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

   1 #include "bcachefs.h"
   2 #include "alloc_background.h"
   3 #include "alloc_foreground.h"
   4 #include "btree_cache.h"
   5 #include "btree_io.h"
   6 #include "btree_update.h"
   7 #include "btree_update_interior.h"
   8 #include "btree_gc.h"
   9 #include "buckets.h"
  10 #include "clock.h"
  11 #include "debug.h"
  12 #include "error.h"
  13 #include "journal_io.h"
  14
  15 #include <linux/kthread.h>
  16 #include <linux/math64.h>
  17 #include <linux/random.h>
  18 #include <linux/rculist.h>
  19 #include <linux/rcupdate.h>
  20 #include <linux/sched/task.h>
  21 #include <linux/sort.h>
  22 #include <trace/events/bcachefs.h>
  23
  24 static void bch2_recalc_oldest_io(struct bch_fs *, struct bch_dev *, int);
  25
  26 /* Ratelimiting/PD controllers */
  27
  28 static void pd_controllers_update(struct work_struct *work)
  29 {
  30         struct bch_fs *c = container_of(to_delayed_work(work),
  31                                            struct bch_fs,
  32                                            pd_controllers_update);
  33         struct bch_dev *ca;
  34         unsigned i;
  35
  36         for_each_member_device(ca, c, i) {
  37                 struct bch_dev_usage stats = bch2_dev_usage_read(c, ca);
  38
  39                 u64 free = bucket_to_sector(ca,
  40                                 __dev_buckets_free(ca, stats)) << 9;
  41                 /*
  42                  * Bytes of internal fragmentation, which can be
  43                  * reclaimed by copy GC
  44                  */
  45                 s64 fragmented = (bucket_to_sector(ca,
  46                                         stats.buckets[BCH_DATA_USER] +
  47                                         stats.buckets[BCH_DATA_CACHED]) -
  48                                   (stats.sectors[BCH_DATA_USER] +
  49                                    stats.sectors[BCH_DATA_CACHED])) << 9;
  50
  51                 fragmented = max(0LL, fragmented);
  52
  53                 bch2_pd_controller_update(&ca->copygc_pd,
  54                                          free, fragmented, -1);
  55         }
  56
  57         schedule_delayed_work(&c->pd_controllers_update,
  58                               c->pd_controllers_update_seconds * HZ);
  59 }
  60
  61 /* Persistent alloc info: */
  62
  63 static unsigned bch_alloc_val_u64s(const struct bch_alloc *a)
  64 {
  65         unsigned bytes = offsetof(struct bch_alloc, data);
  66
  67         if (a->fields & (1 << BCH_ALLOC_FIELD_READ_TIME))
  68                 bytes += 2;
  69         if (a->fields & (1 << BCH_ALLOC_FIELD_WRITE_TIME))
  70                 bytes += 2;
  71
  72         return DIV_ROUND_UP(bytes, sizeof(u64));
  73 }
  74
  75 const char *bch2_alloc_invalid(const struct bch_fs *c, struct bkey_s_c k)
  76 {
  77         if (k.k->p.inode >= c->sb.nr_devices ||
  78             !c->devs[k.k->p.inode])
  79                 return "invalid device";
  80
  81         switch (k.k->type) {
  82         case BCH_ALLOC: {
  83                 struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
  84
  85                 if (bch_alloc_val_u64s(a.v) != bkey_val_u64s(a.k))
  86                         return "incorrect value size";
  87                 break;
  88         }
  89         default:
  90                 return "invalid type";
  91         }
  92
  93         return NULL;
  94 }
  95
  96 int bch2_alloc_to_text(struct bch_fs *c, char *buf,
  97                        size_t size, struct bkey_s_c k)
  98 {
  99         buf[0] = '\0';
 100
 101         switch (k.k->type) {
 102         case BCH_ALLOC:
 103                 break;
 104         }
 105
 106         return 0;
 107 }
 108
 109 static inline unsigned get_alloc_field(const u8 **p, unsigned bytes)
 110 {
 111         unsigned v;
 112
 113         switch (bytes) {
 114         case 1:
 115                 v = **p;
 116                 break;
 117         case 2:
 118                 v = le16_to_cpup((void *) *p);
 119                 break;
 120         case 4:
 121                 v = le32_to_cpup((void *) *p);
 122                 break;
 123         default:
 124                 BUG();
 125         }
 126
 127         *p += bytes;
 128         return v;
 129 }
 130
 131 static inline void put_alloc_field(u8 **p, unsigned bytes, unsigned v)
 132 {
 133         switch (bytes) {
 134         case 1:
 135                 **p = v;
 136                 break;
 137         case 2:
 138                 *((__le16 *) *p) = cpu_to_le16(v);
 139                 break;
 140         case 4:
 141                 *((__le32 *) *p) = cpu_to_le32(v);
 142                 break;
 143         default:
 144                 BUG();
 145         }
 146
 147         *p += bytes;
 148 }
 149
 150 static void bch2_alloc_read_key(struct bch_fs *c, struct bkey_s_c k)
 151 {
 152         struct bch_dev *ca;
 153         struct bkey_s_c_alloc a;
 154         struct bucket_mark new;
 155         struct bucket *g;
 156         const u8 *d;
 157
 158         if (k.k->type != BCH_ALLOC)
 159                 return;
 160
 161         a = bkey_s_c_to_alloc(k);
