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  * 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
  77 #include <linux/preempt.h>
  78 #include <trace/events/bcachefs.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         bch2_fs_usage_acc_to_base(c, 0);
 126         bch2_fs_usage_acc_to_base(c, 1);
 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 *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 *bch2_fs_usage_scratch_get(struct bch_fs *c)
 160 {
 161         struct bch_fs_usage *ret;
 162         unsigned bytes = fs_usage_u64s(c) * sizeof(u64);
 163
 164         ret = kzalloc(bytes, GFP_NOWAIT);
 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 *bch2_fs_usage_read(struct bch_fs *c)
 223 {
 224         struct bch_fs_usage *ret;
 225         unsigned seq, v, u64s = fs_usage_u64s(c);
 226 retry:
 227         ret = kmalloc(u64s * sizeof(u64), GFP_NOFS);
 228         if (unlikely(!ret))
 229                 return NULL;
 230
 231         percpu_down_read(&c->mark_lock);
 232
 233         v = fs_usage_u64s(c);
 234         if (unlikely(u64s != v)) {
 235                 u64s = v;
 236                 percpu_up_read(&c->mark_lock);
 237                 kfree(ret);
 238                 goto retry;
 239         }
 240
 241         do {
 242                 seq = read_seqcount_begin(&c->usage_lock);
 243                 memcpy(ret, c->usage_base, u64s * sizeof(u64));
 244                 acc_u64s_percpu((u64 *) ret, (u64 __percpu *) c->usage[0], u64s);
 245                 acc_u64s_percpu((u64 *) ret, (u64 __percpu *) c->usage[1], u64s);
 246         } while (read_seqcount_retry(&c->usage_lock, seq));
 247
 248         return ret;
 249 }
 250
 251 void bch2_fs_usage_acc_to_base(struct bch_fs *c, unsigned idx)
 252 {
 253         unsigned u64s = fs_usage_u64s(c);
 254
 255         BUG_ON(idx >= 2);
 256
 257         write_seqcount_begin(&c->usage_lock);
 258
 259         acc_u64s_percpu((u64 *) c->usage_base,
 260                         (u64 __percpu *) c->usage[idx], u64s);
 261         percpu_memset(c->usage[idx], 0, u64s * sizeof(u64));
 262
 263         write_seqcount_end(&c->usage_lock);
 264 }
 265
 266 void bch2_fs_usage_to_text(struct printbuf *out,
 267                            struct bch_fs *c,
 268                            struct bch_fs_usage *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->hidden);
 276         pr_buf(out, "data:\t\t\t\t%llu\n",
 277                fs_usage->data);
 278         pr_buf(out, "cached:\t\t\t\t%llu\n",
 279                fs_usage->cached);
 280         pr_buf(out, "reserved:\t\t\t%llu\n",
 281                fs_usage->reserved);
 282         pr_buf(out, "nr_inodes:\t\t\t%llu\n",
 283                fs_usage->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->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->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->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 *fs_usage)
 318 {
 319         return min(fs_usage->hidden +
 320                    fs_usage->btree +
 321                    fs_usage->data +
 322                    reserve_factor(fs_usage->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                 bch2_fs_usage_read_one(c, &c->usage_base->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 *fs_usage,
 393                         struct disk_reservation *disk_res,
 394                         unsigned journal_seq)
 395 {
 396         s64 added = fs_usage->data + fs_usage->reserved;
 397         s64 should_not_have_added;
 398         int ret = 0;
 399
 400         percpu_rwsem_assert_held(&c->mark_lock);
 401
 402         /*
 403          * Not allowed to reduce sectors_available except by getting a
 404          * reservation:
 405          */
 406         should_not_have_added = added - (s64) (disk_res ? disk_res->sectors : 0);
 407         if (WARN_ONCE(should_not_have_added > 0,
 408                       "disk usage increased by %lli without a reservation",
 409                       should_not_have_added)) {
 410                 atomic64_sub(should_not_have_added, &c->sectors_available);
 411                 added -= should_not_have_added;
 412                 ret = -1;
 413         }
 414
 415         if (added > 0) {
 416                 disk_res->sectors               -= added;
 417                 fs_usage->online_reserved       -= added;
 418         }
 419
 420         preempt_disable();
 421         acc_u64s((u64 *) fs_usage_ptr(c, journal_seq, false),
 422                  (u64 *) fs_usage, fs_usage_u64s(c));
 423         preempt_enable();
 424
 425         return ret;
 426 }
 427
 428 static inline void account_bucket(struct bch_fs_usage *fs_usage,
 429                                   struct bch_dev_usage *dev_usage,
 430                                   enum bch_data_type type,
 431                                   int nr, s64 size)
 432 {
 433         if (type == BCH_DATA_SB || type == BCH_DATA_JOURNAL)
 434                 fs_usage->hidden        += size;
 435
 436         dev_usage->buckets[type]        += nr;
 437 }
 438
 439 static void bch2_dev_usage_update(struct bch_fs *c, struct bch_dev *ca,
 440                                   struct bch_fs_usage *fs_usage,
 441                                   struct bucket_mark old, struct bucket_mark new,
 442                                   bool gc)
 443 {
 444         struct bch_dev_usage *dev_usage;
 445
 446         percpu_rwsem_assert_held(&c->mark_lock);
 447
 448         bch2_fs_inconsistent_on(old.data_type && new.data_type &&
 449                                 old.data_type != new.data_type, c,
 450                 "different types of data in same bucket: %s, %s",
 451                 bch2_data_types[old.data_type],
 452                 bch2_data_types[new.data_type]);
 453
 454         preempt_disable();
 455         dev_usage = this_cpu_ptr(ca->usage[gc]);
 456
 457         if (bucket_type(old))
 458                 account_bucket(fs_usage, dev_usage, bucket_type(old),
 459                                -1, -ca->mi.bucket_size);
 460
 461         if (bucket_type(new))
 462                 account_bucket(fs_usage, dev_usage, bucket_type(new),
 463                                1, ca->mi.bucket_size);
 464
 465         dev_usage->buckets_alloc +=
 466                 (int) new.owned_by_allocator - (int) old.owned_by_allocator;
 467         dev_usage->buckets_ec +=
 468                 (int) new.stripe - (int) old.stripe;
 469         dev_usage->buckets_unavailable +=
 470                 is_unavailable_bucket(new) - is_unavailable_bucket(old);
 471
 472         dev_usage->sectors[old.data_type] -= old.dirty_sectors;
 473         dev_usage->sectors[new.data_type] += new.dirty_sectors;
 474         dev_usage->sectors[BCH_DATA_CACHED] +=
 475                 (int) new.cached_sectors - (int) old.cached_sectors;
 476         dev_usage->sectors_fragmented +=
 477                 is_fragmented_bucket(new, ca) - is_fragmented_bucket(old, ca);
 478         preempt_enable();
 479
 480         if (!is_available_bucket(old) && is_available_bucket(new))
 481                 bch2_wake_allocator(ca);
 482 }
 483
 484 void bch2_dev_usage_from_buckets(struct bch_fs *c)
 485 {
 486         struct bch_dev *ca;
 487         struct bucket_mark old = { .v.counter = 0 };
 488         struct bucket_array *buckets;
 489         struct bucket *g;
 490         unsigned i;
 491         int cpu;
 492
 493         c->usage_base->hidden = 0;
 494
 495         for_each_member_device(ca, c, i) {
 496                 for_each_possible_cpu(cpu)
 497                         memset(per_cpu_ptr(ca->usage[0], cpu), 0,
 498                                sizeof(*ca->usage[0]));
 499
 500                 buckets = bucket_array(ca);
 501
 502                 for_each_bucket(g, buckets)
 503                         bch2_dev_usage_update(c, ca, c->usage_base,
 504                                               old, g->mark, false);
 505         }
 506 }
 507
 508 #define bucket_data_cmpxchg(c, ca, fs_usage, g, new, expr)      \
 509 ({                                                              \
 510         struct bucket_mark _old = bucket_cmpxchg(g, new, expr); \
 511                                                                 \
 512         bch2_dev_usage_update(c, ca, fs_usage, _old, new, gc);  \
 513         _old;                                                   \
 514 })
 515
 516 static inline void update_replicas(struct bch_fs *c,
 517                                    struct bch_fs_usage *fs_usage,
 518                                    struct bch_replicas_entry *r,
