fs/bcachefs/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  * Bucket states:
   8  * - free bucket: mark == 0
   9  *   The bucket contains no data and will not be read
  10  *
  11  * - allocator bucket: owned_by_allocator == 1
  12  *   The bucket is on a free list, or it is an open bucket
  13  *
  14  * - cached bucket: owned_by_allocator == 0 &&
  15  *                  dirty_sectors == 0 &&
  16  *                  cached_sectors > 0
  17  *   The bucket contains data but may be safely discarded as there are
  18  *   enough replicas of the data on other cache devices, or it has been
  19  *   written back to the backing device
  20  *
  21  * - dirty bucket: owned_by_allocator == 0 &&
  22  *                 dirty_sectors > 0
  23  *   The bucket contains data that we must not discard (either only copy,
  24  *   or one of the 'main copies' for data requiring multiple replicas)
  25  *
  26  * - metadata bucket: owned_by_allocator == 0 && is_metadata == 1
  27  *   This is a btree node, journal or gen/prio bucket
  28  *
  29  * Lifecycle:
  30  *
  31  * bucket invalidated => bucket on freelist => open bucket =>
  32  *     [dirty bucket =>] cached bucket => bucket invalidated => ...
  33  *
  34  * Note that cache promotion can skip the dirty bucket step, as data
  35  * is copied from a deeper tier to a shallower tier, onto a cached
  36  * bucket.
  37  * Note also that a cached bucket can spontaneously become dirty --
  38  * see below.
  39  *
  40  * Only a traversal of the key space can determine whether a bucket is
  41  * truly dirty or cached.
  42  *
  43  * Transitions:
  44  *
  45  * - free => allocator: bucket was invalidated
  46  * - cached => allocator: bucket was invalidated
  47  *
  48  * - allocator => dirty: open bucket was filled up
  49  * - allocator => cached: open bucket was filled up
  50  * - allocator => metadata: metadata was allocated
  51  *
  52  * - dirty => cached: dirty sectors were copied to a deeper tier
  53  * - dirty => free: dirty sectors were overwritten or moved (copy gc)
  54  * - cached => free: cached sectors were overwritten
  55  *
  56  * - metadata => free: metadata was freed
  57  *
  58  * Oddities:
  59  * - cached => dirty: a device was removed so formerly replicated data
  60  *                    is no longer sufficiently replicated
  61  * - free => cached: cannot happen
  62  * - free => dirty: cannot happen
  63  * - free => metadata: cannot happen
  64  */
  65
  66 #include "bcachefs.h"
  67 #include "alloc_background.h"
  68 #include "bset.h"
  69 #include "btree_gc.h"
  70 #include "btree_update.h"
  71 #include "buckets.h"
  72 #include "ec.h"
  73 #include "error.h"
  74 #include "movinggc.h"
  75 #include "replicas.h"
  76 #include "trace.h"
  77
  78 #include <linux/preempt.h>
  79
  80 /*
  81  * Clear journal_seq_valid for buckets for which it's not needed, to prevent
  82  * wraparound:
  83  */
  84 void bch2_bucket_seq_cleanup(struct bch_fs *c)
  85 {
  86         u64 journal_seq = atomic64_read(&c->journal.seq);
  87         u16 last_seq_ondisk = c->journal.last_seq_ondisk;
  88         struct bch_dev *ca;
  89         struct bucket_array *buckets;
  90         struct bucket *g;
  91         struct bucket_mark m;
  92         unsigned i;
  93
  94         if (journal_seq - c->last_bucket_seq_cleanup <
  95             (1U << (BUCKET_JOURNAL_SEQ_BITS - 2)))
  96                 return;
  97
  98         c->last_bucket_seq_cleanup = journal_seq;
  99
 100         for_each_member_device(ca, c, i) {
 101                 down_read(&ca->bucket_lock);
 102                 buckets = bucket_array(ca);
 103
 104                 for_each_bucket(g, buckets) {
 105                         bucket_cmpxchg(g, m, ({
 106                                 if (!m.journal_seq_valid ||
 107                                     bucket_needs_journal_commit(m, last_seq_ondisk))
 108                                         break;
 109
 110                                 m.journal_seq_valid = 0;
 111                         }));
 112                 }
 113                 up_read(&ca->bucket_lock);
 114         }
 115 }
 116
 117 void bch2_fs_usage_initialize(struct bch_fs *c)
 118 {
 119         struct bch_fs_usage *usage;
 120         unsigned i;
 121
 122         percpu_down_write(&c->mark_lock);
 123         usage = c->usage_base;
 124
 125         for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 126                 bch2_fs_usage_acc_to_base(c, i);
 127
 128         for (i = 0; i < BCH_REPLICAS_MAX; i++)
 129                 usage->reserved += usage->persistent_reserved[i];
 130
 131         for (i = 0; i < c->replicas.nr; i++) {
 132                 struct bch_replicas_entry *e =
 133                         cpu_replicas_entry(&c->replicas, i);
 134
 135                 switch (e->data_type) {
 136                 case BCH_DATA_BTREE:
 137                         usage->btree    += usage->replicas[i];
 138                         break;
 139                 case BCH_DATA_USER:
 140                         usage->data     += usage->replicas[i];
 141                         break;
 142                 case BCH_DATA_CACHED:
 143                         usage->cached   += usage->replicas[i];
 144                         break;
 145                 }
 146         }
 147
 148         percpu_up_write(&c->mark_lock);
 149 }
 150
 151 void bch2_fs_usage_scratch_put(struct bch_fs *c, struct bch_fs_usage_online *fs_usage)
 152 {
 153         if (fs_usage == c->usage_scratch)
 154                 mutex_unlock(&c->usage_scratch_lock);
 155         else
 156                 kfree(fs_usage);
 157 }
 158
 159 struct bch_fs_usage_online *bch2_fs_usage_scratch_get(struct bch_fs *c)
 160 {
 161         struct bch_fs_usage_online *ret;
 162         unsigned bytes = sizeof(struct bch_fs_usage_online) + sizeof(u64) *
 163                 READ_ONCE(c->replicas.nr);
 164         ret = kzalloc(bytes, GFP_NOWAIT|__GFP_NOWARN);
 165         if (ret)
 166                 return ret;
 167
 168         if (mutex_trylock(&c->usage_scratch_lock))
 169                 goto out_pool;
 170
 171         ret = kzalloc(bytes, GFP_NOFS);
 172         if (ret)
 173                 return ret;
 174
 175         mutex_lock(&c->usage_scratch_lock);
 176 out_pool:
 177         ret = c->usage_scratch;
 178         memset(ret, 0, bytes);
 179         return ret;
 180 }
 181
 182 struct bch_dev_usage bch2_dev_usage_read(struct bch_fs *c, struct bch_dev *ca)
 183 {
 184         struct bch_dev_usage ret;
 185
 186         memset(&ret, 0, sizeof(ret));
 187         acc_u64s_percpu((u64 *) &ret,
 188                         (u64 __percpu *) ca->usage[0],
 189                         sizeof(ret) / sizeof(u64));
 190
 191         return ret;
 192 }
 193
 194 static inline struct bch_fs_usage *fs_usage_ptr(struct bch_fs *c,
 195                                                 unsigned journal_seq,
 196                                                 bool gc)
 197 {
 198         return this_cpu_ptr(gc
 199                             ? c->usage_gc
 200                             : c->usage[journal_seq & 1]);
 201 }
 202
 203 u64 bch2_fs_usage_read_one(struct bch_fs *c, u64 *v)
 204 {
 205         ssize_t offset = v - (u64 *) c->usage_base;
 206         unsigned seq;
 207         u64 ret;
 208
 209         BUG_ON(offset < 0 || offset >= fs_usage_u64s(c));
 210         percpu_rwsem_assert_held(&c->mark_lock);
 211
 212         do {
 213                 seq = read_seqcount_begin(&c->usage_lock);
 214                 ret = *v +
 215                         percpu_u64_get((u64 __percpu *) c->usage[0] + offset) +
 216                         percpu_u64_get((u64 __percpu *) c->usage[1] + offset);
 217         } while (read_seqcount_retry(&c->usage_lock, seq));
 218
 219         return ret;
 220 }
 221
 222 struct bch_fs_usage_online *bch2_fs_usage_read(struct bch_fs *c)
 223 {
 224         struct bch_fs_usage_online *ret;
 225         unsigned seq, i, u64s;
 226
 227         percpu_down_read(&c->mark_lock);
 228
 229         ret = kmalloc(sizeof(struct bch_fs_usage_online) +
 230                       sizeof(u64) + c->replicas.nr, GFP_NOFS);
 231         if (unlikely(!ret)) {
 232                 percpu_up_read(&c->mark_lock);
 233                 return NULL;
 234         }
 235
 236         ret->online_reserved = percpu_u64_get(c->online_reserved);
 237
 238         u64s = fs_usage_u64s(c);
 239         do {
 240                 seq = read_seqcount_begin(&c->usage_lock);
 241                 memcpy(&ret->u, c->usage_base, u64s * sizeof(u64));
 242                 for (i = 0; i < ARRAY_SIZE(c->usage); i++)
 243                         acc_u64s_percpu((u64 *) &ret->u, (u64 __percpu *) c->usage[i], u64s);
 244         } while (read_seqcount_retry(&c->usage_lock, seq));
 245
 246         return ret;
 247 }
 248
 249 void bch2_fs_usage_acc_to_base(struct bch_fs *c, unsigned idx)
 250 {
 251         unsigned u64s = fs_usage_u64s(c);
 252
 253         BUG_ON(idx >= ARRAY_SIZE(c->usage));
 254
 255         preempt_disable();
 256         write_seqcount_begin(&c->usage_lock);
 257
 258         acc_u64s_percpu((u64 *) c->usage_base,
 259                         (u64 __percpu *) c->usage[idx], u64s);
 260         percpu_memset(c->usage[idx], 0, u64s * sizeof(u64));
 261
 262         write_seqcount_end(&c->usage_lock);
 263         preempt_enable();
 264 }
 265
 266 void bch2_fs_usage_to_text(struct printbuf *out,
 267                            struct bch_fs *c,
 268                            struct bch_fs_usage_online *fs_usage)
 269 {
 270         unsigned i;
 271
 272         pr_buf(out, "capacity:\t\t\t%llu\n", c->capacity);
 273
 274         pr_buf(out, "hidden:\t\t\t\t%llu\n",
 275                fs_usage->u.hidden);
 276         pr_buf(out, "data:\t\t\t\t%llu\n",
 277                fs_usage->u.data);
 278         pr_buf(out, "cached:\t\t\t\t%llu\n",
 279                fs_usage->u.cached);
 280         pr_buf(out, "reserved:\t\t\t%llu\n",
 281                fs_usage->u.reserved);
 282         pr_buf(out, "nr_inodes:\t\t\t%llu\n",
 283                fs_usage->u.nr_inodes);
