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 "buckets_waiting_for_journal.h"
  15 #include "ec.h"
  16 #include "error.h"
  17 #include "inode.h"
  18 #include "movinggc.h"
  19 #include "recovery.h"
  20 #include "reflink.h"
  21 #include "replicas.h"
  22 #include "subvolume.h"
  23
  24 #include <linux/preempt.h>
  25 #include <trace/events/bcachefs.h>
  26
  27 static inline void fs_usage_data_type_to_base(struct bch_fs_usage *fs_usage,
  28                                               enum bch_data_type data_type,
  29                                               s64 sectors)
  30 {
  31         switch (data_type) {
  32         case BCH_DATA_btree:
  33                 fs_usage->btree         += sectors;
  34                 break;
  35         case BCH_DATA_user:
  36         case BCH_DATA_parity:
  37                 fs_usage->data          += sectors;
  38                 break;
  39         case BCH_DATA_cached:
  40                 fs_usage->cached        += sectors;
  41                 break;
  42         default:
  43                 break;
  44         }
  45 }
  46
  47 void bch2_fs_usage_initialize(struct bch_fs *c)
  48 {
  49         struct bch_fs_usage *usage;
  50         struct bch_dev *ca;
  51         unsigned i;
  52
  53         percpu_down_write(&c->mark_lock);
  54         usage = c->usage_base;
  55
  56         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
  57                 bch2_fs_usage_acc_to_base(c, i);
  58
  59         for (i = 0; i < BCH_REPLICAS_MAX; i++)
  60                 usage->reserved += usage->persistent_reserved[i];
  61
  62         for (i = 0; i < c->replicas.nr; i++) {
  63                 struct bch_replicas_entry *e =
  64                         cpu_replicas_entry(&c->replicas, i);
  65
  66                 fs_usage_data_type_to_base(usage, e->data_type, usage->replicas[i]);
  67         }
  68
  69         for_each_member_device(ca, c, i) {
  70                 struct bch_dev_usage dev = bch2_dev_usage_read(ca);
  71
  72                 usage->hidden += (dev.d[BCH_DATA_sb].buckets +
  73                                   dev.d[BCH_DATA_journal].buckets) *
  74                         ca->mi.bucket_size;
  75         }
  76
  77         percpu_up_write(&c->mark_lock);
  78 }
  79
  80 static inline struct bch_dev_usage *dev_usage_ptr(struct bch_dev *ca,
  81                                                   unsigned journal_seq,
  82                                                   bool gc)
  83 {
  84         BUG_ON(!gc && !journal_seq);
  85
  86         return this_cpu_ptr(gc
  87                             ? ca->usage_gc
  88                             : ca->usage[journal_seq & JOURNAL_BUF_MASK]);
  89 }
  90
  91 struct bch_dev_usage bch2_dev_usage_read(struct bch_dev *ca)
  92 {
  93         struct bch_fs *c = ca->fs;
  94         struct bch_dev_usage ret;
  95         unsigned seq, i, u64s = dev_usage_u64s();
  96
  97         do {
  98                 seq = read_seqcount_begin(&c->usage_lock);
  99                 memcpy(&ret, ca->usage_base, u64s * sizeof(u64));
 100                 for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
 101                         acc_u64s_percpu((u64 *) &ret, (u64 __percpu *) ca->usage[i], u64s);
 102         } while (read_seqcount_retry(&c->usage_lock, seq));
 103
 104         return ret;
 105 }
 106
 107 static inline struct bch_fs_usage *fs_usage_ptr(struct bch_fs *c,
 108                                                 unsigned journal_seq,
 109                                                 bool gc)
 110 {
 111         percpu_rwsem_assert_held(&c->mark_lock);
 112         BUG_ON(!gc && !journal_seq);
 113
 114         return this_cpu_ptr(gc
 115                             ? c->usage_gc
 116                             : c->usage[journal_seq & JOURNAL_BUF_MASK]);
 117 }
 118
 119 u64 bch2_fs_usage_read_one(struct bch_fs *c, u64 *v)
 120 {
 121         ssize_t offset = v - (u64 *) c->usage_base;
 122         unsigned i, seq;
 123         u64 ret;
 124
 125         BUG_ON(offset < 0 || offset >= fs_usage_u64s(c));
 126         percpu_rwsem_assert_held(&c->mark_lock);
 127
 128         do {
 129                 seq = read_seqcount_begin(&c->usage_lock);
 130                 ret = *v;
 131
 132                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 133                         ret += percpu_u64_get((u64 __percpu *) c->usage[i] + offset);
 134         } while (read_seqcount_retry(&c->usage_lock, seq));
 135
 136         return ret;
 137 }
 138
 139 struct bch_fs_usage_online *bch2_fs_usage_read(struct bch_fs *c)
 140 {
 141         struct bch_fs_usage_online *ret;
 142         unsigned seq, i, u64s;
 143
 144         percpu_down_read(&c->mark_lock);
 145
 146         ret = kmalloc(sizeof(struct bch_fs_usage_online) +
 147                       sizeof(u64) * c->replicas.nr, GFP_NOFS);
 148         if (unlikely(!ret)) {
 149                 percpu_up_read(&c->mark_lock);
 150                 return NULL;
 151         }
 152
 153         ret->online_reserved = percpu_u64_get(c->online_reserved);
 154
 155         u64s = fs_usage_u64s(c);
 156         do {
 157                 seq = read_seqcount_begin(&c->usage_lock);
 158                 memcpy(&ret->u, c->usage_base, u64s * sizeof(u64));
 159                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 160                         acc_u64s_percpu((u64 *) &ret->u, (u64 __percpu *) c->usage[i], u64s);
 161         } while (read_seqcount_retry(&c->usage_lock, seq));
 162
 163         return ret;
 164 }
 165
 166 void bch2_fs_usage_acc_to_base(struct bch_fs *c, unsigned idx)
 167 {
 168         struct bch_dev *ca;
 169         unsigned i, u64s = fs_usage_u64s(c);
 170
 171         BUG_ON(idx >= ARRAY_SIZE(c->usage));
 172
 173         preempt_disable();
 174         write_seqcount_begin(&c->usage_lock);
 175
 176         acc_u64s_percpu((u64 *) c->usage_base,
 177                         (u64 __percpu *) c->usage[idx], u64s);
 178         percpu_memset(c->usage[idx], 0, u64s * sizeof(u64));
 179
 180         rcu_read_lock();
 181         for_each_member_device_rcu(ca, c, i, NULL) {
 182                 u64s = dev_usage_u64s();
 183
 184                 acc_u64s_percpu((u64 *) ca->usage_base,
 185                                 (u64 __percpu *) ca->usage[idx], u64s);
 186                 percpu_memset(ca->usage[idx], 0, u64s * sizeof(u64));
 187         }
 188         rcu_read_unlock();
 189
 190         write_seqcount_end(&c->usage_lock);
 191         preempt_enable();
 192 }
 193
 194 void bch2_fs_usage_to_text(struct printbuf *out,
 195                            struct bch_fs *c,
 196                            struct bch_fs_usage_online *fs_usage)
 197 {
 198         unsigned i;
 199
 200         pr_buf(out, "capacity:\t\t\t%llu\n", c->capacity);
 201
 202         pr_buf(out, "hidden:\t\t\t\t%llu\n",
 203                fs_usage->u.hidden);
 204         pr_buf(out, "data:\t\t\t\t%llu\n",
 205                fs_usage->u.data);
 206         pr_buf(out, "cached:\t\t\t\t%llu\n",
 207                fs_usage->u.cached);
 208         pr_buf(out, "reserved:\t\t\t%llu\n",
 209                fs_usage->u.reserved);
 210         pr_buf(out, "nr_inodes:\t\t\t%llu\n",
 211                fs_usage->u.nr_inodes);
 212         pr_buf(out, "online reserved:\t\t%llu\n",
 213                fs_usage->online_reserved);
 214
 215         for (i = 0;
 216              i < ARRAY_SIZE(fs_usage->u.persistent_reserved);
 217              i++) {
 218                 pr_buf(out, "%u replicas:\n", i + 1);
 219                 pr_buf(out, "\treserved:\t\t%llu\n",
 220                        fs_usage->u.persistent_reserved[i]);
 221         }
 222
 223         for (i = 0; i < c->replicas.nr; i++) {
 224                 struct bch_replicas_entry *e =
 225                         cpu_replicas_entry(&c->replicas, i);
 226
 227                 pr_buf(out, "\t");
 228                 bch2_replicas_entry_to_text(out, e);
 229                 pr_buf(out, ":\t%llu\n", fs_usage->u.replicas[i]);
 230         }
 231 }
 232
 233 static u64 reserve_factor(u64 r)
 234 {
 235         return r + (round_up(r, (1 << RESERVE_FACTOR)) >> RESERVE_FACTOR);
 236 }
 237
 238 u64 bch2_fs_sectors_used(struct bch_fs *c, struct bch_fs_usage_online *fs_usage)
 239 {
 240         return min(fs_usage->u.hidden +
 241                    fs_usage->u.btree +
 242                    fs_usage->u.data +
 243                    reserve_factor(fs_usage->u.reserved +
 244                                   fs_usage->online_reserved),
 245                    c->capacity);
 246 }
 247
 248 static struct bch_fs_usage_short
 249 __bch2_fs_usage_read_short(struct bch_fs *c)
 250 {
 251         struct bch_fs_usage_short ret;
 252         u64 data, reserved;
 253
 254         ret.capacity = c->capacity -
 255                 bch2_fs_usage_read_one(c, &c->usage_base->hidden);
 256
 257         data            = bch2_fs_usage_read_one(c, &c->usage_base->data) +
 258                 bch2_fs_usage_read_one(c, &c->usage_base->btree);
 259         reserved        = bch2_fs_usage_read_one(c, &c->usage_base->reserved) +
 260                 percpu_u64_get(c->online_reserved);
 261
 262         ret.used        = min(ret.capacity, data + reserve_factor(reserved));
 263         ret.free        = ret.capacity - ret.used;
 264
 265         ret.nr_inodes   = bch2_fs_usage_read_one(c, &c->usage_base->nr_inodes);
 266
 267         return ret;
 268 }
 269
 270 struct bch_fs_usage_short
 271 bch2_fs_usage_read_short(struct bch_fs *c)
 272 {
 273         struct bch_fs_usage_short ret;
 274
 275         percpu_down_read(&c->mark_lock);
 276         ret = __bch2_fs_usage_read_short(c);
 277         percpu_up_read(&c->mark_lock);
 278
 279         return ret;
 280 }
 281
 282 static inline int is_unavailable_bucket(struct bucket_mark m)
 283 {
 284         return !is_available_bucket(m);
 285 }
 286
 287 static inline int bucket_sectors_fragmented(struct bch_dev *ca,
 288                                             struct bucket_mark m)
 289 {
 290         return m.dirty_sectors
 291                 ? max(0, (int) ca->mi.bucket_size - (int) m.dirty_sectors)
 292                 : 0;
 293 }
 294
 295 static inline int is_stripe_data_bucket(struct bucket_mark m)
 296 {
 297         return m.stripe && m.data_type != BCH_DATA_parity;
 298 }
 299
 300 static inline enum bch_data_type bucket_type(struct bucket_mark m)
 301 {
 302         return m.cached_sectors && !m.dirty_sectors
 303                 ? BCH_DATA_cached
 304                 : m.data_type;
 305 }
 306
 307 static inline void account_bucket(struct bch_fs_usage *fs_usage,
 308                                   struct bch_dev_usage *dev_usage,
 309                                   enum bch_data_type type,
 310                                   int nr, s64 size)
 311 {
 312         if (type == BCH_DATA_sb || type == BCH_DATA_journal)
 313                 fs_usage->hidden        += size;
 314
 315         dev_usage->d[type].buckets      += nr;
 316 }
 317
 318 static void bch2_dev_usage_update(struct bch_fs *c, struct bch_dev *ca,
 319                                   struct bucket_mark old, struct bucket_mark new,
