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 static int __mark_extent(struct btree_trans *trans,
 952                          enum btree_id btree_id, unsigned level,
 953                          struct bkey_s_c k, unsigned flags)
 954 {
 955         u64 journal_seq = trans->journal_res.seq;
 956         struct bch_fs *c = trans->c;
 957         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 958         const union bch_extent_entry *entry;
 959         struct extent_ptr_decoded p;
 960         struct bch_replicas_padded r;
 961         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
 962                 ? BCH_DATA_btree
 963                 : BCH_DATA_user;
 964         s64 sectors = bkey_is_btree_ptr(k.k)
 965                 ? btree_sectors(c)
 966                 : k.k->size;
 967         s64 dirty_sectors = 0;
 968         bool stale;
 969         int ret;
 970
 971         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 972
 973         r.e.data_type   = data_type;
 974         r.e.nr_devs     = 0;
 975         r.e.nr_required = 1;
 976
 977         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 978                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
 979
 980                 if (flags & BTREE_TRIGGER_OVERWRITE)
 981                         disk_sectors = -disk_sectors;
 982
 983                 ret = bch2_mark_pointer(trans, btree_id, level, k, p, disk_sectors, flags);
 984                 if (ret < 0)
 985                         return ret;
 986
 987                 stale = ret > 0;
 988
 989                 if (p.ptr.cached) {
 990                         if (!stale) {
 991                                 ret = update_cached_sectors(c, k, p.ptr.dev,
 992                                                 disk_sectors, journal_seq, true);
 993                                 if (ret) {
 994                                         bch2_fs_fatal_error(c, "%s(): no replicas entry while updating cached sectors",
 995                                                             __func__);
 996                                         return ret;
 997                                 }
 998                         }
 999                 } else if (!p.has_ec) {
1000                         dirty_sectors          += disk_sectors;
1001                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1002                 } else {
1003                         ret = bch2_mark_stripe_ptr(trans, k, p.ec, data_type,
1004                                         disk_sectors, flags);
1005                         if (ret)
1006                                 return ret;
1007
1008                         /*
1009                          * There may be other dirty pointers in this extent, but
1010                          * if so they're not required for mounting if we have an
1011                          * erasure coded pointer in this extent:
1012                          */
1013                         r.e.nr_required = 0;
1014                 }
1015         }
1016
1017         if (r.e.nr_devs) {
1018                 ret = update_replicas(c, k, &r.e, dirty_sectors, journal_seq, true);
1019                 if (ret) {
1020                         struct printbuf buf = PRINTBUF;
1021
1022                         bch2_bkey_val_to_text(&buf, c, k);
1023                         bch2_fs_fatal_error(c, "%s(): no replicas entry for %s", __func__, buf.buf);
1024                         printbuf_exit(&buf);
1025                         return ret;
1026                 }
1027         }
1028
1029         return 0;
1030 }
1031
1032 int bch2_mark_extent(struct btree_trans *trans,
1033                      enum btree_id btree_id, unsigned level,
1034                      struct bkey_s_c old, struct bkey_s_c new,
1035                      unsigned flags)
1036 {
1037         return mem_trigger_run_insert_then_overwrite(__mark_extent, trans, btree_id, level, old, new, flags);
1038 }
1039
1040 int bch2_mark_stripe(struct btree_trans *trans,
1041                      enum btree_id btree_id, unsigned level,
1042                      struct bkey_s_c old, struct bkey_s_c new,
1043                      unsigned flags)
1044 {
1045         bool gc = flags & BTREE_TRIGGER_GC;
1046         u64 journal_seq = trans->journal_res.seq;
1047         struct bch_fs *c = trans->c;
1048         u64 idx = new.k->p.offset;
1049         const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
1050                 ? bkey_s_c_to_stripe(old).v : NULL;
1051         const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
1052                 ? bkey_s_c_to_stripe(new).v : NULL;
1053         unsigned i;
1054         int ret;
1055
1056         BUG_ON(gc && old_s);
1057
1058         if (!gc) {
1059                 struct stripe *m = genradix_ptr(&c->stripes, idx);
1060
1061                 if (!m) {
1062                         struct printbuf buf1 = PRINTBUF;
1063                         struct printbuf buf2 = PRINTBUF;
1064
1065                         bch2_bkey_val_to_text(&buf1, c, old);
1066                         bch2_bkey_val_to_text(&buf2, c, new);
1067                         bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
1068                                             "old %s\n"
1069                                             "new %s", idx, buf1.buf, buf2.buf);
1070                         printbuf_exit(&buf2);
1071                         printbuf_exit(&buf1);
1072                         bch2_inconsistent_error(c);
1073                         return -1;
1074                 }
1075
1076                 if (!new_s) {
1077                         bch2_stripes_heap_del(c, m, idx);
1078
1079                         memset(m, 0, sizeof(*m));
1080                 } else {
1081                         m->sectors      = le16_to_cpu(new_s->sectors);
1082                         m->algorithm    = new_s->algorithm;
1083                         m->nr_blocks    = new_s->nr_blocks;
1084                         m->nr_redundant = new_s->nr_redundant;
1085                         m->blocks_nonempty = 0;
1086
1087                         for (i = 0; i < new_s->nr_blocks; i++)
1088                                 m->blocks_nonempty += !!stripe_blockcount_get(new_s, i);
1089