 519                                    s64 sectors)
 520 {
 521         int idx = bch2_replicas_entry_idx(c, r);
 522
 523         BUG_ON(idx < 0);
 524         BUG_ON(!sectors);
 525
 526         switch (r->data_type) {
 527         case BCH_DATA_BTREE:
 528                 fs_usage->btree         += sectors;
 529                 break;
 530         case BCH_DATA_USER:
 531                 fs_usage->data          += sectors;
 532                 break;
 533         case BCH_DATA_CACHED:
 534                 fs_usage->cached        += sectors;
 535                 break;
 536         }
 537         fs_usage->replicas[idx]         += sectors;
 538 }
 539
 540 static inline void update_cached_sectors(struct bch_fs *c,
 541                                          struct bch_fs_usage *fs_usage,
 542                                          unsigned dev, s64 sectors)
 543 {
 544         struct bch_replicas_padded r;
 545
 546         bch2_replicas_entry_cached(&r.e, dev);
 547
 548         update_replicas(c, fs_usage, &r.e, sectors);
 549 }
 550
 551 static struct replicas_delta_list *
 552 replicas_deltas_realloc(struct btree_trans *trans, unsigned more)
 553 {
 554         struct replicas_delta_list *d = trans->fs_usage_deltas;
 555         unsigned new_size = d ? (d->size + more) * 2 : 128;
 556
 557         if (!d || d->used + more > d->size) {
 558                 d = krealloc(d, sizeof(*d) + new_size, GFP_NOIO|__GFP_ZERO);
 559                 BUG_ON(!d);
 560
 561                 d->size = new_size;
 562                 trans->fs_usage_deltas = d;
 563         }
 564         return d;
 565 }
 566
 567 static inline void update_replicas_list(struct btree_trans *trans,
 568                                         struct bch_replicas_entry *r,
 569                                         s64 sectors)
 570 {
 571         struct replicas_delta_list *d;
 572         struct replicas_delta *n;
 573         unsigned b = replicas_entry_bytes(r) + 8;
 574
 575         d = replicas_deltas_realloc(trans, b);
 576
 577         n = (void *) d->d + d->used;
 578         n->delta = sectors;
 579         memcpy(&n->r, r, replicas_entry_bytes(r));
 580         d->used += b;
 581 }
 582
 583 static inline void update_cached_sectors_list(struct btree_trans *trans,
 584                                               unsigned dev, s64 sectors)
 585 {
 586         struct bch_replicas_padded r;
 587
 588         bch2_replicas_entry_cached(&r.e, dev);
 589
 590         update_replicas_list(trans, &r.e, sectors);
 591 }
 592
 593 void bch2_replicas_delta_list_apply(struct bch_fs *c,
 594                                     struct bch_fs_usage *fs_usage,
 595                                     struct replicas_delta_list *r)
 596 {
 597         struct replicas_delta *d = r->d;
 598         struct replicas_delta *top = (void *) r->d + r->used;
 599
 600         acc_u64s((u64 *) fs_usage,
 601                  (u64 *) &r->fs_usage, sizeof(*fs_usage) / sizeof(u64));
 602
 603         while (d != top) {
 604                 BUG_ON((void *) d > (void *) top);
 605
 606                 update_replicas(c, fs_usage, &d->r, d->delta);
 607
 608                 d = (void *) d + replicas_entry_bytes(&d->r) + 8;
 609         }
 610 }
 611
 612 #define do_mark_fn(fn, c, pos, flags, ...)                              \
 613 ({                                                                      \
 614         int gc, ret = 0;                                                \
 615                                                                         \
 616         percpu_rwsem_assert_held(&c->mark_lock);                        \
 617                                                                         \
 618         for (gc = 0; gc < 2 && !ret; gc++)                              \
 619                 if (!gc == !(flags & BCH_BUCKET_MARK_GC) ||             \
 620                     (gc && gc_visited(c, pos)))                         \
 621                         ret = fn(c, __VA_ARGS__, gc);                   \
 622         ret;                                                            \
 623 })
 624
 625 static int __bch2_invalidate_bucket(struct bch_fs *c, struct bch_dev *ca,
 626                                     size_t b, struct bucket_mark *ret,
 627                                     bool gc)
 628 {
 629         struct bch_fs_usage *fs_usage = fs_usage_ptr(c, 0, gc);
 630         struct bucket *g = __bucket(ca, b, gc);
 631         struct bucket_mark old, new;
 632
 633         old = bucket_data_cmpxchg(c, ca, fs_usage, g, new, ({
 634                 BUG_ON(!is_available_bucket(new));
 635
 636                 new.owned_by_allocator  = true;
 637                 new.dirty               = true;
 638                 new.data_type           = 0;
 639                 new.cached_sectors      = 0;
 640                 new.dirty_sectors       = 0;
 641                 new.gen++;
 642         }));
 643
 644         if (old.cached_sectors)
 645                 update_cached_sectors(c, fs_usage, ca->dev_idx,
 646                                       -((s64) old.cached_sectors));
 647
 648         if (!gc)
 649                 *ret = old;
 650         return 0;
 651 }
 652
 653 void bch2_invalidate_bucket(struct bch_fs *c, struct bch_dev *ca,
 654                             size_t b, struct bucket_mark *old)
 655 {
 656         do_mark_fn(__bch2_invalidate_bucket, c, gc_phase(GC_PHASE_START), 0,
 657                    ca, b, old);
 658
 659         if (!old->owned_by_allocator && old->cached_sectors)
 660                 trace_invalidate(ca, bucket_to_sector(ca, b),
 661                                  old->cached_sectors);
 662 }
 663
 664 static int __bch2_mark_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 665                                     size_t b, bool owned_by_allocator,
 666                                     bool gc)
 667 {
 668         struct bch_fs_usage *fs_usage = fs_usage_ptr(c, 0, gc);
 669         struct bucket *g = __bucket(ca, b, gc);
 670         struct bucket_mark old, new;
 671
 672         old = bucket_data_cmpxchg(c, ca, fs_usage, g, new, ({
 673                 new.owned_by_allocator  = owned_by_allocator;
 674         }));
 675
 676         BUG_ON(!gc &&
 677                !owned_by_allocator && !old.owned_by_allocator);
 678
 679         return 0;
 680 }
 681
 682 void bch2_mark_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 683                             size_t b, bool owned_by_allocator,
 684                             struct gc_pos pos, unsigned flags)
 685 {
 686         preempt_disable();
 687
 688         do_mark_fn(__bch2_mark_alloc_bucket, c, pos, flags,
 689                    ca, b, owned_by_allocator);
 690
 691         preempt_enable();
 692 }
 693
 694 static int bch2_mark_alloc(struct bch_fs *c, struct bkey_s_c k,
 695                            struct bch_fs_usage *fs_usage,
 696                            u64 journal_seq, unsigned flags)
 697 {
 698         bool gc = flags & BCH_BUCKET_MARK_GC;
 699         struct bkey_alloc_unpacked u;
 700         struct bch_dev *ca;
 701         struct bucket *g;
 702         struct bucket_mark old, m;
 703
 704         /*
 705          * alloc btree is read in by bch2_alloc_read, not gc:
 706          */
 707         if ((flags & BCH_BUCKET_MARK_GC) &&
 708             !(flags & BCH_BUCKET_MARK_BUCKET_INVALIDATE))
 709                 return 0;
 710
 711         ca = bch_dev_bkey_exists(c, k.k->p.inode);
 712
 713         if (k.k->p.offset >= ca->mi.nbuckets)
 714                 return 0;
 715
 716         g = __bucket(ca, k.k->p.offset, gc);
 717         u = bch2_alloc_unpack(k);
 718
 719         old = bucket_cmpxchg(g, m, ({
 720                 m.gen                   = u.gen;
 721                 m.data_type             = u.data_type;
 722                 m.dirty_sectors         = u.dirty_sectors;
 723                 m.cached_sectors        = u.cached_sectors;
 724
 725                 if (journal_seq) {
 726                         m.journal_seq_valid     = 1;
 727                         m.journal_seq           = journal_seq;
 728                 }
 729         }));
 730
 731         if (!(flags & BCH_BUCKET_MARK_ALLOC_READ))
 732                 bch2_dev_usage_update(c, ca, fs_usage, old, m, gc);
 733
 734         g->io_time[READ]        = u.read_time;
 735         g->io_time[WRITE]       = u.write_time;
 736         g->oldest_gen           = u.oldest_gen;
 737         g->gen_valid            = 1;
 738
 739         /*
 740          * need to know if we're getting called from the invalidate path or