 320                                   u64 journal_seq, bool gc)
 321 {
 322         struct bch_fs_usage *fs_usage;
 323         struct bch_dev_usage *u;
 324
 325         preempt_disable();
 326         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 327         u = dev_usage_ptr(ca, journal_seq, gc);
 328
 329         if (bucket_type(old))
 330                 account_bucket(fs_usage, u, bucket_type(old),
 331                                -1, -ca->mi.bucket_size);
 332
 333         if (bucket_type(new))
 334                 account_bucket(fs_usage, u, bucket_type(new),
 335                                1, ca->mi.bucket_size);
 336
 337         u->buckets_ec += (int) new.stripe - (int) old.stripe;
 338         u->buckets_unavailable +=
 339                 is_unavailable_bucket(new) - is_unavailable_bucket(old);
 340
 341         u->d[old.data_type].sectors -= old.dirty_sectors;
 342         u->d[new.data_type].sectors += new.dirty_sectors;
 343         u->d[BCH_DATA_cached].sectors +=
 344                 (int) new.cached_sectors - (int) old.cached_sectors;
 345
 346         u->d[old.data_type].fragmented -= bucket_sectors_fragmented(ca, old);
 347         u->d[new.data_type].fragmented += bucket_sectors_fragmented(ca, new);
 348
 349         preempt_enable();
 350
 351         if (!is_available_bucket(old) && is_available_bucket(new))
 352                 bch2_wake_allocator(ca);
 353 }
 354
 355 static inline int __update_replicas(struct bch_fs *c,
 356                                     struct bch_fs_usage *fs_usage,
 357                                     struct bch_replicas_entry *r,
 358                                     s64 sectors)
 359 {
 360         int idx = bch2_replicas_entry_idx(c, r);
 361
 362         if (idx < 0)
 363                 return -1;
 364
 365         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 366         fs_usage->replicas[idx]         += sectors;
 367         return 0;
 368 }
 369
 370 static inline int update_replicas(struct bch_fs *c, struct bkey_s_c k,
 371                         struct bch_replicas_entry *r, s64 sectors,
 372                         unsigned journal_seq, bool gc)
 373 {
 374         struct bch_fs_usage __percpu *fs_usage;
 375         int idx, ret = 0;
 376         char buf[200];
 377
 378         percpu_down_read(&c->mark_lock);
 379
 380         idx = bch2_replicas_entry_idx(c, r);
 381         if (idx < 0 &&
 382             (test_bit(BCH_FS_REBUILD_REPLICAS, &c->flags) ||
 383              fsck_err(c, "no replicas entry\n"
 384                       "  while marking %s",
 385                       (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf)))) {
 386                 percpu_up_read(&c->mark_lock);
 387                 ret = bch2_mark_replicas(c, r);
 388                 if (ret)
 389                         return ret;
 390
 391                 percpu_down_read(&c->mark_lock);
 392                 idx = bch2_replicas_entry_idx(c, r);
 393         }
 394         if (idx < 0) {
 395                 ret = -1;
 396                 goto err;
 397         }
 398
 399         preempt_disable();
 400         fs_usage = fs_usage_ptr(c, journal_seq, gc);
 401         fs_usage_data_type_to_base(fs_usage, r->data_type, sectors);
 402         fs_usage->replicas[idx]         += sectors;
 403         preempt_enable();
 404 err:
 405 fsck_err:
 406         percpu_up_read(&c->mark_lock);
 407         return ret;
 408 }
 409
 410 static inline int update_cached_sectors(struct bch_fs *c,
 411                         struct bkey_s_c k,
 412                         unsigned dev, s64 sectors,
 413                         unsigned journal_seq, bool gc)
 414 {
 415         struct bch_replicas_padded r;
 416
 417         bch2_replicas_entry_cached(&r.e, dev);
 418
 419         return update_replicas(c, k, &r.e, sectors, journal_seq, gc);
 420 }
 421
 422 static struct replicas_delta_list *
 423 replicas_deltas_realloc(struct btree_trans *trans, unsigned more)
 424 {
 425         struct replicas_delta_list *d = trans->fs_usage_deltas;
 426         unsigned new_size = d ? (d->size + more) * 2 : 128;
 427         unsigned alloc_size = sizeof(*d) + new_size;
 428
 429         WARN_ON_ONCE(alloc_size > REPLICAS_DELTA_LIST_MAX);
 430
 431         if (!d || d->used + more > d->size) {
 432                 d = krealloc(d, alloc_size, GFP_NOIO|__GFP_ZERO);
 433
 434                 BUG_ON(!d && alloc_size > REPLICAS_DELTA_LIST_MAX);
 435
 436                 if (!d) {
 437                         d = mempool_alloc(&trans->c->replicas_delta_pool, GFP_NOIO);
 438                         memset(d, 0, REPLICAS_DELTA_LIST_MAX);
 439
 440                         if (trans->fs_usage_deltas)
 441                                 memcpy(d, trans->fs_usage_deltas,
 442                                        trans->fs_usage_deltas->size + sizeof(*d));
 443
 444                         new_size = REPLICAS_DELTA_LIST_MAX - sizeof(*d);
 445                         kfree(trans->fs_usage_deltas);
 446                 }
 447
 448                 d->size = new_size;
 449                 trans->fs_usage_deltas = d;
 450         }
 451         return d;
 452 }
 453
 454 static inline void update_replicas_list(struct btree_trans *trans,
 455                                         struct bch_replicas_entry *r,
 456                                         s64 sectors)
 457 {
 458         struct replicas_delta_list *d;
 459         struct replicas_delta *n;
 460         unsigned b;
 461
 462         if (!sectors)
 463                 return;
 464
 465         b = replicas_entry_bytes(r) + 8;
 466         d = replicas_deltas_realloc(trans, b);
 467
 468         n = (void *) d->d + d->used;
 469         n->delta = sectors;
 470         memcpy(&n->r, r, replicas_entry_bytes(r));
 471         bch2_replicas_entry_sort(&n->r);
 472         d->used += b;
 473 }
 474
 475 static inline void update_cached_sectors_list(struct btree_trans *trans,
 476                                               unsigned dev, s64 sectors)
 477 {
 478         struct bch_replicas_padded r;
 479
 480         bch2_replicas_entry_cached(&r.e, dev);
 481
 482         update_replicas_list(trans, &r.e, sectors);
 483 }
 484
 485 void bch2_mark_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 486                             size_t b, bool owned_by_allocator)
 487 {
 488         struct bucket *g = bucket(ca, b);
 489         struct bucket_mark old, new;
 490
 491         old = bucket_cmpxchg(g, new, ({
 492                 new.owned_by_allocator  = owned_by_allocator;
 493         }));
 494
 495         BUG_ON(owned_by_allocator == old.owned_by_allocator);
 496 }
 497
 498 static int bch2_mark_alloc(struct btree_trans *trans,
 499                            struct bkey_s_c old, struct bkey_s_c new,
 500                            unsigned flags)
 501 {
 502         bool gc = flags & BTREE_TRIGGER_GC;
 503         u64 journal_seq = trans->journal_res.seq;
 504         struct bch_fs *c = trans->c;
 505         struct bkey_alloc_unpacked old_u = bch2_alloc_unpack(old);
 506         struct bkey_alloc_unpacked new_u = bch2_alloc_unpack(new);
 507         struct bch_dev *ca;
 508         struct bucket *g;
 509         struct bucket_mark old_m, m;
 510         int ret = 0;
 511
 512         /*
 513          * alloc btree is read in by bch2_alloc_read, not gc:
 514          */
 515         if ((flags & BTREE_TRIGGER_GC) &&
 516             !(flags & BTREE_TRIGGER_BUCKET_INVALIDATE))
 517                 return 0;
 518
 519         if ((flags & BTREE_TRIGGER_INSERT) &&
 520             !old_u.data_type != !new_u.data_type &&
 521             new.k->type == KEY_TYPE_alloc_v3) {
 522                 struct bch_alloc_v3 *v = (struct bch_alloc_v3 *) new.v;
 523                 u64 old_journal_seq = le64_to_cpu(v->journal_seq);
 524
 525                 BUG_ON(!journal_seq);
 526
 527                 /*
 528                  * If the btree updates referring to a bucket weren't flushed
 529                  * before the bucket became empty again, then the we don't have
 530                  * to wait on a journal flush before we can reuse the bucket:
 531                  */
 532                 new_u.journal_seq = !new_u.data_type &&
 533                         (journal_seq == old_journal_seq ||
 534                          bch2_journal_noflush_seq(&c->journal, old_journal_seq))
 535                         ? 0 : journal_seq;
 536                 v->journal_seq = cpu_to_le64(new_u.journal_seq);
 537         }
 538
 539         if (old_u.data_type && !new_u.data_type && new_u.journal_seq) {
 540                 ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 541                                 c->journal.flushed_seq_ondisk,
 542                                 new_u.dev, new_u.bucket,
 543                                 new_u.journal_seq);
 544                 if (ret) {
 545                         bch2_fs_fatal_error(c,
 546                                 "error setting bucket_needs_journal_commit: %i", ret);
 547                         return ret;
 548                 }
 549         }
 550
 551         ca = bch_dev_bkey_exists(c, new_u.dev);
 552
 553         if (new_u.bucket >= ca->mi.nbuckets)
 554                 return 0;
 555
 556         percpu_down_read(&c->mark_lock);
 557         if (!gc && new_u.gen != old_u.gen)
 558                 *bucket_gen(ca, new_u.bucket) = new_u.gen;
 559
 560         g = __bucket(ca, new_u.bucket, gc);
 561
 562         old_m = bucket_cmpxchg(g, m, ({
 563                 m.gen                   = new_u.gen;
 564                 m.data_type             = new_u.data_type;
 565                 m.dirty_sectors         = new_u.dirty_sectors;
 566                 m.cached_sectors        = new_u.cached_sectors;
 567                 m.stripe                = new_u.stripe != 0;
 568         }));
 569
 570         bch2_dev_usage_update(c, ca, old_m, m, journal_seq, gc);
 571
 572         g->io_time[READ]        = new_u.read_time;
 573         g->io_time[WRITE]       = new_u.write_time;
 574         g->oldest_gen           = new_u.oldest_gen;
 575         g->gen_valid            = 1;
 576         g->stripe               = new_u.stripe;
 577         g->stripe_redundancy    = new_u.stripe_redundancy;
 578         percpu_up_read(&c->mark_lock);
 579
 580         /*
 581          * need to know if we're getting called from the invalidate path or
 582          * not:
 583          */
 584
 585         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 586             old_m.cached_sectors) {
 587                 ret = update_cached_sectors(c, new, ca->dev_idx,
 588                                             -old_m.cached_sectors,
 589                                             journal_seq, gc);
 590                 if (ret) {
 591                         bch2_fs_fatal_error(c, "bch2_mark_alloc(): no replicas entry while updating cached sectors");
 592                         return ret;
 593                 }
 594
 595                 trace_invalidate(ca, bucket_to_sector(ca, new_u.bucket),
 596                                  old_m.cached_sectors);
 597         }
 598
 599         return 0;
 600 }