1090                         if (!old_s)
1091                                 bch2_stripes_heap_insert(c, m, idx);
1092                         else
1093                                 bch2_stripes_heap_update(c, m, idx);
1094                 }
1095         } else {
1096                 struct gc_stripe *m =
1097                         genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
1098
1099                 if (!m) {
1100                         bch_err(c, "error allocating memory for gc_stripes, idx %llu",
1101                                 idx);
1102                         return -BCH_ERR_ENOMEM_mark_stripe;
1103                 }
1104                 /*
1105                  * This will be wrong when we bring back runtime gc: we should
1106                  * be unmarking the old key and then marking the new key
1107                  */
1108                 m->alive        = true;
1109                 m->sectors      = le16_to_cpu(new_s->sectors);
1110                 m->nr_blocks    = new_s->nr_blocks;
1111                 m->nr_redundant = new_s->nr_redundant;
1112
1113                 for (i = 0; i < new_s->nr_blocks; i++)
1114                         m->ptrs[i] = new_s->ptrs[i];
1115
1116                 bch2_bkey_to_replicas(&m->r.e, new);
1117
1118                 /*
1119                  * gc recalculates this field from stripe ptr
1120                  * references:
1121                  */
1122                 memset(m->block_sectors, 0, sizeof(m->block_sectors));
1123
1124                 for (i = 0; i < new_s->nr_blocks; i++) {
1125                         ret = mark_stripe_bucket(trans, new, i, flags);
1126                         if (ret)
1127                                 return ret;
1128                 }
1129
1130                 ret = update_replicas(c, new, &m->r.e,
1131                                       ((s64) m->sectors * m->nr_redundant),
1132                                       journal_seq, gc);
1133                 if (ret) {
1134                         struct printbuf buf = PRINTBUF;
1135
1136                         bch2_bkey_val_to_text(&buf, c, new);
1137                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf.buf);
1138                         printbuf_exit(&buf);
1139                         return ret;
1140                 }
1141         }
1142
1143         return 0;
1144 }
1145
1146 int bch2_mark_inode(struct btree_trans *trans,
1147                     enum btree_id btree_id, unsigned level,
1148                     struct bkey_s_c old, struct bkey_s_c new,
1149                     unsigned flags)
1150 {
1151         struct bch_fs *c = trans->c;
1152         struct bch_fs_usage __percpu *fs_usage;
1153         u64 journal_seq = trans->journal_res.seq;
1154
1155         if (flags & BTREE_TRIGGER_INSERT) {
1156                 struct bch_inode_v3 *v = (struct bch_inode_v3 *) new.v;
1157
1158                 BUG_ON(!journal_seq);
1159                 BUG_ON(new.k->type != KEY_TYPE_inode_v3);
1160
1161                 v->bi_journal_seq = cpu_to_le64(journal_seq);
1162         }
1163
1164         if (flags & BTREE_TRIGGER_GC) {
1165                 percpu_down_read(&c->mark_lock);
1166                 preempt_disable();
1167
1168                 fs_usage = fs_usage_ptr(c, journal_seq, flags & BTREE_TRIGGER_GC);
1169                 fs_usage->nr_inodes += bkey_is_inode(new.k);
1170                 fs_usage->nr_inodes -= bkey_is_inode(old.k);
1171
1172                 preempt_enable();
1173                 percpu_up_read(&c->mark_lock);
1174         }
1175         return 0;
1176 }
1177
1178 static int __mark_reservation(struct btree_trans *trans,
1179                               enum btree_id btree_id, unsigned level,
1180                               struct bkey_s_c k, unsigned flags)
1181 {
1182         struct bch_fs *c = trans->c;
1183         struct bch_fs_usage __percpu *fs_usage;
1184         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1185         s64 sectors = (s64) k.k->size;
1186
1187         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1188
1189         if (flags & BTREE_TRIGGER_OVERWRITE)
1190                 sectors = -sectors;
1191         sectors *= replicas;
1192
1193         percpu_down_read(&c->mark_lock);
1194         preempt_disable();
1195
1196         fs_usage = fs_usage_ptr(c, trans->journal_res.seq, flags & BTREE_TRIGGER_GC);
1197         replicas = clamp_t(unsigned, replicas, 1,
1198                            ARRAY_SIZE(fs_usage->persistent_reserved));
1199
1200         fs_usage->reserved                              += sectors;
1201         fs_usage->persistent_reserved[replicas - 1]     += sectors;
1202
1203         preempt_enable();
1204         percpu_up_read(&c->mark_lock);
1205
1206         return 0;
1207 }
1208
1209 int bch2_mark_reservation(struct btree_trans *trans,
1210                           enum btree_id btree_id, unsigned level,
1211                           struct bkey_s_c old, struct bkey_s_c new,
1212                           unsigned flags)
1213 {
1214         return mem_trigger_run_insert_then_overwrite(__mark_reservation, trans, btree_id, level, old, new, flags);
1215 }
1216
1217 static s64 __bch2_mark_reflink_p(struct btree_trans *trans,
1218                                  struct bkey_s_c_reflink_p p,
1219                                  u64 start, u64 end,
1220                                  u64 *idx, unsigned flags, size_t r_idx)
1221 {
1222         struct bch_fs *c = trans->c;
1223         struct reflink_gc *r;
1224         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1225         u64 next_idx = end;
1226         s64 ret = 0;
1227         struct printbuf buf = PRINTBUF;
1228
1229         if (r_idx >= c->reflink_gc_nr)
1230                 goto not_found;
1231
1232         r = genradix_ptr(&c->reflink_gc_table, r_idx);
1233         next_idx = min(next_idx, r->offset - r->size);
1234         if (*idx < next_idx)
1235                 goto not_found;
1236
1237         BUG_ON((s64) r->refcount + add < 0);
1238
1239         r->refcount += add;
1240         *idx = r->offset;
1241         return 0;
1242 not_found:
1243         if (fsck_err(c, "pointer to missing indirect extent\n"
1244                      "  %s\n"
1245                      "  missing range %llu-%llu",
