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 "inode.h"
  17 #include "movinggc.h"
  18 #include "recovery.h"
  19 #include "reflink.h"
  20 #include "replicas.h"
  21 #include "subvolume.h"
  22
  23 #include <linux/preempt.h>
  24 #include <trace/events/bcachefs.h>
  25
  26 static inline void fs_usage_data_type_to_base(struct bch_fs_usage *fs_usage,
  27                                               enum bch_data_type data_type,
  28                                               s64 sectors)
  29 {
  30         switch (data_type) {
  31         case BCH_DATA_btree:
  32                 fs_usage->btree         += sectors;
  33                 break;
  34         case BCH_DATA_user:
  35         case BCH_DATA_parity:
  36                 fs_usage->data          += sectors;
  37                 break;
  38         case BCH_DATA_cached:
  39                 fs_usage->cached        += sectors;
  40                 break;
  41         default:
  42                 break;
  43         }
  44 }
  45
  46 /*
  47  * Clear journal_seq_valid for buckets for which it's not needed, to prevent
  48  * wraparound:
  49  */
  50 void bch2_bucket_seq_cleanup(struct bch_fs *c)
  51 {
  52         u64 journal_seq = atomic64_read(&c->journal.seq);
  53         u16 last_seq_ondisk = c->journal.flushed_seq_ondisk;
  54         struct bch_dev *ca;
  55         struct bucket_array *buckets;
  56         struct bucket *g;
  57         struct bucket_mark m;
  58         unsigned i;
  59
  60         if (journal_seq - c->last_bucket_seq_cleanup <
  61             (1U << (BUCKET_JOURNAL_SEQ_BITS - 2)))
  62                 return;
  63
  64         c->last_bucket_seq_cleanup = journal_seq;
  65
  66         for_each_member_device(ca, c, i) {
  67                 down_read(&ca->bucket_lock);
  68                 buckets = bucket_array(ca);
  69
  70                 for_each_bucket(g, buckets) {
  71                         bucket_cmpxchg(g, m, ({
  72                                 if (!m.journal_seq_valid ||
  73                                     bucket_needs_journal_commit(m, last_seq_ondisk))
  74                                         break;
  75
  76                                 m.journal_seq_valid = 0;
  77                         }));
  78                 }
  79                 up_read(&ca->bucket_lock);
  80         }
  81 }
  82
  83 void bch2_fs_usage_initialize(struct bch_fs *c)
  84 {
  85         struct bch_fs_usage *usage;
  86         struct bch_dev *ca;
  87         unsigned i;
  88
  89         percpu_down_write(&c->mark_lock);
  90         usage = c->usage_base;
  91
  92         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
  93                 bch2_fs_usage_acc_to_base(c, i);
  94
  95         for (i = 0; i < BCH_REPLICAS_MAX; i++)
  96                 usage->reserved += usage->persistent_reserved[i];
  97
  98         for (i = 0; i < c->replicas.nr; i++) {
  99                 struct bch_replicas_entry *e =
 100                         cpu_replicas_entry(&c->replicas, i);
 101
 102                 fs_usage_data_type_to_base(usage, e->data_type, usage->replicas[i]);
 103         }
 104
 105         for_each_member_device(ca, c, i) {
 106                 struct bch_dev_usage dev = bch2_dev_usage_read(ca);
 107
 108                 usage->hidden += (dev.d[BCH_DATA_sb].buckets +
 109                                   dev.d[BCH_DATA_journal].buckets) *
 110                         ca->mi.bucket_size;
 111         }
 112
 113         percpu_up_write(&c->mark_lock);
 114 }
 115
 116 static inline struct bch_dev_usage *dev_usage_ptr(struct bch_dev *ca,
 117                                                   unsigned journal_seq,
 118                                                   bool gc)
 119 {
 120         BUG_ON(!gc && !journal_seq);
 121
 122         return this_cpu_ptr(gc
 123                             ? ca->usage_gc
 124                             : ca->usage[journal_seq & JOURNAL_BUF_MASK]);
 125 }
 126
 127 struct bch_dev_usage bch2_dev_usage_read(struct bch_dev *ca)
 128 {
 129         struct bch_fs *c = ca->fs;
 130         struct bch_dev_usage ret;
 131         unsigned seq, i, u64s = dev_usage_u64s();
 132
 133         do {
 134                 seq = read_seqcount_begin(&c->usage_lock);
 135                 memcpy(&ret, ca->usage_base, u64s * sizeof(u64));
 136                 for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
 137                         acc_u64s_percpu((u64 *) &ret, (u64 __percpu *) ca->usage[i], u64s);
 138         } while (read_seqcount_retry(&c->usage_lock, seq));
 139
 140         return ret;
 141 }
 142
 143 static inline struct bch_fs_usage *fs_usage_ptr(struct bch_fs *c,
 144                                                 unsigned journal_seq,
 145                                                 bool gc)
 146 {
 147         percpu_rwsem_assert_held(&c->mark_lock);
 148         BUG_ON(!gc && !journal_seq);
 149
 150         return this_cpu_ptr(gc
 151                             ? c->usage_gc
 152                             : c->usage[journal_seq & JOURNAL_BUF_MASK]);
 153 }
 154
 155 u64 bch2_fs_usage_read_one(struct bch_fs *c, u64 *v)
 156 {
 157         ssize_t offset = v - (u64 *) c->usage_base;
 158         unsigned i, seq;
 159         u64 ret;
 160
 161         BUG_ON(offset < 0 || offset >= fs_usage_u64s(c));
 162         percpu_rwsem_assert_held(&c->mark_lock);
 163
 164         do {
 165                 seq = read_seqcount_begin(&c->usage_lock);
 166                 ret = *v;
 167
 168                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 169                         ret += percpu_u64_get((u64 __percpu *) c->usage[i] + offset);
 170         } while (read_seqcount_retry(&c->usage_lock, seq));
 171
 172         return ret;
 173 }
 174
 175 struct bch_fs_usage_online *bch2_fs_usage_read(struct bch_fs *c)
 176 {
 177         struct bch_fs_usage_online *ret;
 178         unsigned seq, i, u64s;
 179
 180         percpu_down_read(&c->mark_lock);
 181
 182         ret = kmalloc(sizeof(struct bch_fs_usage_online) +
 183                       sizeof(u64) * c->replicas.nr, GFP_NOFS);
 184         if (unlikely(!ret)) {
 185                 percpu_up_read(&c->mark_lock);
 186                 return NULL;
 187         }
 188
 189         ret->online_reserved = percpu_u64_get(c->online_reserved);
 190
 191         u64s = fs_usage_u64s(c);
 192         do {
 193                 seq = read_seqcount_begin(&c->usage_lock);
 194                 memcpy(&ret->u, c->usage_base, u64s * sizeof(u64));
 195                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 196                         acc_u64s_percpu((u64 *) &ret->u, (u64 __percpu *) c->usage[i], u64s);
 197         } while (read_seqcount_retry(&c->usage_lock, seq));
 198
 199         return ret;
 200 }
 201
 202 void bch2_fs_usage_acc_to_base(struct bch_fs *c, unsigned idx)
 203 {
 204         struct bch_dev *ca;
 205         unsigned i, u64s = fs_usage_u64s(c);
 206
 207         BUG_ON(idx >= ARRAY_SIZE(c->usage));
 208
 209         preempt_disable();
 210         write_seqcount_begin(&c->usage_lock);
 211
 212         acc_u64s_percpu((u64 *) c->usage_base,
 213                         (u64 __percpu *) c->usage[idx], u64s);
 214         percpu_memset(c->usage[idx], 0, u64s * sizeof(u64));
 215
 216         rcu_read_lock();
 217         for_each_member_device_rcu(ca, c, i, NULL) {
 218                 u64s = dev_usage_u64s();
 219
 220                 acc_u64s_percpu((u64 *) ca->usage_base,
 221                                 (u64 __percpu *) ca->usage[idx], u64s);
 222                 percpu_memset(ca->usage[idx], 0, u64s * sizeof(u64));
 223         }
 224         rcu_read_unlock();
 225
 226         write_seqcount_end(&c->usage_lock);
 227         preempt_enable();
 228 }
 229
 230 void bch2_fs_usage_to_text(struct printbuf *out,
 231                            struct bch_fs *c,
 232                            struct bch_fs_usage_online *fs_usage)
 233 {
 234         unsigned i;
 235
 236         pr_buf(out, "capacity:\t\t\t%llu\n", c->capacity);
 237
 238         pr_buf(out, "hidden:\t\t\t\t%llu\n",
 239                fs_usage->u.hidden);
 240         pr_buf(out, "data:\t\t\t\t%llu\n",
 241                fs_usage->u.data);
 242         pr_buf(out, "cached:\t\t\t\t%llu\n",
 243                fs_usage->u.cached);
 244         pr_buf(out, "reserved:\t\t\t%llu\n",
 245                fs_usage->u.reserved);
 246         pr_buf(out, "nr_inodes:\t\t\t%llu\n",
 247                fs_usage->u.nr_inodes);
 248         pr_buf(out, "online reserved:\t\t%llu\n",
 249                fs_usage->online_reserved);
 250
 251         for (i = 0;
 252              i < ARRAY_SIZE(fs_usage->u.persistent_reserved);
 253              i++) {
 254                 pr_buf(out, "%u replicas:\n", i + 1);
 255                 pr_buf(out, "\treserved:\t\t%llu\n",
 256                        fs_usage->u.persistent_reserved[i]);
 257         }
 258
 259         for (i = 0; i < c->replicas.nr; i++) {
 260                 struct bch_replicas_entry *e =
 261                         cpu_replicas_entry(&c->replicas, i);
 262
 263                 pr_buf(out, "\t");
 264                 bch2_replicas_entry_to_text(out, e);
 265                 pr_buf(out, ":\t%llu\n", fs_usage->u.replicas[i]);
 266         }
 267 }
 268
 269 static u64 reserve_factor(u64 r)
 270 {
 271         return r + (round_up(r, (1 << RESERVE_FACTOR)) >> RESERVE_FACTOR);
 272 }
 273
 274 u64 bch2_fs_sectors_used(struct bch_fs *c, struct bch_fs_usage_online *fs_usage)
 275 {
 276         return min(fs_usage->u.hidden +
 277                    fs_usage->u.btree +
 278                    fs_usage->u.data +
 279                    reserve_factor(fs_usage->u.reserved +
 280                                   fs_usage->online_reserved),
 281                    c->capacity);
 282 }
 283
 284 static struct bch_fs_usage_short
 285 __bch2_fs_usage_read_short(struct bch_fs *c)
 286 {
 287         struct bch_fs_usage_short ret;
 288         u64 data, reserved;
 289
 290         ret.capacity = c->capacity -
 291                 bch2_fs_usage_read_one(c, &c->usage_base->hidden);
 292
 293         data            = bch2_fs_usage_read_one(c, &c->usage_base->data) +
 294                 bch2_fs_usage_read_one(c, &c->usage_base->btree);
 295         reserved        = bch2_fs_usage_read_one(c, &c->usage_base->reserved) +
 296                 percpu_u64_get(c->online_reserved);
 297
 298         ret.used        = min(ret.capacity, data + reserve_factor(reserved));
 299         ret.free        = ret.capacity - ret.used;
 300
 301         ret.nr_inodes   = bch2_fs_usage_read_one(c, &c->usage_base->nr_inodes);
 302
 303         return ret;
 304 }
 305
 306 struct bch_fs_usage_short
 307 bch2_fs_usage_read_short(struct bch_fs *c)
 308 {
 309         struct bch_fs_usage_short ret;
 310
 311         percpu_down_read(&c->mark_lock);
 312         ret = __bch2_fs_usage_read_short(c);
 313         percpu_up_read(&c->mark_lock);
 314
 315         return ret;
 316 }
 317
 318 static inline int is_unavailable_bucket(struct bucket_mark m)
 319 {
 320         return !is_available_bucket(m);