 162         ca = bch_dev_bkey_exists(c, a.k->p.inode);
 163
 164         if (a.k->p.offset >= ca->mi.nbuckets)
 165                 return;
 166
 167         percpu_down_read_preempt_disable(&c->usage_lock);
 168
 169         g = bucket(ca, a.k->p.offset);
 170         bucket_cmpxchg(g, new, ({
 171                 new.gen = a.v->gen;
 172                 new.gen_valid = 1;
 173         }));
 174
 175         d = a.v->data;
 176         if (a.v->fields & (1 << BCH_ALLOC_FIELD_READ_TIME))
 177                 g->io_time[READ] = get_alloc_field(&d, 2);
 178         if (a.v->fields & (1 << BCH_ALLOC_FIELD_WRITE_TIME))
 179                 g->io_time[WRITE] = get_alloc_field(&d, 2);
 180
 181         percpu_up_read_preempt_enable(&c->usage_lock);
 182 }
 183
 184 int bch2_alloc_read(struct bch_fs *c, struct list_head *journal_replay_list)
 185 {
 186         struct journal_replay *r;
 187         struct btree_iter iter;
 188         struct bkey_s_c k;
 189         struct bch_dev *ca;
 190         unsigned i;
 191         int ret;
 192
 193         for_each_btree_key(&iter, c, BTREE_ID_ALLOC, POS_MIN, 0, k) {
 194                 bch2_alloc_read_key(c, k);
 195                 bch2_btree_iter_cond_resched(&iter);
 196         }
 197
 198         ret = bch2_btree_iter_unlock(&iter);
 199         if (ret)
 200                 return ret;
 201
 202         list_for_each_entry(r, journal_replay_list, list) {
 203                 struct bkey_i *k, *n;
 204                 struct jset_entry *entry;
 205
 206                 for_each_jset_key(k, n, entry, &r->j)
 207                         if (entry->btree_id == BTREE_ID_ALLOC)
 208                                 bch2_alloc_read_key(c, bkey_i_to_s_c(k));
 209         }
 210
 211         mutex_lock(&c->bucket_clock[READ].lock);
 212         for_each_member_device(ca, c, i) {
 213                 down_read(&ca->bucket_lock);
 214                 bch2_recalc_oldest_io(c, ca, READ);
 215                 up_read(&ca->bucket_lock);
 216         }
 217         mutex_unlock(&c->bucket_clock[READ].lock);
 218
 219         mutex_lock(&c->bucket_clock[WRITE].lock);
 220         for_each_member_device(ca, c, i) {
 221                 down_read(&ca->bucket_lock);
 222                 bch2_recalc_oldest_io(c, ca, WRITE);
 223                 up_read(&ca->bucket_lock);
 224         }
 225         mutex_unlock(&c->bucket_clock[WRITE].lock);
 226
 227         return 0;
 228 }
 229
 230 static int __bch2_alloc_write_key(struct bch_fs *c, struct bch_dev *ca,
 231                                   size_t b, struct btree_iter *iter,
 232                                   u64 *journal_seq, unsigned flags)
 233 {
 234         struct bucket_mark m;
 235         __BKEY_PADDED(k, DIV_ROUND_UP(sizeof(struct bch_alloc), 8)) alloc_key;
 236         struct bucket *g;
 237         struct bkey_i_alloc *a;
 238         u8 *d;
 239
 240         percpu_down_read_preempt_disable(&c->usage_lock);
 241         g = bucket(ca, b);
 242
 243         m = READ_ONCE(g->mark);
 244         a = bkey_alloc_init(&alloc_key.k);
 245         a->k.p          = POS(ca->dev_idx, b);
 246         a->v.fields     = 0;
 247         a->v.gen        = m.gen;
 248         set_bkey_val_u64s(&a->k, bch_alloc_val_u64s(&a->v));
 249
 250         d = a->v.data;
 251         if (a->v.fields & (1 << BCH_ALLOC_FIELD_READ_TIME))
 252                 put_alloc_field(&d, 2, g->io_time[READ]);
 253         if (a->v.fields & (1 << BCH_ALLOC_FIELD_WRITE_TIME))
 254                 put_alloc_field(&d, 2, g->io_time[WRITE]);
 255         percpu_up_read_preempt_enable(&c->usage_lock);
 256
 257         bch2_btree_iter_cond_resched(iter);
 258
 259         bch2_btree_iter_set_pos(iter, a->k.p);
 260
 261         return bch2_btree_insert_at(c, NULL, journal_seq,
 262                                     BTREE_INSERT_NOFAIL|
 263                                     BTREE_INSERT_USE_RESERVE|
 264                                     BTREE_INSERT_USE_ALLOC_RESERVE|
 265                                     flags,
 266                                     BTREE_INSERT_ENTRY(iter, &a->k_i));
 267 }
 268
 269 int bch2_alloc_replay_key(struct bch_fs *c, struct bpos pos)
 270 {
 271         struct bch_dev *ca;
 272         struct btree_iter iter;
 273         int ret;
 274
 275         if (pos.inode >= c->sb.nr_devices || !c->devs[pos.inode])
 276                 return 0;
 277
 278         ca = bch_dev_bkey_exists(c, pos.inode);
 279
 280         if (pos.offset >= ca->mi.nbuckets)
 281                 return 0;
 282
 283         bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, POS_MIN,
 284                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 285
 286         ret = __bch2_alloc_write_key(c, ca, pos.offset, &iter, NULL, 0);
 287         bch2_btree_iter_unlock(&iter);
 288         return ret;
 289 }
 290
 291 int bch2_alloc_write(struct bch_fs *c)
 292 {
 293         struct bch_dev *ca;
 294         unsigned i;
 295         int ret = 0;
 296
 297         for_each_rw_member(ca, c, i) {
 298                 struct btree_iter iter;
 299                 unsigned long bucket;
 300