1246                      (bch2_bkey_val_to_text(&buf, c, p.s_c), buf.buf),
1247                      *idx, next_idx)) {
1248                 struct bkey_i_error *new;
1249
1250                 new = bch2_trans_kmalloc(trans, sizeof(*new));
1251                 ret = PTR_ERR_OR_ZERO(new);
1252                 if (ret)
1253                         goto err;
1254
1255                 bkey_init(&new->k);
1256                 new->k.type     = KEY_TYPE_error;
1257                 new->k.p                = bkey_start_pos(p.k);
1258                 new->k.p.offset += *idx - start;
1259                 bch2_key_resize(&new->k, next_idx - *idx);
1260                 ret = __bch2_btree_insert(trans, BTREE_ID_extents, &new->k_i,
1261                                           BTREE_TRIGGER_NORUN);
1262         }
1263
1264         *idx = next_idx;
1265 err:
1266 fsck_err:
1267         printbuf_exit(&buf);
1268         return ret;
1269 }
1270
1271 static int __mark_reflink_p(struct btree_trans *trans,
1272                             enum btree_id btree_id, unsigned level,
1273                             struct bkey_s_c k, unsigned flags)
1274 {
1275         struct bch_fs *c = trans->c;
1276         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1277         struct reflink_gc *ref;
1278         size_t l, r, m;
1279         u64 idx = le64_to_cpu(p.v->idx), start = idx;
1280         u64 end = le64_to_cpu(p.v->idx) + p.k->size;
1281         int ret = 0;
1282
1283         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1284
1285         if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix) {
1286                 idx -= le32_to_cpu(p.v->front_pad);
1287                 end += le32_to_cpu(p.v->back_pad);
1288         }
1289
1290         l = 0;
1291         r = c->reflink_gc_nr;
1292         while (l < r) {
1293                 m = l + (r - l) / 2;
1294
1295                 ref = genradix_ptr(&c->reflink_gc_table, m);
1296                 if (ref->offset <= idx)
1297                         l = m + 1;
1298                 else
1299                         r = m;
1300         }
1301
1302         while (idx < end && !ret)
1303                 ret = __bch2_mark_reflink_p(trans, p, start, end,
1304                                             &idx, flags, l++);
1305
1306         return ret;
1307 }
1308
1309 int bch2_mark_reflink_p(struct btree_trans *trans,
1310                         enum btree_id btree_id, unsigned level,
1311                         struct bkey_s_c old, struct bkey_s_c new,
1312                         unsigned flags)
1313 {
1314         return mem_trigger_run_insert_then_overwrite(__mark_reflink_p, trans, btree_id, level, old, new, flags);
1315 }
1316
1317 void bch2_trans_fs_usage_revert(struct btree_trans *trans,
1318                                 struct replicas_delta_list *deltas)
1319 {
1320         struct bch_fs *c = trans->c;
1321         struct bch_fs_usage *dst;
1322         struct replicas_delta *d, *top = (void *) deltas->d + deltas->used;
1323         s64 added = 0;
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         /* revert changes: */
1331         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1332                 switch (d->r.data_type) {
1333                 case BCH_DATA_btree:
1334                 case BCH_DATA_user:
1335                 case BCH_DATA_parity:
1336                         added += d->delta;
1337                 }
1338                 BUG_ON(__update_replicas(c, dst, &d->r, -d->delta));
1339         }
1340
1341         dst->nr_inodes -= deltas->nr_inodes;
1342
1343         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1344                 added                           -= deltas->persistent_reserved[i];
1345                 dst->reserved                   -= deltas->persistent_reserved[i];
1346                 dst->persistent_reserved[i]     -= deltas->persistent_reserved[i];
1347         }
1348
1349         if (added > 0) {
1350                 trans->disk_res->sectors += added;
1351                 this_cpu_add(*c->online_reserved, added);
1352         }
1353
1354         preempt_enable();
1355         percpu_up_read(&c->mark_lock);
1356 }
1357
1358 int bch2_trans_fs_usage_apply(struct btree_trans *trans,
1359                               struct replicas_delta_list *deltas)
1360 {
1361         struct bch_fs *c = trans->c;
1362         static int warned_disk_usage = 0;
1363         bool warn = false;
1364         unsigned disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1365         struct replicas_delta *d = deltas->d, *d2;
1366         struct replicas_delta *top = (void *) deltas->d + deltas->used;
1367         struct bch_fs_usage *dst;
1368         s64 added = 0, should_not_have_added;
1369         unsigned i;
1370
1371         percpu_down_read(&c->mark_lock);
1372         preempt_disable();
1373         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1374
1375         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1376                 switch (d->r.data_type) {
1377                 case BCH_DATA_btree:
1378                 case BCH_DATA_user:
1379                 case BCH_DATA_parity:
1380                         added += d->delta;
1381                 }
1382
1383                 if (__update_replicas(c, dst, &d->r, d->delta))
1384                         goto need_mark;
1385         }
1386
1387         dst->nr_inodes += deltas->nr_inodes;
1388
1389         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1390                 added                           += deltas->persistent_reserved[i];
1391                 dst->reserved                   += deltas->persistent_reserved[i];
1392                 dst->persistent_reserved[i]     += deltas->persistent_reserved[i];
1393         }
1394
1395         /*
1396          * Not allowed to reduce sectors_available except by getting a
1397          * reservation:
1398          */
1399         should_not_have_added = added - (s64) disk_res_sectors;
1400         if (unlikely(should_not_have_added > 0)) {
1401                 u64 old, new, v = atomic64_read(&c->sectors_available);
1402
1403                 do {
1404                         old = v;
1405                         new = max_t(s64, 0, old - should_not_have_added);