 741          * not:
 742          */
 743
 744         if ((flags & BCH_BUCKET_MARK_BUCKET_INVALIDATE) &&
 745             old.cached_sectors) {
 746                 update_cached_sectors(c, fs_usage, ca->dev_idx,
 747                                       -old.cached_sectors);
 748                 trace_invalidate(ca, bucket_to_sector(ca, k.k->p.offset),
 749                                  old.cached_sectors);
 750         }
 751
 752         return 0;
 753 }
 754
 755 #define checked_add(a, b)                                       \
 756 ({                                                              \
 757         unsigned _res = (unsigned) (a) + (b);                   \
 758         bool overflow = _res > U16_MAX;                         \
 759         if (overflow)                                           \
 760                 _res = U16_MAX;                                 \
 761         (a) = _res;                                             \
 762         overflow;                                               \
 763 })
 764
 765 static int __bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 766                                        size_t b, enum bch_data_type type,
 767                                        unsigned sectors, bool gc)
 768 {
 769         struct bucket *g = __bucket(ca, b, gc);
 770         struct bucket_mark old, new;
 771         bool overflow;
 772
 773         BUG_ON(type != BCH_DATA_SB &&
 774                type != BCH_DATA_JOURNAL);
 775
 776         old = bucket_cmpxchg(g, new, ({
 777                 new.dirty       = true;
 778                 new.data_type   = type;
 779                 overflow = checked_add(new.dirty_sectors, sectors);
 780         }));
 781
 782         bch2_fs_inconsistent_on(overflow, c,
 783                 "bucket sector count overflow: %u + %u > U16_MAX",
 784                 old.dirty_sectors, sectors);
 785
 786         if (c)
 787                 bch2_dev_usage_update(c, ca, fs_usage_ptr(c, 0, gc),
 788                                       old, new, gc);
 789
 790         return 0;
 791 }
 792
 793 void bch2_mark_metadata_bucket(struct bch_fs *c, struct bch_dev *ca,
 794                                size_t b, enum bch_data_type type,
 795                                unsigned sectors, struct gc_pos pos,
 796                                unsigned flags)
 797 {
 798         BUG_ON(type != BCH_DATA_SB &&
 799                type != BCH_DATA_JOURNAL);
 800
 801         preempt_disable();
 802
 803         if (likely(c)) {
 804                 do_mark_fn(__bch2_mark_metadata_bucket, c, pos, flags,
 805                            ca, b, type, sectors);
 806         } else {
 807                 __bch2_mark_metadata_bucket(c, ca, b, type, sectors, 0);
 808         }
 809
 810         preempt_enable();
 811 }
 812
 813 static s64 ptr_disk_sectors_delta(struct extent_ptr_decoded p,
 814                                   unsigned offset, s64 delta,
 815                                   unsigned flags)
 816 {
 817         if (flags & BCH_BUCKET_MARK_OVERWRITE_SPLIT) {
 818                 BUG_ON(offset + -delta > p.crc.live_size);
 819
 820                 return -((s64) ptr_disk_sectors(p)) +
 821                         __ptr_disk_sectors(p, offset) +
 822                         __ptr_disk_sectors(p, p.crc.live_size -
 823                                            offset + delta);
 824         } else if (flags & BCH_BUCKET_MARK_OVERWRITE) {
 825                 BUG_ON(offset + -delta > p.crc.live_size);
 826
 827                 return -((s64) ptr_disk_sectors(p)) +
 828                         __ptr_disk_sectors(p, p.crc.live_size +
 829                                            delta);
 830         } else {
 831                 return ptr_disk_sectors(p);
 832         }
 833 }
 834
 835 static void bucket_set_stripe(struct bch_fs *c,
 836                               const struct bch_stripe *v,
 837                               struct bch_fs_usage *fs_usage,
 838                               u64 journal_seq,
 839                               unsigned flags)
 840 {
 841         bool enabled = !(flags & BCH_BUCKET_MARK_OVERWRITE);
 842         bool gc = flags & BCH_BUCKET_MARK_GC;
 843         unsigned i;
 844
 845         for (i = 0; i < v->nr_blocks; i++) {
 846                 const struct bch_extent_ptr *ptr = v->ptrs + i;
 847                 struct bch_dev *ca = bch_dev_bkey_exists(c, ptr->dev);
 848                 struct bucket *g = PTR_BUCKET(ca, ptr, gc);
 849                 struct bucket_mark new, old;
 850
 851                 old = bucket_data_cmpxchg(c, ca, fs_usage, g, new, ({
 852                         new.dirty                       = true;
 853                         new.stripe                      = enabled;
 854                         if (journal_seq) {
 855                                 new.journal_seq_valid   = 1;
 856                                 new.journal_seq         = journal_seq;
 857                         }
 858                 }));
 859
 860                 /*
 861                  * XXX write repair code for these, flag stripe as possibly bad
 862                  */
 863                 if (old.gen != ptr->gen)
 864                         bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 865                                       "stripe with stale pointer");
 866 #if 0
 867                 /*
 868                  * We'd like to check for these, but these checks don't work
 869                  * yet:
 870                  */
 871                 if (old.stripe && enabled)
 872                         bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 873                                       "multiple stripes using same bucket");
 874
 875                 if (!old.stripe && !enabled)
 876                         bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 877                                       "deleting stripe but bucket not marked as stripe bucket");
 878 #endif
 879         }
 880 }
 881
 882 static bool bch2_mark_pointer(struct bch_fs *c,
 883                               struct extent_ptr_decoded p,
 884                               s64 sectors, enum bch_data_type data_type,
 885                               struct bch_fs_usage *fs_usage,
 886                               u64 journal_seq, unsigned flags)
 887 {
 888         bool gc = flags & BCH_BUCKET_MARK_GC;
 889         struct bucket_mark old, new;
 890         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
 891         struct bucket *g = PTR_BUCKET(ca, &p.ptr, gc);
 892         bool overflow;
 893         u64 v;
 894
 895         v = atomic64_read(&g->_mark.v);
 896         do {
 897                 new.v.counter = old.v.counter = v;
 898
 899                 new.dirty = true;
 900
 901                 /*
 902                  * Check this after reading bucket mark to guard against
 903                  * the allocator invalidating a bucket after we've already
 904                  * checked the gen
 905                  */
 906                 if (gen_after(new.gen, p.ptr.gen)) {
 907                         /* XXX write repair code for this */
 908                         if (!p.ptr.cached &&
 909                             test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))
 910                                 bch2_fsck_err(c, FSCK_CAN_IGNORE|FSCK_NEED_FSCK,
 911                                               "stale dirty pointer");
 912                         return true;
 913                 }
 914
 915                 if (!p.ptr.cached)
 916                         overflow = checked_add(new.dirty_sectors, sectors);
 917                 else
 918                         overflow = checked_add(new.cached_sectors, sectors);
 919
 920                 if (!new.dirty_sectors &&
 921                     !new.cached_sectors) {
 922                         new.data_type   = 0;
 923
 924                         if (journal_seq) {
 925                                 new.journal_seq_valid = 1;
 926                                 new.journal_seq = journal_seq;
 927                         }
 928                 } else {
 929                         new.data_type = data_type;
 930                 }
 931
 932                 if (flags & BCH_BUCKET_MARK_NOATOMIC) {
 933                         g->_mark = new;
 934                         break;
 935                 }
 936         } while ((v = atomic64_cmpxchg(&g->_mark.v,
 937                               old.v.counter,
 938                               new.v.counter)) != old.v.counter);
 939
