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

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Code for manipulating bucket marks for garbage collection.
   4  *
   5  * Copyright 2014 Datera, Inc.
   6  */
   7
   8 #include "bcachefs.h"
   9 #include "alloc_background.h"
  10 #include "bset.h"
  11 #include "btree_gc.h"
  12 #include "btree_update.h"
  13 #include "buckets.h"
  14 #include "ec.h"
  15 #include "error.h"
  16 #include "movinggc.h"
  17 #include "reflink.h"
  18 #include "replicas.h"
  19 #include "subvolume.h"
  20
  21 #include <linux/preempt.h>
  22 #include <trace/events/bcachefs.h>
  23
  24 static inline void fs_usage_data_type_to_base(struct bch_fs_usage *fs_usage,
  25                                               enum bch_data_type data_type,
  26                                               s64 sectors)
  27 {
  28         switch (data_type) {
  29         case BCH_DATA_btree:
  30                 fs_usage->btree         += sectors;
  31                 break;
  32         case BCH_DATA_user:
  33         case BCH_DATA_parity:
  34                 fs_usage->data          += sectors;
  35                 break;
  36         case BCH_DATA_cached:
  37                 fs_usage->cached        += sectors;
  38                 break;
  39         default:
  40                 break;
  41         }
  42 }
  43
  44 /*
  45  * Clear journal_seq_valid for buckets for which it's not needed, to prevent
  46  * wraparound:
  47  */
  48 void bch2_bucket_seq_cleanup(struct bch_fs *c)
  49 {
  50         u64 journal_seq = atomic64_read(&c->journal.seq);
  51         u16 last_seq_ondisk = c->journal.last_seq_ondisk;
  52         struct bch_dev *ca;
  53         struct bucket_array *buckets;
  54         struct bucket *g;
  55         struct bucket_mark m;
  56         unsigned i;
  57
  58         if (journal_seq - c->last_bucket_seq_cleanup <
  59             (1U << (BUCKET_JOURNAL_SEQ_BITS - 2)))
  60                 return;
  61
  62         c->last_bucket_seq_cleanup = journal_seq;
  63
  64         for_each_member_device(ca, c, i) {
  65                 down_read(&ca->bucket_lock);
  66                 buckets = bucket_array(ca);
  67
  68                 for_each_bucket(g, buckets) {
  69                         bucket_cmpxchg(g, m, ({
  70                                 if (!m.journal_seq_valid ||
  71                                     bucket_needs_journal_commit(m, last_seq_ondisk))
  72                                         break;
  73
  74                                 m.journal_seq_valid = 0;
  75                         }));
  76                 }
  77                 up_read(&ca->bucket_lock);
  78         }
  79 }
  80
  81 void bch2_fs_usage_initialize(struct bch_fs *c)
  82 {
  83         struct bch_fs_usage *usage;
  84         struct bch_dev *ca;
  85         unsigned i;
  86
  87         percpu_down_write(&c->mark_lock);
  88         usage = c->usage_base;
  89
  90         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
  91                 bch2_fs_usage_acc_to_base(c, i);
  92
  93         for (i = 0; i < BCH_REPLICAS_MAX; i++)
  94                 usage->reserved += usage->persistent_reserved[i];
  95
  96         for (i = 0; i < c->replicas.nr; i++) {
  97                 struct bch_replicas_entry *e =
  98                         cpu_replicas_entry(&c->replicas, i);
  99
 100                 fs_usage_data_type_to_base(usage, e->data_type, usage->replicas[i]);
 101         }
 102
 103         for_each_member_device(ca, c, i) {
 104                 struct bch_dev_usage dev = bch2_dev_usage_read(ca);
 105
 106                 usage->hidden += (dev.d[BCH_DATA_sb].buckets +
 107                                   dev.d[BCH_DATA_journal].buckets) *
 108                         ca->mi.bucket_size;
 109         }
 110
 111         percpu_up_write(&c->mark_lock);
 112 }
 113
 114 static inline struct bch_dev_usage *dev_usage_ptr(struct bch_dev *ca,
 115                                                   unsigned journal_seq,
 116                                                   bool gc)
 117 {
 118         return this_cpu_ptr(gc
 119                             ? ca->usage_gc
 120                             : ca->usage[journal_seq & JOURNAL_BUF_MASK]);
 121 }
 122
 123 struct bch_dev_usage bch2_dev_usage_read(struct bch_dev *ca)
 124 {
 125         struct bch_fs *c = ca->fs;
 126         struct bch_dev_usage ret;
 127         unsigned seq, i, u64s = dev_usage_u64s();
 128
 129         do {
 130                 seq = read_seqcount_begin(&c->usage_lock);
 131                 memcpy(&ret, ca->usage_base, u64s * sizeof(u64));
 132                 for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
 133                         acc_u64s_percpu((u64 *) &ret, (u64 __percpu *) ca->usage[i], u64s);
 134         } while (read_seqcount_retry(&c->usage_lock, seq));
 135
 136         return ret;
 137 }
 138
 139 static inline struct bch_fs_usage *fs_usage_ptr(struct bch_fs *c,
 140                                                 unsigned journal_seq,
 141                                                 bool gc)
 142 {
 143         return this_cpu_ptr(gc
 144                             ? c->usage_gc
 145                             : c->usage[journal_seq & JOURNAL_BUF_MASK]);
 146 }
 147
 148 u64 bch2_fs_usage_read_one(struct bch_fs *c, u64 *v)
 149 {
 150         ssize_t offset = v - (u64 *) c->usage_base;
 151         unsigned i, seq;
 152         u64 ret;
 153
 154         BUG_ON(offset < 0 || offset >= fs_usage_u64s(c));
 155         percpu_rwsem_assert_held(&c->mark_lock);
 156
 157         do {
 158                 seq = read_seqcount_begin(&c->usage_lock);
 159                 ret = *v;
 160
 161                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 162                         ret += percpu_u64_get((u64 __percpu *) c->usage[i] + offset);
 163         } while (read_seqcount_retry(&c->usage_lock, seq));
 164
 165         return ret;
 166 }
 167
 168 struct bch_fs_usage_online *bch2_fs_usage_read(struct bch_fs *c)
 169 {
 170         struct bch_fs_usage_online *ret;
 171         unsigned seq, i, u64s;
 172
 173         percpu_down_read(&c->mark_lock);
 174
 175         ret = kmalloc(sizeof(struct bch_fs_usage_online) +
 176                       sizeof(u64) * c->replicas.nr, GFP_NOFS);
 177         if (unlikely(!ret)) {
 178                 percpu_up_read(&c->mark_lock);
 179                 return NULL;
 180         }
 181
 182         ret->online_reserved = percpu_u64_get(c->online_reserved);
 183
 184         u64s = fs_usage_u64s(c);
 185         do {
 186                 seq = read_seqcount_begin(&c->usage_lock);
 187                 memcpy(&ret->u, c->usage_base, u64s * sizeof(u64));
 188                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 189                         acc_u64s_percpu((u64 *) &ret->u, (u64 __percpu *) c->usage[i], u64s);
 190         } while (read_seqcount_retry(&c->usage_lock, seq));
 191
 192         return ret;
 193 }
 194
 195 void bch2_fs_usage_acc_to_base(struct bch_fs *c, unsigned idx)
 196 {
 197         struct bch_dev *ca;
 198         unsigned i, u64s = fs_usage_u64s(c);
 199
 200         BUG_ON(idx >= ARRAY_SIZE(c->usage));
 201
 202         preempt_disable();
 203         write_seqcount_begin(&c->usage_lock);
 204
 205         acc_u64s_percpu((u64 *) c->usage_base,
 206                         (u64 __percpu *) c->usage[idx], u64s);
 207         percpu_memset(c->usage[idx], 0, u64s * sizeof(u64));
 208
 209         rcu_read_lock();
 210         for_each_member_device_rcu(ca, c, i, NULL) {
 211                 u64s = dev_usage_u64s();
 212
 213                 acc_u64s_percpu((u64 *) ca->usage_base,
 214                                 (u64 __percpu *) ca->usage[idx], u64s);
 215                 percpu_memset(ca->usage[idx], 0, u64s * sizeof(u64));
 216         }
 217         rcu_read_unlock();
 218
 219         write_seqcount_end(&c->usage_lock);
 220         preempt_enable();
 221 }
 222
 223 void bch2_fs_usage_to_text(struct printbuf *out,
 224                            struct bch_fs *c,
 225                            struct bch_fs_usage_online *fs_usage)
 226 {
 227         unsigned i;
 228
 229         pr_buf(out, "capacity:\t\t\t%llu\n", c->capacity);
 230
 231         pr_buf(out, "hidden:\t\t\t\t%llu\n",
 232                fs_usage->u.hidden);
 233         pr_buf(out, "data:\t\t\t\t%llu\n",
 234                fs_usage->u.data);
 235         pr_buf(out, "cached:\t\t\t\t%llu\n",
 236                fs_usage->u.cached);
 237         pr_buf(out, "reserved:\t\t\t%llu\n",
 238                fs_usage->u.reserved);
 239         pr_buf(out, "nr_inodes:\t\t\t%llu\n",
 240                fs_usage->u.nr_inodes);
 241         pr_buf(out, "online reserved:\t\t%llu\n",
 242                fs_usage->online_reserved);
 243
 244         for (i = 0;
 245              i < ARRAY_SIZE(fs_usage->u.persistent_reserved);
 246              i++) {
 247                 pr_buf(out, "%u replicas:\n", i + 1);
 248                 pr_buf(out, "\treserved:\t\t%llu\n",
 249                        fs_usage->u.persistent_reserved[i]);
 250         }
 251
 252         for (i = 0; i < c->replicas.nr; i++) {
 253                 struct bch_replicas_entry *e =
 254                         cpu_replicas_entry(&c->replicas, i);
 255
 256                 pr_buf(out, "\t");
 257                 bch2_replicas_entry_to_text(out, e);
 258                 pr_buf(out, ":\t%llu\n", fs_usage->u.replicas[i]);
 259         }
 260 }
 261
 262 static u64 reserve_factor(u64 r)
 263 {
 264         return r + (round_up(r, (1 << RESERVE_FACTOR)) >> RESERVE_FACTOR);
 265 }
 266
 267 u64 bch2_fs_sectors_used(struct bch_fs *c, struct bch_fs_usage_online *fs_usage)
 268 {
 269         return min(fs_usage->u.hidden +
 270                    fs_usage->u.btree +
 271                    fs_usage->u.data +
 272                    reserve_factor(fs_usage->u.reserved +
 273                                   fs_usage->online_reserved),
 274                    c->capacity);
 275 }
 276
 277 static struct bch_fs_usage_short
 278 __bch2_fs_usage_read_short(struct bch_fs *c)
 279 {
 280         struct bch_fs_usage_short ret;
 281         u64 data, reserved;
 282
 283         ret.capacity = c->capacity -
 284                 bch2_fs_usage_read_one(c, &c->usage_base->hidden);
 285
 286         data            = bch2_fs_usage_read_one(c, &c->usage_base->data) +
 287                 bch2_fs_usage_read_one(c, &c->usage_base->btree);
 288         reserved        = bch2_fs_usage_read_one(c, &c->usage_base->reserved) +
 289                 percpu_u64_get(c->online_reserved);
 290