 301                 bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, POS_MIN,
 302                                      BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 303
 304                 down_read(&ca->bucket_lock);
 305                 for_each_set_bit(bucket, ca->buckets_dirty, ca->mi.nbuckets) {
 306                         ret = __bch2_alloc_write_key(c, ca, bucket,
 307                                                      &iter, NULL, 0);
 308                         if (ret)
 309                                 break;
 310
 311                         clear_bit(bucket, ca->buckets_dirty);
 312                 }
 313                 up_read(&ca->bucket_lock);
 314                 bch2_btree_iter_unlock(&iter);
 315
 316                 if (ret) {
 317                         percpu_ref_put(&ca->io_ref);
 318                         break;
 319                 }
 320         }
 321
 322         return ret;
 323 }
 324
 325 /* Bucket IO clocks: */
 326
 327 static void bch2_recalc_oldest_io(struct bch_fs *c, struct bch_dev *ca, int rw)
 328 {
 329         struct bucket_clock *clock = &c->bucket_clock[rw];
 330         struct bucket_array *buckets = bucket_array(ca);
 331         struct bucket *g;
 332         u16 max_last_io = 0;
 333         unsigned i;
 334
 335         lockdep_assert_held(&c->bucket_clock[rw].lock);
 336
 337         /* Recalculate max_last_io for this device: */
 338         for_each_bucket(g, buckets)
 339                 max_last_io = max(max_last_io, bucket_last_io(c, g, rw));
 340
 341         ca->max_last_bucket_io[rw] = max_last_io;
 342
 343         /* Recalculate global max_last_io: */
 344         max_last_io = 0;
 345
 346         for_each_member_device(ca, c, i)
 347                 max_last_io = max(max_last_io, ca->max_last_bucket_io[rw]);
 348
 349         clock->max_last_io = max_last_io;
 350 }
 351
 352 static void bch2_rescale_bucket_io_times(struct bch_fs *c, int rw)
 353 {
 354         struct bucket_clock *clock = &c->bucket_clock[rw];
 355         struct bucket_array *buckets;
 356         struct bch_dev *ca;
 357         struct bucket *g;
 358         unsigned i;
 359
 360         trace_rescale_prios(c);
 361
 362         for_each_member_device(ca, c, i) {
 363                 down_read(&ca->bucket_lock);
 364                 buckets = bucket_array(ca);
 365
 366                 for_each_bucket(g, buckets)
 367                         g->io_time[rw] = clock->hand -
 368                         bucket_last_io(c, g, rw) / 2;
 369
 370                 bch2_recalc_oldest_io(c, ca, rw);
 371
 372                 up_read(&ca->bucket_lock);
 373         }
 374 }
 375
 376 static void bch2_inc_clock_hand(struct io_timer *timer)
 377 {
 378         struct bucket_clock *clock = container_of(timer,
 379                                                 struct bucket_clock, rescale);
 380         struct bch_fs *c = container_of(clock,
 381                                         struct bch_fs, bucket_clock[clock->rw]);
 382         struct bch_dev *ca;
 383         u64 capacity;
 384         unsigned i;
 385
 386         mutex_lock(&clock->lock);
 387
 388         /* if clock cannot be advanced more, rescale prio */
 389         if (clock->max_last_io >= U16_MAX - 2)
 390                 bch2_rescale_bucket_io_times(c, clock->rw);
 391
 392         BUG_ON(clock->max_last_io >= U16_MAX - 2);
 393
 394         for_each_member_device(ca, c, i)
 395                 ca->max_last_bucket_io[clock->rw]++;
 396         clock->max_last_io++;
 397         clock->hand++;
 398
 399         mutex_unlock(&clock->lock);
 400
 401         capacity = READ_ONCE(c->capacity);
 402
 403         if (!capacity)
 404                 return;
 405
 406         /*
 407          * we only increment when 0.1% of the filesystem capacity has been read
 408          * or written too, this determines if it's time
 409          *
 410          * XXX: we shouldn't really be going off of the capacity of devices in
 411          * RW mode (that will be 0 when we're RO, yet we can still service
 412          * reads)
 413          */
 414         timer->expire += capacity >> 10;
 415
 416         bch2_io_timer_add(&c->io_clock[clock->rw], timer);
 417 }
 418
 419 static void bch2_bucket_clock_init(struct bch_fs *c, int rw)
 420 {
 421         struct bucket_clock *clock = &c->bucket_clock[rw];
 422
 423         clock->hand             = 1;
 424         clock->rw               = rw;
 425         clock->rescale.fn       = bch2_inc_clock_hand;
 426         clock->rescale.expire   = c->capacity >> 10;
 427         mutex_init(&clock->lock);
 428 }
 429
 430 /* Background allocator thread: */
 431
 432 /*
 433  * Scans for buckets to be invalidated, invalidates them, rewrites prios/gens
 434  * (marking them as invalidated on disk), then optionally issues discard
 435  * commands to the newly free buckets, then puts them on the various freelists.
 436  */
 437
 438 #define BUCKET_GC_GEN_MAX       96U
 439
 440 /**
 441  * wait_buckets_available - wait on reclaimable buckets
 442  *
 443  * If there aren't enough available buckets to fill up free_inc, wait until
 444  * there are.