 940         bch2_fs_inconsistent_on(overflow, c,
 941                 "bucket sector count overflow: %u + %lli > U16_MAX",
 942                 !p.ptr.cached
 943                 ? old.dirty_sectors
 944                 : old.cached_sectors, sectors);
 945
 946         bch2_dev_usage_update(c, ca, fs_usage, old, new, gc);
 947
 948         BUG_ON(!gc && bucket_became_unavailable(old, new));
 949
 950         return false;
 951 }
 952
 953 static int bch2_mark_stripe_ptr(struct bch_fs *c,
 954                                 struct bch_extent_stripe_ptr p,
 955                                 enum bch_data_type data_type,
 956                                 struct bch_fs_usage *fs_usage,
 957                                 s64 sectors, unsigned flags)
 958 {
 959         bool gc = flags & BCH_BUCKET_MARK_GC;
 960         struct stripe *m;
 961         unsigned old, new, nr_data;
 962         int blocks_nonempty_delta;
 963         s64 parity_sectors;
 964
 965         BUG_ON(!sectors);
 966
 967         m = genradix_ptr(&c->stripes[gc], p.idx);
 968
 969         spin_lock(&c->ec_stripes_heap_lock);
 970
 971         if (!m || !m->alive) {
 972                 spin_unlock(&c->ec_stripes_heap_lock);
 973                 bch_err_ratelimited(c, "pointer to nonexistent stripe %llu",
 974                                     (u64) p.idx);
 975                 return -EIO;
 976         }
 977
 978         BUG_ON(m->r.e.data_type != data_type);
 979
 980         nr_data = m->nr_blocks - m->nr_redundant;
 981
 982         parity_sectors = DIV_ROUND_UP(abs(sectors) * m->nr_redundant, nr_data);
 983
 984         if (sectors < 0)
 985                 parity_sectors = -parity_sectors;
 986         sectors += parity_sectors;
 987
 988         old = m->block_sectors[p.block];
 989         m->block_sectors[p.block] += sectors;
 990         new = m->block_sectors[p.block];
 991
 992         blocks_nonempty_delta = (int) !!new - (int) !!old;
 993         if (blocks_nonempty_delta) {
 994                 m->blocks_nonempty += blocks_nonempty_delta;
 995
 996                 if (!gc)
 997                         bch2_stripes_heap_update(c, m, p.idx);
 998         }
 999
1000         m->dirty = true;
1001
1002         spin_unlock(&c->ec_stripes_heap_lock);
1003
1004         update_replicas(c, fs_usage, &m->r.e, sectors);
1005
1006         return 0;
1007 }
1008
1009 static int bch2_mark_extent(struct bch_fs *c, struct bkey_s_c k,
1010                             unsigned offset, s64 sectors,
1011                             enum bch_data_type data_type,
1012                             struct bch_fs_usage *fs_usage,
1013                             unsigned journal_seq, unsigned flags)
1014 {
1015         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1016         const union bch_extent_entry *entry;
1017         struct extent_ptr_decoded p;
1018         struct bch_replicas_padded r;
1019         s64 dirty_sectors = 0;
1020         unsigned i;
1021         int ret;
1022
1023         r.e.data_type   = data_type;
1024         r.e.nr_devs     = 0;
1025         r.e.nr_required = 1;
1026
1027         BUG_ON(!sectors);
1028
1029         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1030                 s64 disk_sectors = data_type == BCH_DATA_BTREE
1031                         ? sectors
1032                         : ptr_disk_sectors_delta(p, offset, sectors, flags);
1033                 bool stale = bch2_mark_pointer(c, p, disk_sectors, data_type,
1034                                         fs_usage, journal_seq, flags);
1035
1036                 if (p.ptr.cached) {
1037                         if (disk_sectors && !stale)
1038                                 update_cached_sectors(c, fs_usage, p.ptr.dev,
1039                                                       disk_sectors);
1040                 } else if (!p.ec_nr) {
1041                         dirty_sectors          += disk_sectors;
1042                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1043                 } else {
1044                         for (i = 0; i < p.ec_nr; i++) {
1045                                 ret = bch2_mark_stripe_ptr(c, p.ec[i],
1046                                                 data_type, fs_usage,
1047                                                 disk_sectors, flags);
1048                                 if (ret)
1049                                         return ret;
1050                         }
1051
1052                         r.e.nr_required = 0;
1053                 }
1054         }
1055
1056         if (dirty_sectors)
1057                 update_replicas(c, fs_usage, &r.e, dirty_sectors);
1058
1059         return 0;
1060 }
1061
1062 static int bch2_mark_stripe(struct bch_fs *c, struct bkey_s_c k,
1063                             struct bch_fs_usage *fs_usage,
1064                             u64 journal_seq, unsigned flags)
1065 {
1066         bool gc = flags & BCH_BUCKET_MARK_GC;
1067         struct bkey_s_c_stripe s = bkey_s_c_to_stripe(k);
1068         size_t idx = s.k->p.offset;
1069         struct stripe *m = genradix_ptr(&c->stripes[gc], idx);
1070         unsigned i;
1071
1072         spin_lock(&c->ec_stripes_heap_lock);
1073
1074         if (!m || ((flags & BCH_BUCKET_MARK_OVERWRITE) && !m->alive)) {
1075                 spin_unlock(&c->ec_stripes_heap_lock);
1076                 bch_err_ratelimited(c, "error marking nonexistent stripe %zu",
1077                                     idx);
1078                 return -1;
1079         }
1080
1081         if (!(flags & BCH_BUCKET_MARK_OVERWRITE)) {
1082                 m->sectors      = le16_to_cpu(s.v->sectors);
1083                 m->algorithm    = s.v->algorithm;
1084                 m->nr_blocks    = s.v->nr_blocks;
1085                 m->nr_redundant = s.v->nr_redundant;
1086
1087                 bch2_bkey_to_replicas(&m->r.e, k);
1088
1089                 /*
1090                  * XXX: account for stripes somehow here
1091                  */
1092 #if 0
1093                 update_replicas(c, fs_usage, &m->r.e, stripe_sectors);
1094 #endif
1095
1096                 /* gc recalculates these fields: */
1097                 if (!(flags & BCH_BUCKET_MARK_GC)) {
1098                         for (i = 0; i < s.v->nr_blocks; i++) {
1099                                 m->block_sectors[i] =
1100                                         stripe_blockcount_get(s.v, i);
1101                                 m->blocks_nonempty += !!m->block_sectors[i];
1102                         }
1103                 }
1104
1105                 if (!gc)
1106                         bch2_stripes_heap_update(c, m, idx);
1107                 m->alive        = true;
1108         } else {
1109                 if (!gc)
1110                         bch2_stripes_heap_del(c, m, idx);
1111                 memset(m, 0, sizeof(*m));
1112         }
1113
1114         spin_unlock(&c->ec_stripes_heap_lock);
1115
1116         bucket_set_stripe(c, s.v, fs_usage, 0, flags);
1117         return 0;
1118 }
1119
1120 int bch2_mark_key_locked(struct bch_fs *c,
1121                    struct bkey_s_c k,
1122                    unsigned offset, s64 sectors,
1123                    struct bch_fs_usage *fs_usage,
1124                    u64 journal_seq, unsigned flags)
1125 {
1126         int ret = 0;
1127
1128         preempt_disable();
1129
1130         if (!fs_usage || (flags & BCH_BUCKET_MARK_GC))
1131                 fs_usage = fs_usage_ptr(c, journal_seq,
1132                                         flags & BCH_BUCKET_MARK_GC);
1133
1134         switch (k.k->type) {
1135         case KEY_TYPE_alloc:
1136                 ret = bch2_mark_alloc(c, k, fs_usage, journal_seq, flags);
1137                 break;
1138         case KEY_TYPE_btree_ptr:
1139                 sectors = !(flags & BCH_BUCKET_MARK_OVERWRITE)
1140                         ?  c->opts.btree_node_size
1141                         : -c->opts.btree_node_size;
1142
1143                 ret = bch2_mark_extent(c, k, offset, sectors, BCH_DATA_BTREE,
1144                                 fs_usage, journal_seq, flags);
1145                 break;
1146         case KEY_TYPE_extent:
1147         case KEY_TYPE_reflink_v:
1148                 ret = bch2_mark_extent(c, k, offset, sectors, BCH_DATA_USER,
1149                                 fs_usage, journal_seq, flags);
1150                 break;
1151         case KEY_TYPE_stripe:
1152                 ret = bch2_mark_stripe(c, k, fs_usage, journal_seq, flags);
1153                 break;
1154         case KEY_TYPE_inode:
1155                 if (!(flags & BCH_BUCKET_MARK_OVERWRITE))