 601
 602 #define checked_add(a, b)                                       \
 603 ({                                                              \
 604         unsigned _res = (unsigned) (a) + (b);                   \
 605         bool overflow = _res > U16_MAX;                         \
 606         if (overflow)                                           \
 607                 _res = U16_MAX;                                 \
 608         (a) = _res;                                             \
 609         overflow;                                               \
 610 })
 611
 612 void bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 613                                size_t b, enum bch_data_type data_type,
 614                                unsigned sectors, struct gc_pos pos,
 615                                unsigned flags)
 616 {
 617         struct bucket *g;
 618         struct bucket_mark old, new;
 619         bool overflow;
 620
 621         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 622         BUG_ON(data_type != BCH_DATA_sb &&
 623                data_type != BCH_DATA_journal);
 624
 625         /*
 626          * Backup superblock might be past the end of our normal usable space:
 627          */
 628         if (b >= ca->mi.nbuckets)
 629                 return;
 630
 631         percpu_down_read(&c->mark_lock);
 632         g = gc_bucket(ca, b);
 633         old = bucket_cmpxchg(g, new, ({
 634                 new.data_type   = data_type;
 635                 overflow = checked_add(new.dirty_sectors, sectors);
 636         }));
 637
 638         bch2_fs_inconsistent_on(old.data_type &&
 639                                 old.data_type != data_type, c,
 640                 "different types of data in same bucket: %s, %s",
 641                 bch2_data_types[old.data_type],
 642                 bch2_data_types[data_type]);
 643
 644         bch2_fs_inconsistent_on(overflow, c,
 645                 "bucket %u:%zu gen %u data type %s sector count overflow: %u + %u > U16_MAX",
 646                 ca->dev_idx, b, new.gen,
 647                 bch2_data_types[old.data_type ?: data_type],
 648                 old.dirty_sectors, sectors);
 649
 650         bch2_dev_usage_update(c, ca, old, new, 0, true);
 651         percpu_up_read(&c->mark_lock);
 652 }
 653
 654 static s64 ptr_disk_sectors(s64 sectors, struct extent_ptr_decoded p)
 655 {
 656         EBUG_ON(sectors < 0);
 657
 658         return p.crc.compression_type &&
 659                 p.crc.compression_type != BCH_COMPRESSION_TYPE_incompressible
 660                 ? DIV_ROUND_UP_ULL(sectors * p.crc.compressed_size,
 661                                p.crc.uncompressed_size)
 662                 : sectors;
 663 }
 664
 665 static int check_bucket_ref(struct bch_fs *c,
 666                             struct bkey_s_c k,
 667                             const struct bch_extent_ptr *ptr,
 668                             s64 sectors, enum bch_data_type ptr_data_type,
 669                             u8 bucket_gen, u8 bucket_data_type,
 670                             u16 dirty_sectors, u16 cached_sectors)
 671 {
 672         size_t bucket_nr = PTR_BUCKET_NR(bch_dev_bkey_exists(c, ptr->dev), ptr);
 673         u16 bucket_sectors = !ptr->cached
 674                 ? dirty_sectors
 675                 : cached_sectors;
 676         char buf[200];
 677
 678         if (gen_after(ptr->gen, bucket_gen)) {
 679                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 680                         "bucket %u:%zu gen %u data type %s: ptr gen %u newer than bucket gen\n"
 681                         "while marking %s",
 682                         ptr->dev, bucket_nr, bucket_gen,
 683                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 684                         ptr->gen,
 685                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 686                 return -EIO;
 687         }
 688
 689         if (gen_cmp(bucket_gen, ptr->gen) > BUCKET_GC_GEN_MAX) {
 690                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 691                         "bucket %u:%zu gen %u data type %s: ptr gen %u too stale\n"
 692                         "while marking %s",
 693                         ptr->dev, bucket_nr, bucket_gen,
 694                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 695                         ptr->gen,
 696                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 697                 return -EIO;
 698         }
 699
 700         if (bucket_gen != ptr->gen && !ptr->cached) {
 701                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 702                         "bucket %u:%zu gen %u data type %s: stale dirty ptr (gen %u)\n"
 703                         "while marking %s",
 704                         ptr->dev, bucket_nr, bucket_gen,
 705                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 706                         ptr->gen,
 707                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 708                 return -EIO;
 709         }
 710
 711         if (bucket_gen != ptr->gen)
 712                 return 1;
 713
 714         if (bucket_data_type && ptr_data_type &&
 715             bucket_data_type != ptr_data_type) {
 716                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 717                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s\n"
 718                         "while marking %s",
 719                         ptr->dev, bucket_nr, bucket_gen,
 720                         bch2_data_types[bucket_data_type],
 721                         bch2_data_types[ptr_data_type],
 722                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 723                 return -EIO;
 724         }
 725
 726         if ((unsigned) (bucket_sectors + sectors) > U16_MAX) {
 727                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 728                         "bucket %u:%zu gen %u data type %s sector count overflow: %u + %lli > U16_MAX\n"
 729                         "while marking %s",
 730                         ptr->dev, bucket_nr, bucket_gen,
 731                         bch2_data_types[bucket_data_type ?: ptr_data_type],
 732                         bucket_sectors, sectors,
 733                         (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 734                 return -EIO;
 735         }
 736
 737         return 0;
 738 }
 739
 740 static int mark_stripe_bucket(struct btree_trans *trans,
 741                               struct bkey_s_c k,
 742                               unsigned ptr_idx,
 743                               unsigned flags)
 744 {
 745         struct bch_fs *c = trans->c;
 746         u64 journal_seq = trans->journal_res.seq;
 747         const struct bch_stripe *s = bkey_s_c_to_stripe(k).v;
 748         unsigned nr_data = s->nr_blocks - s->nr_redundant;
 749         bool parity = ptr_idx >= nr_data;
 750         enum bch_data_type data_type = parity ? BCH_DATA_parity : 0;
 751         s64 sectors = parity ? le16_to_cpu(s->sectors) : 0;
 752         const struct bch_extent_ptr *ptr = s->ptrs + ptr_idx;
 753         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 754         struct bucket *g;
 755         struct bucket_mark new, old;
 756         char buf[200];
 757         int ret = 0;
 758
 759         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 760
 761         /* * XXX doesn't handle deletion */
 762
 763         percpu_down_read(&c->mark_lock);
 764         g = PTR_GC_BUCKET(ca, ptr);
 765
 766         if (g->mark.dirty_sectors ||
 767             (g->stripe && g->stripe != k.k->p.offset)) {
 768                 bch2_fs_inconsistent(c,
 769                               "bucket %u:%zu gen %u: multiple stripes using same bucket\n%s",
 770                               ptr->dev, PTR_BUCKET_NR(ca, ptr), g->mark.gen,
 771                               (bch2_bkey_val_to_text(&PBUF(buf), c, k), buf));
 772                 ret = -EINVAL;
 773                 goto err;
 774         }
 775
 776         old = bucket_cmpxchg(g, new, ({
 777                 ret = check_bucket_ref(c, k, ptr, sectors, data_type,
 778                                        new.gen, new.data_type,
 779                                        new.dirty_sectors, new.cached_sectors);
 780                 if (ret)
 781                         goto err;
 782
 783                 new.dirty_sectors += sectors;
 784                 if (data_type)
 785                         new.data_type           = data_type;
 786
 787                 new.stripe = true;
 788         }));
 789
 790         g->stripe               = k.k->p.offset;
 791         g->stripe_redundancy    = s->nr_redundant;
 792
 793         bch2_dev_usage_update(c, ca, old, new, journal_seq, true);
 794 err:
 795         percpu_up_read(&c->mark_lock);
 796
 797         return 0;
 798 }
 799
 800 static int __mark_pointer(struct btree_trans *trans,
 801                           struct bkey_s_c k,
 802                           const struct bch_extent_ptr *ptr,
 803                           s64 sectors, enum bch_data_type ptr_data_type,
 804                           u8 bucket_gen, u8 *bucket_data_type,
 805                           u16 *dirty_sectors, u16 *cached_sectors)
 806 {
 807         u16 *dst_sectors = !ptr->cached
 808                 ? dirty_sectors
 809                 : cached_sectors;
 810         int ret = check_bucket_ref(trans->c, k, ptr, sectors, ptr_data_type,
 811                                    bucket_gen, *bucket_data_type,
 812                                    *dirty_sectors, *cached_sectors);
 813
 814         if (ret)
 815                 return ret;
 816
 817         *dst_sectors += sectors;
 818         *bucket_data_type = *dirty_sectors || *cached_sectors
 819                 ? ptr_data_type : 0;
 820         return 0;
 821 }
 822
 823 static int bch2_mark_pointer(struct btree_trans *trans,
 824                              struct bkey_s_c k,
 825                              struct extent_ptr_decoded p,
 826                              s64 sectors, enum bch_data_type data_type,
 827                              unsigned flags)
 828 {
 829         u64 journal_seq = trans->journal_res.seq;
 830         struct bch_fs *c = trans->c;
 831         struct bucket_mark old, new;
 832         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 833         struct bucket *g;
 834         u8 bucket_data_type;
 835         u64 v;
 836         int ret = 0;
 837
 838         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 839
 840         percpu_down_read(&c->mark_lock);
 841         g = PTR_GC_BUCKET(ca, &p.ptr);
 842
 843         v = atomic64_read(&g->_mark.v);
 844         do {
 845                 new.v.counter = old.v.counter = v;
 846                 bucket_data_type = new.data_type;
 847
 848                 ret = __mark_pointer(trans, k, &p.ptr, sectors,
 849                                      data_type, new.gen,
 850                                      &bucket_data_type,
 851                                      &new.dirty_sectors,
 852                                      &new.cached_sectors);