 291         ret.used        = min(ret.capacity, data + reserve_factor(reserved));
 292         ret.free        = ret.capacity - ret.used;
 293
 294         ret.nr_inodes   = bch2_fs_usage_read_one(c, &c->usage_base->nr_inodes);
 295
 296         return ret;
 297 }
 298
 299 struct bch_fs_usage_short
 300 bch2_fs_usage_read_short(struct bch_fs *c)
 301 {
 302         struct bch_fs_usage_short ret;
 303
 304         percpu_down_read(&c->mark_lock);
 305         ret = __bch2_fs_usage_read_short(c);
 306         percpu_up_read(&c->mark_lock);
 307
 308         return ret;
 309 }
 310
 311 static inline int is_unavailable_bucket(struct bucket_mark m)
 312 {
 313         return !is_available_bucket(m);
 314 }
 315
 316 static inline int bucket_sectors_fragmented(struct bch_dev *ca,
 317                                             struct bucket_mark m)
 318 {
 319         return bucket_sectors_used(m)
 320                 ? max(0, (int) ca->mi.bucket_size - (int) bucket_sectors_used(m))
 321                 : 0;
 322 }
 323
 324 static inline int is_stripe_data_bucket(struct bucket_mark m)
 325 {
 326         return m.stripe && m.data_type != BCH_DATA_parity;
 327 }
 328
 329 static inline enum bch_data_type bucket_type(struct bucket_mark m)
 330 {
 331         return m.cached_sectors && !m.dirty_sectors
 332                 ? BCH_DATA_cached
 333                 : m.data_type;
 334 }
 335
 336 static bool bucket_became_unavailable(struct bucket_mark old,
 337                                       struct bucket_mark new)
 338 {
 339         return is_available_bucket(old) &&
 340                !is_available_bucket(new);
 341 }
 342
 343 static inline void account_bucket(struct bch_fs_usage *fs_usage,
 344                                   struct bch_dev_usage *dev_usage,
 345                                   enum bch_data_type type,
 346                                   int nr, s64 size)
 347 {
 348         if (type == BCH_DATA_sb || type == BCH_DATA_journal)
 349                 fs_usage->hidden        += size;
 350
 351         dev_usage->d[type].buckets      += nr;
 352 }
 353
 354 static void bch2_dev_usage_update(struct bch_fs *c, struct bch_dev *ca,
 355                                   struct bucket_mark old, struct bucket_mark new,
 356                                   u64 journal_seq, bool gc)
 357 {
 358         struct bch_fs_usage *fs_usage;
 359         struct bch_dev_usage *u;
 360
 361         percpu_rwsem_assert_held(&c->mark_lock);
 362
 363         preempt_disable();
 364         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 365         u = dev_usage_ptr(ca, journal_seq, gc);
 366
 367         if (bucket_type(old))
 368                 account_bucket(fs_usage, u, bucket_type(old),
 369                                -1, -ca->mi.bucket_size);
 370
 371         if (bucket_type(new))
 372                 account_bucket(fs_usage, u, bucket_type(new),
 373                                1, ca->mi.bucket_size);
 374
 375         u->buckets_ec += (int) new.stripe - (int) old.stripe;
 376         u->buckets_unavailable +=
 377                 is_unavailable_bucket(new) - is_unavailable_bucket(old);
 378
 379         u->d[old.data_type].sectors -= old.dirty_sectors;
 380         u->d[new.data_type].sectors += new.dirty_sectors;
 381         u->d[BCH_DATA_cached].sectors +=
 382                 (int) new.cached_sectors - (int) old.cached_sectors;
 383
 384         u->d[old.data_type].fragmented -= bucket_sectors_fragmented(ca, old);
 385         u->d[new.data_type].fragmented += bucket_sectors_fragmented(ca, new);
 386
 387         preempt_enable();
 388
 389         if (!is_available_bucket(old) && is_available_bucket(new))
 390                 bch2_wake_allocator(ca);
 391 }
 392
 393 static inline int __update_replicas(struct bch_fs *c,
 394                                     struct bch_fs_usage *fs_usage,
 395                                     struct bch_replicas_entry *r,
 396                                     s64 sectors)
 397 {
 398         int idx = bch2_replicas_entry_idx(c, r);
 399
 400         if (idx < 0)
 401                 return -1;
 402
 403         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 404         fs_usage->replicas[idx]         += sectors;
 405         return 0;
 406 }
 407
 408 static inline int update_replicas(struct bch_fs *c,
 409                         struct bch_replicas_entry *r, s64 sectors,
 410                         unsigned journal_seq, bool gc)
 411 {
 412         struct bch_fs_usage __percpu *fs_usage;
 413         int idx = bch2_replicas_entry_idx(c, r);
 414
 415         if (idx < 0)
 416                 return -1;
 417
 418         preempt_disable();
 419         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 420         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 421         fs_usage->replicas[idx]         += sectors;
 422         preempt_enable();
 423         return 0;
 424 }
 425
 426 static inline int update_cached_sectors(struct bch_fs *c,
 427                         unsigned dev, s64 sectors,
 428                         unsigned journal_seq, bool gc)
 429 {
 430         struct bch_replicas_padded r;
 431
 432         bch2_replicas_entry_cached(&r.e, dev);
 433
 434         return update_replicas(c, &r.e, sectors, journal_seq, gc);
 435 }
 436
 437 static struct replicas_delta_list *
 438 replicas_deltas_realloc(struct btree_trans *trans, unsigned more)
 439 {
 440         struct replicas_delta_list *d = trans->fs_usage_deltas;
 441         unsigned new_size = d ? (d->size + more) * 2 : 128;
 442         unsigned alloc_size = sizeof(*d) + new_size;
 443
 444         WARN_ON_ONCE(alloc_size > REPLICAS_DELTA_LIST_MAX);
 445
 446         if (!d || d->used + more > d->size) {
 447                 d = krealloc(d, alloc_size, GFP_NOIO|__GFP_ZERO);
 448
 449                 BUG_ON(!d && alloc_size > REPLICAS_DELTA_LIST_MAX);
 450
 451                 if (!d) {
 452                         d = mempool_alloc(&trans->c->replicas_delta_pool, GFP_NOIO);
 453                         memset(d, 0, REPLICAS_DELTA_LIST_MAX);
 454
 455                         if (trans->fs_usage_deltas)
 456                                 memcpy(d, trans->fs_usage_deltas,
 457                                        trans->fs_usage_deltas->size + sizeof(*d));
 458
 459                         new_size = REPLICAS_DELTA_LIST_MAX - sizeof(*d);
 460                         kfree(trans->fs_usage_deltas);
 461                 }
 462
 463                 d->size = new_size;
 464                 trans->fs_usage_deltas = d;
 465         }
 466         return d;
 467 }
 468
 469 static inline void 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
 477         if (!sectors)
 478                 return;
 479
 480         b = replicas_entry_bytes(r) + 8;
 481         d = replicas_deltas_realloc(trans, b);
 482
 483         n = (void *) d->d + d->used;
 484         n->delta = sectors;
 485         memcpy(&n->r, r, replicas_entry_bytes(r));
 486         bch2_replicas_entry_sort(&n->r);
 487         d->used += b;
 488 }
 489
 490 static inline void update_cached_sectors_list(struct btree_trans *trans,
 491                                               unsigned dev, s64 sectors)
 492 {
 493         struct bch_replicas_padded r;
 494
 495         bch2_replicas_entry_cached(&r.e, dev);
 496
 497         update_replicas_list(trans, &r.e, sectors);
 498 }
 499
 500 #define do_mark_fn(fn, c, pos, flags, ...)                              \
 501 ({                                                                      \
 502         int gc, ret = 0;                                                \
 503                                                                         \
 504         percpu_rwsem_assert_held(&c->mark_lock);                        \
 505                                                                         \
 506         for (gc = 0; gc < 2 && !ret; gc++)                              \
 507                 if (!gc == !(flags & BTREE_TRIGGER_GC) ||               \
 508                     (gc && gc_visited(c, pos)))                         \
 509                         ret = fn(c, __VA_ARGS__, gc);                   \
 510         ret;                                                            \
 511 })
 512
 513 void bch2_mark_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 514                             size_t b, bool owned_by_allocator)
 515 {
 516         struct bucket *g = bucket(ca, b);
 517         struct bucket_mark old, new;
 518
 519         old = bucket_cmpxchg(g, new, ({
 520                 new.owned_by_allocator  = owned_by_allocator;
 521         }));
 522
 523         BUG_ON(owned_by_allocator == old.owned_by_allocator);
 524 }
 525
 526 static int bch2_mark_alloc(struct bch_fs *c,
 527                            struct bkey_s_c old, struct bkey_s_c new,
 528                            u64 journal_seq, unsigned flags)
 529 {
 530         bool gc = flags & BTREE_TRIGGER_GC;
 531         struct bkey_alloc_unpacked u;
 532         struct bch_dev *ca;
 533         struct bucket *g;
 534         struct bucket_mark old_m, m;
 535
 536         /* We don't do anything for deletions - do we?: */
 537         if (new.k->type != KEY_TYPE_alloc &&
 538             new.k->type != KEY_TYPE_alloc_v2)
 539                 return 0;
 540
 541         /*
 542          * alloc btree is read in by bch2_alloc_read, not gc:
 543          */
 544         if ((flags & BTREE_TRIGGER_GC) &&
 545             !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
 546                 return 0;
 547
 548         ca = bch_dev_bkey_exists(c, new.k->p.inode);
 549
 550         if (new.k->p.offset >= ca->mi.nbuckets)
 551                 return 0;
 552
 553         g = __bucket(ca, new.k->p.offset, gc);
 554         u = bch2_alloc_unpack(new);
 555
 556         old_m = bucket_cmpxchg(g, m, ({
 557                 m.gen                   = u.gen;
 558                 m.data_type             = u.data_type;
 559                 m.dirty_sectors         = u.dirty_sectors;
 560                 m.cached_sectors        = u.cached_sectors;
 561                 m.stripe                = u.stripe != 0;
 562
 563                 if (journal_seq) {
 564                         m.journal_seq_valid     = 1;
 565                         m.journal_seq           = journal_seq;
 566                 }
 567         }));
 568
 569         bch2_dev_usage_update(c, ca, old_m, m, journal_seq, gc);
 570
 571         g->io_time[READ]        = u.read_time;
 572         g->io_time[WRITE]       = u.write_time;
 573         g->oldest_gen           = u.oldest_gen;
 574         g->gen_valid            = 1;
 575         g->stripe               = u.stripe;