 321 }
 322
 323 static inline int bucket_sectors_fragmented(struct bch_dev *ca,
 324                                             struct bucket_mark m)
 325 {
 326         return bucket_sectors_used(m)
 327                 ? max(0, (int) ca->mi.bucket_size - (int) bucket_sectors_used(m))
 328                 : 0;
 329 }
 330
 331 static inline int is_stripe_data_bucket(struct bucket_mark m)
 332 {
 333         return m.stripe && m.data_type != BCH_DATA_parity;
 334 }
 335
 336 static inline enum bch_data_type bucket_type(struct bucket_mark m)
 337 {
 338         return m.cached_sectors && !m.dirty_sectors
 339                 ? BCH_DATA_cached
 340                 : m.data_type;
 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         preempt_disable();
 362         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 363         u = dev_usage_ptr(ca, journal_seq, gc);
 364
 365         if (bucket_type(old))
 366                 account_bucket(fs_usage, u, bucket_type(old),
 367                                -1, -ca->mi.bucket_size);
 368
 369         if (bucket_type(new))
 370                 account_bucket(fs_usage, u, bucket_type(new),
 371                                1, ca->mi.bucket_size);
 372
 373         u->buckets_ec += (int) new.stripe - (int) old.stripe;
 374         u->buckets_unavailable +=
 375                 is_unavailable_bucket(new) - is_unavailable_bucket(old);
 376
 377         u->d[old.data_type].sectors -= old.dirty_sectors;
 378         u->d[new.data_type].sectors += new.dirty_sectors;
 379         u->d[BCH_DATA_cached].sectors +=
 380                 (int) new.cached_sectors - (int) old.cached_sectors;
 381
 382         u->d[old.data_type].fragmented -= bucket_sectors_fragmented(ca, old);
 383         u->d[new.data_type].fragmented += bucket_sectors_fragmented(ca, new);
 384
 385         preempt_enable();
 386
 387         if (!is_available_bucket(old) && is_available_bucket(new))
 388                 bch2_wake_allocator(ca);
 389 }
 390
 391 static inline int __update_replicas(struct bch_fs *c,
 392                                     struct bch_fs_usage *fs_usage,
 393                                     struct bch_replicas_entry *r,
 394                                     s64 sectors)
 395 {
 396         int idx = bch2_replicas_entry_idx(c, r);
 397
 398         if (idx < 0)
 399                 return -1;
 400
 401         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 402         fs_usage->replicas[idx]         += sectors;
 403         return 0;
 404 }
 405
 406 static inline int update_replicas(struct bch_fs *c, struct bkey_s_c k,
 407                         struct bch_replicas_entry *r, s64 sectors,
 408                         unsigned journal_seq, bool gc)
 409 {
 410         struct bch_fs_usage __percpu *fs_usage;
 411         int idx, ret = 0;
 412         char buf[200];
 413
 414         percpu_down_read(&c->mark_lock);
 415
 416         idx = bch2_replicas_entry_idx(c, r);
 417         if (idx < 0 &&
 418             (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
 419              fsck_err(c, "no replicas entry\n"
 420                       "  while marking %s",
 421                       (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf)))) {
 422                 percpu_up_read(&c->mark_lock);
 423                 ret = bch2_mark_replicas(c, r);
 424                 if (ret)
 425                         return ret;
 426
 427                 percpu_down_read(&c->mark_lock);
 428                 idx = bch2_replicas_entry_idx(c, r);
 429         }
 430         if (idx < 0) {
 431                 ret = -1;
 432                 goto err;
 433         }
 434
 435         preempt_disable();
 436         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 437         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 438         fs_usage->replicas[idx]         += sectors;
 439         preempt_enable();
 440 err:
 441 fsck_err:
 442         percpu_up_read(&c->mark_lock);
 443         return ret;
 444 }
 445
 446 static inline int update_cached_sectors(struct bch_fs *c,
 447                         struct bkey_s_c k,
 448                         unsigned dev, s64 sectors,
 449                         unsigned journal_seq, bool gc)
 450 {
 451         struct bch_replicas_padded r;
 452
 453         bch2_replicas_entry_cached(&r.e, dev);
 454
 455         return update_replicas(c, k, &r.e, sectors, journal_seq, gc);
 456 }
 457
 458 static struct replicas_delta_list *
 459 replicas_deltas_realloc(struct btree_trans *trans, unsigned more)
 460 {
 461         struct replicas_delta_list *d = trans->fs_usage_deltas;
 462         unsigned new_size = d ? (d->size + more) * 2 : 128;
 463         unsigned alloc_size = sizeof(*d) + new_size;
 464
 465         WARN_ON_ONCE(alloc_size > REPLICAS_DELTA_LIST_MAX);
 466
 467         if (!d || d->used + more > d->size) {
 468                 d = krealloc(d, alloc_size, GFP_NOIO|__GFP_ZERO);
 469
 470                 BUG_ON(!d && alloc_size > REPLICAS_DELTA_LIST_MAX);
 471
 472                 if (!d) {
 473                         d = mempool_alloc(&trans->c->replicas_delta_pool, GFP_NOIO);
 474                         memset(d, 0, REPLICAS_DELTA_LIST_MAX);
 475
 476                         if (trans->fs_usage_deltas)
 477                                 memcpy(d, trans->fs_usage_deltas,
 478                                        trans->fs_usage_deltas->size + sizeof(*d));
 479
 480                         new_size = REPLICAS_DELTA_LIST_MAX - sizeof(*d);
 481                         kfree(trans->fs_usage_deltas);
 482                 }
 483
 484                 d->size = new_size;
 485                 trans->fs_usage_deltas = d;
 486         }
 487         return d;
 488 }
 489
 490 static inline void update_replicas_list(struct btree_trans *trans,
 491                                         struct bch_replicas_entry *r,
 492                                         s64 sectors)
 493 {
 494         struct replicas_delta_list *d;
 495         struct replicas_delta *n;
 496         unsigned b;
 497
 498         if (!sectors)
 499                 return;
 500
 501         b = replicas_entry_bytes(r) + 8;
 502         d = replicas_deltas_realloc(trans, b);
 503
 504         n = (void *) d->d + d->used;
 505         n->delta = sectors;
 506         memcpy(&n->r, r, replicas_entry_bytes(r));
 507         bch2_replicas_entry_sort(&n->r);
 508         d->used += b;
 509 }
 510
 511 static inline void update_cached_sectors_list(struct btree_trans *trans,
 512                                               unsigned dev, s64 sectors)
 513 {
 514         struct bch_replicas_padded r;
 515
 516         bch2_replicas_entry_cached(&r.e, dev);
 517
 518         update_replicas_list(trans, &r.e, sectors);
 519 }
 520
 521 void bch2_mark_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 522                             size_t b, bool owned_by_allocator)
 523 {
 524         struct bucket *g = bucket(ca, b);
 525         struct bucket_mark old, new;
 526
 527         old = bucket_cmpxchg(g, new, ({
 528                 new.owned_by_allocator  = owned_by_allocator;
 529         }));
 530
 531         BUG_ON(owned_by_allocator == old.owned_by_allocator);
 532 }
 533
 534 static inline u8 bkey_alloc_gen(struct bkey_s_c k)
 535 {
 536         switch (k.k->type) {
 537         case KEY_TYPE_alloc:
 538                 return bkey_s_c_to_alloc(k).v->gen;
 539         case KEY_TYPE_alloc_v2:
 540                 return bkey_s_c_to_alloc_v2(k).v->gen;
 541         case KEY_TYPE_alloc_v3:
 542                 return bkey_s_c_to_alloc_v3(k).v->gen;
 543         default:
 544                 return 0;
 545         }
 546 }
 547
 548 static int bch2_mark_alloc(struct btree_trans *trans,
 549                            struct bkey_s_c old, struct bkey_s_c new,
 550                            unsigned flags)
 551 {
 552         bool gc = flags & BTREE_TRIGGER_GC;
 553         u64 journal_seq = trans->journal_res.seq;
 554         struct bch_fs *c = trans->c;
 555         struct bkey_alloc_unpacked old_u = bch2_alloc_unpack(old);
 556         struct bkey_alloc_unpacked new_u = bch2_alloc_unpack(new);
 557         struct bch_dev *ca;
 558         struct bucket *g;
 559         struct bucket_mark old_m, m;
 560         int ret = 0;
 561
 562         /*
 563          * alloc btree is read in by bch2_alloc_read, not gc:
 564          */
 565         if ((flags & BTREE_TRIGGER_GC) &&
 566             !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
 567                 return 0;
 568
 569         if ((flags & BTREE_TRIGGER_INSERT) &&
 570             !old_u.data_type != !new_u.data_type &&
 571             new.k->type == KEY_TYPE_alloc_v3) {
 572                 struct bch_alloc_v3 *v = (struct bch_alloc_v3 *) new.v;
 573
 574                 BUG_ON(!journal_seq);
 575
 576                 /*
 577                  * If the btree updates referring to a bucket weren't flushed
 578                  * before the bucket became empty again, then the we don't have
 579                  * to wait on a journal flush before we can reuse the bucket:
 580                  */
 581                 v->journal_seq = !new_u.data_type &&
 582                         bch2_journal_noflush_seq(&c->journal, journal_seq)
 583                         ? 0 : cpu_to_le64(journal_seq);
 584         }
 585
 586         ca = bch_dev_bkey_exists(c, new.k->p.inode);
 587
 588         if (new.k->p.offset >= ca->mi.nbuckets)
 589                 return 0;
 590
 591         percpu_down_read(&c->mark_lock);
 592         if (!gc && new_u.gen != bkey_alloc_gen(old))
 593                 *bucket_gen(ca, new.k->p.offset) = new_u.gen;
 594
 595         g = __bucket(ca, new.k->p.offset, gc);
 596
 597         old_m = bucket_cmpxchg(g, m, ({
 598                 m.gen                   = new_u.gen;
 599                 m.data_type             = new_u.data_type;
 600                 m.dirty_sectors         = new_u.dirty_sectors;
 601                 m.cached_sectors        = new_u.cached_sectors;
 602                 m.stripe                = new_u.stripe != 0;
 603
 604                 if (journal_seq) {
 605                         m.journal_seq_valid     = 1;
 606                         m.journal_seq           = journal_seq;
 607                 }
 608         }));
 609
 610         bch2_dev_usage_update(c, ca, old_m, m, journal_seq, gc);
 611
 612         g->io_time[READ]        = new_u.read_time;
 613         g->io_time[WRITE]       = new_u.write_time;
 614         g->oldest_gen           = new_u.oldest_gen;
 615         g->gen_valid            = 1;
 616         g->stripe               = new_u.stripe;
 617         g->stripe_redundancy    = new_u.stripe_redundancy;
 618         percpu_up_read(&c->mark_lock);
 619
 620         /*
 621          * need to know if we're getting called from the invalidate path or
 622          * not:
 623          */
 624
 625         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 626             old_m.cached_sectors) {
 627                 ret = update_cached_sectors(c, new, ca->dev_idx,
 628                                             -old_m.cached_sectors,