 284         pr_buf(out, "online reserved:\t\t%llu\n",
 285                fs_usage->online_reserved);
 286
 287         for (i = 0;
 288              i < ARRAY_SIZE(fs_usage->u.persistent_reserved);
 289              i++) {
 290                 pr_buf(out, "%u replicas:\n", i + 1);
 291                 pr_buf(out, "\treserved:\t\t%llu\n",
 292                        fs_usage->u.persistent_reserved[i]);
 293         }
 294
 295         for (i = 0; i < c->replicas.nr; i++) {
 296                 struct bch_replicas_entry *e =
 297                         cpu_replicas_entry(&c->replicas, i);
 298
 299                 pr_buf(out, "\t");
 300                 bch2_replicas_entry_to_text(out, e);
 301                 pr_buf(out, ":\t%llu\n", fs_usage->u.replicas[i]);
 302         }
 303 }
 304
 305 #define RESERVE_FACTOR  6
 306
 307 static u64 reserve_factor(u64 r)
 308 {
 309         return r + (round_up(r, (1 << RESERVE_FACTOR)) >> RESERVE_FACTOR);
 310 }
 311
 312 static u64 avail_factor(u64 r)
 313 {
 314         return (r << RESERVE_FACTOR) / ((1 << RESERVE_FACTOR) + 1);
 315 }
 316
 317 u64 bch2_fs_sectors_used(struct bch_fs *c, struct bch_fs_usage_online *fs_usage)
 318 {
 319         return min(fs_usage->u.hidden +
 320                    fs_usage->u.btree +
 321                    fs_usage->u.data +
 322                    reserve_factor(fs_usage->u.reserved +
 323                                   fs_usage->online_reserved),
 324                    c->capacity);
 325 }
 326
 327 static struct bch_fs_usage_short
 328 __bch2_fs_usage_read_short(struct bch_fs *c)
 329 {
 330         struct bch_fs_usage_short ret;
 331         u64 data, reserved;
 332
 333         ret.capacity = c->capacity -
 334                 bch2_fs_usage_read_one(c, &c->usage_base->hidden);
 335
 336         data            = bch2_fs_usage_read_one(c, &c->usage_base->data) +
 337                 bch2_fs_usage_read_one(c, &c->usage_base->btree);
 338         reserved        = bch2_fs_usage_read_one(c, &c->usage_base->reserved) +
 339                 percpu_u64_get(c->online_reserved);
 340
 341         ret.used        = min(ret.capacity, data + reserve_factor(reserved));
 342         ret.free        = ret.capacity - ret.used;
 343
 344         ret.nr_inodes   = bch2_fs_usage_read_one(c, &c->usage_base->nr_inodes);
 345
 346         return ret;
 347 }
 348
 349 struct bch_fs_usage_short
 350 bch2_fs_usage_read_short(struct bch_fs *c)
 351 {
 352         struct bch_fs_usage_short ret;
 353
 354         percpu_down_read(&c->mark_lock);
 355         ret = __bch2_fs_usage_read_short(c);
 356         percpu_up_read(&c->mark_lock);
 357
 358         return ret;
 359 }
 360
 361 static inline int is_unavailable_bucket(struct bucket_mark m)
 362 {
 363         return !is_available_bucket(m);
 364 }
 365
 366 static inline int is_fragmented_bucket(struct bucket_mark m,
 367                                        struct bch_dev *ca)
 368 {
 369         if (!m.owned_by_allocator &&
 370             m.data_type == BCH_DATA_USER &&
 371             bucket_sectors_used(m))
 372                 return max_t(int, 0, (int) ca->mi.bucket_size -
 373                              bucket_sectors_used(m));
 374         return 0;
 375 }
 376
 377 static inline enum bch_data_type bucket_type(struct bucket_mark m)
 378 {
 379         return m.cached_sectors && !m.dirty_sectors
 380                 ? BCH_DATA_CACHED
 381                 : m.data_type;
 382 }
 383
 384 static bool bucket_became_unavailable(struct bucket_mark old,
 385                                       struct bucket_mark new)
 386 {
 387         return is_available_bucket(old) &&
 388                !is_available_bucket(new);
 389 }
 390
 391 int bch2_fs_usage_apply(struct bch_fs *c,
 392                         struct bch_fs_usage_online *src,
 393                         struct disk_reservation *disk_res,
 394                         unsigned journal_seq)
 395 {
 396         struct bch_fs_usage *dst = fs_usage_ptr(c, journal_seq, false);
 397         s64 added = src->u.data + src->u.reserved;
 398         s64 should_not_have_added;
 399         int ret = 0;
 400
 401         percpu_rwsem_assert_held(&c->mark_lock);
 402
 403         /*
 404          * Not allowed to reduce sectors_available except by getting a
 405          * reservation:
 406          */
 407         should_not_have_added = added - (s64) (disk_res ? disk_res->sectors : 0);
 408         if (WARN_ONCE(should_not_have_added > 0,
 409                       "disk usage increased by %lli more than reservation of %llu",
 410                       added, disk_res ? disk_res->sectors : 0)) {
 411                 atomic64_sub(should_not_have_added, &c->sectors_available);
 412                 added -= should_not_have_added;
 413                 ret = -1;
 414         }
 415
 416         if (added > 0) {
 417                 disk_res->sectors       -= added;
 418                 src->online_reserved    -= added;
 419         }
 420
 421         this_cpu_add(*c->online_reserved, src->online_reserved);
 422
 423         preempt_disable();
 424         acc_u64s((u64 *) dst, (u64 *) &src->u, fs_usage_u64s(c));
 425         preempt_enable();
 426
 427         return ret;
 428 }
 429
 430 static inline void account_bucket(struct bch_fs_usage *fs_usage,
 431                                   struct bch_dev_usage *dev_usage,
 432                                   enum bch_data_type type,
 433                                   int nr, s64 size)
 434 {
 435         if (type == BCH_DATA_SB || type == BCH_DATA_JOURNAL)
 436                 fs_usage->hidden        += size;
 437
 438         dev_usage->buckets[type]        += nr;
 439 }
 440
 441 static void bch2_dev_usage_update(struct bch_fs *c, struct bch_dev *ca,
 442                                   struct bch_fs_usage *fs_usage,
 443                                   struct bucket_mark old, struct bucket_mark new,
 444                                   bool gc)
 445 {
 446         struct bch_dev_usage *dev_usage;
 447
 448         percpu_rwsem_assert_held(&c->mark_lock);
 449
 450         preempt_disable();
 451         dev_usage = this_cpu_ptr(ca->usage[gc]);
 452
 453         if (bucket_type(old))
 454                 account_bucket(fs_usage, dev_usage, bucket_type(old),
 455                                -1, -ca->mi.bucket_size);
 456
 457         if (bucket_type(new))
 458                 account_bucket(fs_usage, dev_usage, bucket_type(new),
 459                                1, ca->mi.bucket_size);
 460
 461         dev_usage->buckets_ec += (int) new.stripe - (int) old.stripe;
 462         dev_usage->buckets_unavailable +=
 463                 is_unavailable_bucket(new) - is_unavailable_bucket(old);
 464
 465         dev_usage->sectors[old.data_type] -= old.dirty_sectors;
 466         dev_usage->sectors[new.data_type] += new.dirty_sectors;
 467         dev_usage->sectors[BCH_DATA_CACHED] +=
 468                 (int) new.cached_sectors - (int) old.cached_sectors;
 469         dev_usage->sectors_fragmented +=
 470                 is_fragmented_bucket(new, ca) - is_fragmented_bucket(old, ca);
 471         preempt_enable();
 472
 473         if (!is_available_bucket(old) && is_available_bucket(new))
 474                 bch2_wake_allocator(ca);
 475 }
 476
 477 void bch2_dev_usage_from_buckets(struct bch_fs *c)
 478 {
 479         struct bch_dev *ca;
 480         struct bucket_mark old = { .v.counter = 0 };
 481         struct bucket_array *buckets;
 482         struct bucket *g;
 483         unsigned i;
 484         int cpu;
 485
 486         c->usage_base->hidden = 0;
 487
 488         for_each_member_device(ca, c, i) {
 489                 for_each_possible_cpu(cpu)
 490                         memset(per_cpu_ptr(ca->usage[0], cpu), 0,
 491                                sizeof(*ca->usage[0]));
 492
 493                 buckets = bucket_array(ca);
 494
 495                 for_each_bucket(g, buckets)
 496                         bch2_dev_usage_update(c, ca, c->usage_base,
 497                                               old, g->mark, false);
 498         }
 499 }
 500
 501 static inline void update_replicas(struct bch_fs *c,
 502                                    struct bch_fs_usage *fs_usage,
 503                                    struct bch_replicas_entry *r,
 504                                    s64 sectors)
 505 {
 506         int idx = bch2_replicas_entry_idx(c, r);
 507
 508         BUG_ON(idx < 0);
 509
 510         switch (r->data_type) {
 511         case BCH_DATA_BTREE:
 512                 fs_usage->btree         += sectors;
 513                 break;
 514         case BCH_DATA_USER:
 515                 fs_usage->data          += sectors;
 516                 break;
 517         case BCH_DATA_CACHED:
 518                 fs_usage->cached        += sectors;
 519                 break;
 520         }
 521         fs_usage->replicas[idx]         += sectors;
 522 }
 523
 524 static inline void update_cached_sectors(struct bch_fs *c,
 525                                          struct bch_fs_usage *fs_usage,
 526                                          unsigned dev, s64 sectors)
 527 {
 528         struct bch_replicas_padded r;
 529
 530         bch2_replicas_entry_cached(&r.e, dev);
 531
 532         update_replicas(c, fs_usage, &r.e, sectors);
 533 }
 534
 535 static struct replicas_delta_list *
 536 replicas_deltas_realloc(struct btree_trans *trans, unsigned more)
 537 {
 538         struct replicas_delta_list *d = trans->fs_usage_deltas;
 539         unsigned new_size = d ? (d->size + more) * 2 : 128;
 540
 541         if (!d || d->used + more > d->size) {
 542                 d = krealloc(d, sizeof(*d) + new_size, GFP_NOIO|__GFP_ZERO);
 543                 BUG_ON(!d);
 544
 545                 d->size = new_size;
 546                 trans->fs_usage_deltas = d;
 547         }
 548         return d;
 549 }
 550
 551 static inline void update_replicas_list(struct btree_trans *trans,
 552                                         struct bch_replicas_entry *r,
 553                                         s64 sectors)
 554 {
 555         struct replicas_delta_list *d;