 576         g->stripe_redundancy    = u.stripe_redundancy;
 577
 578         /*
 579          * need to know if we're getting called from the invalidate path or
 580          * not:
 581          */
 582
 583         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 584             old_m.cached_sectors) {
 585                 if (update_cached_sectors(c, ca->dev_idx, -old_m.cached_sectors,
 586                                           journal_seq, gc)) {
 587                         bch2_fs_fatal_error(c, "bch2_mark_alloc(): no replicas entry while updating cached sectors");
 588                         return -1;
 589                 }
 590
 591                 trace_invalidate(ca, bucket_to_sector(ca, new.k->p.offset),
 592                                  old_m.cached_sectors);
 593         }
 594
 595         return 0;
 596 }
 597
 598 #define checked_add(a, b)                                       \
 599 ({                                                              \
 600         unsigned _res = (unsigned) (a) + (b);                   \
 601         bool overflow = _res > U16_MAX;                         \
 602         if (overflow)                                           \
 603                 _res = U16_MAX;                                 \
 604         (a) = _res;                                             \
 605         overflow;                                               \
 606 })
 607
 608 static int __bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 609                                        size_t b, enum bch_data_type data_type,
 610                                        unsigned sectors, bool gc)
 611 {
 612         struct bucket *g = __bucket(ca, b, gc);
 613         struct bucket_mark old, new;
 614         bool overflow;
 615
 616         BUG_ON(data_type != BCH_DATA_sb &&
 617                data_type != BCH_DATA_journal);
 618
 619         old = bucket_cmpxchg(g, new, ({
 620                 new.data_type   = data_type;
 621                 overflow = checked_add(new.dirty_sectors, sectors);
 622         }));
 623
 624         bch2_fs_inconsistent_on(old.data_type &&
 625                                 old.data_type != data_type, c,
 626                 "different types of data in same bucket: %s, %s",
 627                 bch2_data_types[old.data_type],
 628                 bch2_data_types[data_type]);
 629
 630         bch2_fs_inconsistent_on(overflow, c,
 631                 "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > U16_MAX",
 632                 ca->dev_idx, b, new.gen,
 633                 bch2_data_types[old.data_type ?: data_type],
 634                 old.dirty_sectors, sectors);
 635
 636         if (c)
 637                 bch2_dev_usage_update(c, ca, old, new, 0, gc);
 638
 639         return 0;
 640 }
 641
 642 void bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 643                                size_t b, enum bch_data_type type,
 644                                unsigned sectors, struct gc_pos pos,
 645                                unsigned flags)
 646 {
 647         BUG_ON(type != BCH_DATA_sb &&
 648                type != BCH_DATA_journal);
 649
 650         /*
 651          * Backup superblock might be past the end of our normal usable space:
 652          */
 653         if (b >= ca->mi.nbuckets)
 654                 return;
 655
 656         if (likely(c)) {
 657                 do_mark_fn(__bch2_mark_metadata_bucket, c, pos, flags,
 658                            ca, b, type, sectors);
 659         } else {
 660                 __bch2_mark_metadata_bucket(c, ca, b, type, sectors, 0);
 661         }
 662 }
 663
 664 static s64 ptr_disk_sectors(s64 sectors, struct extent_ptr_decoded p)
 665 {
 666         EBUG_ON(sectors < 0);
 667
 668         return p.crc.compression_type &&
 669                 p.crc.compression_type != BCH_COMPRESSION_TYPE_incompressible
 670                 ? DIV_ROUND_UP_ULL(sectors * p.crc.compressed_size,
 671                                p.crc.uncompressed_size)
 672                 : sectors;
 673 }
 674
 675 static int check_bucket_ref(struct bch_fs *c, struct bkey_s_c k,
 676                             const struct bch_extent_ptr *ptr,
 677                             s64 sectors, enum bch_data_type ptr_data_type,
 678                             u8 bucket_gen, u8 bucket_data_type,
 679                             u16 dirty_sectors, u16 cached_sectors)
 680 {
 681         size_t bucket_nr = PTR_BUCKET_NR(bch_dev_bkey_exists(c, ptr->dev), ptr);
 682         u16 bucket_sectors = !ptr->cached
 683                 ? dirty_sectors
 684                 : cached_sectors;
 685         char buf[200];
 686
 687         if (gen_after(ptr->gen, bucket_gen)) {
 688                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 689                         "bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\n"
 690                         "while marking %s",
 691                         ptr->dev, bucket_nr, bucket_gen,
 692                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 693                         ptr->gen,
 694                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 695                 return -EIO;
 696         }
 697
 698         if (gen_cmp(bucket_gen, ptr->gen) > BUCKET_GC_GEN_MAX) {
 699                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 700                         "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 701                         "while marking %s",
 702                         ptr->dev, bucket_nr, bucket_gen,
 703                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 704                         ptr->gen,
 705                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 706                 return -EIO;
 707         }
 708
 709         if (bucket_gen != ptr->gen && !ptr->cached) {
 710                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 711                         "bucket %u:%zu gen %u data type %s: stale dirty ptr (gen %u)\n"
 712                         "while marking %s",
 713                         ptr->dev, bucket_nr, bucket_gen,
 714                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 715                         ptr->gen,
 716                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 717                 return -EIO;
 718         }
 719
 720         if (bucket_gen != ptr->gen)
 721                 return 1;
 722
 723         if (bucket_data_type && ptr_data_type &&
 724             bucket_data_type != ptr_data_type) {
 725                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 726                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 727                         "while marking %s",
 728                         ptr->dev, bucket_nr, bucket_gen,
 729                         bch2_data_types[bucket_data_type],
 730                         bch2_data_types[ptr_data_type],
 731                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 732                 return -EIO;
 733         }
 734
 735         if ((unsigned) (bucket_sectors + sectors) > U16_MAX) {
 736                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 737                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U16_MAX\n"
 738                         "while marking %s",
 739                         ptr->dev, bucket_nr, bucket_gen,
 740                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 741                         bucket_sectors, sectors,
 742                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 743                 return -EIO;
 744         }
 745
 746         return 0;
 747 }
 748
 749 static int mark_stripe_bucket(struct bch_fs *c, struct bkey_s_c k,
 750                              unsigned ptr_idx,
 751                              u64 journal_seq, unsigned flags)
 752 {
 753         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 754         unsigned nr_data = s->nr_blocks - s->nr_redundant;
 755         bool parity = ptr_idx >= nr_data;
 756         const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
 757         bool gc = flags & BTREE_TRIGGER_GC;
 758         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 759         struct bucket *g = PTR_BUCKET(ca, ptr, gc);
 760         struct bucket_mark new, old;
 761         char buf[200];
 762         int ret;
 763
 764         if (g->stripe && g->stripe != k.k->p.offset) {
 765                 bch2_fs_inconsistent(c,
 766                               "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
 767                               ptr->dev, PTR_BUCKET_NR(ca, ptr), g->mark.gen,
 768                               (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 769                 return -EINVAL;
 770         }
 771
 772         old = bucket_cmpxchg(g, new, ({
 773                 ret = check_bucket_ref(c, k, ptr, 0, 0, new.gen, new.data_type,
 774                                        new.dirty_sectors, new.cached_sectors);
 775                 if (ret)
 776                         return ret;
 777
 778                 if (parity) {
 779                         new.data_type           = BCH_DATA_parity;
 780                         new.dirty_sectors       = le16_to_cpu(s->sectors);
 781                 }
 782
 783                 if (journal_seq) {
 784                         new.journal_seq_valid   = 1;
 785                         new.journal_seq         = journal_seq;
 786                 }
 787         }));
 788
 789         g->stripe               = k.k->p.offset;
 790         g->stripe_redundancy    = s->nr_redundant;
 791
 792         bch2_dev_usage_update(c, ca, old, new, journal_seq, gc);
 793         return 0;
 794 }
 795
 796 static int __mark_pointer(struct bch_fs *c, struct bkey_s_c k,
 797                           const struct bch_extent_ptr *ptr,
 798                           s64 sectors, enum bch_data_type ptr_data_type,
 799                           u8 bucket_gen, u8 *bucket_data_type,
 800                           u16 *dirty_sectors, u16 *cached_sectors)
 801 {
 802         u16 *dst_sectors = !ptr->cached
 803                 ? dirty_sectors
 804                 : cached_sectors;
 805         int ret = check_bucket_ref(c, k, ptr, sectors, ptr_data_type,
 806                                    bucket_gen, *bucket_data_type,
 807                                    *dirty_sectors, *cached_sectors);
 808
 809         if (ret)
 810                 return ret;
 811
 812         *dst_sectors += sectors;
 813         *bucket_data_type = *dirty_sectors || *cached_sectors
 814                 ? ptr_data_type : 0;
 815         return 0;
 816 }
 817