 853                 if (ret)
 854                         goto err;
 855
 856                 new.data_type = bucket_data_type;
 857
 858                 if (flags & BTREE_TRIGGER_NOATOMIC) {
 859                         g->_mark = new;
 860                         break;
 861                 }
 862         } while ((v = atomic64_cmpxchg(&g->_mark.v,
 863                               old.v.counter,
 864                               new.v.counter)) != old.v.counter);
 865
 866         bch2_dev_usage_update(c, ca, old, new, journal_seq, true);
 867 err:
 868         percpu_up_read(&c->mark_lock);
 869
 870         return ret;
 871 }
 872
 873 static int bch2_mark_stripe_ptr(struct btree_trans *trans,
 874                                 struct bkey_s_c k,
 875                                 struct bch_extent_stripe_ptr p,
 876                                 enum bch_data_type data_type,
 877                                 s64 sectors,
 878                                 unsigned flags)
 879 {
 880         struct bch_fs *c = trans->c;
 881         struct bch_replicas_padded r;
 882         struct gc_stripe *m;
 883
 884         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 885
 886         m = genradix_ptr_alloc(&c->gc_stripes, p.idx, GFP_KERNEL);
 887         if (!m) {
 888                 bch_err(c, "error allocating memory for gc_stripes, idx %llu",
 889                         (u64) p.idx);
 890                 return -ENOMEM;
 891         }
 892
 893         spin_lock(&c->ec_stripes_heap_lock);
 894
 895         if (!m || !m->alive) {
 896                 spin_unlock(&c->ec_stripes_heap_lock);
 897                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
 898                                     (u64) p.idx);
 899                 bch2_inconsistent_error(c);
 900                 return -EIO;
 901         }
 902
 903         m->block_sectors[p.block] += sectors;
 904
 905         r = m->r;
 906         spin_unlock(&c->ec_stripes_heap_lock);
 907
 908         r.e.data_type = data_type;
 909         update_replicas(c, k, &r.e, sectors, trans->journal_res.seq, true);
 910
 911         return 0;
 912 }
 913
 914 static int bch2_mark_extent(struct btree_trans *trans,
 915                             struct bkey_s_c old, struct bkey_s_c new,
 916                             unsigned flags)
 917 {
 918         u64 journal_seq = trans->journal_res.seq;
 919         struct bch_fs *c = trans->c;
 920         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
 921         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
 922         const union bch_extent_entry *entry;
 923         struct extent_ptr_decoded p;
 924         struct bch_replicas_padded r;
 925         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
 926                 ? BCH_DATA_btree
 927                 : BCH_DATA_user;
 928         s64 sectors = bkey_is_btree_ptr(k.k)
 929                 ? btree_sectors(c)
 930                 : k.k->size;
 931         s64 dirty_sectors = 0;
 932         bool stale;
 933         int ret;
 934
 935         BUG_ON(!(flags & BTREE_TRIGGER_GC));
 936
 937         r.e.data_type   = data_type;
 938         r.e.nr_devs     = 0;
 939         r.e.nr_required = 1;
 940
 941         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
 942                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
 943
 944                 if (flags & BTREE_TRIGGER_OVERWRITE)
 945                         disk_sectors = -disk_sectors;
 946
 947                 ret = bch2_mark_pointer(trans, k, p, disk_sectors,
 948                                         data_type, flags);
 949                 if (ret < 0)
 950                         return ret;
 951
 952                 stale = ret > 0;
 953
 954                 if (p.ptr.cached) {
 955                         if (!stale) {
 956                                 ret = update_cached_sectors(c, k, p.ptr.dev,
 957                                                 disk_sectors, journal_seq, true);
 958                                 if (ret) {
 959                                         bch2_fs_fatal_error(c, "bch2_mark_extent(): no replicas entry while updating cached sectors");
 960                                         return ret;
 961                                 }
 962                         }
 963                 } else if (!p.has_ec) {
 964                         dirty_sectors          += disk_sectors;
 965                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
 966                 } else {
 967                         ret = bch2_mark_stripe_ptr(trans, k, p.ec, data_type,
 968                                         disk_sectors, flags);
 969                         if (ret)
 970                                 return ret;
 971
 972                         /*
 973                          * There may be other dirty pointers in this extent, but
 974                          * if so they're not required for mounting if we have an
 975                          * erasure coded pointer in this extent:
 976                          */
 977                         r.e.nr_required = 0;
 978                 }
 979         }
 980
 981         if (r.e.nr_devs) {
 982                 ret = update_replicas(c, k, &r.e, dirty_sectors, journal_seq, true);
 983                 if (ret) {
 984                         char buf[200];
 985
 986                         bch2_bkey_val_to_text(&PBUF(buf), c, k);
 987                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf);
 988                         return ret;
 989                 }
 990         }
 991
 992         return 0;
 993 }
 994
 995 static int bch2_mark_stripe(struct btree_trans *trans,
 996                             struct bkey_s_c old, struct bkey_s_c new,
 997                             unsigned flags)
 998 {
 999         bool gc = flags & BTREE_TRIGGER_GC;
1000         u64 journal_seq = trans->journal_res.seq;
1001         struct bch_fs *c = trans->c;
1002         u64 idx = new.k->p.offset;
1003         const struct bch_stripe *old_s = old.k->type == KEY_TYPE_stripe
1004                 ? bkey_s_c_to_stripe(old).v : NULL;
1005         const struct bch_stripe *new_s = new.k->type == KEY_TYPE_stripe
1006                 ? bkey_s_c_to_stripe(new).v : NULL;
1007         unsigned i;
1008         int ret;
1009
1010         BUG_ON(gc && old_s);
1011
1012         if (!gc) {
1013                 struct stripe *m = genradix_ptr(&c->stripes, idx);
1014
1015                 if (!m || (old_s && !m->alive)) {
1016                         char buf1[200], buf2[200];
1017
1018                         bch2_bkey_val_to_text(&PBUF(buf1), c, old);
1019                         bch2_bkey_val_to_text(&PBUF(buf2), c, new);
1020                         bch_err_ratelimited(c, "error marking nonexistent stripe %llu while marking\n"
1021                                             "old %s\n"
1022                                             "new %s", idx, buf1, buf2);
1023                         bch2_inconsistent_error(c);
1024                         return -1;
1025                 }
1026
1027                 if (!new_s) {
1028                         spin_lock(&c->ec_stripes_heap_lock);
1029                         bch2_stripes_heap_del(c, m, idx);
1030                         spin_unlock(&c->ec_stripes_heap_lock);
1031
1032                         memset(m, 0, sizeof(*m));
1033                 } else {
1034                         m->alive        = true;
1035                         m->sectors      = le16_to_cpu(new_s->sectors);
1036                         m->algorithm    = new_s->algorithm;
1037                         m->nr_blocks    = new_s->nr_blocks;
1038                         m->nr_redundant = new_s->nr_redundant;
1039                         m->blocks_nonempty = 0;
1040
1041                         for (i = 0; i < new_s->nr_blocks; i++)
1042                                 m->blocks_nonempty += !!stripe_blockcount_get(new_s, i);
1043
1044                         spin_lock(&c->ec_stripes_heap_lock);
1045                         bch2_stripes_heap_update(c, m, idx);
1046                         spin_unlock(&c->ec_stripes_heap_lock);
1047                 }
1048         } else {
1049                 struct gc_stripe *m =
1050                         genradix_ptr_alloc(&c->gc_stripes, idx, GFP_KERNEL);
1051
1052                 if (!m) {
1053                         bch_err(c, "error allocating memory for gc_stripes, idx %llu",
1054                                 idx);
1055                         return -ENOMEM;
1056                 }
1057                 /*
1058                  * This will be wrong when we bring back runtime gc: we should
1059                  * be unmarking the old key and then marking the new key
1060                  */
1061                 m->alive        = true;
1062                 m->sectors      = le16_to_cpu(new_s->sectors);
1063                 m->nr_blocks    = new_s->nr_blocks;
1064                 m->nr_redundant = new_s->nr_redundant;
1065
1066                 for (i = 0; i < new_s->nr_blocks; i++)
1067                         m->ptrs[i] = new_s->ptrs[i];
1068
1069                 bch2_bkey_to_replicas(&m->r.e, new);
1070
1071                 /*
1072                  * gc recalculates this field from stripe ptr
1073                  * references:
1074                  */
1075                 memset(m->block_sectors, 0, sizeof(m->block_sectors));
1076
1077                 for (i = 0; i < new_s->nr_blocks; i++) {
1078                         ret = mark_stripe_bucket(trans, new, i, flags);
1079                         if (ret)
1080                                 return ret;
1081                 }
1082
1083                 ret = update_replicas(c, new, &m->r.e,
1084                                       ((s64) m->sectors * m->nr_redundant),
1085                                       journal_seq, gc);
1086                 if (ret) {
1087                         char buf[200];
1088
1089                         bch2_bkey_val_to_text(&PBUF(buf), c, new);
1090                         bch2_fs_fatal_error(c, "no replicas entry for %s", buf);
1091                         return ret;
1092                 }
1093         }
1094
1095         return 0;
1096 }
1097
1098 static int bch2_mark_inode(struct btree_trans *trans,
1099                            struct bkey_s_c old, struct bkey_s_c new,
1100                            unsigned flags)
1101 {
1102         struct bch_fs *c = trans->c;
1103         struct bch_fs_usage __percpu *fs_usage;
1104         u64 journal_seq = trans->journal_res.seq;
1105
1106         if (flags & BTREE_TRIGGER_INSERT) {
1107                 struct bch_inode_v2 *v = (struct bch_inode_v2 *) new.v;
1108
1109                 BUG_ON(!journal_seq);
1110                 BUG_ON(new.k->type != KEY_TYPE_inode_v2);
1111
1112                 v->bi_journal_seq = cpu_to_le64(journal_seq);
1113         }
1114
1115         if (flags & BTREE_TRIGGER_GC) {
1116                 percpu_down_read(&c->mark_lock);
1117                 preempt_disable();
1118
1119                 fs_usage = fs_usage_ptr(c, journal_seq, flags & BTREE_TRIGGER_GC);
1120                 fs_usage->nr_inodes += bkey_is_inode(new.k);