 445  */
 446 static int wait_buckets_available(struct bch_fs *c, struct bch_dev *ca)
 447 {
 448         unsigned long gc_count = c->gc_count;
 449         int ret = 0;
 450
 451         while (1) {
 452                 set_current_state(TASK_INTERRUPTIBLE);
 453                 if (kthread_should_stop()) {
 454                         ret = 1;
 455                         break;
 456                 }
 457
 458                 if (gc_count != c->gc_count)
 459                         ca->inc_gen_really_needs_gc = 0;
 460
 461                 if ((ssize_t) (dev_buckets_available(c, ca) -
 462                                ca->inc_gen_really_needs_gc) >=
 463                     (ssize_t) fifo_free(&ca->free_inc))
 464                         break;
 465
 466                 up_read(&c->gc_lock);
 467                 schedule();
 468                 try_to_freeze();
 469                 down_read(&c->gc_lock);
 470         }
 471
 472         __set_current_state(TASK_RUNNING);
 473         return ret;
 474 }
 475
 476 static bool bch2_can_invalidate_bucket(struct bch_dev *ca,
 477                                        size_t bucket,
 478                                        struct bucket_mark mark)
 479 {
 480         u8 gc_gen;
 481
 482         if (!is_available_bucket(mark))
 483                 return false;
 484
 485         gc_gen = bucket_gc_gen(ca, bucket);
 486
 487         if (gc_gen >= BUCKET_GC_GEN_MAX / 2)
 488                 ca->inc_gen_needs_gc++;
 489
 490         if (gc_gen >= BUCKET_GC_GEN_MAX)
 491                 ca->inc_gen_really_needs_gc++;
 492
 493         return gc_gen < BUCKET_GC_GEN_MAX;
 494 }
 495
 496 /*
 497  * Determines what order we're going to reuse buckets, smallest bucket_key()
 498  * first.
 499  *
 500  *
 501  * - We take into account the read prio of the bucket, which gives us an
 502  *   indication of how hot the data is -- we scale the prio so that the prio
 503  *   farthest from the clock is worth 1/8th of the closest.
 504  *
 505  * - The number of sectors of cached data in the bucket, which gives us an
 506  *   indication of the cost in cache misses this eviction will cause.
 507  *
 508  * - If hotness * sectors used compares equal, we pick the bucket with the
 509  *   smallest bucket_gc_gen() - since incrementing the same bucket's generation
 510  *   number repeatedly forces us to run mark and sweep gc to avoid generation
 511  *   number wraparound.
 512  */
 513
 514 static unsigned long bucket_sort_key(struct bch_fs *c, struct bch_dev *ca,
 515                                      size_t b, struct bucket_mark m)
 516 {
 517         unsigned last_io = bucket_last_io(c, bucket(ca, b), READ);
 518         unsigned max_last_io = ca->max_last_bucket_io[READ];
 519
 520         /*
 521          * Time since last read, scaled to [0, 8) where larger value indicates
 522          * more recently read data:
 523          */
 524         unsigned long hotness = (max_last_io - last_io) * 7 / max_last_io;
 525
 526         /* How much we want to keep the data in this bucket: */
 527         unsigned long data_wantness =
 528                 (hotness + 1) * bucket_sectors_used(m);
 529
 530         unsigned long needs_journal_commit =
 531                 bucket_needs_journal_commit(m, c->journal.last_seq_ondisk);
 532
 533         return  (data_wantness << 9) |
 534                 (needs_journal_commit << 8) |
 535                 (bucket_gc_gen(ca, b) / 16);
 536 }
 537
 538 static inline int bucket_alloc_cmp(alloc_heap *h,
 539                                    struct alloc_heap_entry l,
 540                                    struct alloc_heap_entry r)
 541 {
 542         return (l.key > r.key) - (l.key < r.key) ?:
 543                 (l.nr < r.nr)  - (l.nr  > r.nr) ?:
 544                 (l.bucket > r.bucket) - (l.bucket < r.bucket);
 545 }
 546
 547 static inline int bucket_idx_cmp(const void *_l, const void *_r)
 548 {
 549         const struct alloc_heap_entry *l = _l, *r = _r;
 550
 551         return (l->bucket > r->bucket) - (l->bucket < r->bucket);
 552 }
 553
 554 static void find_reclaimable_buckets_lru(struct bch_fs *c, struct bch_dev *ca)
 555 {
 556         struct bucket_array *buckets;
 557         struct alloc_heap_entry e = { 0 };
 558         size_t b, i, nr = 0;
 559
 560         ca->alloc_heap.used = 0;
 561
 562         mutex_lock(&c->bucket_clock[READ].lock);
 563         down_read(&ca->bucket_lock);
 564
 565         buckets = bucket_array(ca);
 566
 567         bch2_recalc_oldest_io(c, ca, READ);
 568
 569         /*
 570          * Find buckets with lowest read priority, by building a maxheap sorted
 571          * by read priority and repeatedly replacing the maximum element until
 572          * all buckets have been visited.