 818 static int bch2_mark_pointer(struct bch_fs *c, struct bkey_s_c k,
 819                              struct extent_ptr_decoded p,
 820                              s64 sectors, enum bch_data_type data_type,
 821                              u64 journal_seq, unsigned flags)
 822 {
 823         bool gc = flags & BTREE_TRIGGER_GC;
 824         struct bucket_mark old, new;
 825         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 826         struct bucket *g = PTR_BUCKET(ca, &p.ptr, gc);
 827         u8 bucket_data_type;
 828         u64 v;
 829         int ret;
 830
 831         v = atomic64_read(&g->_mark.v);
 832         do {
 833                 new.v.counter = old.v.counter = v;
 834                 bucket_data_type = new.data_type;
 835
 836                 ret = __mark_pointer(c, k, &p.ptr, sectors, data_type, new.gen,
 837                                      &bucket_data_type,
 838                                      &new.dirty_sectors,
 839                                      &new.cached_sectors);
 840                 if (ret)
 841                         return ret;
 842
 843                 new.data_type = bucket_data_type;
 844
 845                 if (journal_seq) {
 846                         new.journal_seq_valid = 1;
 847                         new.journal_seq = journal_seq;
 848                 }
 849
 850                 if (flags & BTREE_TRIGGER_NOATOMIC) {
 851                         g->_mark = new;
 852                         break;
 853                 }
 854         } while ((v = atomic64_cmpxchg(&g->_mark.v,
 855                               old.v.counter,
 856                               new.v.counter)) != old.v.counter);
 857
 858         bch2_dev_usage_update(c, ca, old, new, journal_seq, gc);
 859
 860         BUG_ON(!gc && bucket_became_unavailable(old, new));
 861
 862         return 0;
 863 }
 864
 865 static int bch2_mark_stripe_ptr(struct bch_fs *c,
 866                                 struct bch_extent_stripe_ptr p,
 867                                 enum bch_data_type data_type,
 868                                 s64 sectors,
 869                                 unsigned journal_seq, unsigned flags)
 870 {
 871         bool gc = flags & BTREE_TRIGGER_GC;
 872         struct bch_replicas_padded r;
 873         struct stripe *m;
 874         unsigned i, blocks_nonempty = 0;
 875
 876         m = genradix_ptr(&c->stripes[gc], p.idx);
 877
 878         spin_lock(&c->ec_stripes_heap_lock);
 879
 880         if (!m || !m->alive) {
 881                 spin_unlock(&c->ec_stripes_heap_lock);
 882                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
 883                                     (u64) p.idx);
 884                 bch2_inconsistent_error(c);
 885                 return -EIO;
 886         }
 887
 888         m->block_sectors[p.block] += sectors;
 889
 890         r = m->r;
 891
 892         for (i = 0; i < m->nr_blocks; i++)
 893                 blocks_nonempty += m->block_sectors[i] != 0;
 894
 895         if (m->blocks_nonempty != blocks_nonempty) {
 896                 m->blocks_nonempty = blocks_nonempty;
 897                 if (!gc)
 898                         bch2_stripes_heap_update(c, m, p.idx);
 899         }
 900
 901         spin_unlock(&c->ec_stripes_heap_lock);
 902
 903         r.e.data_type = data_type;
 904         update_replicas(c, &r.e, sectors, journal_seq, gc);
 905
 906         return 0;
 907 }
 908
 909 static int bch2_mark_extent(struct bch_fs *c,
 910                             struct bkey_s_c old, struct bkey_s_c new,
 911                             unsigned journal_seq, unsigned flags)
 912 {
 913         bool gc = flags & BTREE_TRIGGER_GC;
 914         struct bkey_s_c k = flags & BTREE_TRIGGER_INSERT ? new : old;
 915         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 916         const union bch_extent_entry *entry;
 917         struct extent_ptr_decoded p;
 918         struct bch_replicas_padded r;
 919         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
 920                 ? BCH_DATA_btree
 921                 : BCH_DATA_user;
 922         s64 sectors = bkey_is_btree_ptr(k.k)
 923                 ? c->opts.btree_node_size
 924                 : k.k->size;
 925         s64 dirty_sectors = 0;
 926         bool stale;
 927         int ret;
 928
 929         BUG_ON((flags & (BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE)) ==
 930                (BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE));
 931
 932         r.e.data_type   = data_type;
 933         r.e.nr_devs     = 0;
 934         r.e.nr_required = 1;
 935
 936         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 937                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
 938
 939                 if (flags & BTREE_TRIGGER_OVERWRITE)
 940                         disk_sectors = -disk_sectors;
 941
 942                 ret = bch2_mark_pointer(c, k, p, disk_sectors, data_type,
 943                                         journal_seq, flags);
 944                 if (ret < 0)
 945                         return ret;
 946
 947                 stale = ret > 0;
 948
 949                 if (p.ptr.cached) {
 950                         if (!stale)
 951                                 if (update_cached_sectors(c, p.ptr.dev, disk_sectors,
 952                                                           journal_seq, gc)) {
 953                                         bch2_fs_fatal_error(c, "bch2_mark_extent(): no replicas entry while updating cached sectors");
 954                                         return -1;
 955
 956                                 }
 957                 } else if (!p.has_ec) {
 958                         dirty_sectors          += disk_sectors;
 959                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
 960                 } else {
 961                         ret = bch2_mark_stripe_ptr(c, p.ec, data_type,
 962                                         disk_sectors, journal_seq, flags);
 963                         if (ret)
 964                                 return ret;
 965
 966                         /*
 967                          * There may be other dirty pointers in this extent, but
 968                          * if so they're not required for mounting if we have an
 969                          * erasure coded pointer in this extent:
 970                          */
 971                         r.e.nr_required = 0;
 972                 }
 973         }
 974
 975         if (r.e.nr_devs) {
 976                 if (update_replicas(c, &r.e, dirty_sectors, journal_seq, gc)) {
 977                         char buf[200];
 978
 979                         bch2_bkey_val_to_text(&PBUF(buf), c, k);
 980                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf);
 981                         return -1;
 982                 }
 983         }
 984
 985         return 0;
 986 }
 987
 988 static int bch2_mark_stripe(struct bch_fs *c,
 989                         struct bkey_s_c old, struct bkey_s_c new,
 990                         u64 journal_seq, unsigned flags)
 991 {
 992         bool gc = flags & BTREE_TRIGGER_GC;
 993         size_t idx = new.k->p.offset;
 994         const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
 995                 ? bkey_s_c_to_stripe(old).v : NULL;
 996         const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
 997                 ? bkey_s_c_to_stripe(new).v : NULL;
 998         struct stripe *m = genradix_ptr(&c->stripes[gc], idx);
 999         unsigned i;
1000         int ret;
1001
1002         BUG_ON(gc && old_s);
1003
1004         if (!m || (old_s && !m->alive)) {
1005                 bch_err_ratelimited(c, "error marking nonexistent stripe %zu",
1006                                     idx);
1007                 bch2_inconsistent_error(c);
1008                 return -1;
1009         }
1010
1011         if (!new_s) {
1012                 spin_lock(&c->ec_stripes_heap_lock);
1013                 bch2_stripes_heap_del(c, m, idx);
1014                 spin_unlock(&c->ec_stripes_heap_lock);
1015
1016                 memset(m, 0, sizeof(*m));
1017         } else {
1018                 m->alive        = true;
1019                 m->sectors      = le16_to_cpu(new_s->sectors);
1020                 m->algorithm    = new_s->algorithm;
1021                 m->nr_blocks    = new_s->nr_blocks;
1022                 m->nr_redundant = new_s->nr_redundant;
1023                 m->blocks_nonempty = 0;
1024
1025                 for (i = 0; i < new_s->nr_blocks; i++) {
1026                         m->block_sectors[i] =
1027                                 stripe_blockcount_get(new_s, i);
1028                         m->blocks_nonempty += !!m->block_sectors[i];
1029
1030                         m->ptrs[i] = new_s->ptrs[i];
1031                 }
1032
1033                 bch2_bkey_to_replicas(&m->r.e, new);
1034
1035                 if (!gc) {
1036                         spin_lock(&c->ec_stripes_heap_lock);
1037                         bch2_stripes_heap_update(c, m, idx);
1038                         spin_unlock(&c->ec_stripes_heap_lock);
1039                 }
1040         }
1041
1042         if (gc) {
1043                 /*
1044                  * gc recalculates this field from stripe ptr
1045                  * references:
1046                  */
1047                 memset(m->block_sectors, 0, sizeof(m->block_sectors));
1048                 m->blocks_nonempty = 0;
1049
1050                 for (i = 0; i < new_s->nr_blocks; i++) {
1051                         ret = mark_stripe_bucket(c, new, i, journal_seq, flags);
1052                         if (ret)
1053                                 return ret;
1054                 }
1055
1056                 if (update_replicas(c, &m->r.e,
1057                                     ((s64) m->sectors * m->nr_redundant),
1058                                     journal_seq, gc)) {
1059                         char buf[200];
1060
1061                         bch2_bkey_val_to_text(&PBUF(buf), c, new);
1062                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf);
1063                         return -1;
1064                 }
1065         }
1066
1067         return 0;
1068 }
1069
1070 static int bch2_mark_inode(struct bch_fs *c,
1071                         struct bkey_s_c old, struct bkey_s_c new,
1072                         u64 journal_seq, unsigned flags)
1073 {
1074         struct bch_fs_usage __percpu *fs_usage;
1075
1076         preempt_disable();
1077         fs_usage = fs_usage_ptr(c, journal_seq, flags & BTREE_TRIGGER_GC);
1078         fs_usage->nr_inodes += new.k->type == KEY_TYPE_inode;