1406                 } while ((v = atomic64_cmpxchg(&c->sectors_available,
1407                                                old, new)) != old);
1408
1409                 added -= should_not_have_added;
1410                 warn = true;
1411         }
1412
1413         if (added > 0) {
1414                 trans->disk_res->sectors -= added;
1415                 this_cpu_sub(*c->online_reserved, added);
1416         }
1417
1418         preempt_enable();
1419         percpu_up_read(&c->mark_lock);
1420
1421         if (unlikely(warn) && !xchg(&warned_disk_usage, 1))
1422                 bch2_trans_inconsistent(trans,
1423                                         "disk usage increased %lli more than %u sectors reserved)",
1424                                         should_not_have_added, disk_res_sectors);
1425         return 0;
1426 need_mark:
1427         /* revert changes: */
1428         for (d2 = deltas->d; d2 != d; d2 = replicas_delta_next(d2))
1429                 BUG_ON(__update_replicas(c, dst, &d2->r, -d2->delta));
1430
1431         preempt_enable();
1432         percpu_up_read(&c->mark_lock);
1433         return -1;
1434 }
1435
1436 /* trans_mark: */
1437
1438 static inline int bch2_trans_mark_pointer(struct btree_trans *trans,
1439                                    enum btree_id btree_id, unsigned level,
1440                                    struct bkey_s_c k, struct extent_ptr_decoded p,
1441                                    unsigned flags)
1442 {
1443         bool insert = !(flags & BTREE_TRIGGER_OVERWRITE);
1444         struct btree_iter iter;
1445         struct bkey_i_alloc_v4 *a;
1446         struct bpos bucket;
1447         struct bch_backpointer bp;
1448         s64 sectors;
1449         int ret;
1450
1451         bch2_extent_ptr_to_bp(trans->c, btree_id, level, k, p, &bucket, &bp);
1452         sectors = bp.bucket_len;
1453         if (!insert)
1454                 sectors = -sectors;
1455
1456         a = bch2_trans_start_alloc_update(trans, &iter, bucket);
1457         if (IS_ERR(a))
1458                 return PTR_ERR(a);
1459
1460         ret = __mark_pointer(trans, k, &p.ptr, sectors, bp.data_type,
1461                              a->v.gen, &a->v.data_type,
1462                              &a->v.dirty_sectors, &a->v.cached_sectors) ?:
1463                 bch2_trans_update(trans, &iter, &a->k_i, 0);
1464         bch2_trans_iter_exit(trans, &iter);
1465
1466         if (ret)
1467                 return ret;
1468
1469         if (!p.ptr.cached) {
1470                 ret = bch2_bucket_backpointer_mod(trans, bucket, bp, k, insert);
1471                 if (ret)
1472                         return ret;
1473         }
1474
1475         return 0;
1476 }
1477
1478 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1479                         struct extent_ptr_decoded p,
1480                         s64 sectors, enum bch_data_type data_type)
1481 {
1482         struct btree_iter iter;
1483         struct bkey_i_stripe *s;
1484         struct bch_replicas_padded r;
1485         int ret = 0;
1486
1487         s = bch2_bkey_get_mut_typed(trans, &iter,
1488                         BTREE_ID_stripes, POS(0, p.ec.idx),
1489                         BTREE_ITER_WITH_UPDATES, stripe);
1490         ret = PTR_ERR_OR_ZERO(s);
1491         if (unlikely(ret)) {
1492                 bch2_trans_inconsistent_on(bch2_err_matches(ret, ENOENT), trans,
1493                         "pointer to nonexistent stripe %llu",
1494                         (u64) p.ec.idx);
1495                 goto err;
1496         }
1497
1498         if (!bch2_ptr_matches_stripe(&s->v, p)) {
1499                 bch2_trans_inconsistent(trans,
1500                         "stripe pointer doesn't match stripe %llu",
1501                         (u64) p.ec.idx);
1502                 ret = -EIO;
1503                 goto err;
1504         }
1505
1506         stripe_blockcount_set(&s->v, p.ec.block,
1507                 stripe_blockcount_get(&s->v, p.ec.block) +
1508                 sectors);
1509
1510         bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1511         r.e.data_type = data_type;
1512         ret = update_replicas_list(trans, &r.e, sectors);
1513 err:
1514         bch2_trans_iter_exit(trans, &iter);
1515         return ret;
1516 }
1517
1518 static int __trans_mark_extent(struct btree_trans *trans,
1519                                enum btree_id btree_id, unsigned level,
1520                                struct bkey_s_c k, unsigned flags)
1521 {
1522         struct bch_fs *c = trans->c;
1523         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1524         const union bch_extent_entry *entry;
1525         struct extent_ptr_decoded p;
1526         struct bch_replicas_padded r;
1527         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
1528                 ? BCH_DATA_btree
1529                 : BCH_DATA_user;
1530         s64 sectors = bkey_is_btree_ptr(k.k)
1531                 ? btree_sectors(c)
1532                 : k.k->size;
1533         s64 dirty_sectors = 0;
1534         bool stale;
1535         int ret = 0;
1536
1537         r.e.data_type   = data_type;
1538         r.e.nr_devs     = 0;
1539         r.e.nr_required = 1;
1540
1541         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1542                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
1543
1544                 if (flags & BTREE_TRIGGER_OVERWRITE)
1545                         disk_sectors = -disk_sectors;
1546
1547                 ret = bch2_trans_mark_pointer(trans, btree_id, level, k, p, flags);
1548                 if (ret < 0)
1549                         return ret;
1550
1551                 stale = ret > 0;
1552
1553                 if (p.ptr.cached) {
1554                         if (!stale) {
1555                                 ret = update_cached_sectors_list(trans, p.ptr.dev,
1556                                                                  disk_sectors);
1557                                 if (ret)
1558                                         return ret;
1559                         }
1560                 } else if (!p.has_ec) {
1561                         dirty_sectors          += disk_sectors;
1562                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1563                 } else {