 573          */
 574         for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++) {
 575                 struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
 576                 unsigned long key = bucket_sort_key(c, ca, b, m);
 577
 578                 if (!bch2_can_invalidate_bucket(ca, b, m))
 579                         continue;
 580
 581                 if (e.nr && e.bucket + e.nr == b && e.key == key) {
 582                         e.nr++;
 583                 } else {
 584                         if (e.nr)
 585                                 heap_add_or_replace(&ca->alloc_heap, e, -bucket_alloc_cmp);
 586
 587                         e = (struct alloc_heap_entry) {
 588                                 .bucket = b,
 589                                 .nr     = 1,
 590                                 .key    = key,
 591                         };
 592                 }
 593
 594                 cond_resched();
 595         }
 596
 597         if (e.nr)
 598                 heap_add_or_replace(&ca->alloc_heap, e, -bucket_alloc_cmp);
 599
 600         for (i = 0; i < ca->alloc_heap.used; i++)
 601                 nr += ca->alloc_heap.data[i].nr;
 602
 603         while (nr - ca->alloc_heap.data[0].nr >= ALLOC_SCAN_BATCH(ca)) {
 604                 nr -= ca->alloc_heap.data[0].nr;
 605                 heap_pop(&ca->alloc_heap, e, -bucket_alloc_cmp);
 606         }
 607
 608         up_read(&ca->bucket_lock);
 609         mutex_unlock(&c->bucket_clock[READ].lock);
 610 }
 611
 612 static void find_reclaimable_buckets_fifo(struct bch_fs *c, struct bch_dev *ca)
 613 {
 614         struct bucket_array *buckets = bucket_array(ca);
 615         struct bucket_mark m;
 616         size_t b, start;
 617
 618         if (ca->fifo_last_bucket <  ca->mi.first_bucket ||
 619             ca->fifo_last_bucket >= ca->mi.nbuckets)
 620                 ca->fifo_last_bucket = ca->mi.first_bucket;
 621
 622         start = ca->fifo_last_bucket;
 623
 624         do {
 625                 ca->fifo_last_bucket++;
 626                 if (ca->fifo_last_bucket == ca->mi.nbuckets)
 627                         ca->fifo_last_bucket = ca->mi.first_bucket;
 628
 629                 b = ca->fifo_last_bucket;
 630                 m = READ_ONCE(buckets->b[b].mark);
 631
 632                 if (bch2_can_invalidate_bucket(ca, b, m)) {
 633                         struct alloc_heap_entry e = { .bucket = b, .nr = 1, };
 634
 635                         heap_add(&ca->alloc_heap, e, bucket_alloc_cmp);
 636                         if (heap_full(&ca->alloc_heap))
 637                                 break;
 638                 }
 639
 640                 cond_resched();
 641         } while (ca->fifo_last_bucket != start);
 642 }
 643
 644 static void find_reclaimable_buckets_random(struct bch_fs *c, struct bch_dev *ca)
 645 {
 646         struct bucket_array *buckets = bucket_array(ca);
 647         struct bucket_mark m;
 648         size_t checked, i;
 649
 650         for (checked = 0;
 651              checked < ca->mi.nbuckets / 2;
 652              checked++) {
 653                 size_t b = bch2_rand_range(ca->mi.nbuckets -
 654                                            ca->mi.first_bucket) +
 655                         ca->mi.first_bucket;
 656
 657                 m = READ_ONCE(buckets->b[b].mark);
 658
 659                 if (bch2_can_invalidate_bucket(ca, b, m)) {
 660                         struct alloc_heap_entry e = { .bucket = b, .nr = 1, };
 661
 662                         heap_add(&ca->alloc_heap, e, bucket_alloc_cmp);
 663                         if (heap_full(&ca->alloc_heap))
 664                                 break;
 665                 }
 666
 667                 cond_resched();
 668         }
 669
 670         sort(ca->alloc_heap.data,
 671              ca->alloc_heap.used,
 672              sizeof(ca->alloc_heap.data[0]),
 673              bucket_idx_cmp, NULL);
 674
 675         /* remove duplicates: */
 676         for (i = 0; i + 1 < ca->alloc_heap.used; i++)
 677                 if (ca->alloc_heap.data[i].bucket ==
 678                     ca->alloc_heap.data[i + 1].bucket)
 679                         ca->alloc_heap.data[i].nr = 0;
 680 }
 681
 682 static size_t find_reclaimable_buckets(struct bch_fs *c, struct bch_dev *ca)
 683 {
 684         size_t i, nr = 0;
 685
 686         ca->inc_gen_needs_gc                    = 0;
 687
 688         switch (ca->mi.replacement) {
 689         case CACHE_REPLACEMENT_LRU:
 690                 find_reclaimable_buckets_lru(c, ca);
 691                 break;
 692         case CACHE_REPLACEMENT_FIFO:
 693                 find_reclaimable_buckets_fifo(c, ca);
 694                 break;
 695         case CACHE_REPLACEMENT_RANDOM:
 696                 find_reclaimable_buckets_random(c, ca);
 697                 break;
 698         }
 699
 700         heap_resort(&ca->alloc_heap, bucket_alloc_cmp);
 701
 702         for (i = 0; i < ca->alloc_heap.used; i++)
 703                 nr += ca->alloc_heap.data[i].nr;
 704
 705         return nr;
 706 }
 707
 708 static inline long next_alloc_bucket(struct bch_dev *ca)
 709 {
 710         struct alloc_heap_entry e, *top = ca->alloc_heap.data;
 711
 712         while (ca->alloc_heap.used) {
 713                 if (top->nr) {
 714                         size_t b = top->bucket;
 715
 716                         top->bucket++;
 717                         top->nr--;