 629                                             journal_seq, gc);
 630                 if (ret) {
 631                         bch2_fs_fatal_error(c, "bch2_mark_alloc(): no replicas entry while updating cached sectors");
 632                         return ret;
 633                 }
 634
 635                 trace_invalidate(ca, bucket_to_sector(ca, new.k->p.offset),
 636                                  old_m.cached_sectors);
 637         }
 638
 639         return 0;
 640 }
 641
 642 #define checked_add(a, b)                                       \
 643 ({                                                              \
 644         unsigned _res = (unsigned) (a) + (b);                   \
 645         bool overflow = _res > U16_MAX;                         \
 646         if (overflow)                                           \
 647                 _res = U16_MAX;                                 \
 648         (a) = _res;                                             \
 649         overflow;                                               \
 650 })
 651
 652 void bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 653                                size_t b, enum bch_data_type data_type,
 654                                unsigned sectors, struct gc_pos pos,
 655                                unsigned flags)
 656 {
 657         struct bucket *g;
 658         struct bucket_mark old, new;
 659         bool overflow;
 660
 661         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 662         BUG_ON(data_type != BCH_DATA_sb &&
 663                data_type != BCH_DATA_journal);
 664
 665         /*
 666          * Backup superblock might be past the end of our normal usable space:
 667          */
 668         if (b >= ca->mi.nbuckets)
 669                 return;
 670
 671         percpu_down_read(&c->mark_lock);
 672         g = gc_bucket(ca, b);
 673         old = bucket_cmpxchg(g, new, ({
 674                 new.data_type   = data_type;
 675                 overflow = checked_add(new.dirty_sectors, sectors);
 676         }));
 677
 678         bch2_fs_inconsistent_on(old.data_type &&
 679                                 old.data_type != data_type, c,
 680                 "different types of data in same bucket: %s, %s",
 681                 bch2_data_types[old.data_type],
 682                 bch2_data_types[data_type]);
 683
 684         bch2_fs_inconsistent_on(overflow, c,
 685                 "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > U16_MAX",
 686                 ca->dev_idx, b, new.gen,
 687                 bch2_data_types[old.data_type ?: data_type],
 688                 old.dirty_sectors, sectors);
 689
 690         bch2_dev_usage_update(c, ca, old, new, 0, true);
 691         percpu_up_read(&c->mark_lock);
 692 }
 693
 694 static s64 ptr_disk_sectors(s64 sectors, struct extent_ptr_decoded p)
 695 {
 696         EBUG_ON(sectors < 0);
 697
 698         return p.crc.compression_type &&
 699                 p.crc.compression_type != BCH_COMPRESSION_TYPE_incompressible
 700                 ? DIV_ROUND_UP_ULL(sectors * p.crc.compressed_size,
 701                                p.crc.uncompressed_size)
 702                 : sectors;
 703 }
 704
 705 static int check_bucket_ref(struct bch_fs *c,
 706                             struct bkey_s_c k,
 707                             const struct bch_extent_ptr *ptr,
 708                             s64 sectors, enum bch_data_type ptr_data_type,
 709                             u8 bucket_gen, u8 bucket_data_type,
 710                             u16 dirty_sectors, u16 cached_sectors)
 711 {
 712         size_t bucket_nr = PTR_BUCKET_NR(bch_dev_bkey_exists(c, ptr->dev), ptr);
 713         u16 bucket_sectors = !ptr->cached
 714                 ? dirty_sectors
 715                 : cached_sectors;
 716         char buf[200];
 717
 718         if (gen_after(ptr->gen, bucket_gen)) {
 719                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 720                         "bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\n"
 721                         "while marking %s",
 722                         ptr->dev, bucket_nr, bucket_gen,
 723                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 724                         ptr->gen,
 725                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 726                 return -EIO;
 727         }
 728
 729         if (gen_cmp(bucket_gen, ptr->gen) > BUCKET_GC_GEN_MAX) {
 730                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 731                         "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 732                         "while marking %s",
 733                         ptr->dev, bucket_nr, bucket_gen,
 734                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 735                         ptr->gen,
 736                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 737                 return -EIO;
 738         }
 739
 740         if (bucket_gen != ptr->gen && !ptr->cached) {
 741                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 742                         "bucket %u:%zu gen %u data type %s: stale dirty ptr (gen %u)\n"
 743                         "while marking %s",
 744                         ptr->dev, bucket_nr, bucket_gen,
 745                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 746                         ptr->gen,
 747                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 748                 return -EIO;
 749         }
 750
 751         if (bucket_gen != ptr->gen)
 752                 return 1;
 753
 754         if (bucket_data_type && ptr_data_type &&
 755             bucket_data_type != ptr_data_type) {
 756                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 757                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 758                         "while marking %s",
 759                         ptr->dev, bucket_nr, bucket_gen,
 760                         bch2_data_types[bucket_data_type],
 761                         bch2_data_types[ptr_data_type],
 762                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 763                 return -EIO;
 764         }
 765
 766         if ((unsigned) (bucket_sectors + sectors) > U16_MAX) {
 767                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 768                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U16_MAX\n"
 769                         "while marking %s",
 770                         ptr->dev, bucket_nr, bucket_gen,
 771                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 772                         bucket_sectors, sectors,
 773                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 774                 return -EIO;
 775         }
 776
 777         return 0;
 778 }
 779
 780 static int mark_stripe_bucket(struct btree_trans *trans,
 781                               struct bkey_s_c k,
 782                               unsigned ptr_idx,
 783                               u64 journal_seq, unsigned flags)
 784 {
 785         struct bch_fs *c = trans->c;
 786         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 787         unsigned nr_data = s->nr_blocks - s->nr_redundant;
 788         bool parity = ptr_idx >= nr_data;
 789         enum bch_data_type data_type = parity ? BCH_DATA_parity : 0;
 790         s64 sectors = parity ? le16_to_cpu(s->sectors) : 0;
 791         const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
 792         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 793         struct bucket *g;
 794         struct bucket_mark new, old;
 795         char buf[200];
 796         int ret = 0;
 797
 798         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 799
 800         /* * XXX doesn't handle deletion */
 801
 802         percpu_down_read(&c->mark_lock);
 803         g = PTR_GC_BUCKET(ca, ptr);
 804
 805         if (g->mark.dirty_sectors ||
 806             (g->stripe && g->stripe != k.k->p.offset)) {
 807                 bch2_fs_inconsistent(c,
 808                               "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
 809                               ptr->dev, PTR_BUCKET_NR(ca, ptr), g->mark.gen,
 810                               (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 811                 ret = -EINVAL;
 812                 goto err;
 813         }
 814
 815         old = bucket_cmpxchg(g, new, ({
 816                 ret = check_bucket_ref(c, k, ptr, sectors, data_type,
 817                                        new.gen, new.data_type,
 818                                        new.dirty_sectors, new.cached_sectors);
 819                 if (ret)
 820                         goto err;
 821
 822                 new.dirty_sectors += sectors;
 823                 if (data_type)
 824                         new.data_type           = data_type;
 825
 826                 if (journal_seq) {
 827                         new.journal_seq_valid   = 1;
 828                         new.journal_seq         = journal_seq;
 829                 }
 830
 831                 new.stripe = true;
 832         }));
 833
 834         g->stripe               = k.k->p.offset;
 835         g->stripe_redundancy    = s->nr_redundant;
 836
 837         bch2_dev_usage_update(c, ca, old, new, journal_seq, true);
 838 err:
 839         percpu_up_read(&c->mark_lock);
 840
 841         return 0;
 842 }
 843
 844 static int __mark_pointer(struct btree_trans *trans,
 845                           struct bkey_s_c k,
 846                           const struct bch_extent_ptr *ptr,
 847                           s64 sectors, enum bch_data_type ptr_data_type,
 848                           u8 bucket_gen, u8 *bucket_data_type,
 849                           u16 *dirty_sectors, u16 *cached_sectors)
 850 {
 851         u16 *dst_sectors = !ptr->cached
 852                 ? dirty_sectors
 853                 : cached_sectors;
 854         int ret = check_bucket_ref(trans->c, k, ptr, sectors, ptr_data_type,
 855                                    bucket_gen, *bucket_data_type,
 856                                    *dirty_sectors, *cached_sectors);
 857
 858         if (ret)
 859                 return ret;
 860
 861         *dst_sectors += sectors;
 862         *bucket_data_type = *dirty_sectors || *cached_sectors
 863                 ? ptr_data_type : 0;
 864         return 0;
 865 }
 866
 867 static int bch2_mark_pointer(struct btree_trans *trans,
 868                              struct bkey_s_c k,
 869                              struct extent_ptr_decoded p,
 870                              s64 sectors, enum bch_data_type data_type,
 871                              unsigned flags)
 872 {
 873         u64 journal_seq = trans->journal_res.seq;
 874         struct bch_fs *c = trans->c;
 875         struct bucket_mark old, new;
 876         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 877         struct bucket *g;
 878         u8 bucket_data_type;
 879         u64 v;
 880         int ret = 0;
 881
 882         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 883