1564                         ret = bch2_trans_mark_stripe_ptr(trans, p,
1565                                         disk_sectors, data_type);
1566                         if (ret)
1567                                 return ret;
1568
1569                         r.e.nr_required = 0;
1570                 }
1571         }
1572
1573         if (r.e.nr_devs)
1574                 ret = update_replicas_list(trans, &r.e, dirty_sectors);
1575
1576         return ret;
1577 }
1578
1579 int bch2_trans_mark_extent(struct btree_trans *trans,
1580                            enum btree_id btree_id, unsigned level,
1581                            struct bkey_s_c old, struct bkey_i *new,
1582                            unsigned flags)
1583 {
1584         return trigger_run_insert_then_overwrite(__trans_mark_extent, trans, btree_id, level, old, new, flags);
1585 }
1586
1587 static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
1588                                          struct bkey_s_c_stripe s,
1589                                          unsigned idx, bool deleting)
1590 {
1591         struct bch_fs *c = trans->c;
1592         const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
1593         struct btree_iter iter;
1594         struct bkey_i_alloc_v4 *a;
1595         enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
1596                 ? BCH_DATA_parity : 0;
1597         s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
1598         int ret = 0;
1599
1600         if (deleting)
1601                 sectors = -sectors;
1602
1603         a = bch2_trans_start_alloc_update(trans, &iter, PTR_BUCKET_POS(c, ptr));
1604         if (IS_ERR(a))
1605                 return PTR_ERR(a);
1606
1607         ret = check_bucket_ref(trans, s.s_c, ptr, sectors, data_type,
1608                                a->v.gen, a->v.data_type,
1609                                a->v.dirty_sectors, a->v.cached_sectors);
1610         if (ret)
1611                 goto err;
1612
1613         if (!deleting) {
1614                 if (bch2_trans_inconsistent_on(a->v.stripe ||
1615                                                a->v.stripe_redundancy, trans,
1616                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
1617                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1618                                 bch2_data_types[a->v.data_type],
1619                                 a->v.dirty_sectors,
1620                                 a->v.stripe, s.k->p.offset)) {
1621                         ret = -EIO;
1622                         goto err;
1623                 }
1624
1625                 if (bch2_trans_inconsistent_on(data_type && a->v.dirty_sectors, trans,
1626                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
1627                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1628                                 bch2_data_types[a->v.data_type],
1629                                 a->v.dirty_sectors,
1630                                 s.k->p.offset)) {
1631                         ret = -EIO;
1632                         goto err;
1633                 }
1634
1635                 a->v.stripe             = s.k->p.offset;
1636                 a->v.stripe_redundancy  = s.v->nr_redundant;
1637                 a->v.data_type          = BCH_DATA_stripe;
1638         } else {
1639                 if (bch2_trans_inconsistent_on(a->v.stripe != s.k->p.offset ||
1640                                                a->v.stripe_redundancy != s.v->nr_redundant, trans,
1641                                 "bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
1642                                 iter.pos.inode, iter.pos.offset, a->v.gen,
1643                                 s.k->p.offset, a->v.stripe)) {
1644                         ret = -EIO;
1645                         goto err;
1646                 }
1647
1648                 a->v.stripe             = 0;
1649                 a->v.stripe_redundancy  = 0;
1650                 a->v.data_type          = alloc_data_type(a->v, BCH_DATA_user);
1651         }
1652
1653         a->v.dirty_sectors += sectors;
1654         if (data_type)
1655                 a->v.data_type = !deleting ? data_type : 0;
1656
1657         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1658         if (ret)
1659                 goto err;
1660 err:
1661         bch2_trans_iter_exit(trans, &iter);
1662         return ret;
1663 }
1664
1665 int bch2_trans_mark_stripe(struct btree_trans *trans,
1666                            enum btree_id btree_id, unsigned level,
1667                            struct bkey_s_c old, struct bkey_i *new,
1668                            unsigned flags)
1669 {
1670         const struct bch_stripe *old_s = NULL;
1671         struct bch_stripe *new_s = NULL;
1672         struct bch_replicas_padded r;
1673         unsigned i, nr_blocks;
1674         int ret = 0;
1675
1676         if (old.k->type == KEY_TYPE_stripe)
1677                 old_s = bkey_s_c_to_stripe(old).v;
1678         if (new->k.type == KEY_TYPE_stripe)
1679                 new_s = &bkey_i_to_stripe(new)->v;
1680
1681         /*
1682          * If the pointers aren't changing, we don't need to do anything:
1683          */
1684         if (new_s && old_s &&
1685             new_s->nr_blocks    == old_s->nr_blocks &&
1686             new_s->nr_redundant == old_s->nr_redundant &&
1687             !memcmp(old_s->ptrs, new_s->ptrs,
1688                     new_s->nr_blocks * sizeof(struct bch_extent_ptr)))
1689                 return 0;
1690
1691         BUG_ON(new_s && old_s &&
1692                (new_s->nr_blocks        != old_s->nr_blocks ||
1693                 new_s->nr_redundant     != old_s->nr_redundant));
1694
1695         nr_blocks = new_s ? new_s->nr_blocks : old_s->nr_blocks;
1696
1697         if (new_s) {
1698                 s64 sectors = le16_to_cpu(new_s->sectors);
1699
1700                 bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(new));
1701                 ret = update_replicas_list(trans, &r.e, sectors * new_s->nr_redundant);
1702                 if (ret)
1703                         return ret;
1704         }
1705
1706         if (old_s) {
1707                 s64 sectors = -((s64) le16_to_cpu(old_s->sectors));
