 goto err;
1757
1758         refcount = bkey_refcount(k);
1759         if (!refcount) {
1760                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1761                 bch2_trans_inconsistent(trans,
1762                         "nonexistent indirect extent at %llu while marking\n  %s",
1763                         *idx, buf.buf);
1764                 ret = -EIO;
1765                 goto err;
1766         }
1767
1768         if (!*refcount && (flags & BTREE_TRIGGER_OVERWRITE)) {
1769                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1770                 bch2_trans_inconsistent(trans,
1771                         "indirect extent refcount underflow at %llu while marking\n  %s",
1772                         *idx, buf.buf);
1773                 ret = -EIO;
1774                 goto err;
1775         }
1776
1777         if (flags & BTREE_TRIGGER_INSERT) {
1778                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1779                 u64 pad;
1780
1781                 pad = max_t(s64, le32_to_cpu(v->front_pad),
1782                             le64_to_cpu(v->idx) - bkey_start_offset(&k->k));
1783                 BUG_ON(pad > U32_MAX);
1784                 v->front_pad = cpu_to_le32(pad);
1785
1786                 pad = max_t(s64, le32_to_cpu(v->back_pad),
1787                             k->k.p.offset - p.k->size - le64_to_cpu(v->idx));
1788                 BUG_ON(pad > U32_MAX);
1789                 v->back_pad = cpu_to_le32(pad);
1790         }
1791
1792         le64_add_cpu(refcount, add);
1793
1794         bch2_btree_iter_set_pos_to_extent_start(&iter);
1795         ret = bch2_trans_update(trans, &iter, k, 0);
1796         if (ret)
1797                 goto err;
1798
1799         *idx = k->k.p.offset;
1800 err:
1801         bch2_trans_iter_exit(trans, &iter);
1802         printbuf_exit(&buf);
1803         return ret;
1804 }
1805
1806 static int __trans_mark_reflink_p(struct btree_trans *trans,
1807                                 enum btree_id btree_id, unsigned level,
1808                                 struct bkey_s_c k, unsigned flags)
1809 {
1810         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1811         u64 idx, end_idx;
1812         int ret = 0;
1813
1814         idx     = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
1815         end_idx = le64_to_cpu(p.v->idx) + p.k->size +
1816                 le32_to_cpu(p.v->back_pad);
1817
1818         while (idx < end_idx && !ret)
1819                 ret = trans_mark_reflink_p_segment(trans, p, &idx, flags);
1820         return ret;
1821 }
1822
1823 int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1824                               enum btree_id btree_id, unsigned level,
1825                               struct bkey_s_c old,
1826                               struct bkey_i *new,
1827                               unsigned flags)
1828 {
1829         if (flags & BTREE_TRIGGER_INSERT) {
1830                 struct bch_reflink_p *v = &bkey_i_to_reflink_p(new)->v;
1831
1832                 v->front_pad = v->back_pad = 0;
1833         }
1834
1835         return trigger_run_overwrite_then_insert(__trans_mark_reflink_p, trans, btree_id, level, old, new, flags);
1836 }
1837
1838 static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1839                                     struct bch_dev *ca, size_t b,
1840                                     enum bch_data_type type,
1841                                     unsigned sectors)
1842 {
1843         struct bch_fs *c = trans->c;
1844         struct btree_iter iter;
1845         struct bkey_i_alloc_v4 *a;
1846         int ret = 0;
1847
1848         /*
1849          * Backup superblock might be past the end of our normal usable space:
1850          */
1851         if (b >= ca->mi.nbuckets)
1852                 return 0;
1853
1854         a = bch2_trans_start_alloc_update(trans, &iter, POS(ca->dev_idx, b));
1855         if (IS_ERR(a))
1856                 return PTR_ERR(a);
1857
1858         if (a->v.data_type && type && a->v.data_type != type) {
1859                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1860                         "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
1861                         "while marking %s",
1862                         iter.pos.inode, iter.pos.offset, a->v.gen,
1863                         bch2_data_types[a->v.data_type],
1864                         bch2_data_types[type],
1865                         bch2_data_types[type]);
1866                 ret = -EIO;
1867                 goto err;
1868         }
1869
1870         if (a->v.data_type      != type ||
1871             a->v.dirty_sectors  != sectors) {
1872                 a->v.data_type          = type;
1873                 a->v.dirty_sectors      = sectors;
1874                 ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1875         }
1876 err:
1877         bch2_trans_iter_exit(trans, &iter);
1878         return ret;
1879 }
1880
1881 int bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1882                                     struct bch_dev *ca, size_t b,