 556         struct replicas_delta *n;
 557         unsigned b;
 558
 559         if (!sectors)
 560                 return;
 561
 562         b = replicas_entry_bytes(r) + 8;
 563         d = replicas_deltas_realloc(trans, b);
 564
 565         n = (void *) d->d + d->used;
 566         n->delta = sectors;
 567         memcpy((void *) n + offsetof(struct replicas_delta, r),
 568                r, replicas_entry_bytes(r));
 569         d->used += b;
 570 }
 571
 572 static inline void update_cached_sectors_list(struct btree_trans *trans,
 573                                               unsigned dev, s64 sectors)
 574 {
 575         struct bch_replicas_padded r;
 576
 577         bch2_replicas_entry_cached(&r.e, dev);
 578
 579         update_replicas_list(trans, &r.e, sectors);
 580 }
 581
 582 void bch2_replicas_delta_list_apply(struct bch_fs *c,
 583                                     struct bch_fs_usage *fs_usage,
 584                                     struct replicas_delta_list *r)
 585 {
 586         struct replicas_delta *d = r->d;
 587         struct replicas_delta *top = (void *) r->d + r->used;
 588         unsigned i;
 589
 590         fs_usage->nr_inodes += r->nr_inodes;
 591
 592         for (i = 0; i < BCH_REPLICAS_MAX; i++) {
 593                 fs_usage->reserved += r->persistent_reserved[i];
 594                 fs_usage->persistent_reserved[i] += r->persistent_reserved[i];
 595         }
 596
 597         while (d != top) {
 598                 BUG_ON((void *) d > (void *) top);
 599
 600                 update_replicas(c, fs_usage, &d->r, d->delta);
 601
 602                 d = (void *) d + replicas_entry_bytes(&d->r) + 8;
 603         }
 604 }
 605
 606 #define do_mark_fn(fn, c, pos, flags, ...)                              \
 607 ({                                                                      \
 608         int gc, ret = 0;                                                \
 609                                                                         \
 610         percpu_rwsem_assert_held(&c->mark_lock);                        \
 611                                                                         \
 612         for (gc = 0; gc < 2 && !ret; gc++)                              \
 613                 if (!gc == !(flags & BCH_BUCKET_MARK_GC) ||             \
 614                     (gc && gc_visited(c, pos)))                         \
 615                         ret = fn(c, __VA_ARGS__, gc);                   \
 616         ret;                                                            \
 617 })
 618
 619 static int __bch2_invalidate_bucket(struct bch_fs *c, struct bch_dev *ca,
 620                                     size_t b, struct bucket_mark *ret,
 621                                     bool gc)
 622 {
 623         struct bch_fs_usage *fs_usage = fs_usage_ptr(c, 0, gc);
 624         struct bucket *g = __bucket(ca, b, gc);
 625         struct bucket_mark old, new;
 626
 627         old = bucket_cmpxchg(g, new, ({
 628                 BUG_ON(!is_available_bucket(new));
 629
 630                 new.owned_by_allocator  = true;
 631                 new.data_type           = 0;
 632                 new.cached_sectors      = 0;
 633                 new.dirty_sectors       = 0;
 634                 new.gen++;
 635         }));
 636
 637         bch2_dev_usage_update(c, ca, fs_usage, old, new, gc);
 638
 639         if (old.cached_sectors)
 640                 update_cached_sectors(c, fs_usage, ca->dev_idx,
 641                                       -((s64) old.cached_sectors));
 642
 643         if (!gc)
 644                 *ret = old;
 645         return 0;
 646 }
 647
 648 void bch2_invalidate_bucket(struct bch_fs *c, struct bch_dev *ca,
 649                             size_t b, struct bucket_mark *old)
 650 {
 651         do_mark_fn(__bch2_invalidate_bucket, c, gc_phase(GC_PHASE_START), 0,
 652                    ca, b, old);
 653
 654         if (!old->owned_by_allocator && old->cached_sectors)
 655                 trace_invalidate(ca, bucket_to_sector(ca, b),
 656                                  old->cached_sectors);
 657 }
 658
 659 static int __bch2_mark_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 660                                     size_t b, bool owned_by_allocator,
 661                                     bool gc)
 662 {
 663         struct bch_fs_usage *fs_usage = fs_usage_ptr(c, 0, gc);
 664         struct bucket *g = __bucket(ca, b, gc);
 665         struct bucket_mark old, new;
 666
 667         old = bucket_cmpxchg(g, new, ({
 668                 new.owned_by_allocator  = owned_by_allocator;
 669         }));
 670
 671         bch2_dev_usage_update(c, ca, fs_usage, old, new, gc);
 672
 673         BUG_ON(!gc &&
 674                !owned_by_allocator && !old.owned_by_allocator);
 675
 676         return 0;
 677 }
 678
 679 void bch2_mark_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 680                             size_t b, bool owned_by_allocator,
 681                             struct gc_pos pos, unsigned flags)
 682 {
 683         preempt_disable();
 684
 685         do_mark_fn(__bch2_mark_alloc_bucket, c, pos, flags,
 686                    ca, b, owned_by_allocator);
 687
 688         preempt_enable();
 689 }
 690
 691 static int bch2_mark_alloc(struct bch_fs *c, struct bkey_s_c k,
 692                            struct bch_fs_usage *fs_usage,
 693                            u64 journal_seq, unsigned flags)
 694 {
 695         bool gc = flags & BCH_BUCKET_MARK_GC;
 696         struct bkey_alloc_unpacked u;
 697         struct bch_dev *ca;
 698         struct bucket *g;
 699         struct bucket_mark old, m;
 700
 701         /*
 702          * alloc btree is read in by bch2_alloc_read, not gc:
 703          */
 704         if ((flags & BCH_BUCKET_MARK_GC) &&
 705             !(flags & BCH_BUCKET_MARK_BUCKET_INVALIDATE))
 706                 return 0;
 707
 708         ca = bch_dev_bkey_exists(c, k.k->p.inode);
 709
 710         if (k.k->p.offset >= ca->mi.nbuckets)
 711                 return 0;
 712
 713         g = __bucket(ca, k.k->p.offset, gc);
 714         u = bch2_alloc_unpack(k);
 715
 716         old = bucket_cmpxchg(g, m, ({
 717                 m.gen                   = u.gen;
 718                 m.data_type             = u.data_type;
 719                 m.dirty_sectors         = u.dirty_sectors;
 720                 m.cached_sectors        = u.cached_sectors;
 721
 722                 if (journal_seq) {
 723                         m.journal_seq_valid     = 1;
 724                         m.journal_seq           = journal_seq;
 725                 }
 726         }));
 727
 728         if (!(flags & BCH_BUCKET_MARK_ALLOC_READ))
 729                 bch2_dev_usage_update(c, ca, fs_usage, old, m, gc);
 730
 731         g->io_time[READ]        = u.read_time;
 732         g->io_time[WRITE]       = u.write_time;
 733         g->oldest_gen           = u.oldest_gen;
 734         g->gen_valid            = 1;
 735
 736         /*
 737          * need to know if we're getting called from the invalidate path or
 738          * not:
 739          */
 740
 741         if ((flags & BCH_BUCKET_MARK_BUCKET_INVALIDATE) &&
 742             old.cached_sectors) {
 743                 update_cached_sectors(c, fs_usage, ca->dev_idx,
 744                                       -old.cached_sectors);
 745                 trace_invalidate(ca, bucket_to_sector(ca, k.k->p.offset),
 746                                  old.cached_sectors);
 747         }
 748
 749         return 0;
 750 }
 751
 752 #define checked_add(a, b)                                       \
 753 ({                                                              \
 754         unsigned _res = (unsigned) (a) + (b);                   \
 755         bool overflow = _res > U16_MAX;                         \
 756         if (overflow)                                           \
 757                 _res = U16_MAX;                                 \
 758         (a) = _res;                                             \
 759         overflow;                                               \
 760 })
 761
 762 static int __bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 763                                        size_t b, enum bch_data_type type,
 764                                        unsigned sectors, bool gc)
 765 {
 766         struct bucket *g = __bucket(ca, b, gc);
 767         struct bucket_mark old, new;
 768         bool overflow;
 769
 770         BUG_ON(type != BCH_DATA_SB &&
 771                type != BCH_DATA_JOURNAL);
 772
 773         old = bucket_cmpxchg(g, new, ({
 774                 new.data_type   = type;
 775                 overflow = checked_add(new.dirty_sectors, sectors);
 776         }));
 777
 778         bch2_fs_inconsistent_on(old.data_type &&
 779                                 old.data_type != type, c,
 780                 "different types of data in same bucket: %s, %s",
 781                 bch2_data_types[old.data_type],
 782                 bch2_data_types[type]);
 783
 784         bch2_fs_inconsistent_on(overflow, c,
 785                 "bucket sector count overflow: %u + %u > U16_MAX",
 786                 old.dirty_sectors, sectors);
 787
 788         if (c)
 789                 bch2_dev_usage_update(c, ca, fs_usage_ptr(c, 0, gc),
 790                                       old, new, gc);
 791
 792         return 0;
 793 }
 794
 795 void bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 796                                size_t b, enum bch_data_type type,
 797                                unsigned sectors, struct gc_pos pos,
 798                                unsigned flags)
 799 {
 800         BUG_ON(type != BCH_DATA_SB &&
 801                type != BCH_DATA_JOURNAL);
 802
 803         preempt_disable();
 804
 805         if (likely(c)) {
 806                 do_mark_fn(__bch2_mark_metadata_bucket, c, pos, flags,
 807                            ca, b, type, sectors);
 808         } else {
 809                 __bch2_mark_metadata_bucket(c, ca, b, type, sectors, 0);
 810         }
 811
 812         preempt_enable();