 718                         return b;
 719                 }
 720
 721                 heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp);
 722         }
 723
 724         return -1;
 725 }
 726
 727 static bool bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca,
 728                                        size_t bucket, u64 *flush_seq)
 729 {
 730         struct bucket_mark m;
 731
 732         percpu_down_read_preempt_disable(&c->usage_lock);
 733         spin_lock(&c->freelist_lock);
 734
 735         bch2_invalidate_bucket(c, ca, bucket, &m);
 736
 737         verify_not_on_freelist(c, ca, bucket);
 738         BUG_ON(!fifo_push(&ca->free_inc, bucket));
 739
 740         spin_unlock(&c->freelist_lock);
 741
 742         bucket_io_clock_reset(c, ca, bucket, READ);
 743         bucket_io_clock_reset(c, ca, bucket, WRITE);
 744
 745         percpu_up_read_preempt_enable(&c->usage_lock);
 746
 747         if (m.journal_seq_valid) {
 748                 u64 journal_seq = atomic64_read(&c->journal.seq);
 749                 u64 bucket_seq  = journal_seq;
 750
 751                 bucket_seq &= ~((u64) U16_MAX);
 752                 bucket_seq |= m.journal_seq;
 753
 754                 if (bucket_seq > journal_seq)
 755                         bucket_seq -= 1 << 16;
 756
 757                 *flush_seq = max(*flush_seq, bucket_seq);
 758         }
 759
 760         return m.cached_sectors != 0;
 761 }
 762
 763 /*
 764  * Pull buckets off ca->alloc_heap, invalidate them, move them to ca->free_inc:
 765  */
 766 static int bch2_invalidate_buckets(struct bch_fs *c, struct bch_dev *ca)
 767 {
 768         struct btree_iter iter;
 769         u64 journal_seq = 0;
 770         int ret = 0;
 771         long b;
 772
 773         bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, POS(ca->dev_idx, 0),
 774                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 775
 776         /* Only use nowait if we've already invalidated at least one bucket: */
 777         while (!ret &&
 778                !fifo_full(&ca->free_inc) &&
 779                (b = next_alloc_bucket(ca)) >= 0) {
 780                 bool must_flush =
 781                         bch2_invalidate_one_bucket(c, ca, b, &journal_seq);
 782
 783                 ret = __bch2_alloc_write_key(c, ca, b, &iter,
 784                                 must_flush ? &journal_seq : NULL,
 785                                 !fifo_empty(&ca->free_inc) ? BTREE_INSERT_NOWAIT : 0);
 786         }
 787
 788         bch2_btree_iter_unlock(&iter);
 789
 790         /* If we used NOWAIT, don't return the error: */
 791         if (!fifo_empty(&ca->free_inc))
 792                 ret = 0;
 793         if (ret) {
 794                 bch_err(ca, "error invalidating buckets: %i", ret);
 795                 return ret;
 796         }
 797
 798         if (journal_seq)
 799                 ret = bch2_journal_flush_seq(&c->journal, journal_seq);
 800         if (ret) {
 801                 bch_err(ca, "journal error: %i", ret);
 802                 return ret;
 803         }
 804
 805         return 0;
 806 }
 807
 808 static int push_invalidated_bucket(struct bch_fs *c, struct bch_dev *ca, size_t bucket)
 809 {
 810         unsigned i;
 811         int ret = 0;
 812
 813         while (1) {
 814                 set_current_state(TASK_INTERRUPTIBLE);
 815
 816                 spin_lock(&c->freelist_lock);
 817                 for (i = 0; i < RESERVE_NR; i++)
 818                         if (fifo_push(&ca->free[i], bucket)) {
 819                                 fifo_pop(&ca->free_inc, bucket);
 820                                 closure_wake_up(&c->freelist_wait);
 821                                 spin_unlock(&c->freelist_lock);
 822                                 goto out;
 823                         }
 824                 spin_unlock(&c->freelist_lock);
 825
 826                 if ((current->flags & PF_KTHREAD) &&
 827                     kthread_should_stop()) {
 828                         ret = 1;
 829                         break;
 830                 }
 831
 832                 schedule();
 833                 try_to_freeze();
 834         }
 835 out:
 836         __set_current_state(TASK_RUNNING);
 837         return ret;
 838 }
 839
 840 /*
 841  * Pulls buckets off free_inc, discards them (if enabled), then adds them to
 842  * freelists, waiting until there's room if necessary:
 843  */
 844 static int discard_invalidated_buckets(struct bch_fs *c, struct bch_dev *ca)
 845 {
 846         while (!fifo_empty(&ca->free_inc)) {
 847                 size_t bucket = fifo_peek(&ca->free_inc);
 848
 849                 if (ca->mi.discard &&
 850                     blk_queue_discard(bdev_get_queue(ca->disk_sb.bdev)))
 851                         blkdev_issue_discard(ca->disk_sb.bdev,
 852                                              bucket_to_sector(ca, bucket),
 853                                              ca->mi.bucket_size, GFP_NOIO, 0);
 854
 855                 if (push_invalidated_bucket(c, ca, bucket))
 856                         return 1;
 857         }
 858
 859         return 0;
 860 }
 861
 862 /**
 863  * bch_allocator_thread - move buckets from free_inc to reserves
 864  *
 865  * The free_inc FIFO is populated by find_reclaimable_buckets(), and
 866  * the reserves are depleted by bucket allocation. When we run out
 867  * of free_inc, try to invalidate some buckets and write out
 868  * prios and gens.