1708
1709                 bch2_bkey_to_replicas(&r.e, old);
1710                 ret = update_replicas_list(trans, &r.e, sectors * old_s->nr_redundant);
1711                 if (ret)
1712                         return ret;
1713         }
1714
1715         for (i = 0; i < nr_blocks; i++) {
1716                 if (new_s && old_s &&
1717                     !memcmp(&new_s->ptrs[i],
1718                             &old_s->ptrs[i],
1719                             sizeof(new_s->ptrs[i])))
1720                         continue;
1721
1722                 if (new_s) {
1723                         ret = bch2_trans_mark_stripe_bucket(trans,
1724                                         bkey_i_to_s_c_stripe(new), i, false);
1725                         if (ret)
1726                                 break;
1727                 }
1728
1729                 if (old_s) {
1730                         ret = bch2_trans_mark_stripe_bucket(trans,
1731                                         bkey_s_c_to_stripe(old), i, true);
1732                         if (ret)
1733                                 break;
1734                 }
1735         }
1736
1737         return ret;
1738 }
1739
1740 int bch2_trans_mark_inode(struct btree_trans *trans,
1741                           enum btree_id btree_id, unsigned level,
1742                           struct bkey_s_c old,
1743                           struct bkey_i *new,
1744                           unsigned flags)
1745 {
1746         int nr = bkey_is_inode(&new->k) - bkey_is_inode(old.k);
1747
1748         if (nr) {
1749                 int ret = replicas_deltas_realloc(trans, 0);
1750                 struct replicas_delta_list *d = trans->fs_usage_deltas;
1751
1752                 if (ret)
1753                         return ret;
1754
1755                 d->nr_inodes += nr;
1756         }
1757
1758         return 0;
1759 }
1760
1761 static int __trans_mark_reservation(struct btree_trans *trans,
1762                                     enum btree_id btree_id, unsigned level,
1763                                     struct bkey_s_c k, unsigned flags)
1764 {
1765         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1766         s64 sectors = (s64) k.k->size;
1767         struct replicas_delta_list *d;
1768         int ret;
1769
1770         if (flags & BTREE_TRIGGER_OVERWRITE)
1771                 sectors = -sectors;
1772         sectors *= replicas;
1773
1774         ret = replicas_deltas_realloc(trans, 0);
1775         if (ret)
1776                 return ret;
1777
1778         d = trans->fs_usage_deltas;
1779         replicas = clamp_t(unsigned, replicas, 1,
1780                            ARRAY_SIZE(d->persistent_reserved));
1781
1782         d->persistent_reserved[replicas - 1] += sectors;
1783         return 0;
1784 }
1785
1786 int bch2_trans_mark_reservation(struct btree_trans *trans,
1787                                 enum btree_id btree_id, unsigned level,
1788                                 struct bkey_s_c old,
1789                                 struct bkey_i *new,
1790                                 unsigned flags)
1791 {
1792         return trigger_run_insert_then_overwrite(__trans_mark_reservation, trans, btree_id, level, old, new, flags);
1793 }
1794
1795 static int trans_mark_reflink_p_segment(struct btree_trans *trans,
1796                         struct bkey_s_c_reflink_p p,
1797                         u64 *idx, unsigned flags)
1798 {
1799         struct bch_fs *c = trans->c;
1800         struct btree_iter iter;
1801         struct bkey_i *k;
1802         __le64 *refcount;
1803         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1804         struct printbuf buf = PRINTBUF;
1805         int ret;
1806
1807         k = bch2_bkey_get_mut_noupdate(trans, &iter,
1808                         BTREE_ID_reflink, POS(0, *idx),
1809                         BTREE_ITER_WITH_UPDATES);
1810         ret = PTR_ERR_OR_ZERO(k);
1811         if (ret)
1812                 goto err;
1813
1814         refcount = bkey_refcount(k);
1815         if (!refcount) {
1816                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1817                 bch2_trans_inconsistent(trans,
1818                         "nonexistent indirect extent at %llu while marking\n  %s",
1819                         *idx, buf.buf);
1820                 ret = -EIO;
1821                 goto err;
1822         }
1823
1824         if (!*refcount && (flags & BTREE_TRIGGER_OVERWRITE)) {
1825                 bch2_bkey_val_to_text(&buf, c, p.s_c);
1826                 bch2_trans_inconsistent(trans,
1827                         "indirect extent refcount underflow at %llu while marking\n  %s",
1828                         *idx, buf.buf);
1829                 ret = -EIO;
1830                 goto err;
1831         }
1832
1833         if (flags & BTREE_TRIGGER_INSERT) {
1834                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1835                 u64 pad;
1836
1837                 pad = max_t(s64, le32_to_cpu(v->front_pad),
1838                             le64_to_cpu(v->idx) - bkey_start_offset(&k->k));
1839                 BUG_ON(pad > U32_MAX);
1840                 v->front_pad = cpu_to_le32(pad);
1841
1842                 pad = max_t(s64, le32_to_cpu(v->back_pad),
1843                             k->k.p.offset - p.k->size - le64_to_cpu(v->idx));
1844                 BUG_ON(pad > U32_MAX);
1845                 v->back_pad = cpu_to_le32(pad);
1846         }
1847
1848         le64_add_cpu(refcount, add);
1849
1850         bch2_btree_iter_set_pos_to_extent_start(&iter);
1851         ret = bch2_trans_update(trans, &iter, k, 0);
1852         if (ret)
1853                 goto err;
1854
1855         *idx = k->k.p.offset;
1856 err:
1857         bch2_trans_iter_exit(trans, &iter);
1858         printbuf_exit(&buf);
1859         return ret;
1860 }
1861
1862 static int __trans_mark_reflink_p(struct btree_trans *trans,
1863                                 enum btree_id btree_id, unsigned level,
1864                                 struct bkey_s_c k, unsigned flags)
1865 {
1866         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1867         u64 idx, end_idx;
1868         int ret = 0;
1869
1870         idx     = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