 813 }
 814
 815 static s64 disk_sectors_scaled(unsigned n, unsigned d, unsigned sectors)
 816 {
 817         return DIV_ROUND_UP(sectors * n, d);
 818 }
 819
 820 static s64 __ptr_disk_sectors_delta(unsigned old_size,
 821                                     unsigned offset, s64 delta,
 822                                     unsigned flags,
 823                                     unsigned n, unsigned d)
 824 {
 825         BUG_ON(!n || !d);
 826
 827         if (flags & BCH_BUCKET_MARK_OVERWRITE_SPLIT) {
 828                 BUG_ON(offset + -delta > old_size);
 829
 830                 return -disk_sectors_scaled(n, d, old_size) +
 831                         disk_sectors_scaled(n, d, offset) +
 832                         disk_sectors_scaled(n, d, old_size - offset + delta);
 833         } else if (flags & BCH_BUCKET_MARK_OVERWRITE) {
 834                 BUG_ON(offset + -delta > old_size);
 835
 836                 return -disk_sectors_scaled(n, d, old_size) +
 837                         disk_sectors_scaled(n, d, old_size + delta);
 838         } else {
 839                 return  disk_sectors_scaled(n, d, delta);
 840         }
 841 }
 842
 843 static s64 ptr_disk_sectors_delta(struct extent_ptr_decoded p,
 844                                   unsigned offset, s64 delta,
 845                                   unsigned flags)
 846 {
 847         return __ptr_disk_sectors_delta(p.crc.live_size,
 848                                         offset, delta, flags,
 849                                         p.crc.compressed_size,
 850                                         p.crc.uncompressed_size);
 851 }
 852
 853 static void bucket_set_stripe(struct bch_fs *c,
 854                               const struct bch_stripe *v,
 855                               struct bch_fs_usage *fs_usage,
 856                               u64 journal_seq,
 857                               unsigned flags)
 858 {
 859         bool enabled = !(flags & BCH_BUCKET_MARK_OVERWRITE);
 860         bool gc = flags & BCH_BUCKET_MARK_GC;
 861         unsigned i;
 862
 863         for (i = 0; i < v->nr_blocks; i++) {
 864                 const struct bch_extent_ptr *ptr = v->ptrs + i;
 865                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 866                 struct bucket *g = PTR_BUCKET(ca, ptr, gc);
 867                 struct bucket_mark new, old;
 868
 869                 old = bucket_cmpxchg(g, new, ({
 870                         new.stripe                      = enabled;
 871                         if (journal_seq) {
 872                                 new.journal_seq_valid   = 1;
 873                                 new.journal_seq         = journal_seq;
 874                         }
 875                 }));
 876
 877                 bch2_dev_usage_update(c, ca, fs_usage, old, new, gc);
 878
 879                 /*
 880                  * XXX write repair code for these, flag stripe as possibly bad
 881                  */
 882                 if (old.gen != ptr->gen)
 883                         bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 884                                       "stripe with stale pointer");
 885 #if 0
 886                 /*
 887                  * We'd like to check for these, but these checks don't work
 888                  * yet:
 889                  */
 890                 if (old.stripe && enabled)
 891                         bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 892                                       "multiple stripes using same bucket");
 893
 894                 if (!old.stripe && !enabled)
 895                         bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 896                                       "deleting stripe but bucket not marked as stripe bucket");
 897 #endif
 898         }
 899 }
 900
 901 static bool bch2_mark_pointer(struct bch_fs *c,
 902                               struct extent_ptr_decoded p,
 903                               s64 sectors, enum bch_data_type data_type,
 904                               struct bch_fs_usage *fs_usage,
 905                               u64 journal_seq, unsigned flags)
 906 {
 907         bool gc = flags & BCH_BUCKET_MARK_GC;
 908         struct bucket_mark old, new;
 909         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 910         struct bucket *g = PTR_BUCKET(ca, &p.ptr, gc);
 911         bool overflow;
 912         u64 v;
 913
 914         v = atomic64_read(&g->_mark.v);
 915         do {
 916                 new.v.counter = old.v.counter = v;
 917
 918                 /*
 919                  * Check this after reading bucket mark to guard against
 920                  * the allocator invalidating a bucket after we've already
 921                  * checked the gen
 922                  */
 923                 if (gen_after(p.ptr.gen, new.gen)) {
 924                         bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 925                                       "pointer gen in the future");
 926                         return true;
 927                 }
 928
 929                 if (new.gen != p.ptr.gen) {
 930                         /* XXX write repair code for this */
 931                         if (!p.ptr.cached &&
 932                             test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
 933                                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 934                                               "stale dirty pointer");
 935                         return true;
 936                 }
 937
 938                 if (!p.ptr.cached)
 939                         overflow = checked_add(new.dirty_sectors, sectors);
 940                 else
 941                         overflow = checked_add(new.cached_sectors, sectors);
 942
 943                 if (!new.dirty_sectors &&
 944                     !new.cached_sectors) {
 945                         new.data_type   = 0;
 946
 947                         if (journal_seq) {
 948                                 new.journal_seq_valid = 1;
 949                                 new.journal_seq = journal_seq;
 950                         }
 951                 } else {
 952                         new.data_type = data_type;
 953                 }
 954
 955                 if (flags & BCH_BUCKET_MARK_NOATOMIC) {
 956                         g->_mark = new;
 957                         break;
 958                 }
 959         } while ((v = atomic64_cmpxchg(&g->_mark.v,
 960                               old.v.counter,
 961                               new.v.counter)) != old.v.counter);
 962
 963         if (old.data_type && old.data_type != data_type)
 964                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 965                         "bucket %u:%zu gen %u different types of data in same bucket: %s, %s",
 966                         p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr),
 967                         new.gen,
 968                         bch2_data_types[old.data_type],
 969                         bch2_data_types[data_type]);
 970
 971         bch2_fs_inconsistent_on(overflow, c,
 972                 "bucket sector count overflow: %u + %lli > U16_MAX",
 973                 !p.ptr.cached
 974                 ? old.dirty_sectors
 975                 : old.cached_sectors, sectors);
 976
 977         bch2_dev_usage_update(c, ca, fs_usage, old, new, gc);
 978
 979         BUG_ON(!gc && bucket_became_unavailable(old, new));
 980
 981         return false;
 982 }
 983
 984 static int bch2_mark_stripe_ptr(struct bch_fs *c,
 985                                 struct bch_extent_stripe_ptr p,
 986                                 enum bch_data_type data_type,
 987                                 struct bch_fs_usage *fs_usage,
 988                                 s64 sectors, unsigned flags,
 989                                 struct bch_replicas_padded *r,
 990                                 unsigned *nr_data,
 991                                 unsigned *nr_parity)
 992 {
 993         bool gc = flags & BCH_BUCKET_MARK_GC;
 994         struct stripe *m;
 995         unsigned old, new;
 996         int blocks_nonempty_delta;
 997
 998         m = genradix_ptr(&c->stripes[gc], p.idx);
 999
1000         spin_lock(&c->ec_stripes_heap_lock);
1001
1002         if (!m || !m->alive) {
1003                 spin_unlock(&c->ec_stripes_heap_lock);
1004                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
1005                                     (u64) p.idx);
1006                 return -EIO;
1007         }
1008
1009         BUG_ON(m->r.e.data_type != data_type);
1010
1011         *nr_data        = m->nr_blocks - m->nr_redundant;
1012         *nr_parity      = m->nr_redundant;
1013         *r = m->r;
1014
1015         old = m->block_sectors[p.block];
1016         m->block_sectors[p.block] += sectors;
1017         new = m->block_sectors[p.block];
1018
1019         blocks_nonempty_delta = (int) !!new - (int) !!old;
1020         if (blocks_nonempty_delta) {
1021                 m->blocks_nonempty += blocks_nonempty_delta;
1022
1023                 if (!gc)
1024                         bch2_stripes_heap_update(c, m, p.idx);
1025         }
1026
1027         m->dirty = true;
1028
1029         spin_unlock(&c->ec_stripes_heap_lock);
1030
1031         return 0;
1032 }
1033
1034 static int bch2_mark_extent(struct bch_fs *c, struct bkey_s_c k,
1035                             unsigned offset, s64 sectors,
1036                             enum bch_data_type data_type,
1037                             struct bch_fs_usage *fs_usage,
1038                             unsigned journal_seq, unsigned flags)
1039 {
1040         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1041         const union bch_extent_entry *entry;
1042         struct extent_ptr_decoded p;
1043         struct bch_replicas_padded r;
1044         s64 dirty_sectors = 0;
1045         int ret;
1046
1047         r.e.data_type   = data_type;
1048         r.e.nr_devs     = 0;
1049         r.e.nr_required = 1;
1050
1051         BUG_ON(!sectors);
1052