 884         percpu_down_read(&c->mark_lock);
 885         g = PTR_GC_BUCKET(ca, &p.ptr);
 886
 887         v = atomic64_read(&g->_mark.v);
 888         do {
 889                 new.v.counter = old.v.counter = v;
 890                 bucket_data_type = new.data_type;
 891
 892                 ret = __mark_pointer(trans, k, &p.ptr, sectors,
 893                                      data_type, new.gen,
 894                                      &bucket_data_type,
 895                                      &new.dirty_sectors,
 896                                      &new.cached_sectors);
 897                 if (ret)
 898                         goto err;
 899
 900                 new.data_type = bucket_data_type;
 901
 902                 if (journal_seq) {
 903                         new.journal_seq_valid = 1;
 904                         new.journal_seq = journal_seq;
 905                 }
 906
 907                 if (flags & BTREE_TRIGGER_NOATOMIC) {
 908                         g->_mark = new;
 909                         break;
 910                 }
 911         } while ((v = atomic64_cmpxchg(&g->_mark.v,
 912                               old.v.counter,
 913                               new.v.counter)) != old.v.counter);
 914
 915         bch2_dev_usage_update(c, ca, old, new, journal_seq, true);
 916 err:
 917         percpu_up_read(&c->mark_lock);
 918
 919         return ret;
 920 }
 921
 922 static int bch2_mark_stripe_ptr(struct btree_trans *trans,
 923                                 struct bkey_s_c k,
 924                                 struct bch_extent_stripe_ptr p,
 925                                 enum bch_data_type data_type,
 926                                 s64 sectors,
 927                                 unsigned flags)
 928 {
 929         struct bch_fs *c = trans->c;
 930         struct bch_replicas_padded r;
 931         struct gc_stripe *m;
 932
 933         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 934
 935         m = genradix_ptr_alloc(&c->gc_stripes, p.idx, GFP_KERNEL);
 936         if (!m) {
 937                 bch_err(c, "error allocating memory for gc_stripes, idx %llu",
 938                         (u64) p.idx);
 939                 return -ENOMEM;
 940         }
 941
 942         spin_lock(&c->ec_stripes_heap_lock);
 943
 944         if (!m || !m->alive) {
 945                 spin_unlock(&c->ec_stripes_heap_lock);
 946                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
 947                                     (u64) p.idx);
 948                 bch2_inconsistent_error(c);
 949                 return -EIO;
 950         }
 951
 952         m->block_sectors[p.block] += sectors;
 953
 954         r = m->r;
 955         spin_unlock(&c->ec_stripes_heap_lock);
 956
 957         r.e.data_type = data_type;
 958         update_replicas(c, k, &r.e, sectors, trans->journal_res.seq, true);
 959
 960         return 0;
 961 }
 962
 963 static int bch2_mark_extent(struct btree_trans *trans,
 964                             struct bkey_s_c old, struct bkey_s_c new,
 965                             unsigned flags)
 966 {
 967         u64 journal_seq = trans->journal_res.seq;
 968         struct bch_fs *c = trans->c;
 969         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
 970         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 971         const union bch_extent_entry *entry;
 972         struct extent_ptr_decoded p;
 973         struct bch_replicas_padded r;
 974         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
 975                 ? BCH_DATA_btree
 976                 : BCH_DATA_user;
 977         s64 sectors = bkey_is_btree_ptr(k.k)
 978                 ? btree_sectors(c)
 979                 : k.k->size;
 980         s64 dirty_sectors = 0;
 981         bool stale;
 982         int ret;
 983
 984         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 985
 986         r.e.data_type   = data_type;
 987         r.e.nr_devs     = 0;
 988         r.e.nr_required = 1;
 989
 990         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 991                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
 992
 993                 if (flags & BTREE_TRIGGER_OVERWRITE)
 994                         disk_sectors = -disk_sectors;
 995
 996                 ret = bch2_mark_pointer(trans, k, p, disk_sectors,
 997                                         data_type, flags);
 998                 if (ret < 0)
 999                         return ret;
1000
1001                 stale = ret > 0;
1002
1003                 if (p.ptr.cached) {
1004                         if (!stale) {
1005                                 ret = update_cached_sectors(c, k, p.ptr.dev,
1006                                                 disk_sectors, journal_seq, true);
1007                                 if (ret) {
1008                                         bch2_fs_fatal_error(c, "bch2_mark_extent(): no replicas entry while updating cached sectors");
1009                                         return ret;
1010                                 }
1011                         }
1012                 } else if (!p.has_ec) {
1013                         dirty_sectors          += disk_sectors;
1014                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1015                 } else {
1016                         ret = bch2_mark_stripe_ptr(trans, k, p.ec, data_type,
1017                                         disk_sectors, flags);
1018                         if (ret)
1019                                 return ret;
1020
1021                         /*
1022                          * There may be other dirty pointers in this extent, but
1023                          * if so they're not required for mounting if we have an
1024                          * erasure coded pointer in this extent:
1025                          */
1026                         r.e.nr_required = 0;
1027                 }
1028         }
1029
1030         if (r.e.nr_devs) {
1031                 ret = update_replicas(c, k, &r.e, dirty_sectors, journal_seq, true);
1032                 if (ret) {
1033                         char buf[200];
1034
1035                         bch2_bkey_val_to_text(&PBUF(buf), c, k);
1036                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf);
1037                         return ret;
1038                 }
1039         }
1040
1041         return 0;
1042 }
1043
1044 static int bch2_mark_stripe(struct btree_trans *trans,
1045                             struct bkey_s_c old, struct bkey_s_c new,
1046                             unsigned flags)
1047 {
1048         bool gc = flags & BTREE_TRIGGER_GC;
1049         u64 journal_seq = trans->journal_res.seq;
1050         struct bch_fs *c = trans->c;
1051         u64 idx = new.k->p.offset;
1052         const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
1053                 ? bkey_s_c_to_stripe(old).v : NULL;
1054         const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
1055                 ? bkey_s_c_to_stripe(new).v : NULL;
1056         unsigned i;
1057         int ret;
1058
1059         BUG_ON(gc && old_s);
1060
1061         if (!gc) {
1062                 struct stripe *m = genradix_ptr(&c->stripes, idx);
1063
1064                 if (!m || (old_s && !m->alive)) {
1065                         char buf1[200], buf2[200];
1066
1067                         bch2_bkey_val_to_text(&PBUF(buf1), c, old);
1068                         bch2_bkey_val_to_text(&PBUF(buf2), c, new);
1069                         bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
1070                                             "old %s\n"
1071                                             "new %s", idx, buf1, buf2);
1072                         bch2_inconsistent_error(c);
1073                         return -1;
1074                 }
1075
1076                 if (!new_s) {
1077                         spin_lock(&c->ec_stripes_heap_lock);
1078                         bch2_stripes_heap_del(c, m, idx);
1079                         spin_unlock(&c->ec_stripes_heap_lock);
1080
1081                         memset(m, 0, sizeof(*m));
1082                 } else {
1083                         m->alive        = true;
1084                         m->sectors      = le16_to_cpu(new_s->sectors);
1085                         m->algorithm    = new_s->algorithm;
1086                         m->nr_blocks    = new_s->nr_blocks;
1087                         m->nr_redundant = new_s->nr_redundant;
1088                         m->blocks_nonempty = 0;
1089
1090                         for (i = 0; i < new_s->nr_blocks; i++)
1091                                 m->blocks_nonempty += !!stripe_blockcount_get(new_s, i);
1092
1093                         spin_lock(&c->ec_stripes_heap_lock);
1094                         bch2_stripes_heap_update(c, m, idx);
1095                         spin_unlock(&c->ec_stripes_heap_lock);
1096                 }
1097         } else {
1098                 struct gc_stripe *m =
1099                         genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
1100
1101                 if (!m) {
1102                         bch_err(c, "error allocating memory for gc_stripes, idx %llu",
1103                                 idx);
1104                         return -ENOMEM;
1105                 }
1106                 /*
1107                  * This will be wrong when we bring back runtime gc: we should
1108                  * be unmarking the old key and then marking the new key
1109                  */
1110                 m->alive        = true;
1111                 m->sectors      = le16_to_cpu(new_s->sectors);
1112                 m->nr_blocks    = new_s->nr_blocks;
1113                 m->nr_redundant = new_s->nr_redundant;
1114
1115                 for (i = 0; i < new_s->nr_blocks; i++)
1116                         m->ptrs[i] = new_s->ptrs[i];
1117
1118                 bch2_bkey_to_replicas(&m->r.e, new);
1119
1120                 /*
1121                  * gc recalculates this field from stripe ptr
1122                  * references:
1123                  */
1124                 memset(m->block_sectors, 0, sizeof(m->block_sectors));
1125
1126                 for (i = 0; i < new_s->nr_blocks; i++) {
1127                         ret = mark_stripe_bucket(trans, new, i, journal_seq, flags);
1128                         if (ret)
1129                                 return ret;
1130                 }
1131
1132                 ret = update_replicas(c, new, &m->r.e,
1133                                       ((s64) m->sectors * m->nr_redundant),
1134                                       journal_seq, gc);
1135                 if (ret) {
1136                         char buf[200];
1137
1138                         bch2_bkey_val_to_text(&PBUF(buf), c, new);
1139                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf);
1140                         return ret;
1141                 }
1142         }
1143
1144         return 0;
1145 }
1146
1147 static int bch2_mark_inode(struct btree_trans *trans,
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_v2 *v = (struct bch_inode_v2 *) new.v;