1121                 fs_usage->nr_inodes -= bkey_is_inode(old.k);
1122
1123                 preempt_enable();
1124                 percpu_up_read(&c->mark_lock);
1125         }
1126         return 0;
1127 }
1128
1129 static int bch2_mark_reservation(struct btree_trans *trans,
1130                                  struct bkey_s_c old, struct bkey_s_c new,
1131                                  unsigned flags)
1132 {
1133         struct bch_fs *c = trans->c;
1134         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1135         struct bch_fs_usage __percpu *fs_usage;
1136         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1137         s64 sectors = (s64) k.k->size;
1138
1139         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1140
1141         if (flags & BTREE_TRIGGER_OVERWRITE)
1142                 sectors = -sectors;
1143         sectors *= replicas;
1144
1145         percpu_down_read(&c->mark_lock);
1146         preempt_disable();
1147
1148         fs_usage = fs_usage_ptr(c, trans->journal_res.seq, flags & BTREE_TRIGGER_GC);
1149         replicas = clamp_t(unsigned, replicas, 1,
1150                            ARRAY_SIZE(fs_usage->persistent_reserved));
1151
1152         fs_usage->reserved                              += sectors;
1153         fs_usage->persistent_reserved[replicas - 1]     += sectors;
1154
1155         preempt_enable();
1156         percpu_up_read(&c->mark_lock);
1157
1158         return 0;
1159 }
1160
1161 static s64 __bch2_mark_reflink_p(struct bch_fs *c, struct bkey_s_c_reflink_p p,
1162                                  u64 *idx, unsigned flags, size_t r_idx)
1163 {
1164         struct reflink_gc *r;
1165         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1166         s64 ret = 0;
1167
1168         if (r_idx >= c->reflink_gc_nr)
1169                 goto not_found;
1170
1171         r = genradix_ptr(&c->reflink_gc_table, r_idx);
1172         if (*idx < r->offset - r->size)
1173                 goto not_found;
1174
1175         BUG_ON((s64) r->refcount + add < 0);
1176
1177         r->refcount += add;
1178         *idx = r->offset;
1179         return 0;
1180 not_found:
1181         *idx = U64_MAX;
1182         ret = -EIO;
1183
1184         /*
1185          * XXX: we're replacing the entire reflink pointer with an error
1186          * key, we should just be replacing the part that was missing:
1187          */
1188         if (fsck_err(c, "%llu:%llu len %u points to nonexistent indirect extent %llu",
1189                      p.k->p.inode, p.k->p.offset, p.k->size, *idx)) {
1190                 struct bkey_i_error new;
1191
1192                 bkey_init(&new.k);
1193                 new.k.type      = KEY_TYPE_error;
1194                 new.k.p         = p.k->p;
1195                 new.k.size      = p.k->size;
1196                 ret = bch2_journal_key_insert(c, BTREE_ID_extents, 0, &new.k_i);
1197         }
1198 fsck_err:
1199         return ret;
1200 }
1201
1202 static int bch2_mark_reflink_p(struct btree_trans *trans,
1203                                struct bkey_s_c old, struct bkey_s_c new,
1204                                unsigned flags)
1205 {
1206         struct bch_fs *c = trans->c;
1207         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1208         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1209         struct reflink_gc *ref;
1210         size_t l, r, m;
1211         u64 idx = le64_to_cpu(p.v->idx);
1212         u64 end = le64_to_cpu(p.v->idx) + p.k->size;
1213         int ret = 0;
1214
1215         BUG_ON(!(flags & BTREE_TRIGGER_GC));
1216
1217         if (c->sb.version >= bcachefs_metadata_version_reflink_p_fix) {
1218                 idx -= le32_to_cpu(p.v->front_pad);
1219                 end += le32_to_cpu(p.v->back_pad);
1220         }
1221
1222         l = 0;
1223         r = c->reflink_gc_nr;
1224         while (l < r) {
1225                 m = l + (r - l) / 2;
1226
1227                 ref = genradix_ptr(&c->reflink_gc_table, m);
1228                 if (ref->offset <= idx)
1229                         l = m + 1;
1230                 else
1231                         r = m;
1232         }
1233
1234         while (idx < end && !ret)
1235                 ret = __bch2_mark_reflink_p(c, p, &idx, flags, l++);
1236
1237         return ret;
1238 }
1239
1240 int bch2_mark_key(struct btree_trans *trans,
1241                   struct bkey_s_c old,
1242                   struct bkey_s_c new,
1243                   unsigned flags)
1244 {
1245         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1246
1247         switch (k.k->type) {
1248         case KEY_TYPE_alloc:
1249         case KEY_TYPE_alloc_v2:
1250         case KEY_TYPE_alloc_v3:
1251                 return bch2_mark_alloc(trans, old, new, flags);
1252         case KEY_TYPE_btree_ptr:
1253         case KEY_TYPE_btree_ptr_v2:
1254         case KEY_TYPE_extent:
1255         case KEY_TYPE_reflink_v:
1256                 return bch2_mark_extent(trans, old, new, flags);
1257         case KEY_TYPE_stripe:
1258                 return bch2_mark_stripe(trans, old, new, flags);
1259         case KEY_TYPE_inode:
1260         case KEY_TYPE_inode_v2:
1261                 return bch2_mark_inode(trans, old, new, flags);
1262         case KEY_TYPE_reservation:
1263                 return bch2_mark_reservation(trans, old, new, flags);
1264         case KEY_TYPE_reflink_p:
1265                 return bch2_mark_reflink_p(trans, old, new, flags);
1266         case KEY_TYPE_snapshot:
1267                 return bch2_mark_snapshot(trans, old, new, flags);
1268         default:
1269                 return 0;
1270         }
1271 }
1272
1273 int bch2_mark_update(struct btree_trans *trans, struct btree_path *path,
1274                      struct bkey_i *new, unsigned flags)
1275 {
1276         struct bkey             _deleted = KEY(0, 0, 0);
1277         struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
1278         struct bkey_s_c         old;
1279         struct bkey             unpacked;
1280         int ret;
1281
1282         _deleted.p = path->pos;
1283
1284         if (unlikely(flags & BTREE_TRIGGER_NORUN))
1285                 return 0;
1286
1287         if (!btree_node_type_needs_gc(path->btree_id))
1288                 return 0;
1289
1290         old = bch2_btree_path_peek_slot(path, &unpacked);
1291
1292         if (old.k->type == new->k.type &&
1293             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
1294                 ret   = bch2_mark_key(trans, old, bkey_i_to_s_c(new),
1295                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
1296         } else {
1297                 ret   = bch2_mark_key(trans, deleted, bkey_i_to_s_c(new),
1298                                 BTREE_TRIGGER_INSERT|flags) ?:
1299                         bch2_mark_key(trans, old, deleted,
1300                                 BTREE_TRIGGER_OVERWRITE|flags);
1301         }
1302
1303         return ret;
1304 }
1305
1306 static noinline __cold
1307 void fs_usage_apply_warn(struct btree_trans *trans,
1308                          unsigned disk_res_sectors,
1309                          s64 should_not_have_added)
1310 {
1311         struct bch_fs *c = trans->c;
1312         struct btree_insert_entry *i;
1313         char buf[200];
1314
1315         bch_err(c, "disk usage increased %lli more than %u sectors reserved",
1316                 should_not_have_added, disk_res_sectors);
1317
1318         trans_for_each_update(trans, i) {
1319                 pr_err("while inserting");
1320                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
1321                 pr_err("%s", buf);
1322                 pr_err("overlapping with");
1323
1324                 if (!i->cached) {
1325                         struct bkey u;
1326                         struct bkey_s_c k = bch2_btree_path_peek_slot(i->path, &u);
1327
1328                         bch2_bkey_val_to_text(&PBUF(buf), c, k);
1329                         pr_err("%s", buf);
1330                 } else {
1331                         struct bkey_cached *ck = (void *) i->path->l[0].b;
1332
1333                         if (ck->valid) {
1334                                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(ck->k));
1335                                 pr_err("%s", buf);
1336                         }
1337                 }
1338         }
1339         __WARN();
1340 }
1341
1342 int bch2_trans_fs_usage_apply(struct btree_trans *trans,
1343                               struct replicas_delta_list *deltas)
1344 {
1345         struct bch_fs *c = trans->c;
1346         static int warned_disk_usage = 0;
1347         bool warn = false;
1348         unsigned disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1349         struct replicas_delta *d = deltas->d, *d2;
1350         struct replicas_delta *top = (void *) deltas->d + deltas->used;
1351         struct bch_fs_usage *dst;
1352         s64 added = 0, should_not_have_added;
1353         unsigned i;
1354
1355         percpu_down_read(&c->mark_lock);
1356         preempt_disable();
1357         dst = fs_usage_ptr(c, trans->journal_res.seq, false);
1358
1359         for (d = deltas->d; d != top; d = replicas_delta_next(d)) {
1360                 switch (d->r.data_type) {
1361                 case BCH_DATA_btree:
1362                 case BCH_DATA_user:
1363                 case BCH_DATA_parity:
1364                         added += d->delta;
1365                 }
1366
1367                 if (__update_replicas(c, dst, &d->r, d->delta))
1368                         goto need_mark;
1369         }
1370
1371         dst->nr_inodes += deltas->nr_inodes;
1372
1373         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
1374                 added                           += deltas->persistent_reserved[i];
1375                 dst->reserved                   += deltas->persistent_reserved[i];
1376                 dst->persistent_reserved[i]     += deltas->persistent_reserved[i];
1377         }
1378
1379         /*
1380          * Not allowed to reduce sectors_available except by getting a
1381          * reservation:
1382          */
1383         should_not_have_added = added - (s64) disk_res_sectors;
1384         if (unlikely(should_not_have_added > 0)) {
1385                 u64 old, new, v = atomic64_read(&c->sectors_available);
1386
1387                 do {
1388                         old = v;
1389                         new = max_t(s64, 0, old - should_not_have_added);
1390                 } while ((v = atomic64_cmpxchg(&c->sectors_available,
1391                                                old, new)) != old);
1392
1393                 added -= should_not_have_added;
1394                 warn = true;
1395         }
1396
1397         if (added > 0) {
1398                 trans->disk_res->sectors -= added;
1399                 this_cpu_sub(*c->online_reserved, added);
1400         }
1401
1402         preempt_enable();
1403         percpu_up_read(&c->mark_lock);
1404
1405         if (unlikely(warn) && !xchg(&warned_disk_usage, 1))
1406                 fs_usage_apply_warn(trans, disk_res_sectors, should_not_have_added);
1407         return 0;
1408 need_mark:
1409         /* revert changes: */
1410         for (d2 = deltas->d; d2 != d; d2 = replicas_delta_next(d2))
1411                 BUG_ON(__update_replicas(c, dst, &d2->r, -d2->delta));
1412
1413         preempt_enable();
1414         percpu_up_read(&c->mark_lock);
1415         return -1;
1416 }
1417
1418 /* trans_mark: */
1419
1420 static int bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
1421                               const struct bch_extent_ptr *ptr,
1422                               struct bkey_alloc_unpacked *u)
1423 {
1424         struct bch_fs *c = trans->c;
1425         struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
1426         struct bkey_s_c k;
1427         int ret;
1428
1429         bch2_trans_iter_init(trans, iter, BTREE_ID_alloc,
1430                              POS(ptr->dev, PTR_BUCKET_NR(ca, ptr)),
1431                              BTREE_ITER_WITH_UPDATES|
1432                              BTREE_ITER_CACHED|
1433                              BTREE_ITER_INTENT);
1434         k = bch2_btree_iter_peek_slot(iter);
1435         ret = bkey_err(k);
1436         if (ret) {
1437                 bch2_trans_iter_exit(trans, iter);
1438                 return ret;
1439         }
1440
1441         *u = bch2_alloc_unpack(k);
1442         return 0;
1443 }
1444
1445 static int bch2_trans_mark_pointer(struct btree_trans *trans,
1446                         struct bkey_s_c k, struct extent_ptr_decoded p,
1447                         s64 sectors, enum bch_data_type data_type)
1448 {
1449         struct btree_iter iter;
1450         struct bkey_alloc_unpacked u;
1451         int ret;
1452
1453         ret = bch2_trans_start_alloc_update(trans, &iter, &p.ptr, &u);
1454         if (ret)
1455                 return ret;
1456
1457         ret = __mark_pointer(trans, k, &p.ptr, sectors, data_type,
1458                              u.gen, &u.data_type,
1459                              &u.dirty_sectors, &u.cached_sectors);
1460         if (ret)
1461                 goto out;
1462
1463         ret = bch2_alloc_write(trans, &iter, &u, 0);
1464         if (ret)
1465                 goto out;
1466 out:
1467         bch2_trans_iter_exit(trans, &iter);
1468         return ret;
1469 }
1470
1471 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1472                         struct extent_ptr_decoded p,
1473                         s64 sectors, enum bch_data_type data_type)
1474 {
1475         struct bch_fs *c = trans->c;
1476         struct btree_iter iter;
1477         struct bkey_s_c k;
1478         struct bkey_i_stripe *s;
1479         struct bch_replicas_padded r;
1480         int ret = 0;
1481
1482         bch2_trans_iter_init(trans, &iter, BTREE_ID_stripes, POS(0, p.ec.idx),
1483                              BTREE_ITER_INTENT|
1484                              BTREE_ITER_WITH_UPDATES);
1485         k = bch2_btree_iter_peek_slot(&iter);
1486         ret = bkey_err(k);
1487         if (ret)
1488                 goto err;
1489
1490         if (k.k->type != KEY_TYPE_stripe) {
1491                 bch2_fs_inconsistent(c,
1492                         "pointer to nonexistent stripe %llu",
1493                         (u64) p.ec.idx);
1494                 bch2_inconsistent_error(c);
1495                 ret = -EIO;
1496                 goto err;
1497         }
1498
1499         if (!bch2_ptr_matches_stripe(bkey_s_c_to_stripe(k).v, p)) {
1500                 bch2_fs_inconsistent(c,
1501                         "stripe pointer doesn't match stripe %llu",
1502                         (u64) p.ec.idx);
1503                 ret = -EIO;
1504                 goto err;
1505         }
1506
1507         s = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1508         ret = PTR_ERR_OR_ZERO(s);
1509         if (ret)
1510                 goto err;
1511
1512         bkey_reassemble(&s->k_i, k);
1513         stripe_blockcount_set(&s->v, p.ec.block,
1514                 stripe_blockcount_get(&s->v, p.ec.block) +
1515                 sectors);
1516
1517         ret = bch2_trans_update(trans, &iter, &s->k_i, 0);
1518         if (ret)
1519                 goto err;
1520
1521         bch2_bkey_to_replicas(&r.e, bkey_i_to_s_c(&s->k_i));
1522         r.e.data_type = data_type;
1523         update_replicas_list(trans, &r.e, sectors);
1524 err:
1525         bch2_trans_iter_exit(trans, &iter);
1526         return ret;
1527 }
1528
1529 static int bch2_trans_mark_extent(struct btree_trans *trans,
1530                         struct bkey_s_c k, unsigned flags)
1531 {
1532         struct bch_fs *c = trans->c;
1533         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1534         const union bch_extent_entry *entry;
1535         struct extent_ptr_decoded p;
1536         struct bch_replicas_padded r;
1537         enum bch_data_type data_type = bkey_is_btree_ptr(k.k)
1538                 ? BCH_DATA_btree
1539                 : BCH_DATA_user;
1540         s64 sectors = bkey_is_btree_ptr(k.k)
1541                 ? btree_sectors(c)
1542                 : k.k->size;
1543         s64 dirty_sectors = 0;
1544         bool stale;
1545         int ret;
1546
1547         r.e.data_type   = data_type;
1548         r.e.nr_devs     = 0;
1549         r.e.nr_required = 1;
1550
1551         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1552                 s64 disk_sectors = ptr_disk_sectors(sectors, p);
1553
1554                 if (flags & BTREE_TRIGGER_OVERWRITE)
1555                         disk_sectors = -disk_sectors;
1556
1557                 ret = bch2_trans_mark_pointer(trans, k, p,
1558                                         disk_sectors, data_type);
1559                 if (ret < 0)
1560                         return ret;
1561
1562                 stale = ret > 0;
1563
1564                 if (p.ptr.cached) {
1565                         if (!stale)
1566                                 update_cached_sectors_list(trans, p.ptr.dev,
1567                                                            disk_sectors);
1568                 } else if (!p.has_ec) {
1569                         dirty_sectors          += disk_sectors;
1570                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1571                 } else {
1572                         ret = bch2_trans_mark_stripe_ptr(trans, p,
1573                                         disk_sectors, data_type);
1574                         if (ret)
1575                                 return ret;
1576
1577                         r.e.nr_required = 0;
1578                 }
1579         }
1580
1581         if (r.e.nr_devs)
1582                 update_replicas_list(trans, &r.e, dirty_sectors);
1583
1584         return 0;
1585 }
1586
1587 static int bch2_trans_mark_stripe_bucket(struct btree_trans *trans,
1588                                          struct bkey_s_c_stripe s,
1589                                          unsigned idx, bool deleting)
1590 {
1591         struct bch_fs *c = trans->c;
1592         const struct bch_extent_ptr *ptr = &s.v->ptrs[idx];
1593         struct btree_iter iter;
1594         struct bkey_alloc_unpacked u;
1595         enum bch_data_type data_type = idx >= s.v->nr_blocks - s.v->nr_redundant
1596                 ? BCH_DATA_parity : 0;
1597         s64 sectors = data_type ? le16_to_cpu(s.v->sectors) : 0;
1598         int ret = 0;
1599
1600         if (deleting)
1601                 sectors = -sectors;
1602
1603         ret = bch2_trans_start_alloc_update(trans, &iter, ptr, &u);
1604         if (ret)
1605                 return ret;
1606
1607         ret = check_bucket_ref(c, s.s_c, ptr, sectors, data_type,
1608                                u.gen, u.data_type,
1609                                u.dirty_sectors, u.cached_sectors);
1610         if (ret)
1611                 goto err;
1612
1613         if (!deleting) {
1614                 if (bch2_fs_inconsistent_on(u.stripe ||
1615                                             u.stripe_redundancy, c,
1616                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: multiple stripes using same bucket (%u, %llu)",
1617                                 iter.pos.inode, iter.pos.offset, u.gen,
1618                                 bch2_data_types[u.data_type],
1619                                 u.dirty_sectors,
1620                                 u.stripe, s.k->p.offset)) {
1621                         ret = -EIO;
1622                         goto err;
1623                 }
1624
1625                 if (bch2_fs_inconsistent_on(data_type && u.dirty_sectors, c,
1626                                 "bucket %llu:%llu gen %u data type %s dirty_sectors %u: data already in stripe bucket %llu",
1627                                 iter.pos.inode, iter.pos.offset, u.gen,
1628                                 bch2_data_types[u.data_type],
1629                                 u.dirty_sectors,
1630                                 s.k->p.offset)) {
1631                         ret = -EIO;
1632                         goto err;
1633                 }
1634
1635                 u.stripe                = s.k->p.offset;
1636                 u.stripe_redundancy     = s.v->nr_redundant;
1637         } else {
1638                 if (bch2_fs_inconsistent_on(u.stripe != s.k->p.offset ||
1639                                             u.stripe_redundancy != s.v->nr_redundant, c,
1640                                 "bucket %llu:%llu gen %u: not marked as stripe when deleting stripe %llu (got %u)",
1641                                 iter.pos.inode, iter.pos.offset, u.gen,
1642                                 s.k->p.offset, u.stripe)) {
1643                         ret = -EIO;
1644                         goto err;
1645                 }
1646
1647                 u.stripe                = 0;
1648                 u.stripe_redundancy     = 0;
1649         }
1650
1651         u.dirty_sectors += sectors;
1652         if (data_type)
1653                 u.data_type = !deleting ? data_type : 0;
1654
1655         ret = bch2_alloc_write(trans, &iter, &u, 0);
1656         if (ret)
1657                 goto err;
1658 err:
1659         bch2_trans_iter_exit(trans, &iter);
1660         return ret;
1661 }
1662
1663 static int bch2_trans_mark_stripe(struct btree_trans *trans,
1664                                   struct bkey_s_c old, struct bkey_s_c new,
1665                                   unsigned flags)
1666 {
1667         struct bkey_s_c_stripe old_s = { .k = NULL };
1668         struct bkey_s_c_stripe new_s = { .k = NULL };
1669         struct bch_replicas_padded r;
1670         unsigned i, nr_blocks;
1671         int ret = 0;
1672
1673         if (old.k->type == KEY_TYPE_stripe)
1674                 old_s = bkey_s_c_to_stripe(old);
1675         if (new.k->type == KEY_TYPE_stripe)
1676                 new_s = bkey_s_c_to_stripe(new);
1677
1678         /*
1679          * If the pointers aren't changing, we don't need to do anything:
1680          */
1681         if (new_s.k && old_s.k &&
1682             new_s.v->nr_blocks          == old_s.v->nr_blocks &&
1683             new_s.v->nr_redundant       == old_s.v->nr_redundant &&
1684             !memcmp(old_s.v->ptrs, new_s.v->ptrs,
1685                     new_s.v->nr_blocks * sizeof(struct bch_extent_ptr)))