1053         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1054                 s64 disk_sectors = data_type == BCH_DATA_BTREE
1055                         ? sectors
1056                         : ptr_disk_sectors_delta(p, offset, sectors, flags);
1057                 bool stale = bch2_mark_pointer(c, p, disk_sectors, data_type,
1058                                                fs_usage, journal_seq, flags);
1059
1060                 if (p.ptr.cached) {
1061                         if (!stale)
1062                                 update_cached_sectors(c, fs_usage, p.ptr.dev,
1063                                                       disk_sectors);
1064                 } else if (!p.has_ec) {
1065                         dirty_sectors          += disk_sectors;
1066                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1067                 } else {
1068                         struct bch_replicas_padded ec_r;
1069                         unsigned nr_data, nr_parity;
1070                         s64 parity_sectors;
1071
1072                         ret = bch2_mark_stripe_ptr(c, p.ec, data_type,
1073                                         fs_usage, disk_sectors, flags,
1074                                         &ec_r, &nr_data, &nr_parity);
1075                         if (ret)
1076                                 return ret;
1077
1078                         parity_sectors =
1079                                 __ptr_disk_sectors_delta(p.crc.live_size,
1080                                         offset, sectors, flags,
1081                                         p.crc.compressed_size * nr_parity,
1082                                         p.crc.uncompressed_size * nr_data);
1083
1084                         update_replicas(c, fs_usage, &ec_r.e,
1085                                         disk_sectors + parity_sectors);
1086
1087                         /*
1088                          * There may be other dirty pointers in this extent, but
1089                          * if so they're not required for mounting if we have an
1090                          * erasure coded pointer in this extent:
1091                          */
1092                         r.e.nr_required = 0;
1093                 }
1094         }
1095
1096         if (r.e.nr_devs)
1097                 update_replicas(c, fs_usage, &r.e, dirty_sectors);
1098
1099         return 0;
1100 }
1101
1102 static int bch2_mark_stripe(struct bch_fs *c, struct bkey_s_c k,
1103                             struct bch_fs_usage *fs_usage,
1104                             u64 journal_seq, unsigned flags)
1105 {
1106         bool gc = flags & BCH_BUCKET_MARK_GC;
1107         struct bkey_s_c_stripe s = bkey_s_c_to_stripe(k);
1108         size_t idx = s.k->p.offset;
1109         struct stripe *m = genradix_ptr(&c->stripes[gc], idx);
1110         unsigned i;
1111
1112         spin_lock(&c->ec_stripes_heap_lock);
1113
1114         if (!m || ((flags & BCH_BUCKET_MARK_OVERWRITE) && !m->alive)) {
1115                 spin_unlock(&c->ec_stripes_heap_lock);
1116                 bch_err_ratelimited(c, "error marking nonexistent stripe %zu",
1117                                     idx);
1118                 return -1;
1119         }
1120
1121         if (!(flags & BCH_BUCKET_MARK_OVERWRITE)) {
1122                 m->sectors      = le16_to_cpu(s.v->sectors);
1123                 m->algorithm    = s.v->algorithm;
1124                 m->nr_blocks    = s.v->nr_blocks;
1125                 m->nr_redundant = s.v->nr_redundant;
1126
1127                 bch2_bkey_to_replicas(&m->r.e, k);
1128
1129                 /*
1130                  * XXX: account for stripes somehow here
1131                  */
1132 #if 0
1133                 update_replicas(c, fs_usage, &m->r.e, stripe_sectors);
1134 #endif
1135
1136                 /* gc recalculates these fields: */
1137                 if (!(flags & BCH_BUCKET_MARK_GC)) {
1138                         for (i = 0; i < s.v->nr_blocks; i++) {
1139                                 m->block_sectors[i] =
1140                                         stripe_blockcount_get(s.v, i);
1141                                 m->blocks_nonempty += !!m->block_sectors[i];
1142                         }
1143                 }
1144
1145                 if (!gc)
1146                         bch2_stripes_heap_update(c, m, idx);
1147                 m->alive        = true;
1148         } else {
1149                 if (!gc)
1150                         bch2_stripes_heap_del(c, m, idx);
1151                 memset(m, 0, sizeof(*m));
1152         }
1153
1154         spin_unlock(&c->ec_stripes_heap_lock);
1155
1156         bucket_set_stripe(c, s.v, fs_usage, 0, flags);
1157         return 0;
1158 }
1159
1160 int bch2_mark_key_locked(struct bch_fs *c,
1161                    struct bkey_s_c k,
1162                    unsigned offset, s64 sectors,
1163                    struct bch_fs_usage *fs_usage,
1164                    u64 journal_seq, unsigned flags)
1165 {
1166         int ret = 0;
1167
1168         preempt_disable();
1169
1170         if (!fs_usage || (flags & BCH_BUCKET_MARK_GC))
1171                 fs_usage = fs_usage_ptr(c, journal_seq,
1172                                         flags & BCH_BUCKET_MARK_GC);
1173
1174         switch (k.k->type) {
1175         case KEY_TYPE_alloc:
1176                 ret = bch2_mark_alloc(c, k, fs_usage, journal_seq, flags);
1177                 break;
1178         case KEY_TYPE_btree_ptr:
1179                 sectors = !(flags & BCH_BUCKET_MARK_OVERWRITE)
1180                         ?  c->opts.btree_node_size
1181                         : -c->opts.btree_node_size;
1182
1183                 ret = bch2_mark_extent(c, k, offset, sectors, BCH_DATA_BTREE,
1184                                 fs_usage, journal_seq, flags);
1185                 break;
1186         case KEY_TYPE_extent:
1187         case KEY_TYPE_reflink_v:
1188                 ret = bch2_mark_extent(c, k, offset, sectors, BCH_DATA_USER,
1189                                 fs_usage, journal_seq, flags);
1190                 break;
1191         case KEY_TYPE_stripe:
1192                 ret = bch2_mark_stripe(c, k, fs_usage, journal_seq, flags);
1193                 break;
1194         case KEY_TYPE_inode:
1195                 if (!(flags & BCH_BUCKET_MARK_OVERWRITE))
1196                         fs_usage->nr_inodes++;
1197                 else
1198                         fs_usage->nr_inodes--;
1199                 break;
1200         case KEY_TYPE_reservation: {
1201                 unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1202
1203                 sectors *= replicas;
1204                 replicas = clamp_t(unsigned, replicas, 1,
1205                                    ARRAY_SIZE(fs_usage->persistent_reserved));
1206
1207                 fs_usage->reserved                              += sectors;
1208                 fs_usage->persistent_reserved[replicas - 1]     += sectors;
1209                 break;
1210         }
1211         }
1212
1213         preempt_enable();
1214
1215         return ret;
1216 }
1217
1218 int bch2_mark_key(struct bch_fs *c, struct bkey_s_c k,
1219                   unsigned offset, s64 sectors,
1220                   struct bch_fs_usage *fs_usage,
1221                   u64 journal_seq, unsigned flags)
1222 {
1223         int ret;
1224
1225         percpu_down_read(&c->mark_lock);
1226         ret = bch2_mark_key_locked(c, k, offset, sectors,
1227                                    fs_usage, journal_seq, flags);
1228         percpu_up_read(&c->mark_lock);
1229
1230         return ret;
1231 }
1232
1233 inline int bch2_mark_overwrite(struct btree_trans *trans,
1234                                struct btree_iter *iter,
1235                                struct bkey_s_c old,
1236                                struct bkey_i *new,
1237                                struct bch_fs_usage *fs_usage,
1238                                unsigned flags)
1239 {
1240         struct bch_fs           *c = trans->c;
1241         struct btree            *b = iter->l[0].b;
1242         unsigned                offset = 0;
1243         s64                     sectors = 0;
1244
1245         flags |= BCH_BUCKET_MARK_OVERWRITE;
1246
1247         if (btree_node_is_extents(b)
1248             ? bkey_cmp(new->k.p, bkey_start_pos(old.k)) <= 0
1249             : bkey_cmp(new->k.p, old.k->p))
1250                 return 0;
1251
1252         if (btree_node_is_extents(b)) {
1253                 switch (bch2_extent_overlap(&new->k, old.k)) {
1254                 case BCH_EXTENT_OVERLAP_ALL:
1255                         offset = 0;
1256                         sectors = -((s64) old.k->size);
1257                         break;
1258                 case BCH_EXTENT_OVERLAP_BACK:
1259                         offset = bkey_start_offset(&new->k) -
1260                                 bkey_start_offset(old.k);
1261                         sectors = bkey_start_offset(&new->k) -
1262                                 old.k->p.offset;
1263                         break;
1264                 case BCH_EXTENT_OVERLAP_FRONT:
1265                         offset = 0;
1266                         sectors = bkey_start_offset(old.k) -
1267                                 new->k.p.offset;
1268                         break;
1269                 case BCH_EXTENT_OVERLAP_MIDDLE:
1270                         offset = bkey_start_offset(&new->k) -
1271                                 bkey_start_offset(old.k);
1272                         sectors = -((s64) new->k.size);
1273                         flags |= BCH_BUCKET_MARK_OVERWRITE_SPLIT;
1274                         break;
1275                 }
1276
1277                 BUG_ON(sectors >= 0);
1278         }
1279
1280         return bch2_mark_key_locked(c, old, offset, sectors, fs_usage,
1281                                     trans->journal_res.seq, flags) ?: 1;
1282 }
1283
1284 int bch2_mark_update(struct btree_trans *trans,
1285                      struct btree_insert_entry *insert,
1286                      struct bch_fs_usage *fs_usage,
1287                      unsigned flags)