1157
1158                 BUG_ON(!journal_seq);
1159                 BUG_ON(new.k->type != KEY_TYPE_inode_v2);
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 bch2_mark_reservation(struct btree_trans *trans,
1179                                  struct bkey_s_c old, struct bkey_s_c new,
1180                                  unsigned flags)
1181 {
1182         struct bch_fs *c = trans->c;
1183         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1184         struct bch_fs_usage __percpu *fs_usage;
1185         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1186         s64 sectors = (s64) k.k->size;
1187
1188         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1189
1190         if (flags & BTREE_TRIGGER_OVERWRITE)
1191                 sectors = -sectors;
1192         sectors *= replicas;
1193
1194         percpu_down_read(&c->mark_lock);
1195         preempt_disable();
1196
1197         fs_usage = fs_usage_ptr(c, trans->journal_res.seq, flags & BTREE_TRIGGER_GC);
1198         replicas = clamp_t(unsigned, replicas, 1,
1199                            ARRAY_SIZE(fs_usage->persistent_reserved));
1200
1201         fs_usage->reserved                              += sectors;
1202         fs_usage->persistent_reserved[replicas - 1]     += sectors;
1203
1204         preempt_enable();
1205         percpu_up_read(&c->mark_lock);
1206
1207         return 0;
1208 }
1209
1210 static s64 __bch2_mark_reflink_p(struct bch_fs *c, struct bkey_s_c_reflink_p p,
1211                                  u64 *idx, unsigned flags, size_t r_idx)
1212 {
1213         struct reflink_gc *r;
1214         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1215         s64 ret = 0;
1216
1217         if (r_idx >= c->reflink_gc_nr)
1218                 goto not_found;
1219
1220         r = genradix_ptr(&c->reflink_gc_table, r_idx);
1221         if (*idx < r->offset - r->size)
1222                 goto not_found;
1223
1224         BUG_ON((s64) r->refcount + add < 0);
1225
1226         r->refcount += add;
1227         *idx = r->offset;
1228         return 0;
1229 not_found:
1230         *idx = U64_MAX;
1231         ret = -EIO;
1232
1233         /*
1234          * XXX: we're replacing the entire reflink pointer with an error
1235          * key, we should just be replacing the part that was missing:
1236          */
1237         if (fsck_err(c, "%llu:%llu len %u points to nonexistent indirect extent %llu",
1238                      p.k->p.inode, p.k->p.offset, p.k->size, *idx)) {
1239                 struct bkey_i_error new;
1240
1241                 bkey_init(&new.k);
1242                 new.k.type      = KEY_TYPE_error;
1243                 new.k.p         = p.k->p;
1244                 new.k.size      = p.k->size;
1245                 ret = bch2_journal_key_insert(c, BTREE_ID_extents, 0, &new.k_i);
1246         }
1247 fsck_err:
1248         return ret;
1249 }
1250
1251 static int bch2_mark_reflink_p(struct btree_trans *trans,
1252                                struct bkey_s_c old, struct bkey_s_c new,
1253                                unsigned flags)
1254 {
1255         struct bch_fs *c = trans->c;
1256         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1257         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1258         struct reflink_gc *ref;
1259         size_t l, r, m;
1260         u64 idx = le64_to_cpu(p.v->idx);
1261         u64 end = le64_to_cpu(p.v->idx) + p.k->size;
1262         int ret = 0;
1263
1264         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1265
1266         if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix) {
1267                 idx -= le32_to_cpu(p.v->front_pad);
1268                 end += le32_to_cpu(p.v->back_pad);
1269         }
1270
1271         l = 0;
1272         r = c->reflink_gc_nr;
1273         while (l < r) {
1274                 m = l + (r - l) / 2;
1275
1276                 ref = genradix_ptr(&c->reflink_gc_table, m);
1277                 if (ref->offset <= idx)
1278                         l = m + 1;
1279                 else
1280                         r = m;
1281         }
1282
1283         while (idx < end && !ret)
1284                 ret = __bch2_mark_reflink_p(c, p, &idx, flags, l++);
1285
1286         return ret;
1287 }
1288
1289 int bch2_mark_key(struct btree_trans *trans,
1290                   struct bkey_s_c old,
1291                   struct bkey_s_c new,
1292                   unsigned flags)
1293 {
1294         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1295
1296         switch (k.k->type) {
1297         case KEY_TYPE_alloc:
1298         case KEY_TYPE_alloc_v2:
1299         case KEY_TYPE_alloc_v3:
1300                 return bch2_mark_alloc(trans, old, new, flags);
1301         case KEY_TYPE_btree_ptr:
1302         case KEY_TYPE_btree_ptr_v2:
1303         case KEY_TYPE_extent:
1304         case KEY_TYPE_reflink_v:
1305                 return bch2_mark_extent(trans, old, new, flags);
1306         case KEY_TYPE_stripe:
1307                 return bch2_mark_stripe(trans, old, new, flags);
1308         case KEY_TYPE_inode:
1309         case KEY_TYPE_inode_v2:
1310                 return bch2_mark_inode(trans, old, new, flags);
1311         case KEY_TYPE_reservation:
1312                 return bch2_mark_reservation(trans, old, new, flags);
1313         case KEY_TYPE_reflink_p:
1314                 return bch2_mark_reflink_p(trans, old, new, flags);
1315         case KEY_TYPE_snapshot:
1316                 return bch2_mark_snapshot(trans, old, new, flags);
1317         default:
1318                 return 0;
1319         }
1320 }
1321
1322 int bch2_mark_update(struct btree_trans *trans, struct btree_path *path,
1323                      struct bkey_i *new, unsigned flags)
1324 {
1325         struct bkey             _deleted = KEY(0, 0, 0);
1326         struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
1327         struct bkey_s_c         old;
1328         struct bkey             unpacked;
1329         int ret;
1330
1331         _deleted.p = path->pos;
1332
1333         if (unlikely(flags & BTREE_TRIGGER_NORUN))
1334                 return 0;
1335
1336         if (!btree_node_type_needs_gc(path->btree_id))
1337                 return 0;
1338
1339         old = bch2_btree_path_peek_slot(path, &unpacked);
1340
1341         if (old.k->type == new->k.type &&
1342             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
1343                 ret   = bch2_mark_key(trans, old, bkey_i_to_s_c(new),
1344                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
1345         } else {
1346                 ret   = bch2_mark_key(trans, deleted, bkey_i_to_s_c(new),
1347                                 BTREE_TRIGGER_INSERT|flags) ?:
1348                         bch2_mark_key(trans, old, deleted,
1349                                 BTREE_TRIGGER_OVERWRITE|flags);
1350         }
1351
1352         return ret;
1353 }
1354
1355 static noinline __cold
1356 void fs_usage_apply_warn(struct btree_trans *trans,
1357                          unsigned disk_res_sectors,
1358                          s64 should_not_have_added)
1359 {
1360         struct bch_fs *c = trans->c;
1361         struct btree_insert_entry *i;
1362         char buf[200];
1363
1364         bch_err(c, "disk usage increased %lli more than %u sectors reserved",
1365                 should_not_have_added, disk_res_sectors);
1366
1367         trans_for_each_update(trans, i) {
1368                 pr_err("while inserting");
1369                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
1370                 pr_err("%s", buf);
1371                 pr_err("overlapping with");
1372
1373                 if (!i->cached) {
1374                         struct bkey u;
1375                         struct bkey_s_c k = bch2_btree_path_peek_slot(i->path, &u);
1376
1377                         bch2_bkey_val_to_text(&PBUF(buf), c, k);
1378                         pr_err("%s", buf);
1379                 } else {
1380                         struct bkey_cached *ck = (void *) i->path->l[0].b;
1381
1382                         if (ck->valid) {
1383                                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(ck->k));
1384                                 pr_err("%s", buf);
1385                         }
1386                 }
1387         }
1388         __WARN();
1389 }
1390
1391 int bch2_trans_fs_usage_apply(struct btree_trans *trans,
1392                               struct replicas_delta_list *deltas)
1393 {
1394         struct bch_fs *c = trans->c;
1395         static int warned_disk_usage = 0;
1396         bool warn = false;
1397         unsigned disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1398         struct replicas_delta *d = deltas->d, *d2;
1399         struct replicas_delta *top = (void *) deltas->d + deltas->used;
1400         struct bch_fs_usage *dst;
1401         s64 added = 0, should_not_have_added;
1402         unsigned i;
1403
1404         percpu_down_read(&c->mark_lock);
1405         preempt_disable();
1406         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1407
1408         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1409                 switch (d->r.data_type) {
1410                 case BCH_DATA_btree:
1411                 case BCH_DATA_user:
1412                 case BCH_DATA_parity:
1413                         added += d->delta;
1414                 }
1415
1416                 if (__update_replicas(c, dst, &d->r, d->delta))
1417                         goto need_mark;
1418         }
1419
1420         dst->nr_inodes += deltas->nr_inodes;
1421
1422         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1423                 added                           += deltas->persistent_reserved[i];
1424                 dst->reserved                   += deltas->persistent_reserved[i];
1425                 dst->persistent_reserved[i]     += deltas->persistent_reserved[i];
1426         }
1427
1428         /*
1429          * Not allowed to reduce sectors_available except by getting a
1430          * reservation:
1431          */
1432         should_not_have_added = added - (s64) disk_res_sectors;
1433         if (unlikely(should_not_have_added > 0)) {
1434                 u64 old, new, v = atomic64_read(&c->sectors_available);
1435
1436                 do {
1437                         old = v;
1438                         new = max_t(s64, 0, old - should_not_have_added);
1439                 } while ((v = atomic64_cmpxchg(&c->sectors_available,
1440                                                old, new)) != old);
1441
1442                 added -= should_not_have_added;
1443                 warn = true;
1444         }
1445
1446         if (added > 0) {
1447                 trans->disk_res->sectors -= added;
1448                 this_cpu_sub(*c->online_reserved, added);
1449         }
1450
1451         preempt_enable();
1452         percpu_up_read(&c->mark_lock);
1453
1454         if (unlikely(warn) && !xchg(&warned_disk_usage, 1))
1455                 fs_usage_apply_warn(trans, disk_res_sectors, should_not_have_added);
1456         return 0;
1457 need_mark:
1458         /* revert changes: */