1079         fs_usage->nr_inodes -= old.k->type == KEY_TYPE_inode;
1080         preempt_enable();
1081         return 0;
1082 }
1083
1084 static int bch2_mark_reservation(struct bch_fs *c,
1085                         struct bkey_s_c old, struct bkey_s_c new,
1086                         u64 journal_seq, unsigned flags)
1087 {
1088         struct bkey_s_c k = flags & BTREE_TRIGGER_INSERT ? new : old;
1089         struct bch_fs_usage __percpu *fs_usage;
1090         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1091         s64 sectors = (s64) k.k->size;
1092
1093         if (flags & BTREE_TRIGGER_OVERWRITE)
1094                 sectors = -sectors;
1095         sectors *= replicas;
1096
1097         preempt_disable();
1098         fs_usage = fs_usage_ptr(c, journal_seq, flags & BTREE_TRIGGER_GC);
1099         replicas = clamp_t(unsigned, replicas, 1,
1100                            ARRAY_SIZE(fs_usage->persistent_reserved));
1101
1102         fs_usage->reserved                              += sectors;
1103         fs_usage->persistent_reserved[replicas - 1]     += sectors;
1104         preempt_enable();
1105
1106         return 0;
1107 }
1108
1109 static s64 __bch2_mark_reflink_p(struct bch_fs *c, struct bkey_s_c_reflink_p p,
1110                                  u64 idx, unsigned flags, size_t *r_idx)
1111 {
1112         struct reflink_gc *r;
1113         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1114
1115         while (1) {
1116                 if (*r_idx >= c->reflink_gc_nr)
1117                         goto not_found;
1118                 r = genradix_ptr(&c->reflink_gc_table, *r_idx);
1119                 BUG_ON(!r);
1120
1121                 if (idx < r->offset)
1122                         break;
1123                 (*r_idx)++;
1124         }
1125
1126         BUG_ON((s64) r->refcount + add < 0);
1127
1128         r->refcount += add;
1129         return r->offset - idx;
1130 not_found:
1131         bch2_fs_inconsistent(c,
1132                 "%llu:%llu len %u points to nonexistent indirect extent %llu",
1133                 p.k->p.inode, p.k->p.offset, p.k->size, idx);
1134         bch2_inconsistent_error(c);
1135         return -EIO;
1136 }
1137
1138 static int bch2_mark_reflink_p(struct bch_fs *c,
1139                         struct bkey_s_c old, struct bkey_s_c new,
1140                         u64 journal_seq, unsigned flags)
1141 {
1142         struct bkey_s_c k = flags & BTREE_TRIGGER_INSERT ? new : old;
1143         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1144         struct reflink_gc *ref;
1145         size_t l, r, m;
1146         u64 idx = le64_to_cpu(p.v->idx);
1147         unsigned sectors = p.k->size;
1148         s64 ret = 0;
1149
1150         BUG_ON((flags & (BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE)) ==
1151                (BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE));
1152
1153         l = 0;
1154         r = c->reflink_gc_nr;
1155         while (l < r) {
1156                 m = l + (r - l) / 2;
1157
1158                 ref = genradix_ptr(&c->reflink_gc_table, m);
1159                 if (ref->offset <= idx)
1160                         l = m + 1;
1161                 else
1162                         r = m;
1163         }
1164
1165         while (sectors) {
1166                 ret = __bch2_mark_reflink_p(c, p, idx, flags, &l);
1167                 if (ret < 0)
1168                         return ret;
1169
1170                 ret = min_t(s64, ret, sectors);
1171                 idx     += ret;
1172                 sectors -= ret;
1173         }
1174
1175         return 0;
1176 }
1177
1178 static int bch2_mark_key_locked(struct bch_fs *c,
1179                    struct bkey_s_c old,
1180                    struct bkey_s_c new,
1181                    u64 journal_seq, unsigned flags)
1182 {
1183         struct bkey_s_c k = flags & BTREE_TRIGGER_INSERT ? new : old;
1184
1185         BUG_ON(!(flags & (BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE)));
1186
1187         switch (k.k->type) {
1188         case KEY_TYPE_alloc:
1189         case KEY_TYPE_alloc_v2:
1190                 return bch2_mark_alloc(c, old, new, journal_seq, flags);
1191         case KEY_TYPE_btree_ptr:
1192         case KEY_TYPE_btree_ptr_v2:
1193         case KEY_TYPE_extent:
1194         case KEY_TYPE_reflink_v:
1195                 return bch2_mark_extent(c, old, new, journal_seq, flags);
1196         case KEY_TYPE_stripe:
1197                 return bch2_mark_stripe(c, old, new, journal_seq, flags);
1198         case KEY_TYPE_inode:
1199                 return bch2_mark_inode(c, old, new, journal_seq, flags);
1200         case KEY_TYPE_reservation:
1201                 return bch2_mark_reservation(c, old, new, journal_seq, flags);
1202         case KEY_TYPE_reflink_p:
1203                 return bch2_mark_reflink_p(c, old, new, journal_seq, flags);
1204         case KEY_TYPE_snapshot:
1205                 return bch2_mark_snapshot(c, old, new, journal_seq, flags);
1206         default:
1207                 return 0;
1208         }
1209 }
1210
1211 int bch2_mark_key(struct bch_fs *c, struct bkey_s_c new, unsigned flags)
1212 {
1213         struct bkey deleted = KEY(0, 0, 0);
1214         struct bkey_s_c old = (struct bkey_s_c) { &deleted, NULL };
1215         int ret;
1216
1217         percpu_down_read(&c->mark_lock);
1218         ret = bch2_mark_key_locked(c, old, new, 0, flags);
1219         percpu_up_read(&c->mark_lock);
1220
1221         return ret;
1222 }
1223
1224 int bch2_mark_update(struct btree_trans *trans, struct btree_path *path,
1225                      struct bkey_i *new, unsigned flags)
1226 {
1227         struct bch_fs           *c = trans->c;
1228         struct bkey             _deleted = KEY(0, 0, 0);
1229         struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
1230         struct bkey_s_c         old;
1231         struct bkey             unpacked;
1232         int ret;
1233
1234         if (unlikely(flags & BTREE_TRIGGER_NORUN))
1235                 return 0;
1236
1237         if (!btree_node_type_needs_gc(path->btree_id))
1238                 return 0;
1239
1240         old = bch2_btree_path_peek_slot(path, &unpacked);
1241
1242         if (old.k->type == new->k.type &&
1243             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
1244                 ret   = bch2_mark_key_locked(c, old, bkey_i_to_s_c(new),
1245                                 trans->journal_res.seq,
1246                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
1247         } else {
1248                 ret   = bch2_mark_key_locked(c, deleted, bkey_i_to_s_c(new),
1249                                 trans->journal_res.seq,
1250                                 BTREE_TRIGGER_INSERT|flags) ?:
1251                         bch2_mark_key_locked(c, old, deleted,
1252                                 trans->journal_res.seq,
1253                                 BTREE_TRIGGER_OVERWRITE|flags);
1254         }
1255
1256         return ret;
1257 }
1258
1259 static noinline __cold
1260 void fs_usage_apply_warn(struct btree_trans *trans,
1261                          unsigned disk_res_sectors,
1262                          s64 should_not_have_added)
1263 {
1264         struct bch_fs *c = trans->c;
1265         struct btree_insert_entry *i;
1266         char buf[200];
1267
1268         bch_err(c, "disk usage increased %lli more than %u sectors reserved",
1269                 should_not_have_added, disk_res_sectors);
1270
1271         trans_for_each_update(trans, i) {
1272                 pr_err("while inserting");
1273                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
1274                 pr_err("%s", buf);
1275                 pr_err("overlapping with");
1276
1277                 if (!i->cached) {
1278                         struct bkey u;
1279                         struct bkey_s_c k = bch2_btree_path_peek_slot(i->path, &u);
1280
1281                         bch2_bkey_val_to_text(&PBUF(buf), c, k);
1282                         pr_err("%s", buf);
1283                 } else {
1284                         struct bkey_cached *ck = (void *) i->path->l[0].b;
1285
1286                         if (ck->valid) {
1287                                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(ck->k));
1288                                 pr_err("%s", buf);
1289                         }
1290                 }
1291         }
1292         __WARN();
1293 }
1294
1295 void bch2_trans_fs_usage_apply(struct btree_trans *trans,
1296                                struct replicas_delta_list *deltas)
1297 {
1298         struct bch_fs *c = trans->c;
1299         static int warned_disk_usage = 0;
1300         bool warn = false;
1301         unsigned disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1302         struct replicas_delta *d = deltas->d;
1303         struct replicas_delta *top = (void *) deltas->d + deltas->used;
1304         struct bch_fs_usage *dst;
1305         s64 added = 0, should_not_have_added;
1306         unsigned i;
1307
1308         percpu_rwsem_assert_held(&c->mark_lock);
1309
1310         preempt_disable();
1311         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1312
1313         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1314                 switch (d->r.data_type) {
1315                 case BCH_DATA_btree:
1316                 case BCH_DATA_user:
1317                 case BCH_DATA_parity:
1318                         added += d->delta;
1319                 }
1320
1321                 BUG_ON(__update_replicas(c, dst, &d->r, d->delta));
1322         }
1323
1324         dst->nr_inodes += deltas->nr_inodes;
1325
1326         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1327                 added                           += deltas->persistent_reserved[i];
1328                 dst->reserved                   += deltas->persistent_reserved[i];
1329                 dst->persistent_reserved[i]     += deltas->persistent_reserved[i];
1330         }
1331
1332         /*
1333          * Not allowed to reduce sectors_available except by getting a
1334          * reservation:
1335          */
1336         should_not_have_added = added - (s64) disk_res_sectors;
1337         if (unlikely(should_not_have_added > 0)) {
1338                 u64 old, new, v = atomic64_read(&c->sectors_available);
1339
1340                 do {
1341                         old = v;
1342                         new = max_t(s64, 0, old - should_not_have_added);
1343                 } while ((v = atomic64_cmpxchg(&c->sectors_available,
1344                                                old, new)) != old);
1345
1346                 added -= should_not_have_added;