1288 {
1289         struct bch_fs           *c = trans->c;
1290         struct btree_iter       *iter = insert->iter;
1291         struct btree            *b = iter->l[0].b;
1292         struct btree_node_iter  node_iter = iter->l[0].iter;
1293         struct bkey_packed      *_k;
1294         int ret = 0;
1295
1296         if (!btree_node_type_needs_gc(iter->btree_id))
1297                 return 0;
1298
1299         bch2_mark_key_locked(c, bkey_i_to_s_c(insert->k),
1300                 0, insert->k->k.size,
1301                 fs_usage, trans->journal_res.seq,
1302                 BCH_BUCKET_MARK_INSERT|flags);
1303
1304         if (unlikely(trans->flags & BTREE_INSERT_NOMARK_OVERWRITES))
1305                 return 0;
1306
1307         /*
1308          * For non extents, we only mark the new key, not the key being
1309          * overwritten - unless we're actually deleting:
1310          */
1311         if ((iter->btree_id == BTREE_ID_ALLOC ||
1312              iter->btree_id == BTREE_ID_EC) &&
1313             !bkey_deleted(&insert->k->k))
1314                 return 0;
1315
1316         while ((_k = bch2_btree_node_iter_peek_filter(&node_iter, b,
1317                                                       KEY_TYPE_discard))) {
1318                 struct bkey             unpacked;
1319                 struct bkey_s_c         k = bkey_disassemble(b, _k, &unpacked);
1320
1321                 ret = bch2_mark_overwrite(trans, iter, k, insert->k,
1322                                           fs_usage, flags);
1323                 if (ret <= 0)
1324                         break;
1325
1326                 bch2_btree_node_iter_advance(&node_iter, b);
1327         }
1328
1329         return ret;
1330 }
1331
1332 void bch2_trans_fs_usage_apply(struct btree_trans *trans,
1333                                struct bch_fs_usage_online *fs_usage)
1334 {
1335         struct bch_fs *c = trans->c;
1336         struct btree_insert_entry *i;
1337         static int warned_disk_usage = 0;
1338         u64 disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1339         char buf[200];
1340
1341         if (!bch2_fs_usage_apply(c, fs_usage, trans->disk_res,
1342                                  trans->journal_res.seq) ||
1343             warned_disk_usage ||
1344             xchg(&warned_disk_usage, 1))
1345                 return;
1346
1347         bch_err(c, "disk usage increased more than %llu sectors reserved",
1348                 disk_res_sectors);
1349
1350         trans_for_each_update(trans, i) {
1351                 struct btree_iter       *iter = i->iter;
1352                 struct btree            *b = iter->l[0].b;
1353                 struct btree_node_iter  node_iter = iter->l[0].iter;
1354                 struct bkey_packed      *_k;
1355
1356                 pr_err("while inserting");
1357                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
1358                 pr_err("%s", buf);
1359                 pr_err("overlapping with");
1360
1361                 node_iter = iter->l[0].iter;
1362                 while ((_k = bch2_btree_node_iter_peek_filter(&node_iter, b,
1363                                                         KEY_TYPE_discard))) {
1364                         struct bkey             unpacked;
1365                         struct bkey_s_c         k;
1366
1367                         k = bkey_disassemble(b, _k, &unpacked);
1368
1369                         if (btree_node_is_extents(b)
1370                             ? bkey_cmp(i->k->k.p, bkey_start_pos(k.k)) <= 0
1371                             : bkey_cmp(i->k->k.p, k.k->p))
1372                                 break;
1373
1374                         bch2_bkey_val_to_text(&PBUF(buf), c, k);
1375                         pr_err("%s", buf);
1376
1377                         bch2_btree_node_iter_advance(&node_iter, b);
1378                 }
1379         }
1380 }
1381
1382 /* trans_mark: */
1383
1384 static int trans_get_key(struct btree_trans *trans,
1385                          enum btree_id btree_id, struct bpos pos,
1386                          struct btree_iter **iter,
1387                          struct bkey_s_c *k)
1388 {
1389         struct btree_insert_entry *i;
1390         int ret;
1391
1392         trans_for_each_update(trans, i)
1393                 if (i->iter->btree_id == btree_id &&
1394                     (btree_node_type_is_extents(btree_id)
1395                      ? bkey_cmp(pos, bkey_start_pos(&i->k->k)) >= 0 &&
1396                        bkey_cmp(pos, i->k->k.p) < 0
1397                      : !bkey_cmp(pos, i->iter->pos))) {
1398                         *iter   = i->iter;
1399                         *k      = bkey_i_to_s_c(i->k);
1400                         return 1;
1401                 }
1402
1403         *iter = bch2_trans_get_iter(trans, btree_id, pos,
1404                                     BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
1405         if (IS_ERR(*iter))
1406                 return PTR_ERR(*iter);
1407
1408         *k = bch2_btree_iter_peek_slot(*iter);
1409         ret = bkey_err(*k);
1410         if (ret)
1411                 bch2_trans_iter_put(trans, *iter);
1412         return ret;
1413 }
1414
1415 static void *trans_update_key(struct btree_trans *trans,
1416                               struct btree_iter *iter,
1417                               unsigned u64s)
1418 {
1419         struct btree_insert_entry *i;
1420         struct bkey_i *new_k;
1421
1422         new_k = bch2_trans_kmalloc(trans, u64s * sizeof(u64));
1423         if (IS_ERR(new_k))
1424                 return new_k;
1425
1426         bkey_init(&new_k->k);
1427         new_k->k.p = iter->pos;
1428
1429         trans_for_each_update(trans, i)
1430                 if (i->iter == iter) {
1431                         i->k = new_k;
1432                         return new_k;
1433                 }
1434
1435         bch2_trans_update(trans, iter, new_k);
1436         return new_k;
1437 }
1438
1439 static int bch2_trans_mark_pointer(struct btree_trans *trans,
1440                         struct extent_ptr_decoded p,
1441                         s64 sectors, enum bch_data_type data_type)
1442 {
1443         struct bch_fs *c = trans->c;
1444         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
1445         struct btree_iter *iter;
1446         struct bkey_s_c k;
1447         struct bkey_alloc_unpacked u;
1448         struct bkey_i_alloc *a;
1449         unsigned old;
1450         bool overflow;
1451         int ret;
1452
1453         ret = trans_get_key(trans, BTREE_ID_ALLOC,
1454                             POS(p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr)),
1455                             &iter, &k);
1456         if (ret < 0)
1457                 return ret;
1458
1459         if (!ret) {
1460                 /*
1461                  * During journal replay, and if gc repairs alloc info at
1462                  * runtime, the alloc info in the btree might not be up to date
1463                  * yet - so, trust the in memory mark:
1464                  */
1465                 struct bucket *g;
1466                 struct bucket_mark m;
1467
1468                 percpu_down_read(&c->mark_lock);
1469                 g       = bucket(ca, iter->pos.offset);
1470                 m       = READ_ONCE(g->mark);
1471                 u       = alloc_mem_to_key(g, m);
1472                 percpu_up_read(&c->mark_lock);
1473         } else {
1474                 /*
1475                  * Unless we're already updating that key:
1476                  */
1477                 if (k.k->type != KEY_TYPE_alloc) {
1478                         bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1479                                       "pointer to nonexistent bucket %llu:%llu",
1480                                       iter->pos.inode, iter->pos.offset);
1481                         ret = -1;
1482                         goto out;
1483                 }
1484
1485                 u = bch2_alloc_unpack(k);
1486         }
1487
1488         if (gen_after(u.gen, p.ptr.gen)) {
1489                 ret = 1;
1490                 goto out;
1491         }
1492
1493         if (u.data_type && u.data_type != data_type) {
1494                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
1495                         "bucket %llu:%llu gen %u different types of data in same bucket: %s, %s",
1496                         iter->pos.inode, iter->pos.offset,
1497                         u.gen,
1498                         bch2_data_types[u.data_type],
1499                         bch2_data_types[data_type]);
1500                 ret = -1;
1501                 goto out;
1502         }
1503
1504         if (!p.ptr.cached) {
1505                 old = u.dirty_sectors;
1506                 overflow = checked_add(u.dirty_sectors, sectors);
1507         } else {
1508                 old = u.cached_sectors;
1509                 overflow = checked_add(u.cached_sectors, sectors);
1510         }
1511
1512         u.data_type = u.dirty_sectors || u.cached_sectors
1513                 ? data_type : 0;
1514
1515         bch2_fs_inconsistent_on(overflow, c,
1516                 "bucket sector count overflow: %u + %lli > U16_MAX",
1517                 old, sectors);
1518         BUG_ON(overflow);
1519
1520         a = trans_update_key(trans, iter, BKEY_ALLOC_U64s_MAX);
1521         ret = PTR_ERR_OR_ZERO(a);
1522         if (ret)
1523                 goto out;
1524
1525         bkey_alloc_init(&a->k_i);
1526         a->k.p = iter->pos;
1527         bch2_alloc_pack(a, u);
1528 out:
1529         bch2_trans_iter_put(trans, iter);
1530         return ret;
1531 }
1532
1533 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1534                         struct bch_extent_stripe_ptr p,
1535                         s64 sectors, enum bch_data_type data_type,
1536                         struct bch_replicas_padded *r,
1537                         unsigned *nr_data,
1538                         unsigned *nr_parity)
1539 {
1540         struct bch_fs *c = trans->c;
1541         struct btree_iter *iter;