1686                 return 0;
1687
1688         BUG_ON(new_s.k && old_s.k &&
1689                (new_s.v->nr_blocks      != old_s.v->nr_blocks ||
1690                 new_s.v->nr_redundant   != old_s.v->nr_redundant));
1691
1692         nr_blocks = new_s.k ? new_s.v->nr_blocks : old_s.v->nr_blocks;
1693
1694         if (new_s.k) {
1695                 s64 sectors = le16_to_cpu(new_s.v->sectors);
1696
1697                 bch2_bkey_to_replicas(&r.e, new);
1698                 update_replicas_list(trans, &r.e, sectors * new_s.v->nr_redundant);
1699         }
1700
1701         if (old_s.k) {
1702                 s64 sectors = -((s64) le16_to_cpu(old_s.v->sectors));
1703
1704                 bch2_bkey_to_replicas(&r.e, old);
1705                 update_replicas_list(trans, &r.e, sectors * old_s.v->nr_redundant);
1706         }
1707
1708         for (i = 0; i < nr_blocks; i++) {
1709                 if (new_s.k && old_s.k &&
1710                     !memcmp(&new_s.v->ptrs[i],
1711                             &old_s.v->ptrs[i],
1712                             sizeof(new_s.v->ptrs[i])))
1713                         continue;
1714
1715                 if (new_s.k) {
1716                         ret = bch2_trans_mark_stripe_bucket(trans, new_s, i, false);
1717                         if (ret)
1718                                 break;
1719                 }
1720
1721                 if (old_s.k) {
1722                         ret = bch2_trans_mark_stripe_bucket(trans, old_s, i, true);
1723                         if (ret)
1724                                 break;
1725                 }
1726         }
1727
1728         return ret;
1729 }
1730
1731 static int bch2_trans_mark_inode(struct btree_trans *trans,
1732                                  struct bkey_s_c old,
1733                                  struct bkey_s_c new,
1734                                  unsigned flags)
1735 {
1736         int nr = bkey_is_inode(new.k) - bkey_is_inode(old.k);
1737
1738         if (nr) {
1739                 struct replicas_delta_list *d =
1740                         replicas_deltas_realloc(trans, 0);
1741                 d->nr_inodes += nr;
1742         }
1743
1744         return 0;
1745 }
1746
1747 static int bch2_trans_mark_reservation(struct btree_trans *trans,
1748                                        struct bkey_s_c k, unsigned flags)
1749 {
1750         unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1751         s64 sectors = (s64) k.k->size;
1752         struct replicas_delta_list *d;
1753
1754         if (flags & BTREE_TRIGGER_OVERWRITE)
1755                 sectors = -sectors;
1756         sectors *= replicas;
1757
1758         d = replicas_deltas_realloc(trans, 0);
1759
1760         replicas = clamp_t(unsigned, replicas, 1,
1761                            ARRAY_SIZE(d->persistent_reserved));
1762
1763         d->persistent_reserved[replicas - 1] += sectors;
1764         return 0;
1765 }
1766
1767 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1768                         struct bkey_s_c_reflink_p p,
1769                         u64 *idx, unsigned flags)
1770 {
1771         struct bch_fs *c = trans->c;
1772         struct btree_iter iter;
1773         struct bkey_s_c k;
1774         struct bkey_i *n;
1775         __le64 *refcount;
1776         int add = !(flags & BTREE_TRIGGER_OVERWRITE) ? 1 : -1;
1777         char buf[200];
1778         int ret;
1779
1780         bch2_trans_iter_init(trans, &iter, BTREE_ID_reflink, POS(0, *idx),
1781                              BTREE_ITER_INTENT|
1782                              BTREE_ITER_WITH_UPDATES);
1783         k = bch2_btree_iter_peek_slot(&iter);
1784         ret = bkey_err(k);
1785         if (ret)
1786                 goto err;
1787
1788         n = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1789         ret = PTR_ERR_OR_ZERO(n);
1790         if (ret)
1791                 goto err;
1792
1793         bkey_reassemble(n, k);
1794
1795         refcount = bkey_refcount(n);
1796         if (!refcount) {
1797                 bch2_bkey_val_to_text(&PBUF(buf), c, p.s_c);
1798                 bch2_fs_inconsistent(c,
1799                         "nonexistent indirect extent at %llu while marking\n  %s",
1800                         *idx, buf);
1801                 ret = -EIO;
1802                 goto err;
1803         }
1804
1805         if (!*refcount && (flags & BTREE_TRIGGER_OVERWRITE)) {
1806                 bch2_bkey_val_to_text(&PBUF(buf), c, p.s_c);
1807                 bch2_fs_inconsistent(c,
1808                         "indirect extent refcount underflow at %llu while marking\n  %s",
1809                         *idx, buf);
1810                 ret = -EIO;
1811                 goto err;
1812         }
1813
1814         if (flags & BTREE_TRIGGER_INSERT) {
1815                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1816                 u64 pad;
1817
1818                 pad = max_t(s64, le32_to_cpu(v->front_pad),
1819                             le64_to_cpu(v->idx) - bkey_start_offset(k.k));
1820                 BUG_ON(pad > U32_MAX);
1821                 v->front_pad = cpu_to_le32(pad);
1822
1823                 pad = max_t(s64, le32_to_cpu(v->back_pad),
1824                             k.k->p.offset - p.k->size - le64_to_cpu(v->idx));
1825                 BUG_ON(pad > U32_MAX);
1826                 v->back_pad = cpu_to_le32(pad);
1827         }
1828
1829         le64_add_cpu(refcount, add);
1830
1831         if (!*refcount) {
1832                 n->k.type = KEY_TYPE_deleted;
1833                 set_bkey_val_u64s(&n->k, 0);
1834         }
1835
1836         bch2_btree_iter_set_pos_to_extent_start(&iter);
1837         ret = bch2_trans_update(trans, &iter, n, 0);
1838         if (ret)
1839                 goto err;
1840
1841         *idx = k.k->p.offset;
1842 err:
1843         bch2_trans_iter_exit(trans, &iter);
1844         return ret;
1845 }
1846
1847 static int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1848                                      struct bkey_s_c k, unsigned flags)
1849 {
1850         struct bkey_s_c_reflink_p p = bkey_s_c_to_reflink_p(k);
1851         u64 idx, end_idx;
1852         int ret = 0;
1853
1854         if (flags & BTREE_TRIGGER_INSERT) {
1855                 struct bch_reflink_p *v = (struct bch_reflink_p *) p.v;
1856
1857                 v->front_pad = v->back_pad = 0;
1858         }
1859
1860         idx     = le64_to_cpu(p.v->idx) - le32_to_cpu(p.v->front_pad);
1861         end_idx = le64_to_cpu(p.v->idx) + p.k->size +
1862                 le32_to_cpu(p.v->back_pad);
1863
1864         while (idx < end_idx && !ret)
1865                 ret = __bch2_trans_mark_reflink_p(trans, p, &idx, flags);
1866
1867         return ret;
1868 }
1869
1870 int bch2_trans_mark_key(struct btree_trans *trans, struct bkey_s_c old,
1871                         struct bkey_s_c new, unsigned flags)
1872 {
1873         struct bkey_s_c k = flags & BTREE_TRIGGER_OVERWRITE ? old: new;
1874
1875         switch (k.k->type) {
1876         case KEY_TYPE_btree_ptr:
1877         case KEY_TYPE_btree_ptr_v2:
1878         case KEY_TYPE_extent:
1879         case KEY_TYPE_reflink_v:
1880                 return bch2_trans_mark_extent(trans, k, flags);
1881         case KEY_TYPE_stripe:
1882                 return bch2_trans_mark_stripe(trans, old, new, flags);
1883         case KEY_TYPE_inode:
1884         case KEY_TYPE_inode_v2:
1885                 return bch2_trans_mark_inode(trans, old, new, flags);
1886         case KEY_TYPE_reservation:
1887                 return bch2_trans_mark_reservation(trans, k, flags);
1888         case KEY_TYPE_reflink_p:
1889                 return bch2_trans_mark_reflink_p(trans, k, flags);
1890         default:
1891                 return 0;
1892         }
1893 }
1894
1895 static int __bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1896                                     struct bch_dev *ca, size_t b,
1897                                     enum bch_data_type type,
1898                                     unsigned sectors)
1899 {
1900         struct bch_fs *c = trans->c;
1901         struct btree_iter iter;
1902         struct bkey_alloc_unpacked u;
1903         struct bch_extent_ptr ptr = {
1904                 .dev = ca->dev_idx,
1905                 .offset = bucket_to_sector(ca, b),
1906         };
1907         int ret = 0;
1908
1909         /*
1910          * Backup superblock might be past the end of our normal usable space:
1911          */
1912         if (b >= ca->mi.nbuckets)
1913                 return 0;
1914
1915         ret = bch2_trans_start_alloc_update(trans, &iter, &ptr, &u);
1916         if (ret)
1917                 return ret;
1918
1919         if (u.data_type && u.data_type != type) {
1920                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1921                         "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s\n"
1922                         "while marking %s",
1923                         iter.pos.inode, iter.pos.offset, u.gen,
1924                         bch2_data_types[u.data_type],
1925                         bch2_data_types[type],
1926                         bch2_data_types[type]);
1927                 ret = -EIO;
1928                 goto out;
1929         }
1930
1931         u.data_type     = type;
1932         u.dirty_sectors = sectors;
1933
1934         ret = bch2_alloc_write(trans, &iter, &u, 0);
1935         if (ret)
1936                 goto out;
1937 out:
1938         bch2_trans_iter_exit(trans, &iter);
1939         return ret;
1940 }
1941
1942 int bch2_trans_mark_metadata_bucket(struct btree_trans *trans,
1943                                     struct bch_dev *ca, size_t b,
1944                                     enum bch_data_type type,
1945                                     unsigned sectors)
1946 {
1947         return __bch2_trans_do(trans, NULL, NULL, 0,
1948                         __bch2_trans_mark_metadata_bucket(trans, ca, b, type, sectors));
1949 }
1950
1951 static int bch2_trans_mark_metadata_sectors(struct btree_trans *trans,
1952                                             struct bch_dev *ca,
1953                                             u64 start, u64 end,
1954                                             enum bch_data_type type,
1955                                             u64 *bucket, unsigned *bucket_sectors)
1956 {
1957         do {
1958                 u64 b = sector_to_bucket(ca, start);
1959                 unsigned sectors =
1960                         min_t(u64, bucket_to_sector(ca, b + 1), end) - start;
1961
1962                 if (b != *bucket && *bucket_sectors) {
1963                         int ret = bch2_trans_mark_metadata_bucket(trans, ca, *bucket,
1964                                                                   type, *bucket_sectors);
1965                         if (ret)
1966                                 return ret;
1967
1968                         *bucket_sectors = 0;
1969                 }
1970
1971                 *bucket         = b;
1972                 *bucket_sectors += sectors;
1973                 start += sectors;
1974         } while (start < end);
1975
1976         return 0;
1977 }
1978