1156                         fs_usage->nr_inodes++;
1157                 else
1158                         fs_usage->nr_inodes--;
1159                 break;
1160         case KEY_TYPE_reservation: {
1161                 unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1162
1163                 sectors *= replicas;
1164                 replicas = clamp_t(unsigned, replicas, 1,
1165                                    ARRAY_SIZE(fs_usage->persistent_reserved));
1166
1167                 fs_usage->reserved                              += sectors;
1168                 fs_usage->persistent_reserved[replicas - 1]     += sectors;
1169                 break;
1170         }
1171         }
1172
1173         preempt_enable();
1174
1175         return ret;
1176 }
1177
1178 int bch2_mark_key(struct bch_fs *c, struct bkey_s_c k,
1179                   unsigned offset, s64 sectors,
1180                   struct bch_fs_usage *fs_usage,
1181                   u64 journal_seq, unsigned flags)
1182 {
1183         int ret;
1184
1185         percpu_down_read(&c->mark_lock);
1186         ret = bch2_mark_key_locked(c, k, offset, sectors,
1187                                    fs_usage, journal_seq, flags);
1188         percpu_up_read(&c->mark_lock);
1189
1190         return ret;
1191 }
1192
1193 inline int bch2_mark_overwrite(struct btree_trans *trans,
1194                                struct btree_iter *iter,
1195                                struct bkey_s_c old,
1196                                struct bkey_i *new,
1197                                struct bch_fs_usage *fs_usage,
1198                                unsigned flags)
1199 {
1200         struct bch_fs           *c = trans->c;
1201         struct btree            *b = iter->l[0].b;
1202         unsigned                offset = 0;
1203         s64                     sectors = 0;
1204
1205         flags |= BCH_BUCKET_MARK_OVERWRITE;
1206
1207         if (btree_node_is_extents(b)
1208             ? bkey_cmp(new->k.p, bkey_start_pos(old.k)) <= 0
1209             : bkey_cmp(new->k.p, old.k->p))
1210                 return 0;
1211
1212         if (btree_node_is_extents(b)) {
1213                 switch (bch2_extent_overlap(&new->k, old.k)) {
1214                 case BCH_EXTENT_OVERLAP_ALL:
1215                         offset = 0;
1216                         sectors = -((s64) old.k->size);
1217                         break;
1218                 case BCH_EXTENT_OVERLAP_BACK:
1219                         offset = bkey_start_offset(&new->k) -
1220                                 bkey_start_offset(old.k);
1221                         sectors = bkey_start_offset(&new->k) -
1222                                 old.k->p.offset;
1223                         break;
1224                 case BCH_EXTENT_OVERLAP_FRONT:
1225                         offset = 0;
1226                         sectors = bkey_start_offset(old.k) -
1227                                 new->k.p.offset;
1228                         break;
1229                 case BCH_EXTENT_OVERLAP_MIDDLE:
1230                         offset = bkey_start_offset(&new->k) -
1231                                 bkey_start_offset(old.k);
1232                         sectors = -((s64) new->k.size);
1233                         flags |= BCH_BUCKET_MARK_OVERWRITE_SPLIT;
1234                         break;
1235                 }
1236
1237                 BUG_ON(sectors >= 0);
1238         }
1239
1240         return bch2_mark_key_locked(c, old, offset, sectors, fs_usage,
1241                                     trans->journal_res.seq, flags) ?: 1;
1242 }
1243
1244 int bch2_mark_update(struct btree_trans *trans,
1245                      struct btree_insert_entry *insert,
1246                      struct bch_fs_usage *fs_usage,
1247                      unsigned flags)
1248 {
1249         struct bch_fs           *c = trans->c;
1250         struct btree_iter       *iter = insert->iter;
1251         struct btree            *b = iter->l[0].b;
1252         struct btree_node_iter  node_iter = iter->l[0].iter;
1253         struct bkey_packed      *_k;
1254         int ret = 0;
1255
1256         if (!btree_node_type_needs_gc(iter->btree_id))
1257                 return 0;
1258
1259         if (!(trans->flags & BTREE_INSERT_NOMARK_INSERT))
1260                 bch2_mark_key_locked(c, bkey_i_to_s_c(insert->k),
1261                         0, insert->k->k.size,
1262                         fs_usage, trans->journal_res.seq,
1263                         BCH_BUCKET_MARK_INSERT|flags);
1264
1265         if (unlikely(trans->flags & BTREE_INSERT_NOMARK_OVERWRITES))
1266                 return 0;
1267
1268         /*
1269          * For non extents, we only mark the new key, not the key being
1270          * overwritten - unless we're actually deleting:
1271          */
1272         if ((iter->btree_id == BTREE_ID_ALLOC ||
1273              iter->btree_id == BTREE_ID_EC) &&
1274             !bkey_deleted(&insert->k->k))
1275                 return 0;
1276
1277         while ((_k = bch2_btree_node_iter_peek_filter(&node_iter, b,
1278                                                       KEY_TYPE_discard))) {
1279                 struct bkey             unpacked;
1280                 struct bkey_s_c         k = bkey_disassemble(b, _k, &unpacked);
1281
1282                 ret = bch2_mark_overwrite(trans, iter, k, insert->k,
1283                                           fs_usage, flags);
1284                 if (ret <= 0)
1285                         break;
1286
1287                 bch2_btree_node_iter_advance(&node_iter, b);
1288         }
1289
1290         return ret;
1291 }
1292
1293 void bch2_trans_fs_usage_apply(struct btree_trans *trans,
1294                                struct bch_fs_usage *fs_usage)
1295 {
1296         struct bch_fs *c = trans->c;
1297         struct btree_insert_entry *i;
1298         static int warned_disk_usage = 0;
1299         u64 disk_res_sectors = trans->disk_res ? trans->disk_res->sectors : 0;
1300         char buf[200];
1301
1302         if (!bch2_fs_usage_apply(c, fs_usage, trans->disk_res,
1303                                  trans->journal_res.seq) ||
1304             warned_disk_usage ||
1305             xchg(&warned_disk_usage, 1))
1306                 return;
1307
1308         bch_err(c, "disk usage increased more than %llu sectors reserved",
1309                 disk_res_sectors);
1310
1311         trans_for_each_update_iter(trans, i) {
1312                 struct btree_iter       *iter = i->iter;
1313                 struct btree            *b = iter->l[0].b;
1314                 struct btree_node_iter  node_iter = iter->l[0].iter;
1315                 struct bkey_packed      *_k;
1316
1317                 pr_err("while inserting");
1318                 bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
1319                 pr_err("%s", buf);
1320                 pr_err("overlapping with");
1321
1322                 node_iter = iter->l[0].iter;
1323                 while ((_k = bch2_btree_node_iter_peek_filter(&node_iter, b,
1324                                                         KEY_TYPE_discard))) {
1325                         struct bkey             unpacked;
1326                         struct bkey_s_c         k;
1327
1328                         k = bkey_disassemble(b, _k, &unpacked);
1329
1330                         if (btree_node_is_extents(b)
1331                             ? bkey_cmp(i->k->k.p, bkey_start_pos(k.k)) <= 0
1332                             : bkey_cmp(i->k->k.p, k.k->p))
1333                                 break;
1334
1335                         bch2_bkey_val_to_text(&PBUF(buf), c, k);
1336                         pr_err("%s", buf);
1337
1338                         bch2_btree_node_iter_advance(&node_iter, b);
1339                 }
1340         }
1341 }
1342
1343 /* trans_mark: */
1344
1345 static int trans_get_key(struct btree_trans *trans,
1346                          enum btree_id btree_id, struct bpos pos,
1347                          struct btree_iter **iter,
1348                          struct bkey_s_c *k)
1349 {
1350         struct btree_insert_entry *i;
1351         int ret;
1352
1353         for (i = trans->updates;
1354              i < trans->updates + trans->nr_updates;
1355              i++)
1356                 if (!i->deferred &&
1357                     i->iter->btree_id == btree_id &&
1358                     (btree_node_type_is_extents(btree_id)
1359                      ? bkey_cmp(pos, bkey_start_pos(&i->k->k)) >= 0 &&
1360                        bkey_cmp(pos, i->k->k.p) < 0