1459         for (d2 = deltas->d; d2 != d; d2 = replicas_delta_next(d2))
1460                 BUG_ON(__update_replicas(c, dst, &d2->r, -d2->delta));
1461
1462         preempt_enable();
1463         percpu_up_read(&c->mark_lock);
1464         return -1;
1465 }
1466
1467 /* trans_mark: */
1468
1469 static int bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
1470                               const struct bch_extent_ptr *ptr,
1471                               struct bkey_alloc_unpacked *u)
1472 {
1473         struct bch_fs *c = trans->c;
1474         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1475         struct bpos pos = POS(ptr->dev, PTR_BUCKET_NR(ca, ptr));
1476         struct bkey_i *update = btree_trans_peek_updates(trans, BTREE_ID_alloc, pos);
1477         int ret;
1478
1479         bch2_trans_iter_init(trans, iter, BTREE_ID_alloc, pos,
1480                              BTREE_ITER_CACHED|
1481                              BTREE_ITER_CACHED_NOFILL|
1482                              BTREE_ITER_INTENT);
1483         ret = bch2_btree_iter_traverse(iter);
1484         if (ret) {
1485                 bch2_trans_iter_exit(trans, iter);
1486                 return ret;
1487         }
1488
1489         *u = update && !bpos_cmp(update->k.p, pos)
1490                 ? bch2_alloc_unpack(bkey_i_to_s_c(update))
1491                 : alloc_mem_to_key(c, iter);
1492
1493         return 0;
1494 }
1495
1496 static int bch2_trans_mark_pointer(struct btree_trans *trans,
1497                         struct bkey_s_c k, struct extent_ptr_decoded p,
1498                         s64 sectors, enum bch_data_type data_type)
1499 {
1500         struct btree_iter iter;
1501         struct bkey_alloc_unpacked u;
1502         int ret;
1503
1504         ret = bch2_trans_start_alloc_update(trans, &iter, &p.ptr, &u);
1505         if (ret)
1506                 return ret;
1507
1508         ret = __mark_pointer(trans, k, &p.ptr, sectors, data_type,
1509                              u.gen, &u.data_type,
1510                              &u.dirty_sectors, &u.cached_sectors);
1511         if (ret)
1512                 goto out;
1513
1514         ret = bch2_alloc_write(trans, &iter, &u, 0);
1515         if (ret)
1516                 goto out;
1517 out:
1518         bch2_trans_iter_exit(trans, &iter);
1519         return ret;
1520 }
1521
1522 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1523                         struct extent_ptr_decoded p,
1524                         s64 sectors, enum bch_data_type data_type)
1525 {
1526         struct bch_fs *c = trans->c;
1527         struct btree_iter iter;
1528         struct bkey_s_c k;
1529         struct bkey_i_stripe *s;
1530         struct bch_replicas_padded r;
1531         int ret = 0;
1532
1533         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes, POS(0, p.ec.idx),
1534                              BTREE_ITER_INTENT|
1535                              BTREE_ITER_WITH_UPDATES);
1536         k = bch2_btree_iter_peek_slot(&iter);
1537         ret = bkey_err(k);
1538         if (ret)
1539                 goto err;
1540
1541         if (k.k->type != KEY_TYPE_stripe) {
1542                 bch2_fs_inconsistent(c,
1543                         "pointer to nonexistent stripe %llu",
1544                         (u64) p.ec.idx);
1545                 bch2_inconsistent_error(c);
1546                 ret = -EIO;
1547                 goto err;
1548         }
1549
1550         if (!bch2_ptr_matches_stripe(bkey_s_c_to_stripe(k).v, p)) {
1551                 bch2_fs_inconsistent(c,
1552                         "stripe pointer doesn't match stripe %llu",
1553                         (u64) p.ec.idx);
1554                 ret = -EIO;
1555                 goto err;
1556         }
1557
1558         s = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1559         ret = PTR_ERR_OR_ZERO(s);
1560         if (ret)
1561                 goto err;
1562
1563         bkey_reassemble(&s->k_i, k);
1564         stripe_blockcount_set(&s->v, p.ec.block,
1565                 stripe_blockcount_get(&s->v, p.ec.block) +
1566                 sectors);
1567
1568         ret = bch2_trans_update(trans, &iter, &s->k_i, 0);
1569         if (ret)
1570                 goto err;
1571
1572         bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1573         r.e.data_type = data_type;
1574         update_replicas_list(trans, &r.e, sectors);
1575 err:
1576         bch2_trans_iter_exit(trans, &iter);
1577         return ret;
1578 }
1579
1580 static int bch2_trans_mark_extent(struct btree_trans *trans,
1581                         struct bkey_s_c k, unsigned flags)
1582 {
1583         struct bch_fs *c = trans->c;
1584         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1585         const union bch_extent_entry *entry;
1586         struct extent_ptr_decoded p;
1587         struct bch_replicas_padded r;
1588         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
1589                 ? BCH_DATA_btree
1590                 : BCH_DATA_user;
1591         s64 sectors = bkey_is_btree_ptr(k.k)
1592                 ? btree_sectors(c)
1593                 : k.k->size;
1594         s64 dirty_sectors = 0;
1595         bool stale;
1596         int ret;
1597
1598         r.e.data_type   = data_type;
1599         r.e.nr_devs     = 0;
1600         r.e.nr_required = 1;
1601
1602         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1603                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
1604
1605                 if (flags & BTREE_TRIGGER_OVERWRITE)
1606                         disk_sectors = -disk_sectors;
1607
1608                 ret = bch2_trans_mark_pointer(trans, k, p,
1609                                         disk_sectors, data_type);
1610                 if (ret < 0)
1611                         return ret;
1612
1613                 stale = ret > 0;
1614
1615                 if (p.ptr.cached) {
1616                         if (!stale)
1617                                 update_cached_sectors_list(trans, p.ptr.dev,
1618                                                            disk_sectors);
1619                 } else if (!p.has_ec) {
1620                         dirty_sectors          += disk_sectors;
1621                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1622                 } else {
1623                         ret = bch2_trans_mark_stripe_ptr(trans, p,
1624                                         disk_sectors, data_type);
1625                         if (ret)
1626                                 return ret;
1627
1628                         r.e.nr_required = 0;
1629                 }
1630         }
1631
1632         if (r.e.nr_devs)
1633                 update_replicas_list(trans, &r.e, dirty_sectors);
1634
1635         return 0;
1636 }
1637
1638 static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
1639                                          struct bkey_s_c_stripe s,
1640                                          unsigned idx, bool deleting)
1641 {
1642         struct bch_fs *c = trans->c;
1643         const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
1644         struct btree_iter iter;
1645         struct bkey_alloc_unpacked u;
1646         enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
1647                 ? BCH_DATA_parity : 0;
1648         s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
1649         int ret = 0;
1650
1651         if (deleting)
1652                 sectors = -sectors;
1653
1654         ret = bch2_trans_start_alloc_update(trans, &iter, ptr, &u);
1655         if (ret)
1656                 return ret;
1657
1658         ret = check_bucket_ref(c, s.s_c, ptr, sectors, data_type,
1659                                u.gen, u.data_type,
1660                                u.dirty_sectors, u.cached_sectors);
1661         if (ret)
1662                 goto err;
1663
1664         if (!deleting) {
1665                 if (bch2_fs_inconsistent_on(u.stripe ||
1666                                             u.stripe_redundancy, c,
1667                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
1668                                 iter.pos.inode, iter.pos.offset, u.gen,
1669                                 bch2_data_types[u.data_type],
1670                                 u.dirty_sectors,
1671                                 u.stripe, s.k->p.offset)) {
1672                         ret = -EIO;
1673                         goto err;
1674                 }
1675
1676                 if (bch2_fs_inconsistent_on(data_type && u.dirty_sectors, c,
1677                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
1678                                 iter.pos.inode, iter.pos.offset, u.gen,
1679                                 bch2_data_types[u.data_type],
1680                                 u.dirty_sectors,
1681                                 s.k->p.offset)) {
1682                         ret = -EIO;
1683                         goto err;
1684                 }
1685
1686                 u.stripe                = s.k->p.offset;
1687                 u.stripe_redundancy     = s.v->nr_redundant;
1688         } else {
1689                 if (bch2_fs_inconsistent_on(u.stripe != s.k->p.offset ||
1690                                             u.stripe_redundancy != s.v->nr_redundant, c,
1691                                 "bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
1692                                 iter.pos.inode, iter.pos.offset, u.gen,
1693                                 s.k->p.offset, u.stripe)) {
1694                         ret = -EIO;
1695                         goto err;
1696                 }
1697
1698                 u.stripe                = 0;
1699                 u.stripe_redundancy     = 0;
1700         }
1701
1702         u.dirty_sectors += sectors;
1703         if (data_type)
1704                 u.data_type = !deleting ? data_type : 0;
1705
1706         ret = bch2_alloc_write(trans, &iter, &u, 0);
1707         if (ret)
1708                 goto err;
1709 err:
1710         bch2_trans_iter_exit(trans, &iter);
1711         return ret;
1712 }
1713
1714 static int bch2_trans_mark_stripe(struct btree_trans *trans,
1715                                   struct bkey_s_c old, struct bkey_s_c new,
1716                                   unsigned flags)
1717 {
1718         struct bkey_s_c_stripe old_s = { .k = NULL };
1719         struct bkey_s_c_stripe new_s = { .k = NULL };
1720         struct bch_replicas_padded r;
1721         unsigned i, nr_blocks;
1722         int ret = 0;
1723
1724         if (old.k->type == KEY_TYPE_stripe)
1725                 old_s = bkey_s_c_to_stripe(old);
1726         if (new.k->type == KEY_TYPE_stripe)
1727                 new_s = bkey_s_c_to_stripe(new);
1728
1729         /*
1730          * If the pointers aren't changing, we don't need to do anything:
1731          */
1732         if (new_s.k && old_s.k &&
1733             new_s.v->nr_blocks          == old_s.v->nr_blocks &&
1734             new_s.v->nr_redundant       == old_s.v->nr_redundant &&
1735             !memcmp(old_s.v->ptrs, new_s.v->ptrs,
1736                     new_s.v->nr_blocks * sizeof(struct bch_extent_ptr)))