1871         end_idx = le64_to_cpu(p.v->idx) + p.k->size +
1872                 le32_to_cpu(p.v->back_pad);
1873
1874         while (idx < end_idx && !ret)
1875                 ret = trans_mark_reflink_p_segment(trans, p, &idx, flags);
1876         return ret;
1877 }
1878
1879 int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1880                               enum btree_id btree_id, unsigned level,
1881                               struct bkey_s_c old,
1882                               struct bkey_i *new,
1883                               unsigned flags)
1884 {
1885         if (flags & BTREE_TRIGGER_INSERT) {
1886                 struct bch_reflink_p *v = &bkey_i_to_reflink_p(new)->v;
1887
1888                 v->front_pad = v->back_pad = 0;
1889         }
1890
1891         return trigger_run_insert_then_overwrite(__trans_mark_reflink_p, trans, btree_id, level, old, new, flags);
1892 }
1893
1894 static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1895                                     struct bch_dev *ca, size_t b,
1896                                     enum bch_data_type type,
1897                                     unsigned sectors)
1898 {
1899         struct bch_fs *c = trans->c;
1900         struct btree_iter iter;
1901         struct bkey_i_alloc_v4 *a;
1902         int ret = 0;
1903
1904         /*
1905          * Backup superblock might be past the end of our normal usable space:
1906          */
1907         if (b >= ca->mi.nbuckets)
1908                 return 0;
1909
1910         a = bch2_trans_start_alloc_update(trans, &iter, POS(ca->dev_idx, b));
1911         if (IS_ERR(a))
1912                 return PTR_ERR(a);
1913
1914         if (a->v.data_type && type && a->v.data_type != type) {
1915                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1916                         "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
1917                         "while marking %s",
1918                         iter.pos.inode, iter.pos.offset, a->v.gen,
1919                         bch2_data_types[a->v.data_type],
1920                         bch2_data_types[type],
1921                         bch2_data_types[type]);
1922                 ret = -EIO;
1923                 goto out;
1924         }
1925
1926         a->v.data_type          = type;
1927         a->v.dirty_sectors      = sectors;
1928
1929         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1930         if (ret)
1931                 goto out;
1932 out:
1933         bch2_trans_iter_exit(trans, &iter);
1934         return ret;
1935 }
1936
1937 int bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1938                                     struct bch_dev *ca, size_t b,
1939                                     enum bch_data_type type,
1940                                     unsigned sectors)
1941 {
1942         return commit_do(trans, NULL, NULL, 0,
1943                         __bch2_trans_mark_metadata_bucket(trans, ca, b, type, sectors));
1944 }
1945
1946 static int bch2_trans_mark_metadata_sectors(struct btree_trans *trans,
1947                                             struct bch_dev *ca,
1948                                             u64 start, u64 end,
1949                                             enum bch_data_type type,
1950                                             u64 *bucket, unsigned *bucket_sectors)
1951 {
1952         do {
1953                 u64 b = sector_to_bucket(ca, start);
1954                 unsigned sectors =
1955                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1956
1957                 if (b != *bucket && *bucket_sectors) {
1958                         int ret = bch2_trans_mark_metadata_bucket(trans, ca, *bucket,
1959                                                                   type, *bucket_sectors);
1960                         if (ret)
1961                                 return ret;
1962
1963                         *bucket_sectors = 0;
1964                 }
1965
1966                 *bucket         = b;
1967                 *bucket_sectors += sectors;
1968                 start += sectors;
1969         } while (start < end);
1970
1971         return 0;
1972 }
1973
1974 static int __bch2_trans_mark_dev_sb(struct btree_trans *trans,
1975                                     struct bch_dev *ca)
1976 {
1977         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1978         u64 bucket = 0;
1979         unsigned i, bucket_sectors = 0;
1980         int ret;
1981
1982         for (i = 0; i < layout->nr_superblocks; i++) {
1983                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1984
1985                 if (offset == BCH_SB_SECTOR) {
1986                         ret = bch2_trans_mark_metadata_sectors(trans, ca,
1987                                                 0, BCH_SB_SECTOR,
1988                                                 BCH_DATA_sb, &bucket, &bucket_sectors);
1989                         if (ret)
1990                                 return ret;
1991                 }
1992
1993                 ret = bch2_trans_mark_metadata_sectors(trans, ca, offset,
1994                                       offset + (1 << layout->sb_max_size_bits),
1995                                       BCH_DATA_sb, &bucket, &bucket_sectors);
1996                 if (ret)
1997                         return ret;
1998         }
1999
2000         if (bucket_sectors) {
2001                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
2002                                 bucket, BCH_DATA_sb, bucket_sectors);
2003                 if (ret)
2004                         return ret;
2005         }
2006
2007         for (i = 0; i < ca->journal.nr; i++) {
2008                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
2009                                 ca->journal.buckets[i],
2010                                 BCH_DATA_journal, ca->mi.bucket_size);
2011                 if (ret)
2012                         return ret;
2013         }
2014
2015         return 0;
2016 }
2017
2018 int bch2_trans_mark_dev_sb(struct bch_fs *c, struct bch_dev *ca)
2019 {
2020         int ret = bch2_trans_run(c, __bch2_trans_mark_dev_sb(&trans, ca));
2021         if (ret)
2022                 bch_err_fn(c, ret);
2023         return ret;