1361                      : !bkey_cmp(pos, i->iter->pos))) {
1362                         *iter   = i->iter;
1363                         *k      = bkey_i_to_s_c(i->k);
1364                         return 0;
1365                 }
1366
1367         *iter = __bch2_trans_get_iter(trans, btree_id, pos,
1368                                    BTREE_ITER_SLOTS|BTREE_ITER_INTENT, 0);
1369         if (IS_ERR(*iter))
1370                 return PTR_ERR(*iter);
1371
1372         bch2_trans_iter_free_on_commit(trans, *iter);
1373
1374         *k = bch2_btree_iter_peek_slot(*iter);
1375         ret = bkey_err(*k);
1376         if (ret)
1377                 bch2_trans_iter_put(trans, *iter);
1378         return ret;
1379 }
1380
1381 static void *trans_update_key(struct btree_trans *trans,
1382                               struct btree_iter *iter,
1383                               unsigned u64s)
1384 {
1385         struct bkey_i *new_k;
1386         unsigned i;
1387
1388         new_k = bch2_trans_kmalloc(trans, u64s * sizeof(u64));
1389         if (IS_ERR(new_k))
1390                 return new_k;
1391
1392         bkey_init(&new_k->k);
1393         new_k->k.p = iter->pos;
1394
1395         for (i = 0; i < trans->nr_updates; i++)
1396                 if (!trans->updates[i].deferred &&
1397                     trans->updates[i].iter == iter) {
1398                         trans->updates[i].k = new_k;
1399                         return new_k;
1400                 }
1401
1402         bch2_trans_update(trans, ((struct btree_insert_entry) {
1403                 .iter = iter,
1404                 .k = new_k,
1405                 .triggered = true,
1406         }));
1407
1408         return new_k;
1409 }
1410
1411 static int bch2_trans_mark_pointer(struct btree_trans *trans,
1412                         struct extent_ptr_decoded p,
1413                         s64 sectors, enum bch_data_type data_type)
1414 {
1415         struct bch_fs *c = trans->c;
1416         struct bch_dev *ca = bch_dev_bkey_exists(c, p.ptr.dev);
1417         struct btree_iter *iter;
1418         struct bkey_s_c k;
1419         struct bkey_alloc_unpacked u;
1420         struct bkey_i_alloc *a;
1421         bool overflow;
1422         int ret;
1423
1424         ret = trans_get_key(trans, BTREE_ID_ALLOC,
1425                             POS(p.ptr.dev, PTR_BUCKET_NR(ca, &p.ptr)),
1426                             &iter, &k);
1427         if (ret)
1428                 return ret;
1429
1430         if (k.k->type != KEY_TYPE_alloc) {
1431                 bch_err_ratelimited(c, "pointer to nonexistent bucket %u:%zu",
1432                                     p.ptr.dev,
1433                                     PTR_BUCKET_NR(ca, &p.ptr));
1434                 ret = -1;
1435                 goto out;
1436         }
1437
1438         u = bch2_alloc_unpack(k);
1439
1440         if (gen_after(u.gen, p.ptr.gen)) {
1441                 ret = 1;
1442                 goto out;
1443         }
1444
1445         if (!p.ptr.cached)
1446                 overflow = checked_add(u.dirty_sectors, sectors);
1447         else
1448                 overflow = checked_add(u.cached_sectors, sectors);
1449
1450         u.data_type = u.dirty_sectors || u.cached_sectors
1451                 ? data_type : 0;
1452
1453         bch2_fs_inconsistent_on(overflow, c,
1454                 "bucket sector count overflow: %u + %lli > U16_MAX",
1455                 !p.ptr.cached
1456                 ? u.dirty_sectors
1457                 : u.cached_sectors, sectors);
1458
1459         a = trans_update_key(trans, iter, BKEY_ALLOC_U64s_MAX);
1460         ret = PTR_ERR_OR_ZERO(a);
1461         if (ret)
1462                 goto out;
1463
1464         bkey_alloc_init(&a->k_i);
1465         a->k.p = iter->pos;
1466         bch2_alloc_pack(a, u);
1467 out:
1468         bch2_trans_iter_put(trans, iter);
1469         return ret;
1470 }
1471
1472 static int bch2_trans_mark_stripe_ptr(struct btree_trans *trans,
1473                         struct bch_extent_stripe_ptr p,
1474                         s64 sectors, enum bch_data_type data_type)
1475 {
1476         struct bch_fs *c = trans->c;
1477         struct bch_replicas_padded r;
1478         struct btree_iter *iter;
1479         struct bkey_i *new_k;
1480         struct bkey_s_c k;
1481         struct bkey_s_stripe s;
1482         unsigned nr_data;
1483         s64 parity_sectors;
1484         int ret = 0;
1485
1486         ret = trans_get_key(trans, BTREE_ID_EC, POS(0, p.idx), &iter, &k);
1487         if (ret)
1488                 return ret;
1489
1490         if (k.k->type != KEY_TYPE_stripe) {
1491                 bch2_fs_inconsistent(c,
1492                         "pointer to nonexistent stripe %llu",
1493                         (u64) p.idx);
1494                 ret = -EIO;
1495                 goto out;
1496         }
1497
1498         new_k = trans_update_key(trans, iter, k.k->u64s);
1499         ret = PTR_ERR_OR_ZERO(new_k);
1500         if (ret)
1501                 goto out;
1502
1503         bkey_reassemble(new_k, k);
1504         s = bkey_i_to_s_stripe(new_k);
1505
1506         nr_data = s.v->nr_blocks - s.v->nr_redundant;
1507
1508         parity_sectors = DIV_ROUND_UP(abs(sectors) * s.v->nr_redundant, nr_data);
1509
1510         if (sectors < 0)
1511                 parity_sectors = -parity_sectors;
1512
1513         stripe_blockcount_set(s.v, p.block,
1514                 stripe_blockcount_get(s.v, p.block) +
1515                 sectors + parity_sectors);
1516
1517         bch2_bkey_to_replicas(&r.e, s.s_c);
1518
1519         update_replicas_list(trans, &r.e, sectors);
1520 out:
1521         bch2_trans_iter_put(trans, iter);
1522         return ret;
1523 }
1524
1525 static int bch2_trans_mark_extent(struct btree_trans *trans,
1526                         struct bkey_s_c k, unsigned offset,
1527                         s64 sectors, unsigned flags,
1528                         enum bch_data_type data_type)
1529 {
1530         struct bkey_ptrs_c ptrs = bch2_bkey_ptrs_c(k);
1531         const union bch_extent_entry *entry;
1532         struct extent_ptr_decoded p;
1533         struct bch_replicas_padded r;
1534         s64 dirty_sectors = 0;
1535         bool stale;
1536         unsigned i;
1537         int ret;
1538
1539         r.e.data_type   = data_type;
1540         r.e.nr_devs     = 0;
1541         r.e.nr_required = 1;
1542
1543         BUG_ON(!sectors);
1544
1545         bkey_for_each_ptr_decode(k.k, ptrs, p, entry) {
1546                 s64 disk_sectors = data_type == BCH_DATA_BTREE
1547                         ? sectors
1548                         : ptr_disk_sectors_delta(p, offset, sectors, flags);
1549
1550                 /*
1551                  * can happen due to rounding with compressed extents:
1552                  */
1553                 if (!disk_sectors)
1554                         continue;
1555
1556                 ret = bch2_trans_mark_pointer(trans, p, disk_sectors,
1557                                               data_type);
1558                 if (ret < 0)
1559                         return ret;
1560
1561                 stale = ret > 0;
1562
1563                 if (p.ptr.cached) {
1564                         if (disk_sectors && !stale)
1565                                 update_cached_sectors_list(trans, p.ptr.dev,
1566                                                            disk_sectors);
1567                 } else if (!p.ec_nr) {
1568                         dirty_sectors          += disk_sectors;
1569                         r.e.devs[r.e.nr_devs++] = p.ptr.dev;
1570                 } else {
1571                         for (i = 0; i < p.ec_nr; i++) {
1572                                 ret = bch2_trans_mark_stripe_ptr(trans, p.ec[i],
1573                                                 disk_sectors, data_type);
1574                                 if (ret)
1575                                         return ret;
1576                         }
1577
1578                         r.e.nr_required = 0;
1579                 }
1580         }
1581
1582         if (dirty_sectors)
1583                 update_replicas_list(trans, &r.e, dirty_sectors);
1584
1585         return 0;
1586 }
1587
1588 static int __bch2_trans_mark_reflink_p(struct btree_trans *trans,
1589                         struct bkey_s_c_reflink_p p,
1590                         u64 idx, unsigned sectors,
1591                         unsigned flags)
1592 {
1593         struct bch_fs *c = trans->c;
1594         struct btree_iter *iter;
1595         struct bkey_i *new_k;
1596         struct bkey_s_c k;
1597         struct bkey_i_reflink_v *r_v;