 869  */
 870 static int bch2_allocator_thread(void *arg)
 871 {
 872         struct bch_dev *ca = arg;
 873         struct bch_fs *c = ca->fs;
 874         size_t nr;
 875         int ret;
 876
 877         set_freezable();
 878
 879         while (1) {
 880                 cond_resched();
 881
 882                 pr_debug("discarding %zu invalidated buckets",
 883                          fifo_used(&ca->free_inc));
 884
 885                 ret = discard_invalidated_buckets(c, ca);
 886                 if (ret)
 887                         goto stop;
 888
 889                 down_read(&c->gc_lock);
 890
 891                 ret = bch2_invalidate_buckets(c, ca);
 892                 if (ret) {
 893                         up_read(&c->gc_lock);
 894                         goto stop;
 895                 }
 896
 897                 if (!fifo_empty(&ca->free_inc)) {
 898                         up_read(&c->gc_lock);
 899                         continue;
 900                 }
 901
 902                 pr_debug("free_inc now empty");
 903
 904                 do {
 905                         if (test_bit(BCH_FS_GC_FAILURE, &c->flags)) {
 906                                 up_read(&c->gc_lock);
 907                                 bch_err(ca, "gc failure");
 908                                 goto stop;
 909                         }
 910
 911                         /*
 912                          * Find some buckets that we can invalidate, either
 913                          * they're completely unused, or only contain clean data
 914                          * that's been written back to the backing device or
 915                          * another cache tier
 916                          */
 917
 918                         pr_debug("scanning for reclaimable buckets");
 919
 920                         nr = find_reclaimable_buckets(c, ca);
 921
 922                         pr_debug("found %zu buckets", nr);
 923
 924                         trace_alloc_batch(ca, nr, ca->alloc_heap.size);
 925
 926                         if ((ca->inc_gen_needs_gc >= ALLOC_SCAN_BATCH(ca) ||
 927                              ca->inc_gen_really_needs_gc) &&
 928                             c->gc_thread) {
 929                                 atomic_inc(&c->kick_gc);
 930                                 wake_up_process(c->gc_thread);
 931                         }
 932
 933                         /*
 934                          * If we found any buckets, we have to invalidate them
 935                          * before we scan for more - but if we didn't find very
 936                          * many we may want to wait on more buckets being
 937                          * available so we don't spin:
 938                          */
 939                         if (!nr ||
 940                             (nr < ALLOC_SCAN_BATCH(ca) &&
 941                              !fifo_full(&ca->free[RESERVE_MOVINGGC]))) {
 942                                 ca->allocator_blocked = true;
 943                                 closure_wake_up(&c->freelist_wait);
 944
 945                                 ret = wait_buckets_available(c, ca);
 946                                 if (ret) {
 947                                         up_read(&c->gc_lock);
 948                                         goto stop;
 949                                 }
 950                         }
 951                 } while (!nr);
 952
 953                 ca->allocator_blocked = false;
 954                 up_read(&c->gc_lock);
 955
 956                 pr_debug("%zu buckets to invalidate", nr);
 957
 958                 /*
 959                  * alloc_heap is now full of newly-invalidated buckets: next,
 960                  * write out the new bucket gens:
 961                  */
 962         }
 963
 964 stop:
 965         pr_debug("alloc thread stopping (ret %i)", ret);
 966         return 0;
 967 }
 968
 969 /* Startup/shutdown (ro/rw): */
 970
 971 void bch2_recalc_capacity(struct bch_fs *c)
 972 {
 973         struct bch_dev *ca;
 974         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
 975         unsigned long ra_pages = 0;
 976         unsigned i, j;
 977
 978         lockdep_assert_held(&c->state_lock);
 979
 980         for_each_online_member(ca, c, i) {
 981                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_bdi;
 982
 983                 ra_pages += bdi->ra_pages;
 984         }
 985
 986         bch2_set_ra_pages(c, ra_pages);
 987
 988         for_each_rw_member(ca, c, i) {
 989                 u64 dev_reserve = 0;
 990
 991                 /*
 992                  * We need to reserve buckets (from the number
 993                  * of currently available buckets) against
 994                  * foreground writes so that mainly copygc can
 995                  * make forward progress.
 996                  *
 997                  * We need enough to refill the various reserves
 998                  * from scratch - copygc will use its entire
 999                  * reserve all at once, then run against when
1000                  * its reserve is refilled (from the formerly
1001                  * available buckets).
1002                  *
1003                  * This reserve is just used when considering if
1004                  * allocations for foreground writes must wait -
1005                  * not -ENOSPC calculations.
1006                  */
1007                 for (j = 0; j < RESERVE_NONE; j++)
1008                         dev_reserve += ca->free[j].size;
1009
1010                 dev_reserve += ca->free_inc.size;
1011
1012                 dev_reserve += ARRAY_SIZE(c->write_points);
1013
1014                 dev_reserve += 1;       /* btree write point */
1015                 dev_reserve += 1;       /* copygc write point */
1016                 dev_reserve += 1;       /* rebalance write point */
1017                 dev_reserve += WRITE_POINT_COUNT;
1018
1019                 dev_reserve *= ca->mi.bucket_size;
1020
1021                 ca->copygc_threshold = dev_reserve;
1022
1023                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
1024                                              ca->mi.first_bucket);
1025
1026                 reserved_sectors += dev_reserve * 2;
1027         }
1028
1029         gc_reserve = c->opts.gc_reserve_bytes
1030                 ? c->opts.gc_reserve_bytes >> 9
1031                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
1032
1033         reserved_sectors = max(gc_reserve, reserved_sectors);
1034
1035         reserved_sectors = min(reserved_sectors, capacity);
1036
1037         c->capacity = capacity - reserved_sectors;
1038
1039         if (c->capacity) {
1040                 bch2_io_timer_add(&c->io_clock[READ],
1041                                  &c->bucket_clock[READ].rescale);
1042                 bch2_io_timer_add(&c->io_clock[WRITE],
1043                                  &c->bucket_clock[WRITE].rescale);
1044         } else {
1045                 bch2_io_timer_del(&c->io_clock[READ],
1046                                  &c->bucket_clock[READ].rescale);
1047                 bch2_io_timer_del(&c->io_clock[WRITE],
1048                                  &c->bucket_clock[WRITE].rescale);
1049         }
1050
1051         /* Wake up case someone was waiting for buckets */
1052         closure_wake_up(&c->freelist_wait);
1053 }
1054