1883                                     enum bch_data_type type,
1884                                     unsigned sectors)
1885 {
1886         return commit_do(trans, NULL, NULL, 0,
1887                         __bch2_trans_mark_metadata_bucket(trans, ca, b, type, sectors));
1888 }
1889
1890 static int bch2_trans_mark_metadata_sectors(struct btree_trans *trans,
1891                                             struct bch_dev *ca,
1892                                             u64 start, u64 end,
1893                                             enum bch_data_type type,
1894                                             u64 *bucket, unsigned *bucket_sectors)
1895 {
1896         do {
1897                 u64 b = sector_to_bucket(ca, start);
1898                 unsigned sectors =
1899                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1900
1901                 if (b != *bucket && *bucket_sectors) {
1902                         int ret = bch2_trans_mark_metadata_bucket(trans, ca, *bucket,
1903                                                                   type, *bucket_sectors);
1904                         if (ret)
1905                                 return ret;
1906
1907                         *bucket_sectors = 0;
1908                 }
1909
1910                 *bucket         = b;
1911                 *bucket_sectors += sectors;
1912                 start += sectors;
1913         } while (start < end);
1914
1915         return 0;
1916 }
1917
1918 static int __bch2_trans_mark_dev_sb(struct btree_trans *trans,
1919                                     struct bch_dev *ca)
1920 {
1921         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1922         u64 bucket = 0;
1923         unsigned i, bucket_sectors = 0;
1924         int ret;
1925
1926         for (i = 0; i < layout->nr_superblocks; i++) {
1927                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1928
1929                 if (offset == BCH_SB_SECTOR) {
1930                         ret = bch2_trans_mark_metadata_sectors(trans, ca,
1931                                                 0, BCH_SB_SECTOR,
1932                                                 BCH_DATA_sb, &bucket, &bucket_sectors);
1933                         if (ret)
1934                                 return ret;
1935                 }
1936
1937                 ret = bch2_trans_mark_metadata_sectors(trans, ca, offset,
1938                                       offset + (1 << layout->sb_max_size_bits),
1939                                       BCH_DATA_sb, &bucket, &bucket_sectors);
1940                 if (ret)
1941                         return ret;
1942         }
1943
1944         if (bucket_sectors) {
1945                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1946                                 bucket, BCH_DATA_sb, bucket_sectors);
1947                 if (ret)
1948                         return ret;
1949         }
1950
1951         for (i = 0; i < ca->journal.nr; i++) {
1952                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1953                                 ca->journal.buckets[i],
1954                                 BCH_DATA_journal, ca->mi.bucket_size);
1955                 if (ret)
1956                         return ret;
1957         }
1958
1959         return 0;
1960 }
1961
1962 int bch2_trans_mark_dev_sb(struct bch_fs *c, struct bch_dev *ca)
1963 {
1964         int ret = bch2_trans_run(c, __bch2_trans_mark_dev_sb(trans, ca));
1965
1966         if (ret)
1967                 bch_err_fn(c, ret);
1968         return ret;
1969 }
1970
1971 int bch2_trans_mark_dev_sbs(struct bch_fs *c)
1972 {
1973         struct bch_dev *ca;
1974         unsigned i;
1975
1976         for_each_online_member(ca, c, i) {
1977                 int ret = bch2_trans_mark_dev_sb(c, ca);
1978                 if (ret) {
1979                         percpu_ref_put(&ca->ref);
1980                         return ret;
1981                 }
1982         }
1983
1984         return 0;
1985 }
1986
1987 /* Disk reservations: */
1988
1989 #define SECTORS_CACHE   1024
1990
1991 int __bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
1992                               u64 sectors, int flags)
1993 {
1994         struct bch_fs_pcpu *pcpu;
1995         u64 old, v, get;
1996         s64 sectors_available;
1997         int ret;
1998
1999         percpu_down_read(&c->mark_lock);
2000         preempt_disable();
2001         pcpu = this_cpu_ptr(c->pcpu);
2002
2003         if (sectors <= pcpu->sectors_available)
2004                 goto out;
2005
2006         v = atomic64_read(&c->sectors_available);
2007         do {
2008                 old = v;
2009                 get = min((u64) sectors + SECTORS_CACHE, old);
2010
2011                 if (get < sectors) {
2012                         preempt_enable();
2013                         goto recalculate;
2014                 }
2015         } while ((v = atomic64_cmpxchg(&c->sectors_available,
2016                                        old, old - get)) != old);
2017
2018         pcpu->sectors_available         += get;
2019
2020 out:
2021         pcpu->sectors_available         -= sectors;