1347                 warn = true;
1348         }
1349
1350         if (added > 0) {
1351                 trans->disk_res->sectors -= added;
1352                 this_cpu_sub(*c->online_reserved, added);
1353         }
1354
1355         preempt_enable();
1356
1357         if (unlikely(warn) && !xchg(&warned_disk_usage, 1))
1358                 fs_usage_apply_warn(trans, disk_res_sectors, should_not_have_added);
1359 }
1360
1361 /* trans_mark: */
1362
1363 static struct bkey_alloc_buf *
1364 bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
1365                               const struct bch_extent_ptr *ptr,
1366                               struct bkey_alloc_unpacked *u)
1367 {
1368         struct bch_fs *c = trans->c;
1369         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1370         struct bpos pos = POS(ptr->dev, PTR_BUCKET_NR(ca, ptr));
1371         struct bucket *g;
1372         struct bkey_alloc_buf *a;
1373         struct bkey_i *update = btree_trans_peek_updates(trans, BTREE_ID_alloc, pos);
1374         int ret;
1375
1376         a = bch2_trans_kmalloc(trans, sizeof(struct bkey_alloc_buf));
1377         if (IS_ERR(a))
1378                 return a;
1379
1380         bch2_trans_iter_init(trans, iter, BTREE_ID_alloc, pos,
1381                              BTREE_ITER_CACHED|
1382                              BTREE_ITER_CACHED_NOFILL|
1383                              BTREE_ITER_INTENT);
1384         ret = bch2_btree_iter_traverse(iter);
1385         if (ret) {
1386                 bch2_trans_iter_exit(trans, iter);
1387                 return ERR_PTR(ret);
1388         }
1389
1390         if (update && !bpos_cmp(update->k.p, pos)) {
1391                 *u = bch2_alloc_unpack(bkey_i_to_s_c(update));
1392         } else {
1393                 percpu_down_read(&c->mark_lock);
1394                 g = bucket(ca, pos.offset);
1395                 *u = alloc_mem_to_key(iter, g, READ_ONCE(g->mark));
1396                 percpu_up_read(&c->mark_lock);
1397         }
1398
1399         return a;
1400 }
1401
1402 static int bch2_trans_mark_pointer(struct btree_trans *trans,
1403                         struct bkey_s_c k, struct extent_ptr_decoded p,
1404                         s64 sectors, enum bch_data_type data_type)
1405 {
1406         struct bch_fs *c = trans->c;
1407         struct btree_iter iter;
1408         struct bkey_alloc_unpacked u;
1409         struct bkey_alloc_buf *a;
1410         int ret;
1411
1412         a = bch2_trans_start_alloc_update(trans, &iter, &p.ptr, &u);
1413         if (IS_ERR(a))
1414                 return PTR_ERR(a);
1415
1416         ret = __mark_pointer(c, k, &p.ptr, sectors, data_type, u.gen, &u.data_type,
1417                              &u.dirty_sectors, &u.cached_sectors);
1418         if (ret)
1419                 goto out;
1420
1421         bch2_alloc_pack(c, a, u);
1422         bch2_trans_update(trans, &iter, &a->k, 0);
1423 out:
1424         bch2_trans_iter_exit(trans, &iter);
1425         return ret;
1426 }
1427
1428 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1429                         struct extent_ptr_decoded p,
1430                         s64 sectors, enum bch_data_type data_type)
1431 {
1432         struct bch_fs *c = trans->c;
1433         struct btree_iter iter;
1434         struct bkey_s_c k;
1435         struct bkey_i_stripe *s;
1436         struct bch_replicas_padded r;
1437         int ret = 0;
1438
1439         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes, POS(0, p.ec.idx),
1440                              BTREE_ITER_INTENT|
1441                              BTREE_ITER_WITH_UPDATES);
1442         k = bch2_btree_iter_peek_slot(&iter);
1443         ret = bkey_err(k);
1444         if (ret)
1445                 goto err;
1446
1447         if (k.k->type != KEY_TYPE_stripe) {
1448                 bch2_fs_inconsistent(c,
1449                         "pointer to nonexistent stripe %llu",
1450                         (u64) p.ec.idx);
1451                 bch2_inconsistent_error(c);
1452                 ret = -EIO;
1453                 goto err;
1454         }
1455
1456         if (!bch2_ptr_matches_stripe(bkey_s_c_to_stripe(k).v, p)) {
1457                 bch2_fs_inconsistent(c,
1458                         "stripe pointer doesn't match stripe %llu",
1459                         (u64) p.ec.idx);
1460                 ret = -EIO;
1461                 goto err;
1462         }
1463
1464         s = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1465         ret = PTR_ERR_OR_ZERO(s);
1466         if (ret)
1467                 goto err;
1468
1469         bkey_reassemble(&s->k_i, k);
1470         stripe_blockcount_set(&s->v, p.ec.block,
1471                 stripe_blockcount_get(&s->v, p.ec.block) +
1472                 sectors);
1473         bch2_trans_update(trans, &iter, &s->k_i, 0);
1474
1475         bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1476         r.e.data_type = data_type;
1477         update_replicas_list(trans, &r.e, sectors);
1478 err:
1479         bch2_trans_iter_exit(trans, &iter);
1480         return ret;
1481 }
1482
1483 static int bch2_trans_mark_extent(struct btree_trans *trans,
1484                         struct bkey_s_c k, unsigned flags)
1485 {
1486         struct bch_fs *c = trans->c;
1487         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1488         const union bch_extent_entry *entry;
1489         struct extent_ptr_decoded p;
1490         struct bch_replicas_padded r;
1491         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
1492                 ? BCH_DATA_btree
1493                 : BCH_DATA_user;
1494         s64 sectors = bkey_is_btree_ptr(k.k)
1495                 ? c->opts.btree_node_size
1496                 : k.k->size;
1497         s64 dirty_sectors = 0;
1498         bool stale;
1499         int ret;
1500
1501         BUG_ON((flags & (BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE)) ==
1502                (BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE));
1503
1504         r.e.data_type   = data_type;
1505         r.e.nr_devs     = 0;
1506         r.e.nr_required = 1;
1507
1508         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1509                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
1510
1511                 if (flags & BTREE_TRIGGER_OVERWRITE)
1512                         disk_sectors = -disk_sectors;
1513
1514                 ret = bch2_trans_mark_pointer(trans, k, p,
1515                                         disk_sectors, data_type);
1516                 if (ret < 0)
1517                         return ret;
1518
1519                 stale = ret > 0;
1520
1521                 if (p.ptr.cached) {
1522                         if (!stale)
1523                                 update_cached_sectors_list(trans, p.ptr.dev,
1524                                                            disk_sectors);
1525                 } else if (!p.has_ec) {
1526                         dirty_sectors          += disk_sectors;
1527                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1528                 } else {
1529                         ret = bch2_trans_mark_stripe_ptr(trans, p,
1530                                         disk_sectors, data_type);
1531                         if (ret)
1532                                 return ret;
1533
1534                         r.e.nr_required = 0;
1535                 }
1536         }
1537
1538         if (r.e.nr_devs)
1539                 update_replicas_list(trans, &r.e, dirty_sectors);
1540
1541         return 0;
1542 }
1543
1544 static int bch2_trans_mark_stripe_alloc_ref(struct btree_trans *trans,
1545                                             struct bkey_s_c_stripe s,
1546                                             unsigned idx, bool deleting)
1547 {
1548         struct bch_fs *c = trans->c;
1549         const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
1550         struct bkey_alloc_buf *a;
1551         struct btree_iter iter;
1552         struct bkey_alloc_unpacked u;
1553         bool parity = idx >= s.v->nr_blocks - s.v->nr_redundant;
1554         int ret = 0;
1555
1556         a = bch2_trans_start_alloc_update(trans, &iter, ptr, &u);
1557         if (IS_ERR(a))
1558                 return PTR_ERR(a);
1559
1560         if (parity) {
1561                 s64 sectors = le16_to_cpu(s.v->sectors);
1562
1563                 if (deleting)
1564                         sectors = -sectors;
1565
1566                 u.dirty_sectors += sectors;
1567                 u.data_type = u.dirty_sectors
1568                         ? BCH_DATA_parity
1569                         : 0;
1570         }
1571
1572         if (!deleting) {
1573                 if (bch2_fs_inconsistent_on(u.stripe && u.stripe != s.k->p.offset, c,
1574                                 "bucket %llu:%llu gen %u: multiple stripes using same bucket (%u, %llu)",
1575                                 iter.pos.inode, iter.pos.offset, u.gen,
1576                                 u.stripe, s.k->p.offset)) {
1577                         ret = -EIO;
1578                         goto err;
1579                 }
1580
1581                 u.stripe                = s.k->p.offset;
1582                 u.stripe_redundancy     = s.v->nr_redundant;
1583         } else {
1584                 u.stripe                = 0;
1585                 u.stripe_redundancy     = 0;
1586         }
1587
1588         bch2_alloc_pack(c, a, u);
1589         bch2_trans_update(trans, &iter, &a->k, 0);
1590 err:
1591         bch2_trans_iter_exit(trans, &iter);
1592         return ret;
1593 }
1594
1595 static int bch2_trans_mark_stripe(struct btree_trans *trans,
1596                                   struct bkey_s_c old, struct bkey_s_c new,
1597                                   unsigned flags)
1598 {
1599         struct bkey_s_c_stripe old_s = { .k = NULL };
1600         struct bkey_s_c_stripe new_s = { .k = NULL };
1601         struct bch_replicas_padded r;
1602         unsigned i;
1603         int ret = 0;
1604
1605         if (old.k->type == KEY_TYPE_stripe)
1606                 old_s = bkey_s_c_to_stripe(old);
1607         if (new.k->type == KEY_TYPE_stripe)
1608                 new_s = bkey_s_c_to_stripe(new);
1609
1610         /*
1611          * If the pointers aren't changing, we don't need to do anything:
1612          */
1613         if (new_s.k && old_s.k &&
1614             new_s.v->nr_blocks          == old_s.v->nr_blocks &&
1615             new_s.v->nr_redundant       == old_s.v->nr_redundant &&
1616             !memcmp(old_s.v->ptrs, new_s.v->ptrs,
1617                     new_s.v->nr_blocks * sizeof(struct bch_extent_ptr)))
1618                 return 0;
1619
1620         if (new_s.k) {