2024 }
2025
2026 /* Disk reservations: */
2027
2028 #define SECTORS_CACHE   1024
2029
2030 int __bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
2031                               u64 sectors, int flags)
2032 {
2033         struct bch_fs_pcpu *pcpu;
2034         u64 old, v, get;
2035         s64 sectors_available;
2036         int ret;
2037
2038         percpu_down_read(&c->mark_lock);
2039         preempt_disable();
2040         pcpu = this_cpu_ptr(c->pcpu);
2041
2042         if (sectors <= pcpu->sectors_available)
2043                 goto out;
2044
2045         v = atomic64_read(&c->sectors_available);
2046         do {
2047                 old = v;
2048                 get = min((u64) sectors + SECTORS_CACHE, old);
2049
2050                 if (get < sectors) {
2051                         preempt_enable();
2052                         goto recalculate;
2053                 }
2054         } while ((v = atomic64_cmpxchg(&c->sectors_available,
2055                                        old, old - get)) != old);
2056
2057         pcpu->sectors_available         += get;
2058
2059 out:
2060         pcpu->sectors_available         -= sectors;
2061         this_cpu_add(*c->online_reserved, sectors);
2062         res->sectors                    += sectors;
2063
2064         preempt_enable();
2065         percpu_up_read(&c->mark_lock);
2066         return 0;
2067
2068 recalculate:
2069         mutex_lock(&c->sectors_available_lock);
2070
2071         percpu_u64_set(&c->pcpu->sectors_available, 0);
2072         sectors_available = avail_factor(__bch2_fs_usage_read_short(c).free);
2073
2074         if (sectors <= sectors_available ||
2075             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
2076                 atomic64_set(&c->sectors_available,
2077                              max_t(s64, 0, sectors_available - sectors));
2078                 this_cpu_add(*c->online_reserved, sectors);
2079                 res->sectors                    += sectors;
2080                 ret = 0;
2081         } else {
2082                 atomic64_set(&c->sectors_available, sectors_available);
2083                 ret = -BCH_ERR_ENOSPC_disk_reservation;
2084         }
2085
2086         mutex_unlock(&c->sectors_available_lock);
2087         percpu_up_read(&c->mark_lock);
2088
2089         return ret;
2090 }
2091
2092 /* Startup/shutdown: */
2093
2094 static void bucket_gens_free_rcu(struct rcu_head *rcu)
2095 {
2096         struct bucket_gens *buckets =
2097                 container_of(rcu, struct bucket_gens, rcu);
2098
2099         kvpfree(buckets, sizeof(*buckets) + buckets->nbuckets);
2100 }
2101
2102 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
2103 {
2104         struct bucket_gens *bucket_gens = NULL, *old_bucket_gens = NULL;
2105         unsigned long *buckets_nouse = NULL;
2106         bool resize = ca->bucket_gens != NULL;
2107         int ret;
2108
2109         if (!(bucket_gens       = kvpmalloc(sizeof(struct bucket_gens) + nbuckets,
2110                                             GFP_KERNEL|__GFP_ZERO))) {
2111                 ret = -BCH_ERR_ENOMEM_bucket_gens;
2112                 goto err;
2113         }
2114
2115         if ((c->opts.buckets_nouse &&
2116              !(buckets_nouse    = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2117                                             sizeof(unsigned long),
2118                                             GFP_KERNEL|__GFP_ZERO)))) {
2119                 ret = -BCH_ERR_ENOMEM_buckets_nouse;
2120                 goto err;
2121         }
2122
2123         bucket_gens->first_bucket = ca->mi.first_bucket;
2124         bucket_gens->nbuckets   = nbuckets;
2125
2126         bch2_copygc_stop(c);
2127
2128         if (resize) {
2129                 down_write(&c->gc_lock);
2130                 down_write(&ca->bucket_lock);
2131                 percpu_down_write(&c->mark_lock);
2132         }
2133
2134         old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
2135
2136         if (resize) {
2137                 size_t n = min(bucket_gens->nbuckets, old_bucket_gens->nbuckets);
2138
2139                 memcpy(bucket_gens->b,
2140                        old_bucket_gens->b,
2141                        n);
2142                 if (buckets_nouse)
2143                         memcpy(buckets_nouse,
2144                                ca->buckets_nouse,
2145                                BITS_TO_LONGS(n) * sizeof(unsigned long));
2146         }
2147
2148         rcu_assign_pointer(ca->bucket_gens, bucket_gens);
2149         bucket_gens     = old_bucket_gens;
2150
2151         swap(ca->buckets_nouse, buckets_nouse);
2152
2153         nbuckets = ca->mi.nbuckets;
2154
2155         if (resize) {
2156                 percpu_up_write(&c->mark_lock);
2157                 up_write(&ca->bucket_lock);
2158                 up_write(&c->gc_lock);
2159         }
2160
2161         ret = 0;
2162 err:
2163         kvpfree(buckets_nouse,
2164                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2165         if (bucket_gens)
2166                 call_rcu(&bucket_gens->rcu, bucket_gens_free_rcu);
2167
2168         return ret;
2169 }
2170
2171 void bch2_dev_buckets_free(struct bch_dev *ca)
2172 {
2173         unsigned i;
2174
2175         kvpfree(ca->buckets_nouse,
2176                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2177         kvpfree(rcu_dereference_protected(ca->bucket_gens, 1),
2178                 sizeof(struct bucket_gens) + ca->mi.nbuckets);
2179
2180         for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
2181                 free_percpu(ca->usage[i]);
2182         kfree(ca->usage_base);
2183 }
2184
2185 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2186 {
2187         unsigned i;
2188
2189         ca->usage_base = kzalloc(sizeof(struct bch_dev_usage), GFP_KERNEL);
2190         if (!ca->usage_base)
2191                 return -BCH_ERR_ENOMEM_usage_init;
2192
2193         for (i = 0; i < ARRAY_SIZE(ca->usage); i++) {
2194                 ca->usage[i] = alloc_percpu(struct bch_dev_usage);
2195                 if (!ca->usage[i])
2196                         return -BCH_ERR_ENOMEM_usage_init;
2197         }
2198
2199         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);
2200 }