1979 static int __bch2_trans_mark_dev_sb(struct btree_trans *trans,
1980                                     struct bch_dev *ca)
1981 {
1982         struct bch_sb_layout *layout = &ca->disk_sb.sb->layout;
1983         u64 bucket = 0;
1984         unsigned i, bucket_sectors = 0;
1985         int ret;
1986
1987         for (i = 0; i < layout->nr_superblocks; i++) {
1988                 u64 offset = le64_to_cpu(layout->sb_offset[i]);
1989
1990                 if (offset == BCH_SB_SECTOR) {
1991                         ret = bch2_trans_mark_metadata_sectors(trans, ca,
1992                                                 0, BCH_SB_SECTOR,
1993                                                 BCH_DATA_sb, &bucket, &bucket_sectors);
1994                         if (ret)
1995                                 return ret;
1996                 }
1997
1998                 ret = bch2_trans_mark_metadata_sectors(trans, ca, offset,
1999                                       offset + (1 << layout->sb_max_size_bits),
2000                                       BCH_DATA_sb, &bucket, &bucket_sectors);
2001                 if (ret)
2002                         return ret;
2003         }
2004
2005         if (bucket_sectors) {
2006                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
2007                                 bucket, BCH_DATA_sb, bucket_sectors);
2008                 if (ret)
2009                         return ret;
2010         }
2011
2012         for (i = 0; i < ca->journal.nr; i++) {
2013                 ret = bch2_trans_mark_metadata_bucket(trans, ca,
2014                                 ca->journal.buckets[i],
2015                                 BCH_DATA_journal, ca->mi.bucket_size);
2016                 if (ret)
2017                         return ret;
2018         }
2019
2020         return 0;
2021 }
2022
2023 int bch2_trans_mark_dev_sb(struct bch_fs *c, struct bch_dev *ca)
2024 {
2025         return bch2_trans_do(c, NULL, NULL, BTREE_INSERT_LAZY_RW,
2026                         __bch2_trans_mark_dev_sb(&trans, ca));
2027 }
2028
2029 /* Disk reservations: */
2030
2031 #define SECTORS_CACHE   1024
2032
2033 int bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
2034                               u64 sectors, int flags)
2035 {
2036         struct bch_fs_pcpu *pcpu;
2037         u64 old, v, get;
2038         s64 sectors_available;
2039         int ret;
2040
2041         percpu_down_read(&c->mark_lock);
2042         preempt_disable();
2043         pcpu = this_cpu_ptr(c->pcpu);
2044
2045         if (sectors <= pcpu->sectors_available)
2046                 goto out;
2047
2048         v = atomic64_read(&c->sectors_available);
2049         do {
2050                 old = v;
2051                 get = min((u64) sectors + SECTORS_CACHE, old);
2052
2053                 if (get < sectors) {
2054                         preempt_enable();
2055                         goto recalculate;
2056                 }
2057         } while ((v = atomic64_cmpxchg(&c->sectors_available,
2058                                        old, old - get)) != old);
2059
2060         pcpu->sectors_available         += get;
2061
2062 out:
2063         pcpu->sectors_available         -= sectors;
2064         this_cpu_add(*c->online_reserved, sectors);
2065         res->sectors                    += sectors;
2066
2067         preempt_enable();
2068         percpu_up_read(&c->mark_lock);
2069         return 0;
2070
2071 recalculate:
2072         mutex_lock(&c->sectors_available_lock);
2073
2074         percpu_u64_set(&c->pcpu->sectors_available, 0);
2075         sectors_available = avail_factor(__bch2_fs_usage_read_short(c).free);
2076
2077         if (sectors <= sectors_available ||
2078             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
2079                 atomic64_set(&c->sectors_available,
2080                              max_t(s64, 0, sectors_available - sectors));
2081                 this_cpu_add(*c->online_reserved, sectors);
2082                 res->sectors                    += sectors;
2083                 ret = 0;
2084         } else {
2085                 atomic64_set(&c->sectors_available, sectors_available);
2086                 ret = -ENOSPC;
2087         }
2088
2089         mutex_unlock(&c->sectors_available_lock);
2090         percpu_up_read(&c->mark_lock);
2091
2092         return ret;
2093 }
2094
2095 /* Startup/shutdown: */
2096
2097 static void buckets_free_rcu(struct rcu_head *rcu)
2098 {
2099         struct bucket_array *buckets =
2100                 container_of(rcu, struct bucket_array, rcu);
2101
2102         kvpfree(buckets,
2103                 sizeof(*buckets) +
2104                 buckets->nbuckets * sizeof(struct bucket));
2105 }
2106
2107 static void bucket_gens_free_rcu(struct rcu_head *rcu)
2108 {
2109         struct bucket_gens *buckets =
2110                 container_of(rcu, struct bucket_gens, rcu);
2111
2112         kvpfree(buckets, sizeof(*buckets) + buckets->nbuckets);
2113 }
2114
2115 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
2116 {
2117         struct bucket_array *buckets = NULL, *old_buckets = NULL;
2118         struct bucket_gens *bucket_gens = NULL, *old_bucket_gens = NULL;
2119         unsigned long *buckets_nouse = NULL;
2120         alloc_fifo      free[RESERVE_NR];
2121         alloc_fifo      free_inc;
2122         alloc_heap      alloc_heap;
2123
2124         size_t btree_reserve    = DIV_ROUND_UP(BTREE_NODE_RESERVE,
2125                              ca->mi.bucket_size / btree_sectors(c));
2126         /* XXX: these should be tunable */
2127         size_t reserve_none     = max_t(size_t, 1, nbuckets >> 9);
2128         size_t copygc_reserve   = max_t(size_t, 2, nbuckets >> 6);
2129         size_t free_inc_nr      = max(max_t(size_t, 1, nbuckets >> 12),
2130                                       btree_reserve * 2);
2131         bool resize = ca->buckets[0] != NULL;
2132         int ret = -ENOMEM;
2133         unsigned i;
2134
2135         memset(&free,           0, sizeof(free));
2136         memset(&free_inc,       0, sizeof(free_inc));
2137         memset(&alloc_heap,     0, sizeof(alloc_heap));
2138
2139         if (!(buckets           = kvpmalloc(sizeof(struct bucket_array) +
2140                                             nbuckets * sizeof(struct bucket),
2141                                             GFP_KERNEL|__GFP_ZERO)) ||
2142             !(bucket_gens       = kvpmalloc(sizeof(struct bucket_gens) + nbuckets,
2143                                             GFP_KERNEL|__GFP_ZERO)) ||
2144             !(buckets_nouse     = kvpmalloc(BITS_TO_LONGS(nbuckets) *
2145                                             sizeof(unsigned long),
2146                                             GFP_KERNEL|__GFP_ZERO)) ||
2147             !init_fifo(&free[RESERVE_MOVINGGC],
2148                        copygc_reserve, GFP_KERNEL) ||
2149             !init_fifo(&free[RESERVE_NONE], reserve_none, GFP_KERNEL) ||
2150             !init_fifo(&free_inc,       free_inc_nr, GFP_KERNEL) ||
2151             !init_heap(&alloc_heap,     ALLOC_SCAN_BATCH(ca) << 1, GFP_KERNEL))
2152                 goto err;
2153
2154         buckets->first_bucket   = ca->mi.first_bucket;
2155         buckets->nbuckets       = nbuckets;
2156         bucket_gens->first_bucket = ca->mi.first_bucket;
2157         bucket_gens->nbuckets   = nbuckets;
2158
2159         bch2_copygc_stop(c);
2160
2161         if (resize) {
2162                 down_write(&c->gc_lock);
2163                 down_write(&ca->bucket_lock);
2164                 percpu_down_write(&c->mark_lock);
2165         }
2166
2167         old_buckets = bucket_array(ca);
2168         old_bucket_gens = rcu_dereference_protected(ca->bucket_gens, 1);
2169
2170         if (resize) {
2171                 size_t n = min(buckets->nbuckets, old_buckets->nbuckets);
2172
2173                 memcpy(buckets->b,
2174                        old_buckets->b,
2175                        n * sizeof(struct bucket));
2176                 memcpy(bucket_gens->b,
2177                        old_bucket_gens->b,
2178                        n);
2179                 memcpy(buckets_nouse,
2180                        ca->buckets_nouse,
2181                        BITS_TO_LONGS(n) * sizeof(unsigned long));
2182         }
2183
2184         rcu_assign_pointer(ca->buckets[0], buckets);
2185         rcu_assign_pointer(ca->bucket_gens, bucket_gens);
2186         buckets         = old_buckets;
2187         bucket_gens     = old_bucket_gens;
2188
2189         swap(ca->buckets_nouse, buckets_nouse);
2190
2191         if (resize) {
2192                 percpu_up_write(&c->mark_lock);
2193                 up_write(&c->gc_lock);
2194         }
2195
2196         spin_lock(&c->freelist_lock);
2197         for (i = 0; i < RESERVE_NR; i++) {
2198                 fifo_move(&free[i], &ca->free[i]);
2199                 swap(ca->free[i], free[i]);
2200         }
2201         fifo_move(&free_inc, &ca->free_inc);
2202         swap(ca->free_inc, free_inc);
2203         spin_unlock(&c->freelist_lock);
2204
2205         /* with gc lock held, alloc_heap can't be in use: */
2206         swap(ca->alloc_heap, alloc_heap);
2207
2208         nbuckets = ca->mi.nbuckets;
2209
2210         if (resize)
2211                 up_write(&ca->bucket_lock);
2212
2213         ret = 0;
2214 err:
2215         free_heap(&alloc_heap);
2216         free_fifo(&free_inc);
2217         for (i = 0; i < RESERVE_NR; i++)
2218                 free_fifo(&free[i]);
2219         kvpfree(buckets_nouse,
2220                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2221         if (bucket_gens)
2222                 call_rcu(&bucket_gens->rcu, bucket_gens_free_rcu);
2223         if (buckets)
2224                 call_rcu(&buckets->rcu, buckets_free_rcu);
2225
2226         return ret;
2227 }
2228
2229 void bch2_dev_buckets_free(struct bch_dev *ca)
2230 {
2231         unsigned i;
2232
2233         free_heap(&ca->alloc_heap);
2234         free_fifo(&ca->free_inc);
2235         for (i = 0; i < RESERVE_NR; i++)
2236                 free_fifo(&ca->free[i]);
2237         kvpfree(ca->buckets_nouse,
2238                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2239         kvpfree(rcu_dereference_protected(ca->bucket_gens, 1),
2240                 sizeof(struct bucket_gens) + ca->mi.nbuckets);
2241         kvpfree(rcu_dereference_protected(ca->buckets[0], 1),
2242                 sizeof(struct bucket_array) +
2243                 ca->mi.nbuckets * sizeof(struct bucket));
2244
2245         for (i = 0; i < ARRAY_SIZE(ca->usage); i++)
2246                 free_percpu(ca->usage[i]);
2247         kfree(ca->usage_base);
2248 }
2249
2250 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2251 {
2252         unsigned i;
2253
2254         ca->usage_base = kzalloc(sizeof(struct bch_dev_usage), GFP_KERNEL);
2255         if (!ca->usage_base)
2256                 return -ENOMEM;
2257
2258         for (i = 0; i < ARRAY_SIZE(ca->usage); i++) {
2259                 ca->usage[i] = alloc_percpu(struct bch_dev_usage);
2260                 if (!ca->usage[i])
2261                         return -ENOMEM;
2262         }
2263
2264         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);;
2265 }