1621                 s64 sectors = le16_to_cpu(new_s.v->sectors);
1622
1623                 bch2_bkey_to_replicas(&r.e, new);
1624                 update_replicas_list(trans, &r.e, sectors * new_s.v->nr_redundant);
1625
1626                 for (i = 0; i < new_s.v->nr_blocks; i++) {
1627                         ret = bch2_trans_mark_stripe_alloc_ref(trans, new_s,
1628                                                                i, false);
1629                         if (ret)
1630                                 return ret;
1631                 }
1632         }
1633
1634         if (old_s.k) {
1635                 s64 sectors = -((s64) le16_to_cpu(old_s.v->sectors));
1636
1637                 bch2_bkey_to_replicas(&r.e, old);
1638                 update_replicas_list(trans, &r.e, sectors * old_s.v->nr_redundant);
1639
1640                 for (i = 0; i < old_s.v->nr_blocks; i++) {
1641                         ret = bch2_trans_mark_stripe_alloc_ref(trans, old_s,
1642                                                                i, true);
1643                         if (ret)
1644                                 return ret;
1645                 }
1646         }
1647
1648         return ret;
1649 }
1650
1651 static int bch2_trans_mark_inode(struct btree_trans *trans,
1652                                  struct bkey_s_c old,
1653                                  struct bkey_s_c new,
1654                                  unsigned flags)
1655 {
1656         int nr = (new.k->type == KEY_TYPE_inode) -
1657                 (old.k->type == KEY_TYPE_inode);
1658
1659         if (nr) {
1660                 struct replicas_delta_list *d =
1661                         replicas_deltas_realloc(trans, 0);
1662                 d->nr_inodes += nr;
1663         }
1664
1665         return 0;
1666 }
1667
1668 static int bch2_trans_mark_reservation(struct btree_trans *trans,
1669                                        struct bkey_s_c k, unsigned flags)
1670 {
1671         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1672         s64 sectors = (s64) k.k->size;
1673         struct replicas_delta_list *d;
1674
1675         BUG_ON((flags & (BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE)) ==
1676                (BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE));
1677
1678         if (flags & BTREE_TRIGGER_OVERWRITE)
1679                 sectors = -sectors;
1680         sectors *= replicas;
1681
1682         d = replicas_deltas_realloc(trans, 0);
1683
1684         replicas = clamp_t(unsigned, replicas, 1,
1685                            ARRAY_SIZE(d->persistent_reserved));
1686
1687         d->persistent_reserved[replicas - 1] += sectors;
1688         return 0;
1689 }
1690
1691 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1692                         struct bkey_s_c_reflink_p p,
1693                         u64 idx, unsigned flags)
1694 {
1695         struct bch_fs *c = trans->c;
1696         struct btree_iter iter;
1697         struct bkey_s_c k;
1698         struct bkey_i *n;
1699         __le64 *refcount;
1700         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1701         s64 ret;
1702
1703         bch2_trans_iter_init(trans, &iter, BTREE_ID_reflink, POS(0, idx),
1704                              BTREE_ITER_INTENT|
1705                              BTREE_ITER_WITH_UPDATES);
1706         k = bch2_btree_iter_peek_slot(&iter);
1707         ret = bkey_err(k);
1708         if (ret)
1709                 goto err;
1710
1711         n = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1712         ret = PTR_ERR_OR_ZERO(n);
1713         if (ret)
1714                 goto err;
1715
1716         bkey_reassemble(n, k);
1717
1718         refcount = bkey_refcount(n);
1719         if (!refcount) {
1720                 bch2_fs_inconsistent(c,
1721                         "%llu:%llu len %u points to nonexistent indirect extent %llu",
1722                         p.k->p.inode, p.k->p.offset, p.k->size, idx);
1723                 bch2_inconsistent_error(c);
1724                 ret = -EIO;
1725                 goto err;
1726         }
1727
1728         BUG_ON(!*refcount && (flags & BTREE_TRIGGER_OVERWRITE));
1729         le64_add_cpu(refcount, add);
1730
1731         if (!*refcount) {
1732                 n->k.type = KEY_TYPE_deleted;
1733                 set_bkey_val_u64s(&n->k, 0);
1734         }
1735
1736         bch2_btree_iter_set_pos_to_extent_start(&iter);
1737         ret = bch2_trans_update(trans, &iter, n, 0);
1738         if (ret)
1739                 goto err;
1740
1741         ret = k.k->p.offset - idx;
1742 err:
1743         bch2_trans_iter_exit(trans, &iter);
1744         return ret;
1745 }
1746
1747 static int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1748                                      struct bkey_s_c k, unsigned flags)
1749 {
1750         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1751         u64 idx = le64_to_cpu(p.v->idx);
1752         unsigned sectors = p.k->size;
1753         s64 ret = 0;
1754
1755         while (sectors) {
1756                 ret = __bch2_trans_mark_reflink_p(trans, p, idx, flags);
1757                 if (ret < 0)
1758                         return ret;
1759
1760                 ret = min_t(s64, ret, sectors);
1761                 idx     += ret;
1762                 sectors -= ret;
1763         }
1764
1765         return 0;
1766 }
1767
1768 int bch2_trans_mark_key(struct btree_trans *trans, struct bkey_s_c old,
1769                         struct bkey_s_c new, unsigned flags)
1770 {
1771         struct bkey_s_c k = flags & BTREE_TRIGGER_INSERT ? new : old;
1772
1773         BUG_ON(!(flags & (BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE)));
1774
1775         switch (k.k->type) {
1776         case KEY_TYPE_btree_ptr:
1777         case KEY_TYPE_btree_ptr_v2:
1778         case KEY_TYPE_extent:
1779         case KEY_TYPE_reflink_v:
1780                 return bch2_trans_mark_extent(trans, k, flags);
1781         case KEY_TYPE_stripe:
1782                 return bch2_trans_mark_stripe(trans, old, new, flags);
1783         case KEY_TYPE_inode:
1784                 return bch2_trans_mark_inode(trans, old, new, flags);
1785         case KEY_TYPE_reservation:
1786                 return bch2_trans_mark_reservation(trans, k, flags);
1787         case KEY_TYPE_reflink_p:
1788                 return bch2_trans_mark_reflink_p(trans, k, flags);
1789         default:
1790                 return 0;
1791         }
1792 }
1793
1794 int bch2_trans_mark_update(struct btree_trans *trans,
1795                            struct btree_path *path,
1796                            struct bkey_i *new,
1797                            unsigned flags)
1798 {
1799         struct bkey             _deleted = KEY(0, 0, 0);
1800         struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
1801         struct bkey_s_c         old;
1802         struct bkey             unpacked;
1803         int ret;
1804
1805         if (unlikely(flags & BTREE_TRIGGER_NORUN))
1806                 return 0;
1807
1808         if (!btree_node_type_needs_gc(path->btree_id))
1809                 return 0;
1810
1811         old = bch2_btree_path_peek_slot(path, &unpacked);
1812
1813         if (old.k->type == new->k.type &&
1814             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
1815                 ret   = bch2_trans_mark_key(trans, old, bkey_i_to_s_c(new),
1816                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
1817         } else {
1818                 ret   = bch2_trans_mark_key(trans, deleted, bkey_i_to_s_c(new),
1819                                 BTREE_TRIGGER_INSERT|flags) ?:
1820                         bch2_trans_mark_key(trans, old, deleted,
1821                                 BTREE_TRIGGER_OVERWRITE|flags);
1822         }
1823
1824         return ret;
1825 }
1826
1827 static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1828                                     struct bch_dev *ca, size_t b,
1829                                     enum bch_data_type type,
1830                                     unsigned sectors)
1831 {
1832         struct bch_fs *c = trans->c;
1833         struct btree_iter iter;
1834         struct bkey_alloc_unpacked u;
1835         struct bkey_alloc_buf *a;
1836         struct bch_extent_ptr ptr = {
1837                 .dev = ca->dev_idx,
1838                 .offset = bucket_to_sector(ca, b),
1839         };
1840         int ret = 0;
1841
1842         /*
1843          * Backup superblock might be past the end of our normal usable space:
1844          */
1845         if (b >= ca->mi.nbuckets)
1846                 return 0;
1847
1848         a = bch2_trans_start_alloc_update(trans, &iter, &ptr, &u);
1849         if (IS_ERR(a))
1850                 return PTR_ERR(a);
1851
1852         if (u.data_type && u.data_type != type) {
1853                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1854                         "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
1855                         "while marking %s",
1856                         iter.pos.inode, iter.pos.offset, u.gen,
1857                         bch2_data_types[u.data_type],
1858                         bch2_data_types[type],
1859                         bch2_data_types[type]);
1860                 ret = -EIO;
1861                 goto out;
1862         }
1863
1864         u.data_type     = type;
1865         u.dirty_sectors = sectors;
1866
1867         bch2_alloc_pack(c, a, u);
1868         bch2_trans_update(trans, &iter, &a->k, 0);
1869 out:
1870         bch2_trans_iter_exit(trans, &iter);
1871         return ret;
1872 }
1873
1874 int bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1875                                     struct bch_dev *ca, size_t b,
1876                                     enum bch_data_type type,
1877                                     unsigned sectors)
1878 {
1879         return __bch2_trans_do(trans, NULL, NULL, 0,
1880                         __bch2_trans_mark_metadata_bucket(trans, ca, b, type, sectors));
1881 }
1882
1883 static int bch2_trans_mark_metadata_sectors(struct btree_trans *trans,
1884                                             struct bch_dev *ca,
1885                                             u64 start, u64 end,
1886                                             enum bch_data_type type,
1887                                             u64 *bucket, unsigned *bucket_sectors)
1888 {
1889         do {
1890                 u64 b = sector_to_bucket(ca, start);
1891                 unsigned sectors =
1892                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1893
1894                 if (b != *bucket && *bucket_sectors) {