1598         s64 ret;
1599
1600         ret = trans_get_key(trans, BTREE_ID_REFLINK,
1601                             POS(0, idx), &iter, &k);
1602         if (ret)
1603                 return ret;
1604
1605         if (k.k->type != KEY_TYPE_reflink_v) {
1606                 bch2_fs_inconsistent(c,
1607                         "%llu:%llu len %u points to nonexistent indirect extent %llu",
1608                         p.k->p.inode, p.k->p.offset, p.k->size, idx);
1609                 ret = -EIO;
1610                 goto err;
1611         }
1612
1613         if ((flags & BCH_BUCKET_MARK_OVERWRITE) &&
1614             (bkey_start_offset(k.k) < idx ||
1615              k.k->p.offset > idx + sectors))
1616                 goto out;
1617
1618         bch2_btree_iter_set_pos(iter, bkey_start_pos(k.k));
1619         BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
1620
1621         new_k = trans_update_key(trans, iter, k.k->u64s);
1622         ret = PTR_ERR_OR_ZERO(new_k);
1623         if (ret)
1624                 goto err;
1625
1626         bkey_reassemble(new_k, k);
1627         r_v = bkey_i_to_reflink_v(new_k);
1628
1629         le64_add_cpu(&r_v->v.refcount,
1630                      !(flags & BCH_BUCKET_MARK_OVERWRITE) ? 1 : -1);
1631
1632         if (!r_v->v.refcount) {
1633                 r_v->k.type = KEY_TYPE_deleted;
1634                 set_bkey_val_u64s(&r_v->k, 0);
1635         }
1636 out:
1637         ret = k.k->p.offset - idx;
1638 err:
1639         bch2_trans_iter_put(trans, iter);
1640         return ret;
1641 }
1642
1643 static int bch2_trans_mark_reflink_p(struct btree_trans *trans,
1644                         struct bkey_s_c_reflink_p p, unsigned offset,
1645                         s64 sectors, unsigned flags)
1646 {
1647         u64 idx = le64_to_cpu(p.v->idx) + offset;
1648         s64 ret = 0;
1649
1650         sectors = abs(sectors);
1651         BUG_ON(offset + sectors > p.k->size);
1652
1653         while (sectors) {
1654                 ret = __bch2_trans_mark_reflink_p(trans, p, idx, sectors, flags);
1655                 if (ret < 0)
1656                         break;
1657
1658                 idx += ret;
1659                 sectors = max_t(s64, 0LL, sectors - ret);
1660                 ret = 0;
1661         }
1662
1663         return ret;
1664 }
1665
1666 int bch2_trans_mark_key(struct btree_trans *trans, struct bkey_s_c k,
1667                         unsigned offset, s64 sectors, unsigned flags)
1668 {
1669         struct replicas_delta_list *d;
1670         struct bch_fs *c = trans->c;
1671
1672         switch (k.k->type) {
1673         case KEY_TYPE_btree_ptr:
1674                 sectors = !(flags & BCH_BUCKET_MARK_OVERWRITE)
1675                         ?  c->opts.btree_node_size
1676                         : -c->opts.btree_node_size;
1677
1678                 return bch2_trans_mark_extent(trans, k, offset, sectors,
1679                                               flags, BCH_DATA_BTREE);
1680         case KEY_TYPE_extent:
1681         case KEY_TYPE_reflink_v:
1682                 return bch2_trans_mark_extent(trans, k, offset, sectors,
1683                                               flags, BCH_DATA_USER);
1684         case KEY_TYPE_inode:
1685                 d = replicas_deltas_realloc(trans, 0);
1686
1687                 if (!(flags & BCH_BUCKET_MARK_OVERWRITE))
1688                         d->fs_usage.nr_inodes++;
1689                 else
1690                         d->fs_usage.nr_inodes--;
1691                 return 0;
1692         case KEY_TYPE_reservation: {
1693                 unsigned replicas = bkey_s_c_to_reservation(k).v->nr_replicas;
1694
1695                 d = replicas_deltas_realloc(trans, 0);
1696
1697                 sectors *= replicas;
1698                 replicas = clamp_t(unsigned, replicas, 1,
1699                                    ARRAY_SIZE(d->fs_usage.persistent_reserved));
1700
1701                 d->fs_usage.reserved                            += sectors;
1702                 d->fs_usage.persistent_reserved[replicas - 1]   += sectors;
1703                 return 0;
1704         }
1705         case KEY_TYPE_reflink_p:
1706                 return bch2_trans_mark_reflink_p(trans,
1707                                         bkey_s_c_to_reflink_p(k),
1708                                         offset, sectors, flags);
1709         default:
1710                 return 0;
1711         }
1712 }
1713
1714 int bch2_trans_mark_update(struct btree_trans *trans,
1715                            struct btree_iter *iter,
1716                            struct bkey_i *insert)
1717 {
1718         struct btree            *b = iter->l[0].b;
1719         struct btree_node_iter  node_iter = iter->l[0].iter;
1720         struct bkey_packed      *_k;
1721         int ret;
1722
1723         if (!btree_node_type_needs_gc(iter->btree_id))
1724                 return 0;
1725
1726         ret = bch2_trans_mark_key(trans, bkey_i_to_s_c(insert),
1727                         0, insert->k.size, BCH_BUCKET_MARK_INSERT);
1728         if (ret)
1729                 return ret;
1730
1731         while ((_k = bch2_btree_node_iter_peek_filter(&node_iter, b,
1732                                                       KEY_TYPE_discard))) {
1733                 struct bkey             unpacked;
1734                 struct bkey_s_c         k;
1735                 unsigned                offset = 0;
1736                 s64                     sectors = 0;
1737                 unsigned                flags = BCH_BUCKET_MARK_OVERWRITE;
1738
1739                 k = bkey_disassemble(b, _k, &unpacked);
1740
1741                 if (btree_node_is_extents(b)
1742                     ? bkey_cmp(insert->k.p, bkey_start_pos(k.k)) <= 0
1743                     : bkey_cmp(insert->k.p, k.k->p))
1744                         break;
1745
1746                 if (btree_node_is_extents(b)) {
1747                         switch (bch2_extent_overlap(&insert->k, k.k)) {
1748                         case BCH_EXTENT_OVERLAP_ALL:
1749                                 offset = 0;
1750                                 sectors = -((s64) k.k->size);
1751                                 break;
1752                         case BCH_EXTENT_OVERLAP_BACK:
1753                                 offset = bkey_start_offset(&insert->k) -
1754                                         bkey_start_offset(k.k);
1755                                 sectors = bkey_start_offset(&insert->k) -
1756                                         k.k->p.offset;
1757                                 break;
1758                         case BCH_EXTENT_OVERLAP_FRONT:
1759                                 offset = 0;
1760                                 sectors = bkey_start_offset(k.k) -
1761                                         insert->k.p.offset;
1762                                 break;
1763                         case BCH_EXTENT_OVERLAP_MIDDLE:
1764                                 offset = bkey_start_offset(&insert->k) -
1765                                         bkey_start_offset(k.k);
1766                                 sectors = -((s64) insert->k.size);
1767                                 flags |= BCH_BUCKET_MARK_OVERWRITE_SPLIT;
1768                                 break;
1769                         }
1770
1771                         BUG_ON(sectors >= 0);
1772                 }
1773
1774                 ret = bch2_trans_mark_key(trans, k, offset, sectors, flags);
1775                 if (ret)
1776                         return ret;
1777
1778                 bch2_btree_node_iter_advance(&node_iter, b);
1779         }
1780
1781         return 0;
1782 }
1783
1784 /* Disk reservations: */
1785
1786 static u64 bch2_recalc_sectors_available(struct bch_fs *c)
1787 {
1788         percpu_u64_set(&c->pcpu->sectors_available, 0);
1789
1790         return avail_factor(__bch2_fs_usage_read_short(c).free);
1791 }
1792
1793 void __bch2_disk_reservation_put(struct bch_fs *c, struct disk_reservation *res)
1794 {
1795         percpu_down_read(&c->mark_lock);
1796         this_cpu_sub(c->usage[0]->online_reserved,
1797                      res->sectors);
1798         percpu_up_read(&c->mark_lock);
1799
1800         res->sectors = 0;
1801 }
1802
1803 #define SECTORS_CACHE   1024
1804
1805 int bch2_disk_reservation_add(struct bch_fs *c, struct disk_reservation *res,
1806                               unsigned sectors, int flags)
1807 {
1808         struct bch_fs_pcpu *pcpu;
1809         u64 old, v, get;
1810         s64 sectors_available;
1811         int ret;
1812