1542         struct bkey_i *new_k;
1543         struct bkey_s_c k;
1544         struct bkey_s_stripe s;
1545         int ret = 0;
1546
1547         ret = trans_get_key(trans, BTREE_ID_EC, POS(0, p.idx), &iter, &k);
1548         if (ret < 0)
1549                 return ret;
1550
1551         if (k.k->type != KEY_TYPE_stripe) {
1552                 bch2_fs_inconsistent(c,
1553                         "pointer to nonexistent stripe %llu",
1554                         (u64) p.idx);
1555                 ret = -EIO;
1556                 goto out;
1557         }
1558
1559         new_k = trans_update_key(trans, iter, k.k->u64s);
1560         ret = PTR_ERR_OR_ZERO(new_k);
1561         if (ret)
1562                 goto out;
1563
1564         bkey_reassemble(new_k, k);
1565         s = bkey_i_to_s_stripe(new_k);
1566
1567         stripe_blockcount_set(s.v, p.block,
1568                 stripe_blockcount_get(s.v, p.block) +
1569                 sectors);
1570
1571         *nr_data        = s.v->nr_blocks - s.v->nr_redundant;
1572         *nr_parity      = s.v->nr_redundant;
1573         bch2_bkey_to_replicas(&r->e, s.s_c);
1574 out:
1575         bch2_trans_iter_put(trans, iter);
1576         return ret;
1577 }
1578
1579 static int bch2_trans_mark_extent(struct btree_trans *trans,
1580                         struct bkey_s_c k, unsigned offset,
1581                         s64 sectors, unsigned flags,
1582                         enum bch_data_type data_type)
1583 {
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         s64 dirty_sectors = 0;
1589         bool stale;
1590         int ret;
1591
1592         r.e.data_type   = data_type;
1593         r.e.nr_devs     = 0;
1594         r.e.nr_required = 1;
1595
1596         BUG_ON(!sectors);
1597
1598         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1599                 s64 disk_sectors = data_type == BCH_DATA_BTREE
1600                         ? sectors
1601                         : ptr_disk_sectors_delta(p, offset, sectors, flags);
1602
1603                 ret = bch2_trans_mark_pointer(trans, p, disk_sectors,
1604                                               data_type);
1605                 if (ret < 0)
1606                         return ret;
1607
1608                 stale = ret > 0;
1609
1610                 if (p.ptr.cached) {
1611                         if (!stale)
1612                                 update_cached_sectors_list(trans, p.ptr.dev,
1613                                                            disk_sectors);
1614                 } else if (!p.has_ec) {
1615                         dirty_sectors          += disk_sectors;
1616                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1617                 } else {
1618                         struct bch_replicas_padded ec_r;
1619                         unsigned nr_data, nr_parity;
1620                         s64 parity_sectors;
1621
1622                         ret = bch2_trans_mark_stripe_ptr(trans, p.ec,
1623                                         disk_sectors, data_type,
1624                                         &ec_r, &nr_data, &nr_parity);
1625                         if (ret)
1626                                 return ret;
1627
1628                         parity_sectors =
1629                                 __ptr_disk_sectors_delta(p.crc.live_size,
1630                                         offset, sectors, flags,
1631                                         p.crc.compressed_size * nr_parity,
1632                                         p.crc.uncompressed_size * nr_data);
1633
1634                         update_replicas_list(trans, &ec_r.e,
1635                                              disk_sectors + parity_sectors);
1636
1637                         r.e.nr_required = 0;
1638                 }
1639         }
1640
1641         if (r.e.nr_devs)
1642                 update_replicas_list(trans, &r.e, dirty_sectors);
1643
1644         return 0;
1645 }
1646
1647 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1648                         struct bkey_s_c_reflink_p p,
1649                         u64 idx, unsigned sectors,
1650                         unsigned flags)
1651 {
1652         struct bch_fs *c = trans->c;
1653         struct btree_iter *iter;
1654         struct bkey_i *new_k;
1655         struct bkey_s_c k;
1656         struct bkey_i_reflink_v *r_v;
1657         s64 ret;
1658
1659         ret = trans_get_key(trans, BTREE_ID_REFLINK,
1660                             POS(0, idx), &iter, &k);
1661         if (ret < 0)
1662                 return ret;
1663
1664         if (k.k->type != KEY_TYPE_reflink_v) {
1665                 bch2_fs_inconsistent(c,
1666                         "%llu:%llu len %u points to nonexistent indirect extent %llu",
1667                         p.k->p.inode, p.k->p.offset, p.k->size, idx);
1668                 ret = -EIO;
1669                 goto err;
1670         }
1671
1672         if ((flags & BCH_BUCKET_MARK_OVERWRITE) &&
1673             (bkey_start_offset(k.k) < idx ||
1674              k.k->p.offset > idx + sectors))
1675                 goto out;
1676
1677         bch2_btree_iter_set_pos(iter, bkey_start_pos(k.k));
1678         BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
1679
1680         new_k = trans_update_key(trans, iter, k.k->u64s);
1681         ret = PTR_ERR_OR_ZERO(new_k);
1682         if (ret)
1683                 goto err;
1684
1685         bkey_reassemble(new_k, k);
1686         r_v = bkey_i_to_reflink_v(new_k);
1687
1688         le64_add_cpu(&r_v->v.refcount,
1689                      !(flags & BCH_BUCKET_MARK_OVERWRITE) ? 1 : -1);
1690
1691         if (!r_v->v.refcount) {
1692                 r_v->k.type = KEY_TYPE_deleted;
1693                 set_bkey_val_u64s(&r_v->k, 0);
1694         }
1695 out:
1696         ret = k.k->p.offset - idx;
1697 err:
1698         bch2_trans_iter_put(trans, iter);
1699         return ret;
1700 }
1701
1702 static int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1703                         struct bkey_s_c_reflink_p p, unsigned offset,
1704                         s64 sectors, unsigned flags)
1705 {
1706         u64 idx = le64_to_cpu(p.v->idx) + offset;
1707         s64 ret = 0;
1708
1709         sectors = abs(sectors);
1710         BUG_ON(offset + sectors > p.k->size);
1711
1712         while (sectors) {
1713                 ret = __bch2_trans_mark_reflink_p(trans, p, idx, sectors, flags);
1714                 if (ret < 0)
1715                         break;
1716
1717                 idx += ret;
1718                 sectors = max_t(s64, 0LL, sectors - ret);
1719                 ret = 0;
1720         }
1721
1722         return ret;
1723 }
1724
1725 int bch2_trans_mark_key(struct btree_trans *trans, struct bkey_s_c k,
1726                         unsigned offset, s64 sectors, unsigned flags)
1727 {
1728         struct replicas_delta_list *d;
1729         struct bch_fs *c = trans->c;
1730
1731         switch (k.k->type) {
1732         case KEY_TYPE_btree_ptr:
1733                 sectors = !(flags & BCH_BUCKET_MARK_OVERWRITE)
1734                         ?  c->opts.btree_node_size
1735                         : -c->opts.btree_node_size;
1736
1737                 return bch2_trans_mark_extent(trans, k, offset, sectors,
1738                                               flags, BCH_DATA_BTREE);
1739         case KEY_TYPE_extent:
1740         case KEY_TYPE_reflink_v:
1741                 return bch2_trans_mark_extent(trans, k, offset, sectors,
1742                                               flags, BCH_DATA_USER);
1743         case KEY_TYPE_inode:
1744                 d = replicas_deltas_realloc(trans, 0);
1745
1746                 if (!(flags & BCH_BUCKET_MARK_OVERWRITE))
1747                         d->nr_inodes++;
1748                 else
1749                         d->nr_inodes--;
1750                 return 0;
1751         case KEY_TYPE_reservation: {
1752                 unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1753
1754                 d = replicas_deltas_realloc(trans, 0);
1755
1756                 sectors *= replicas;
1757                 replicas = clamp_t(unsigned, replicas, 1,
1758                                    ARRAY_SIZE(d->persistent_reserved));
1759
1760                 d->persistent_reserved[replicas - 1] += sectors;
1761                 return 0;
1762         }
1763         case KEY_TYPE_reflink_p:
1764                 return bch2_trans_mark_reflink_p(trans,
1765                                         bkey_s_c_to_reflink_p(k),
1766                                         offset, sectors, flags);
1767         default:
1768                 return 0;
1769         }
1770 }
1771
1772 int bch2_trans_mark_update(struct btree_trans *trans,
1773                            struct btree_iter *iter,
1774                            struct bkey_i *insert)
1775 {
1776         struct btree            *b = iter->l[0].b;
1777         struct btree_node_iter  node_iter = iter->l[0].iter;
1778         struct bkey_packed      *_k;
1779         int ret;
1780
1781         if (!btree_node_type_needs_gc(iter->btree_id))
1782                 return 0;
1783
1784         ret = bch2_trans_mark_key(trans, bkey_i_to_s_c(insert),
1785                         0, insert->k.size, BCH_BUCKET_MARK_INSERT);
1786         if (ret)
1787                 return ret;
1788
1789         if (unlikely(trans->flags & BTREE_INSERT_NOMARK_OVERWRITES))
1790                 return 0;
1791
1792         while ((_k = bch2_btree_node_iter_peek_filter(&node_iter, b,
1793                                                       KEY_TYPE_discard))) {
1794                 struct bkey             unpacked;
1795                 struct bkey_s_c         k;
1796                 unsigned                offset = 0;
1797                 s64                     sectors = 0;
1798                 unsigned                flags = BCH_BUCKET_MARK_OVERWRITE;
1799
1800                 k = bkey_disassemble(b, _k, &unpacked);
1801
1802                 if (btree_node_is_extents(b)
1803                     ? bkey_cmp(insert->k.p, bkey_start_pos(k.k)) <= 0
1804                     : bkey_cmp(insert->k.p, k.k->p))