1895                         int ret = bch2_trans_mark_metadata_bucket(trans, ca, *bucket,
1896                                                                   type, *bucket_sectors);
1897                         if (ret)
1898                                 return ret;
1899
1900                         *bucket_sectors = 0;
1901                 }
1902
1903                 *bucket         = b;
1904                 *bucket_sectors += sectors;
1905                 start += sectors;
1906         } while (start < end);
1907
1908         return 0;
1909 }
1910
1911 static int __bch2_trans_mark_dev_sb(struct btree_trans *trans,
1912                                     struct bch_dev *ca)
1913 {
1914         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1915         u64 bucket = 0;
1916         unsigned i, bucket_sectors = 0;
1917         int ret;
1918
1919         for (i = 0; i < layout->nr_superblocks; i++) {
1920                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1921
1922                 if (offset == BCH_SB_SECTOR) {
1923                         ret = bch2_trans_mark_metadata_sectors(trans, ca,
1924                                                 0, BCH_SB_SECTOR,
1925                                                 BCH_DATA_sb, &bucket, &bucket_sectors);
1926                         if (ret)
1927                                 return ret;
1928                 }
1929
1930                 ret = bch2_trans_mark_metadata_sectors(trans, ca, offset,
1931                                       offset + (1 << layout->sb_max_size_bits),
1932                                       BCH_DATA_sb, &bucket, &bucket_sectors);
1933                 if (ret)
1934                         return ret;
1935         }
1936
1937         if (bucket_sectors) {
1938                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1939                                 bucket, BCH_DATA_sb, bucket_sectors);
1940                 if (ret)
1941                         return ret;
1942         }
1943
1944         for (i = 0; i < ca->journal.nr; i++) {
1945                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
1946                                 ca->journal.buckets[i],
1947                                 BCH_DATA_journal, ca->mi.bucket_size);
1948                 if (ret)
1949                         return ret;
1950         }
1951
1952         return 0;
1953 }
1954
1955 int bch2_trans_mark_dev_sb(struct bch_fs *c, struct bch_dev *ca)
1956 {
1957         return bch2_trans_do(c, NULL, NULL, BTREE_INSERT_LAZY_RW,
1958                         __bch2_trans_mark_dev_sb(&trans, ca));
1959 }
1960
1961 /* Disk reservations: */
1962
1963 #define SECTORS_CACHE   1024
1964
1965 int bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
1966                               u64 sectors, int flags)
1967 {
1968         struct bch_fs_pcpu *pcpu;
1969         u64 old, v, get;
1970         s64 sectors_available;
1971         int ret;
1972
1973         percpu_down_read(&c->mark_lock);
1974         preempt_disable();
1975         pcpu = this_cpu_ptr(c->pcpu);
1976
1977         if (sectors <= pcpu->sectors_available)
1978                 goto out;
1979
1980         v = atomic64_read(&c->sectors_available);
1981         do {
1982                 old = v;
1983                 get = min((u64) sectors + SECTORS_CACHE, old);
1984
1985                 if (get < sectors) {
1986                         preempt_enable();
1987                         goto recalculate;
1988                 }
1989         } while ((v = atomic64_cmpxchg(&c->sectors_available,
1990                                        old, old - get)) != old);
1991
1992         pcpu->sectors_available         += get;
1993
1994 out:
1995         pcpu->sectors_available         -= sectors;
1996         this_cpu_add(*c->online_reserved, sectors);
1997         res->sectors                    += sectors;
1998
1999         preempt_enable();
2000         percpu_up_read(&c->mark_lock);
2001         return 0;
2002
2003 recalculate:
2004         mutex_lock(&c->sectors_available_lock);
2005
2006         percpu_u64_set(&c->pcpu->sectors_available, 0);
2007         sectors_available = avail_factor(__bch2_fs_usage_read_short(c).free);
2008
2009         if (sectors <= sectors_available ||
2010             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
2011                 atomic64_set(&c->sectors_available,
2012                              max_t(s64, 0, sectors_available - sectors));
2013                 this_cpu_add(*c->online_reserved, sectors);
2014                 res->sectors                    += sectors;
2015                 ret = 0;
2016         } else {
2017                 atomic64_set(&c->sectors_available, sectors_available);
2018                 ret = -ENOSPC;
2019         }
2020
2021         mutex_unlock(&c->sectors_available_lock);
2022         percpu_up_read(&c->mark_lock);
2023
2024         return ret;
2025 }
2026
2027 /* Startup/shutdown: */
2028
2029 static void buckets_free_rcu(struct rcu_head *rcu)
2030 {
2031         struct bucket_array *buckets =
2032                 container_of(rcu, struct bucket_array, rcu);
2033
2034         kvpfree(buckets,
2035                 sizeof(struct bucket_array) +
2036                 buckets->nbuckets * sizeof(struct bucket));
2037 }
2038
2039 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
2040 {
2041         struct bucket_array *buckets = NULL, *old_buckets = NULL;
2042         unsigned long *buckets_nouse = NULL;
2043         alloc_fifo      free[RESERVE_NR];
2044         alloc_fifo      free_inc;
2045         alloc_heap      alloc_heap;
2046
2047         size_t btree_reserve    = DIV_ROUND_UP(BTREE_NODE_RESERVE,
2048                              ca->mi.bucket_size / c->opts.btree_node_size);
2049         /* XXX: these should be tunable */
2050         size_t reserve_none     = max_t(size_t, 1, nbuckets >> 9);
2051         size_t copygc_reserve   = max_t(size_t, 2, nbuckets >> 6);
2052         size_t free_inc_nr      = max(max_t(size_t, 1, nbuckets >> 12),
2053                                       btree_reserve * 2);
2054         bool resize = ca->buckets[0] != NULL;
2055         int ret = -ENOMEM;
2056         unsigned i;
2057
2058         memset(&free,           0, sizeof(free));
2059         memset(&free_inc,       0, sizeof(free_inc));
2060         memset(&alloc_heap,     0, sizeof(alloc_heap));
2061
2062         if (!(buckets           = kvpmalloc(sizeof(struct bucket_array) +
2063                                             nbuckets * sizeof(struct bucket),
2064                                             GFP_KERNEL|__GFP_ZERO)) ||
2065             !(buckets_nouse     = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2066                                             sizeof(unsigned long),
2067                                             GFP_KERNEL|__GFP_ZERO)) ||
2068             !init_fifo(&free[RESERVE_MOVINGGC],
2069                        copygc_reserve, GFP_KERNEL) ||
2070             !init_fifo(&free[RESERVE_NONE], reserve_none, GFP_KERNEL) ||
2071             !init_fifo(&free_inc,       free_inc_nr, GFP_KERNEL) ||
2072             !init_heap(&alloc_heap,     ALLOC_SCAN_BATCH(ca) << 1, GFP_KERNEL))
2073                 goto err;
2074
2075         buckets->first_bucket   = ca->mi.first_bucket;
2076         buckets->nbuckets       = nbuckets;
2077
2078         bch2_copygc_stop(c);
2079
2080         if (resize) {
2081                 down_write(&c->gc_lock);
2082                 down_write(&ca->bucket_lock);
2083                 percpu_down_write(&c->mark_lock);
2084         }
2085
2086         old_buckets = bucket_array(ca);
2087
2088         if (resize) {
2089                 size_t n = min(buckets->nbuckets, old_buckets->nbuckets);
2090
2091                 memcpy(buckets->b,
2092                        old_buckets->b,
2093                        n * sizeof(struct bucket));
2094                 memcpy(buckets_nouse,
2095                        ca->buckets_nouse,
2096                        BITS_TO_LONGS(n) * sizeof(unsigned long));
2097         }
2098
2099         rcu_assign_pointer(ca->buckets[0], buckets);
2100         buckets = old_buckets;
2101
2102         swap(ca->buckets_nouse, buckets_nouse);
2103
2104         if (resize) {
2105                 percpu_up_write(&c->mark_lock);
2106                 up_write(&c->gc_lock);
2107         }
2108
2109         spin_lock(&c->freelist_lock);
2110         for (i = 0; i < RESERVE_NR; i++) {
2111                 fifo_move(&free[i], &ca->free[i]);
2112                 swap(ca->free[i], free[i]);
2113         }
2114         fifo_move(&free_inc, &ca->free_inc);
2115         swap(ca->free_inc, free_inc);
2116         spin_unlock(&c->freelist_lock);
2117
2118         /* with gc lock held, alloc_heap can't be in use: */
2119         swap(ca->alloc_heap, alloc_heap);
2120
2121         nbuckets = ca->mi.nbuckets;
2122
2123         if (resize)
2124                 up_write(&ca->bucket_lock);
2125
2126         ret = 0;
2127 err:
2128         free_heap(&alloc_heap);
2129         free_fifo(&free_inc);
2130         for (i = 0; i < RESERVE_NR; i++)
2131                 free_fifo(&free[i]);
2132         kvpfree(buckets_nouse,
2133                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2134         if (buckets)
2135                 call_rcu(&old_buckets->rcu, buckets_free_rcu);
2136
2137         return ret;
2138 }
2139
2140 void bch2_dev_buckets_free(struct bch_dev *ca)
2141 {
2142         unsigned i;
2143
2144         free_heap(&ca->alloc_heap);
2145         free_fifo(&ca->free_inc);
2146         for (i = 0; i < RESERVE_NR; i++)
2147                 free_fifo(&ca->free[i]);
2148         kvpfree(ca->buckets_nouse,
2149                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2150         kvpfree(rcu_dereference_protected(ca->buckets[0], 1),
2151                 sizeof(struct bucket_array) +
2152                 ca->mi.nbuckets * sizeof(struct bucket));
2153
2154         for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
2155                 free_percpu(ca->usage[i]);
2156         kfree(ca->usage_base);
2157 }
2158
2159 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2160 {
2161         unsigned i;
2162
2163         ca->usage_base = kzalloc(sizeof(struct bch_dev_usage), GFP_KERNEL);
2164         if (!ca->usage_base)
2165                 return -ENOMEM;
2166
2167         for (i = 0; i < ARRAY_SIZE(ca->usage); i++) {
2168                 ca->usage[i] = alloc_percpu(struct bch_dev_usage);
2169                 if (!ca->usage[i])
2170                         return -ENOMEM;
2171         }
2172
2173         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);;
2174 }