1813         percpu_down_read(&c->mark_lock);
1814         preempt_disable();
1815         pcpu = this_cpu_ptr(c->pcpu);
1816
1817         if (sectors <= pcpu->sectors_available)
1818                 goto out;
1819
1820         v = atomic64_read(&c->sectors_available);
1821         do {
1822                 old = v;
1823                 get = min((u64) sectors + SECTORS_CACHE, old);
1824
1825                 if (get < sectors) {
1826                         preempt_enable();
1827                         percpu_up_read(&c->mark_lock);
1828                         goto recalculate;
1829                 }
1830         } while ((v = atomic64_cmpxchg(&c->sectors_available,
1831                                        old, old - get)) != old);
1832
1833         pcpu->sectors_available         += get;
1834
1835 out:
1836         pcpu->sectors_available         -= sectors;
1837         this_cpu_add(c->usage[0]->online_reserved, sectors);
1838         res->sectors                    += sectors;
1839
1840         preempt_enable();
1841         percpu_up_read(&c->mark_lock);
1842         return 0;
1843
1844 recalculate:
1845         percpu_down_write(&c->mark_lock);
1846
1847         sectors_available = bch2_recalc_sectors_available(c);
1848
1849         if (sectors <= sectors_available ||
1850             (flags & BCH_DISK_RESERVATION_NOFAIL)) {
1851                 atomic64_set(&c->sectors_available,
1852                              max_t(s64, 0, sectors_available - sectors));
1853                 this_cpu_add(c->usage[0]->online_reserved, sectors);
1854                 res->sectors                    += sectors;
1855                 ret = 0;
1856         } else {
1857                 atomic64_set(&c->sectors_available, sectors_available);
1858                 ret = -ENOSPC;
1859         }
1860
1861         percpu_up_write(&c->mark_lock);
1862
1863         return ret;
1864 }
1865
1866 /* Startup/shutdown: */
1867
1868 static void buckets_free_rcu(struct rcu_head *rcu)
1869 {
1870         struct bucket_array *buckets =
1871                 container_of(rcu, struct bucket_array, rcu);
1872
1873         kvpfree(buckets,
1874                 sizeof(struct bucket_array) +
1875                 buckets->nbuckets * sizeof(struct bucket));
1876 }
1877
1878 int bch2_dev_buckets_resize(struct bch_fs *c, struct bch_dev *ca, u64 nbuckets)
1879 {
1880         struct bucket_array *buckets = NULL, *old_buckets = NULL;
1881         unsigned long *buckets_nouse = NULL;
1882         unsigned long *buckets_written = NULL;
1883         alloc_fifo      free[RESERVE_NR];
1884         alloc_fifo      free_inc;
1885         alloc_heap      alloc_heap;
1886         copygc_heap     copygc_heap;
1887
1888         size_t btree_reserve    = DIV_ROUND_UP(BTREE_NODE_RESERVE,
1889                              ca->mi.bucket_size / c->opts.btree_node_size);
1890         /* XXX: these should be tunable */
1891         size_t reserve_none     = max_t(size_t, 1, nbuckets >> 9);
1892         size_t copygc_reserve   = max_t(size_t, 2, nbuckets >> 7);
1893         size_t free_inc_nr      = max(max_t(size_t, 1, nbuckets >> 12),
1894                                       btree_reserve * 2);
1895         bool resize = ca->buckets[0] != NULL,
1896              start_copygc = ca->copygc_thread != NULL;
1897         int ret = -ENOMEM;
1898         unsigned i;
1899
1900         memset(&free,           0, sizeof(free));
1901         memset(&free_inc,       0, sizeof(free_inc));
1902         memset(&alloc_heap,     0, sizeof(alloc_heap));
1903         memset(&copygc_heap,    0, sizeof(copygc_heap));
1904
1905         if (!(buckets           = kvpmalloc(sizeof(struct bucket_array) +
1906                                             nbuckets * sizeof(struct bucket),
1907                                             GFP_KERNEL|__GFP_ZERO)) ||
1908             !(buckets_nouse     = kvpmalloc(BITS_TO_LONGS(nbuckets) *
1909                                             sizeof(unsigned long),
1910                                             GFP_KERNEL|__GFP_ZERO)) ||
1911             !(buckets_written   = kvpmalloc(BITS_TO_LONGS(nbuckets) *
1912                                             sizeof(unsigned long),
1913                                             GFP_KERNEL|__GFP_ZERO)) ||
1914             !init_fifo(&free[RESERVE_BTREE], btree_reserve, GFP_KERNEL) ||
1915             !init_fifo(&free[RESERVE_MOVINGGC],
1916                        copygc_reserve, GFP_KERNEL) ||
1917             !init_fifo(&free[RESERVE_NONE], reserve_none, GFP_KERNEL) ||
1918             !init_fifo(&free_inc,       free_inc_nr, GFP_KERNEL) ||
1919             !init_heap(&alloc_heap,     ALLOC_SCAN_BATCH(ca) << 1, GFP_KERNEL) ||
1920             !init_heap(&copygc_heap,    copygc_reserve, GFP_KERNEL))
1921                 goto err;
1922
1923         buckets->first_bucket   = ca->mi.first_bucket;
1924         buckets->nbuckets       = nbuckets;
1925
1926         bch2_copygc_stop(ca);
1927
1928         if (resize) {
1929                 down_write(&c->gc_lock);
1930                 down_write(&ca->bucket_lock);
1931                 percpu_down_write(&c->mark_lock);
1932         }
1933
1934         old_buckets = bucket_array(ca);
1935
1936         if (resize) {
1937                 size_t n = min(buckets->nbuckets, old_buckets->nbuckets);
1938
1939                 memcpy(buckets->b,
1940                        old_buckets->b,
1941                        n * sizeof(struct bucket));
1942                 memcpy(buckets_nouse,
1943                        ca->buckets_nouse,
1944                        BITS_TO_LONGS(n) * sizeof(unsigned long));
1945                 memcpy(buckets_written,
1946                        ca->buckets_written,
1947                        BITS_TO_LONGS(n) * sizeof(unsigned long));
1948         }
1949
1950         rcu_assign_pointer(ca->buckets[0], buckets);
1951         buckets = old_buckets;
1952
1953         swap(ca->buckets_nouse, buckets_nouse);
1954         swap(ca->buckets_written, buckets_written);
1955
1956         if (resize)
1957                 percpu_up_write(&c->mark_lock);
1958
1959         spin_lock(&c->freelist_lock);
1960         for (i = 0; i < RESERVE_NR; i++) {
1961                 fifo_move(&free[i], &ca->free[i]);
1962                 swap(ca->free[i], free[i]);
1963         }
1964         fifo_move(&free_inc, &ca->free_inc);
1965         swap(ca->free_inc, free_inc);
1966         spin_unlock(&c->freelist_lock);
1967
1968         /* with gc lock held, alloc_heap can't be in use: */
1969         swap(ca->alloc_heap, alloc_heap);
1970
1971         /* and we shut down copygc: */
1972         swap(ca->copygc_heap, copygc_heap);
1973
1974         nbuckets = ca->mi.nbuckets;
1975
1976         if (resize) {
1977                 up_write(&ca->bucket_lock);
1978                 up_write(&c->gc_lock);
1979         }
1980
1981         if (start_copygc &&
1982             bch2_copygc_start(c, ca))
1983                 bch_err(ca, "error restarting copygc thread");
1984
1985         ret = 0;
1986 err:
1987         free_heap(&copygc_heap);
1988         free_heap(&alloc_heap);
1989         free_fifo(&free_inc);
1990         for (i = 0; i < RESERVE_NR; i++)
1991                 free_fifo(&free[i]);
1992         kvpfree(buckets_nouse,
1993                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
1994         kvpfree(buckets_written,
1995                 BITS_TO_LONGS(nbuckets) * sizeof(unsigned long));
1996         if (buckets)
1997                 call_rcu(&old_buckets->rcu, buckets_free_rcu);
1998
1999         return ret;
2000 }
2001
2002 void bch2_dev_buckets_free(struct bch_dev *ca)
2003 {
2004         unsigned i;
2005
2006         free_heap(&ca->copygc_heap);
2007         free_heap(&ca->alloc_heap);
2008         free_fifo(&ca->free_inc);
2009         for (i = 0; i < RESERVE_NR; i++)
2010                 free_fifo(&ca->free[i]);
2011         kvpfree(ca->buckets_written,
2012                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2013         kvpfree(ca->buckets_nouse,
2014                 BITS_TO_LONGS(ca->mi.nbuckets) * sizeof(unsigned long));
2015         kvpfree(rcu_dereference_protected(ca->buckets[0], 1),
2016                 sizeof(struct bucket_array) +
2017                 ca->mi.nbuckets * sizeof(struct bucket));
2018
2019         free_percpu(ca->usage[0]);
2020 }
2021
2022 int bch2_dev_buckets_alloc(struct bch_fs *c, struct bch_dev *ca)
2023 {
2024         if (!(ca->usage[0] = alloc_percpu(struct bch_dev_usage)))
2025                 return -ENOMEM;
2026
2027         return bch2_dev_buckets_resize(c, ca, ca->mi.nbuckets);;
2028 }