1805                         break;
1806
1807                 if (btree_node_is_extents(b)) {
1808                         switch (bch2_extent_overlap(&insert->k, k.k)) {
1809                         case BCH_EXTENT_OVERLAP_ALL:
1810                                 offset = 0;
1811                                 sectors = -((s64) k.k->size);
1812                                 break;
1813                         case BCH_EXTENT_OVERLAP_BACK:
1814                                 offset = bkey_start_offset(&insert->k) -
1815                                         bkey_start_offset(k.k);
1816                                 sectors = bkey_start_offset(&insert->k) -
1817                                         k.k->p.offset;
1818                                 break;
1819                         case BCH_EXTENT_OVERLAP_FRONT:
1820                                 offset = 0;
1821                                 sectors = bkey_start_offset(k.k) -
1822                                         insert->k.p.offset;
1823                                 break;
1824                         case BCH_EXTENT_OVERLAP_MIDDLE:
1825                                 offset = bkey_start_offset(&insert->k) -
1826                                         bkey_start_offset(k.k);
1827                                 sectors = -((s64) insert->k.size);
1828                                 flags |= BCH_BUCKET_MARK_OVERWRITE_SPLIT;
1829                                 break;
1830                         }
1831
1832                         BUG_ON(sectors >= 0);
1833                 }
1834
1835                 ret = bch2_trans_mark_key(trans, k, offset, sectors, flags);
1836                 if (ret)
1837                         return ret;
1838
1839                 bch2_btree_node_iter_advance(&node_iter, b);
1840         }
1841
1842         return 0;
1843 }
1844
1845 /* Disk reservations: */
1846
1847 #define SECTORS_CACHE   1024
1848
1849 int bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
1850                               unsigned sectors, int flags)
1851 {
1852         struct bch_fs_pcpu *pcpu;
1853         u64 old, v, get;
1854         s64 sectors_available;
1855         int ret;
1856
1857         percpu_down_read(&c->mark_lock);
1858         preempt_disable();
1859         pcpu = this_cpu_ptr(c->pcpu);
1860
1861         if (sectors <= pcpu->sectors_available)
1862                 goto out;
1863
1864         v = atomic64_read(&c->sectors_available);
1865         do {
1866                 old = v;
1867                 get = min((u64) sectors + SECTORS_CACHE, old);
1868
1869                 if (get < sectors) {
1870                         preempt_enable();
1871                         goto recalculate;
1872                 }
1873         } while ((v = atomic64_cmpxchg(&c->sectors_available,
1874                                        old, old - get)) != old);
1875
1876         pcpu->sectors_available         += get;
1877
1878 out:
1879         pcpu->sectors_available         -= sectors;
1880         this_cpu_add(*c->online_reserved, sectors);
1881         res->sectors                    += sectors;
1882
1883         preempt_enable();
1884         percpu_up_read(&c->mark_lock);
1885         return 0;
1886
1887 recalculate:
1888         mutex_lock(&c->sectors_available_lock);
1889
1890         percpu_u64_set(&c->pcpu->sectors_available, 0);
1891         sectors_available = avail_factor(__bch2_fs_usage_read_short(c).free);
1892
1893         if (sectors <= sectors_available ||
1894             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
1895                 atomic64_set(&c->sectors_available,
1896                              max_t(s64, 0, sectors_available - sectors));
1897                 this_cpu_add(*c->online_reserved, sectors);
1898                 res->sectors                    += sectors;
1899                 ret = 0;
1900         } else {
1901                 atomic64_set(&c->sectors_available, sectors_available);
1902                 ret = -ENOSPC;
1903         }
1904
1905         mutex_unlock(&c->sectors_available_lock);
1906         percpu_up_read(&c->mark_lock);
1907
1908         return ret;
1909 }
1910
1911 /* Startup/shutdown: */
1912
1913 static void buckets_free_rcu(struct rcu_head *rcu)
1914 {
1915         struct bucket_array *buckets =
1916                 container_of(rcu, struct bucket_array, rcu);
1917
1918         kvpfree(buckets,
1919                 sizeof(struct bucket_array) +
1920                 buckets->nbuckets * sizeof(struct bucket));
1921 }
1922
1923 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
1924 {
1925         struct bucket_array *buckets = NULL, *old_buckets = NULL;
1926         unsigned long *buckets_nouse = NULL;
1927         alloc_fifo      free[RESERVE_NR];
1928         alloc_fifo      free_inc;
1929         alloc_heap      alloc_heap;
1930         copygc_heap     copygc_heap;
1931
1932         size_t btree_reserve    = DIV_ROUND_UP(BTREE_NODE_RESERVE,
1933                              ca->mi.bucket_size / c->opts.btree_node_size);
1934         /* XXX: these should be tunable */
1935         size_t reserve_none     = max_t(size_t, 1, nbuckets >> 9);
1936         size_t copygc_reserve   = max_t(size_t, 2, nbuckets >> 7);
1937         size_t free_inc_nr      = max(max_t(size_t, 1, nbuckets >> 12),
1938                                       btree_reserve * 2);
1939         bool resize = ca->buckets[0] != NULL,
1940              start_copygc = ca->copygc_thread != NULL;
1941         int ret = -ENOMEM;
1942         unsigned i;
1943
1944         memset(&free,           0, sizeof(free));
1945         memset(&free_inc,       0, sizeof(free_inc));
1946         memset(&alloc_heap,     0, sizeof(alloc_heap));
1947         memset(&copygc_heap,    0, sizeof(copygc_heap));
1948
1949         if (!(buckets           = kvpmalloc(sizeof(struct bucket_array) +
1950                                             nbuckets * sizeof(struct bucket),
1951                                             GFP_KERNEL|__GFP_ZERO)) ||
1952             !(buckets_nouse     = kvpmalloc(BITS_TO_LONGS(nbuckets) *
1953                                             sizeof(unsigned long),
1954                                             GFP_KERNEL|__GFP_ZERO)) ||
1955             !init_fifo(&free[RESERVE_BTREE], btree_reserve, GFP_KERNEL) ||
1956             !init_fifo(&free[RESERVE_MOVINGGC],
1957                        copygc_reserve, GFP_KERNEL) ||
1958             !init_fifo(&free[RESERVE_NONE], reserve_none, GFP_KERNEL) ||
1959             !init_fifo(&free_inc,       free_inc_nr, GFP_KERNEL) ||
1960             !init_heap(&alloc_heap,     ALLOC_SCAN_BATCH(ca) << 1, GFP_KERNEL) ||
1961             !init_heap(&copygc_heap,    copygc_reserve, GFP_KERNEL))
1962                 goto err;
1963
1964         buckets->first_bucket   = ca->mi.first_bucket;
1965         buckets->nbuckets       = nbuckets;
1966
1967         bch2_copygc_stop(ca);
1968
1969         if (resize) {
1970                 down_write(&c->gc_lock);
1971                 down_write(&ca->bucket_lock);
1972                 percpu_down_write(&c->mark_lock);
1973         }
1974
1975         old_buckets = bucket_array(ca);
1976
1977         if (resize) {
1978                 size_t n = min(buckets->nbuckets, old_buckets->nbuckets);
1979
1980                 memcpy(buckets->b,
1981                        old_buckets->b,
1982                        n * sizeof(struct bucket));
1983                 memcpy(buckets_nouse,
1984                        ca->buckets_nouse,
1985                        BITS_TO_LONGS(n) * sizeof(unsigned long));
1986         }
1987
1988         rcu_assign_pointer(ca->buckets[0], buckets);
1989         buckets = old_buckets;
1990
1991         swap(ca->buckets_nouse, buckets_nouse);
1992
1993         if (resize)
1994                 percpu_up_write(&c->mark_lock);
1995
1996         spin_lock(&c->freelist_lock);
1997         for (i = 0; i < RESERVE_NR; i++) {
1998                 fifo_move(&free[i], &ca->free[i]);
1999                 swap(ca->free[i], free[i]);
2000         }
2001         fifo_move(&free_inc, &ca->free_inc);
2002         swap(ca->free_inc, free_inc);
2003         spin_unlock(&c->freelist_lock);
2004
2005         /* with gc lock held, alloc_heap can't be in use: */
2006         swap(ca->alloc_heap, alloc_heap);
2007
2008         /* and we shut down copygc: */
2009         swap(ca->copygc_heap, copygc_heap);
2010
2011         nbuckets = ca->mi.nbuckets;
2012
2013         if (resize) {
2014                 up_write(&ca->bucket_lock);
2015                 up_write(&c->gc_lock);
2016         }
2017
2018         if (start_copygc &&
2019             bch2_copygc_start(c, ca))
2020                 bch_err(ca, "error restarting copygc thread");
2021
2022         ret = 0;
2023 err:
2024         free_heap(&copygc_heap);
2025         free_heap(&alloc_heap);
2026         free_fifo(&free_inc);
2027         for (i = 0; i < RESERVE_NR; i++)
2028                 free_fifo(&free[i]);
2029         kvpfree(buckets_nouse,
2030                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
2031         if (buckets)
2032                 call_rcu(&old_buckets->rcu, buckets_free_rcu);
2033
2034         return ret;
2035 }
2036
2037 void bch2_dev_buckets_free(struct bch_dev *ca)
2038 {
2039         unsigned i;
2040
2041         free_heap(&ca->copygc_heap);
2042         free_heap(&ca->alloc_heap);
2043         free_fifo(&ca->free_inc);
2044         for (i = 0; i < RESERVE_NR; i++)
2045                 free_fifo(&ca->free[i]);
2046         kvpfree(ca->buckets_nouse,
2047                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2048         kvpfree(rcu_dereference_protected(ca->buckets[0], 1),
2049                 sizeof(struct bucket_array) +
2050                 ca->mi.nbuckets * sizeof(struct bucket));
2051
2052         free_percpu(ca->usage[0]);
2053 }
2054
2055 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2056 {
2057         if (!(ca->usage[0] = alloc_percpu(struct bch_dev_usage)))
2058                 return -ENOMEM;
2059
2060         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);;
2061 }