1737                 return 0;
1738
1739         BUG_ON(new_s.k && old_s.k &&
1740                (new_s.v->nr_blocks      != old_s.v->nr_blocks ||
1741                 new_s.v->nr_redundant   != old_s.v->nr_redundant));
1742
1743         nr_blocks = new_s.k ? new_s.v->nr_blocks : old_s.v->nr_blocks;
1744
1745         if (new_s.k) {
1746                 s64 sectors = le16_to_cpu(new_s.v->sectors);
1747
1748                 bch2_bkey_to_replicas(&r.e, new);
1749                 update_replicas_list(trans, &r.e, sectors * new_s.v->nr_redundant);
1750         }
1751
1752         if (old_s.k) {
1753                 s64 sectors = -((s64) le16_to_cpu(old_s.v->sectors));
1754
1755                 bch2_bkey_to_replicas(&r.e, old);
1756                 update_replicas_list(trans, &r.e, sectors * old_s.v->nr_redundant);
1757         }
1758
1759         for (i = 0; i < nr_blocks; i++) {
1760                 if (new_s.k && old_s.k &&
1761                     !memcmp(&new_s.v->ptrs[i],
1762                             &old_s.v->ptrs[i],
1763                             sizeof(new_s.v->ptrs[i])))
1764                         continue;
1765
1766                 if (new_s.k) {
1767                         ret = bch2_trans_mark_stripe_bucket(trans, new_s, i, false);
1768                         if (ret)
1769                                 break;
1770                 }
1771
1772                 if (old_s.k) {
1773                         ret = bch2_trans_mark_stripe_bucket(trans, old_s, i, true);
1774                         if (ret)
1775                                 break;
1776                 }
1777         }
1778
1779         return ret;
1780 }
1781
1782 static int bch2_trans_mark_inode(struct btree_trans *trans,
1783                                  struct bkey_s_c old,
1784                                  struct bkey_s_c new,
1785                                  unsigned flags)
1786 {
1787         int nr = bkey_is_inode(new.k) - bkey_is_inode(old.k);
1788
1789         if (nr) {
1790                 struct replicas_delta_list *d =
1791                         replicas_deltas_realloc(trans, 0);
1792                 d->nr_inodes += nr;
1793         }
1794
1795         return 0;
1796 }
1797
1798 static int bch2_trans_mark_reservation(struct btree_trans *trans,
1799                                        struct bkey_s_c k, unsigned flags)
1800 {
1801         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1802         s64 sectors = (s64) k.k->size;
1803         struct replicas_delta_list *d;
1804
1805         if (flags & BTREE_TRIGGER_OVERWRITE)
1806                 sectors = -sectors;
1807         sectors *= replicas;
1808
1809         d = replicas_deltas_realloc(trans, 0);
1810
1811         replicas = clamp_t(unsigned, replicas, 1,
1812                            ARRAY_SIZE(d->persistent_reserved));
1813
1814         d->persistent_reserved[replicas - 1] += sectors;
1815         return 0;
1816 }
1817
1818 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1819                         struct bkey_s_c_reflink_p p,
1820                         u64 *idx, unsigned flags)
1821 {
1822         struct bch_fs *c = trans->c;
1823         struct btree_iter iter;
1824         struct bkey_s_c k;
1825         struct bkey_i *n;
1826         __le64 *refcount;
1827         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1828         char buf[200];
1829         int ret;
1830
1831         bch2_trans_iter_init(trans, &iter, BTREE_ID_reflink, POS(0, *idx),
1832                              BTREE_ITER_INTENT|
1833                              BTREE_ITER_WITH_UPDATES);
1834         k = bch2_btree_iter_peek_slot(&iter);
1835         ret = bkey_err(k);
1836         if (ret)
1837                 goto err;
1838
1839         n = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1840         ret = PTR_ERR_OR_ZERO(n);
1841         if (ret)
1842                 goto err;
1843
1844         bkey_reassemble(n, k);
1845
1846         refcount = bkey_refcount(n);
1847         if (!refcount) {
1848                 bch2_bkey_val_to_text(&PBUF(buf), c, p.s_c);
1849                 bch2_fs_inconsistent(c,
1850                         "nonexistent indirect extent at %llu while marking\n  %s",
1851                         *idx, buf);
1852                 ret = -EIO;
1853                 goto err;
1854         }
1855
1856         if (!*refcount && (flags & BTREE_TRIGGER_OVERWRITE)) {
1857                 bch2_bkey_val_to_text(&PBUF(buf), c, p.s_c);
1858                 bch2_fs_inconsistent(c,
1859                         "indirect extent refcount underflow at %llu while marking\n  %s",
1860                         *idx, buf);
1861                 ret = -EIO;
1862                 goto err;
1863         }
1864
1865         if (flags & BTREE_TRIGGER_INSERT) {
1866                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1867                 u64 pad;
1868
1869                 pad = max_t(s64, le32_to_cpu(v->front_pad),
1870                             le64_to_cpu(v->idx) - bkey_start_offset(k.k));
1871                 BUG_ON(pad > U32_MAX);
1872                 v->front_pad = cpu_to_le32(pad);
1873
1874                 pad = max_t(s64, le32_to_cpu(v->back_pad),
1875                             k.k->p.offset - p.k->size - le64_to_cpu(v->idx));
1876                 BUG_ON(pad > U32_MAX);
1877                 v->back_pad = cpu_to_le32(pad);
1878         }
1879
1880         le64_add_cpu(refcount, add);
1881
1882         if (!*refcount) {
1883                 n->k.type = KEY_TYPE_deleted;
1884                 set_bkey_val_u64s(&n->k, 0);
1885         }
1886
1887         bch2_btree_iter_set_pos_to_extent_start(&iter);
1888         ret = bch2_trans_update(trans, &iter, n, 0);
1889         if (ret)
1890                 goto err;
1891
1892         *idx = k.k->p.offset;
1893 err:
1894         bch2_trans_iter_exit(trans, &iter);
1895         return ret;
1896 }
1897
1898 static int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1899                                      struct bkey_s_c k, unsigned flags)
1900 {
1901         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1902         u64 idx, end_idx;
1903         int ret = 0;
1904
1905         if (flags & BTREE_TRIGGER_INSERT) {
1906                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1907
1908                 v->front_pad = v->back_pad = 0;
1909         }
1910
1911         idx     = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
1912         end_idx = le64_to_cpu(p.v->idx) + p.k->size +
1913                 le32_to_cpu(p.v->back_pad);
1914
1915         while (idx < end_idx && !ret)
1916                 ret = __bch2_trans_mark_reflink_p(trans, p, &idx, flags);
1917
1918         return ret;
1919 }
1920
1921 int bch2_trans_mark_key(struct btree_trans *trans, struct bkey_s_c old,
1922                         struct bkey_s_c new, unsigned flags)
1923 {
1924         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1925
1926         switch (k.k->type) {
1927         case KEY_TYPE_btree_ptr:
1928         case KEY_TYPE_btree_ptr_v2:
1929         case KEY_TYPE_extent:
1930         case KEY_TYPE_reflink_v:
1931                 return bch2_trans_mark_extent(trans, k, flags);
1932         case KEY_TYPE_stripe:
1933                 return bch2_trans_mark_stripe(trans, old, new, flags);
1934         case KEY_TYPE_inode:
1935         case KEY_TYPE_inode_v2:
1936                 return bch2_trans_mark_inode(trans, old, new, flags);
1937         case KEY_TYPE_reservation:
1938                 return bch2_trans_mark_reservation(trans, k, flags);
1939         case KEY_TYPE_reflink_p:
1940                 return bch2_trans_mark_reflink_p(trans, k, flags);
1941         default:
1942                 return 0;
1943         }
1944 }
1945
1946 static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1947                                     struct bch_dev *ca, size_t b,
1948                                     enum bch_data_type type,
1949                                     unsigned sectors)
1950 {
1951         struct bch_fs *c = trans->c;
1952         struct btree_iter iter;
1953         struct bkey_alloc_unpacked u;
1954         struct bch_extent_ptr ptr = {
1955                 .dev = ca->dev_idx,
1956                 .offset = bucket_to_sector(ca, b),
1957         };
1958         int ret = 0;
1959
1960         /*
1961          * Backup superblock might be past the end of our normal usable space:
1962          */
1963         if (b >= ca->mi.nbuckets)
1964                 return 0;
1965
1966         ret = bch2_trans_start_alloc_update(trans, &iter, &ptr, &u);
1967         if (ret)
1968                 return ret;
1969
1970         if (u.data_type && u.data_type != type) {
1971                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1972                         "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
1973                         "while marking %s",
1974                         iter.pos.inode, iter.pos.offset, u.gen,
1975                         bch2_data_types[u.data_type],
1976                         bch2_data_types[type],
1977                         bch2_data_types[type]);
1978                 ret = -EIO;
1979                 goto out;
1980         }
1981
1982         u.data_type     = type;
1983         u.dirty_sectors = sectors;
1984
1985         ret = bch2_alloc_write(trans, &iter, &u, 0);
1986         if (ret)
1987                 goto out;
1988 out:
1989         bch2_trans_iter_exit(trans, &iter);
1990         return ret;
1991 }
1992
1993 int bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1994                                     struct bch_dev *ca, size_t b,
1995                                     enum bch_data_type type,
1996                                     unsigned sectors)
1997 {
1998         return __bch2_trans_do(trans, NULL, NULL, 0,
1999                         __bch2_trans_mark_metadata_bucket(trans, ca, b, type, sectors));
2000 }
2001
2002 static int bch2_trans_mark_metadata_sectors(struct btree_trans *trans,
2003                                             struct bch_dev *ca,
2004                                             u64 start, u64 end,
2005                                             enum bch_data_type type,
2006                                             u64 *bucket, unsigned *bucket_sectors)
2007 {
2008         do {
2009                 u64 b = sector_to_bucket(ca, start);
2010                 unsigned sectors =
2011                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
2012
2013                 if (b != *bucket && *bucket_sectors) {
2014                         int ret = bch2_trans_mark_metadata_bucket(trans, ca, *bucket,
2015                                                                   type, *bucket_sectors);
2016                         if (ret)
2017                                 return ret;
2018
2019                         *bucket_sectors = 0;
2020                 }
2021
2022                 *bucket         = b;
2023                 *bucket_sectors += sectors;
2024                 start += sectors;
2025         } while (start < end);
2026
2027         return 0;