2022         this_cpu_add(*c->online_reserved, sectors);
2023         res->sectors                    += sectors;
2024
2025         preempt_enable();
2026         percpu_up_read(&c->mark_lock);
2027         return 0;
2028
2029 recalculate:
2030         mutex_lock(&c->sectors_available_lock);
2031
2032         percpu_u64_set(&c->pcpu->sectors_available, 0);
2033         sectors_available = avail_factor(__bch2_fs_usage_read_short(c).free);
2034
2035         if (sectors <= sectors_available ||
2036             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
2037                 atomic64_set(&c->sectors_available,
2038                              max_t(s64, 0, sectors_available - sectors));
2039                 this_cpu_add(*c->online_reserved, sectors);
2040                 res->sectors                    += sectors;
2041                 ret = 0;
2042         } else {
2043                 atomic64_set(&c->sectors_available, sectors_available);
2044                 ret = -BCH_ERR_ENOSPC_disk_reservation;
2045         }
2046
2047         mutex_unlock(&c->sectors_available_lock);
2048         percpu_up_read(&c->mark_lock);
2049
2050         return ret;
2051 }
2052
2053 /* Startup/shutdown: */
2054
2055 static void bucket_gens_free_rcu(struct rcu_head *rcu)
2056 {
2057         struct bucket_gens *buckets =
2058                 container_of(rcu, struct bucket_gens, rcu);
2059
2060         kvpfree(buckets, sizeof(*buckets) + buckets->nbuckets);
2061 }
2062
2063 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
2064 {
2065         struct bucket_gens *bucket_gens = NULL, *old_bucket_gens = NULL;
2066         unsigned long *buckets_nouse = NULL;
2067         bool resize = ca->bucket_gens != NULL;
2068         int ret;
2069
2070         if (!(bucket_gens       = kvpmalloc(sizeof(struct bucket_gens) + nbuckets,
2071                                             GFP_KERNEL|__GFP_ZERO))) {
2072                 ret = -BCH_ERR_ENOMEM_bucket_gens;
2073                 goto err;
2074         }
2075
2076         if ((c->opts.buckets_nouse &&
2077              !(buckets_nouse    = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2078                                             sizeof(unsigned long),
2079                                             GFP_KERNEL|__GFP_ZERO)))) {
2080                 ret = -BCH_ERR_ENOMEM_buckets_nouse;
2081                 goto err;
2082         }
2083
2084         bucket_gens->first_bucket = ca->mi.first_bucket;
2085         bucket_gens->nbuckets   = nbuckets;
2086
2087         bch2_copygc_stop(c);
2088
2089         if (resize) {
2090                 down_write(&c->gc_lock);
2091                 down_write(&ca->bucket_lock);
2092                 percpu_down_write(&c->mark_lock);
2093         }
2094
2095         old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
2096
2097         if (resize) {
2098                 size_t n = min(bucket_gens->nbuckets, old_bucket_gens->nbuckets);
2099
2100                 memcpy(bucket_gens->b,
2101                        old_bucket_gens->b,
2102                        n);
2103                 if (buckets_nouse)
2104                         memcpy(buckets_nouse,
2105                                ca->buckets_nouse,
2106                                BITS_TO_LONGS(n) * sizeof(unsigned long));
2107         }
2108
2109         rcu_assign_pointer(ca->bucket_gens, bucket_gens);
2110         bucket_gens     = old_bucket_gens;
2111
2112         swap(ca->buckets_nouse, buckets_nouse);
2113
2114         nbuckets = ca->mi.nbuckets;
2115
2116         if (resize) {
2117                 percpu_up_write(&c->mark_lock);
2118                 up_write(&ca->bucket_lock);
2119                 up_write(&c->gc_lock);
2120         }
2121
2122         ret = 0;
2123 err:
2124         kvpfree(buckets_nouse,
2125                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2126         if (bucket_gens)
2127                 call_rcu(&bucket_gens->rcu, bucket_gens_free_rcu);
2128
2129         return ret;
2130 }
2131
2132 void bch2_dev_buckets_free(struct bch_dev *ca)
2133 {
2134         unsigned i;
2135
2136         kvpfree(ca->buckets_nouse,
2137                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2138         kvpfree(rcu_dereference_protected(ca->bucket_gens, 1),
2139                 sizeof(struct bucket_gens) + ca->mi.nbuckets);
2140
2141         for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
2142                 free_percpu(ca->usage[i]);
2143         kfree(ca->usage_base);
2144 }
2145
2146 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2147 {
2148         unsigned i;
2149
2150         ca->usage_base = kzalloc(sizeof(struct bch_dev_usage), GFP_KERNEL);
2151         if (!ca->usage_base)
2152                 return -BCH_ERR_ENOMEM_usage_init;
2153
2154         for (i = 0; i < ARRAY_SIZE(ca->usage); i++) {
2155                 ca->usage[i] = alloc_percpu(struct bch_dev_usage);
2156                 if (!ca->usage[i])
2157                         return -BCH_ERR_ENOMEM_usage_init;
2158         }
2159
2160         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);
2161 }