2028 }
2029
2030 static int __bch2_trans_mark_dev_sb(struct btree_trans *trans,
2031                                     struct bch_dev *ca)
2032 {
2033         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
2034         u64 bucket = 0;
2035         unsigned i, bucket_sectors = 0;
2036         int ret;
2037
2038         for (i = 0; i < layout->nr_superblocks; i++) {
2039                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
2040
2041                 if (offset == BCH_SB_SECTOR) {
2042                         ret = bch2_trans_mark_metadata_sectors(trans, ca,
2043                                                 0, BCH_SB_SECTOR,
2044                                                 BCH_DATA_sb, &bucket, &bucket_sectors);
2045                         if (ret)
2046                                 return ret;
2047                 }
2048
2049                 ret = bch2_trans_mark_metadata_sectors(trans, ca, offset,
2050                                       offset + (1 << layout->sb_max_size_bits),
2051                                       BCH_DATA_sb, &bucket, &bucket_sectors);
2052                 if (ret)
2053                         return ret;
2054         }
2055
2056         if (bucket_sectors) {
2057                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
2058                                 bucket, BCH_DATA_sb, bucket_sectors);
2059                 if (ret)
2060                         return ret;
2061         }
2062
2063         for (i = 0; i < ca->journal.nr; i++) {
2064                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
2065                                 ca->journal.buckets[i],
2066                                 BCH_DATA_journal, ca->mi.bucket_size);
2067                 if (ret)
2068                         return ret;
2069         }
2070
2071         return 0;
2072 }
2073
2074 int bch2_trans_mark_dev_sb(struct bch_fs *c, struct bch_dev *ca)
2075 {
2076         return bch2_trans_do(c, NULL, NULL, BTREE_INSERT_LAZY_RW,
2077                         __bch2_trans_mark_dev_sb(&trans, ca));
2078 }
2079
2080 /* Disk reservations: */
2081
2082 #define SECTORS_CACHE   1024
2083
2084 int bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
2085                               u64 sectors, int flags)
2086 {
2087         struct bch_fs_pcpu *pcpu;
2088         u64 old, v, get;
2089         s64 sectors_available;
2090         int ret;
2091
2092         percpu_down_read(&c->mark_lock);
2093         preempt_disable();
2094         pcpu = this_cpu_ptr(c->pcpu);
2095
2096         if (sectors <= pcpu->sectors_available)
2097                 goto out;
2098
2099         v = atomic64_read(&c->sectors_available);
2100         do {
2101                 old = v;
2102                 get = min((u64) sectors + SECTORS_CACHE, old);
2103
2104                 if (get < sectors) {
2105                         preempt_enable();
2106                         goto recalculate;
2107                 }
2108         } while ((v = atomic64_cmpxchg(&c->sectors_available,
2109                                        old, old - get)) != old);
2110
2111         pcpu->sectors_available         += get;
2112
2113 out:
2114         pcpu->sectors_available         -= sectors;
2115         this_cpu_add(*c->online_reserved, sectors);
2116         res->sectors                    += sectors;
2117
2118         preempt_enable();
2119         percpu_up_read(&c->mark_lock);
2120         return 0;
2121
2122 recalculate:
2123         mutex_lock(&c->sectors_available_lock);
2124
2125         percpu_u64_set(&c->pcpu->sectors_available, 0);
2126         sectors_available = avail_factor(__bch2_fs_usage_read_short(c).free);
2127
2128         if (sectors <= sectors_available ||
2129             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
2130                 atomic64_set(&c->sectors_available,
2131                              max_t(s64, 0, sectors_available - sectors));
2132                 this_cpu_add(*c->online_reserved, sectors);
2133                 res->sectors                    += sectors;
2134                 ret = 0;
2135         } else {
2136                 atomic64_set(&c->sectors_available, sectors_available);
2137                 ret = -ENOSPC;
2138         }
2139
2140         mutex_unlock(&c->sectors_available_lock);
2141         percpu_up_read(&c->mark_lock);
2142
2143         return ret;
2144 }
2145
2146 /* Startup/shutdown: */
2147
2148 static void buckets_free_rcu(struct rcu_head *rcu)
2149 {
2150         struct bucket_array *buckets =
2151                 container_of(rcu, struct bucket_array, rcu);
2152
2153         kvpfree(buckets,
2154                 sizeof(struct bucket_array) +
2155                 buckets->nbuckets * sizeof(struct bucket));
2156 }
2157
2158 static void bucket_gens_free_rcu(struct rcu_head *rcu)
2159 {
2160         struct bucket_gens *buckets =
2161                 container_of(rcu, struct bucket_gens, rcu);
2162
2163         kvpfree(buckets, sizeof(struct bucket_array) + buckets->nbuckets);
2164 }
2165
2166 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
2167 {
2168         struct bucket_array *buckets = NULL, *old_buckets = NULL;
2169         struct bucket_gens *bucket_gens = NULL, *old_bucket_gens = NULL;
2170         unsigned long *buckets_nouse = NULL;
2171         alloc_fifo      free[RESERVE_NR];
2172         alloc_fifo      free_inc;
2173         alloc_heap      alloc_heap;
2174
2175         size_t btree_reserve    = DIV_ROUND_UP(BTREE_NODE_RESERVE,
2176                              ca->mi.bucket_size / btree_sectors(c));
2177         /* XXX: these should be tunable */
2178         size_t reserve_none     = max_t(size_t, 1, nbuckets >> 9);
2179         size_t copygc_reserve   = max_t(size_t, 2, nbuckets >> 6);
2180         size_t free_inc_nr      = max(max_t(size_t, 1, nbuckets >> 12),
2181                                       btree_reserve * 2);
2182         bool resize = ca->buckets[0] != NULL;
2183         int ret = -ENOMEM;
2184         unsigned i;
2185
2186         memset(&free,           0, sizeof(free));
2187         memset(&free_inc,       0, sizeof(free_inc));
2188         memset(&alloc_heap,     0, sizeof(alloc_heap));
2189
2190         if (!(buckets           = kvpmalloc(sizeof(struct bucket_array) +
2191                                             nbuckets * sizeof(struct bucket),
2192                                             GFP_KERNEL|__GFP_ZERO)) ||
2193             !(bucket_gens       = kvpmalloc(sizeof(struct bucket_gens) + nbuckets,
2194                                             GFP_KERNEL|__GFP_ZERO)) ||
2195             !(buckets_nouse     = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2196                                             sizeof(unsigned long),
2197                                             GFP_KERNEL|__GFP_ZERO)) ||
2198             !init_fifo(&free[RESERVE_MOVINGGC],
2199                        copygc_reserve, GFP_KERNEL) ||
2200             !init_fifo(&free[RESERVE_NONE], reserve_none, GFP_KERNEL) ||
2201             !init_fifo(&free_inc,       free_inc_nr, GFP_KERNEL) ||
2202             !init_heap(&alloc_heap,     ALLOC_SCAN_BATCH(ca) << 1, GFP_KERNEL))
2203                 goto err;
2204
2205         buckets->first_bucket   = ca->mi.first_bucket;
2206         buckets->nbuckets       = nbuckets;
2207         bucket_gens->first_bucket = ca->mi.first_bucket;
2208         bucket_gens->nbuckets   = nbuckets;
2209
2210         bch2_copygc_stop(c);
2211
2212         if (resize) {
2213                 down_write(&c->gc_lock);
2214                 down_write(&ca->bucket_lock);
2215                 percpu_down_write(&c->mark_lock);
2216         }
2217
2218         old_buckets = bucket_array(ca);
2219         old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
2220
2221         if (resize) {
2222                 size_t n = min(buckets->nbuckets, old_buckets->nbuckets);
2223
2224                 memcpy(buckets->b,
2225                        old_buckets->b,
2226                        n * sizeof(struct bucket));
2227                 memcpy(bucket_gens->b,
2228                        old_bucket_gens->b,
2229                        n);
2230                 memcpy(buckets_nouse,
2231                        ca->buckets_nouse,
2232                        BITS_TO_LONGS(n) * sizeof(unsigned long));
2233         }
2234
2235         rcu_assign_pointer(ca->buckets[0], buckets);
2236         rcu_assign_pointer(ca->bucket_gens, bucket_gens);
2237         buckets         = old_buckets;
2238         bucket_gens     = old_bucket_gens;
2239
2240         swap(ca->buckets_nouse, buckets_nouse);
2241
2242         if (resize) {
2243                 percpu_up_write(&c->mark_lock);
2244                 up_write(&c->gc_lock);
2245         }
2246
2247         spin_lock(&c->freelist_lock);
2248         for (i = 0; i < RESERVE_NR; i++) {
2249                 fifo_move(&free[i], &ca->free[i]);
2250                 swap(ca->free[i], free[i]);
2251         }
2252         fifo_move(&free_inc, &ca->free_inc);
2253         swap(ca->free_inc, free_inc);
2254         spin_unlock(&c->freelist_lock);
2255
2256         /* with gc lock held, alloc_heap can't be in use: */
2257         swap(ca->alloc_heap, alloc_heap);
2258
2259         nbuckets = ca->mi.nbuckets;
2260
2261         if (resize)
2262                 up_write(&ca->bucket_lock);
2263
2264         ret = 0;
2265 err:
2266         free_heap(&alloc_heap);
2267         free_fifo(&free_inc);
2268         for (i = 0; i < RESERVE_NR; i++)
2269                 free_fifo(&free[i]);
2270         kvpfree(buckets_nouse,
2271                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2272         if (bucket_gens)
2273                 call_rcu(&old_buckets->rcu, bucket_gens_free_rcu);
2274         if (buckets)
2275                 call_rcu(&old_buckets->rcu, buckets_free_rcu);
2276
2277         return ret;
2278 }
2279
2280 void bch2_dev_buckets_free(struct bch_dev *ca)
2281 {
2282         unsigned i;
2283
2284         free_heap(&ca->alloc_heap);
2285         free_fifo(&ca->free_inc);
2286         for (i = 0; i < RESERVE_NR; i++)
2287                 free_fifo(&ca->free[i]);
2288         kvpfree(ca->buckets_nouse,
2289                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2290         kvpfree(rcu_dereference_protected(ca->buckets[0], 1),
2291                 sizeof(struct bucket_array) +
2292                 ca->mi.nbuckets * sizeof(struct bucket));
2293
2294         for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
2295                 free_percpu(ca->usage[i]);
2296         kfree(ca->usage_base);
2297 }
2298
2299 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2300 {
2301         unsigned i;
2302
2303         ca->usage_base = kzalloc(sizeof(struct bch_dev_usage), GFP_KERNEL);
2304         if (!ca->usage_base)
2305                 return -ENOMEM;
2306
2307         for (i = 0; i < ARRAY_SIZE(ca->usage); i++) {
2308                 ca->usage[i] = alloc_percpu(struct bch_dev_usage);
2309                 if (!ca->usage[i])
2310                         return -ENOMEM;
2311         }
2312
2313         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);;
2314 }