1055 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
1056 {
1057         struct open_bucket *ob;
1058         bool ret = false;
1059
1060         for (ob = c->open_buckets;
1061              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1062              ob++) {
1063                 spin_lock(&ob->lock);
1064                 if (ob->valid && !ob->on_partial_list &&
1065                     ob->ptr.dev == ca->dev_idx)
1066                         ret = true;
1067                 spin_unlock(&ob->lock);
1068         }
1069
1070         return ret;
1071 }
1072
1073 /* device goes ro: */
1074 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
1075 {
1076         unsigned i;
1077
1078         BUG_ON(ca->alloc_thread);
1079
1080         /* First, remove device from allocation groups: */
1081
1082         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1083                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
1084
1085         /*
1086          * Capacity is calculated based off of devices in allocation groups:
1087          */
1088         bch2_recalc_capacity(c);
1089
1090         /* Next, close write points that point to this device... */
1091         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1092                 bch2_writepoint_stop(c, ca, &c->write_points[i]);
1093
1094         bch2_writepoint_stop(c, ca, &ca->copygc_write_point);
1095         bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
1096         bch2_writepoint_stop(c, ca, &c->btree_write_point);
1097
1098         mutex_lock(&c->btree_reserve_cache_lock);
1099         while (c->btree_reserve_cache_nr) {
1100                 struct btree_alloc *a =
1101                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1102
1103                 bch2_open_buckets_put(c, &a->ob);
1104         }
1105         mutex_unlock(&c->btree_reserve_cache_lock);
1106
1107         /*
1108          * Wake up threads that were blocked on allocation, so they can notice
1109          * the device can no longer be removed and the capacity has changed:
1110          */
1111         closure_wake_up(&c->freelist_wait);
1112
1113         /*
1114          * journal_res_get() can block waiting for free space in the journal -
1115          * it needs to notice there may not be devices to allocate from anymore:
1116          */
1117         wake_up(&c->journal.wait);
1118
1119         /* Now wait for any in flight writes: */
1120
1121         closure_wait_event(&c->open_buckets_wait,
1122                            !bch2_dev_has_open_write_point(c, ca));
1123 }
1124
1125 /* device goes rw: */
1126 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
1127 {
1128         unsigned i;
1129
1130         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1131                 if (ca->mi.data_allowed & (1 << i))
1132                         set_bit(ca->dev_idx, c->rw_devs[i].d);
1133 }
1134
1135 /* stop allocator thread: */
1136 void bch2_dev_allocator_stop(struct bch_dev *ca)
1137 {
1138         struct task_struct *p;
1139
1140         p = rcu_dereference_protected(ca->alloc_thread, 1);
1141         ca->alloc_thread = NULL;
1142
1143         /*
1144          * We need an rcu barrier between setting ca->alloc_thread = NULL and
1145          * the thread shutting down to avoid bch2_wake_allocator() racing:
1146          *
1147          * XXX: it would be better to have the rcu barrier be asynchronous
1148          * instead of blocking us here
1149          */
1150         synchronize_rcu();
1151
1152         if (p) {
1153                 kthread_stop(p);
1154                 put_task_struct(p);
1155         }
1156 }
1157
1158 /* start allocator thread: */
1159 int bch2_dev_allocator_start(struct bch_dev *ca)
1160 {
1161         struct task_struct *p;
1162
1163         /*
1164          * allocator thread already started?
1165          */
1166         if (ca->alloc_thread)
1167                 return 0;
1168
1169         p = kthread_create(bch2_allocator_thread, ca,
1170                            "bch_alloc[%s]", ca->name);
1171         if (IS_ERR(p))
1172                 return PTR_ERR(p);
1173
1174         get_task_struct(p);
1175         rcu_assign_pointer(ca->alloc_thread, p);
1176         wake_up_process(p);
1177         return 0;
1178 }
1179
1180 static void flush_held_btree_writes(struct bch_fs *c)
1181 {
1182         struct bucket_table *tbl;
1183         struct rhash_head *pos;
1184         struct btree *b;
1185         bool flush_updates;
1186         size_t i, nr_pending_updates;
1187
1188         clear_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags);
1189 again:
1190         pr_debug("flushing dirty btree nodes");
1191         cond_resched();
1192
1193         flush_updates = false;
1194         nr_pending_updates = bch2_btree_interior_updates_nr_pending(c);
1195
1196         rcu_read_lock();
1197         for_each_cached_btree(b, c, tbl, i, pos)
1198                 if (btree_node_dirty(b) && (!b->written || b->level)) {
1199                         if (btree_node_may_write(b)) {
1200                                 rcu_read_unlock();
1201                                 btree_node_lock_type(c, b, SIX_LOCK_read);
1202                                 bch2_btree_node_write(c, b, SIX_LOCK_read);
1203                                 six_unlock_read(&b->lock);
1204                                 goto again;
1205                         } else {
1206                                 flush_updates = true;
1207                         }
1208                 }
1209         rcu_read_unlock();
1210
1211         if (c->btree_roots_dirty)
1212                 bch2_journal_meta(&c->journal);
1213
1214         /*
1215          * This is ugly, but it's needed to flush btree node writes
1216          * without spinning...
1217          */
1218         if (flush_updates) {
1219                 closure_wait_event(&c->btree_interior_update_wait,
1220                                    bch2_btree_interior_updates_nr_pending(c) <
1221                                    nr_pending_updates);
1222                 goto again;
1223         }
1224
1225 }
1226
1227 static void allocator_start_issue_discards(struct bch_fs *c)
1228 {
1229         struct bch_dev *ca;
1230         unsigned dev_iter;
1231         size_t bu;
1232
1233         for_each_rw_member(ca, c, dev_iter)
1234                 while (fifo_pop(&ca->free_inc, bu))
1235                         blkdev_issue_discard(ca->disk_sb.bdev,
1236                                              bucket_to_sector(ca, bu),
1237                                              ca->mi.bucket_size, GFP_NOIO, 0);
1238 }
1239
1240 static int __bch2_fs_allocator_start(struct bch_fs *c)
1241 {
1242         struct bch_dev *ca;
1243         unsigned dev_iter;
1244         u64 journal_seq = 0;
1245         long bu;
1246         bool invalidating_data = false;
1247         int ret = 0;
1248
1249         if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
1250                 return -1;
1251
1252         if (test_alloc_startup(c)) {
1253                 invalidating_data = true;
1254                 goto not_enough;
1255         }
1256
1257         /* Scan for buckets that are already invalidated: */
1258         for_each_rw_member(ca, c, dev_iter) {
1259                 struct btree_iter iter;
1260                 struct bucket_mark m;
1261                 struct bkey_s_c k;
1262
1263                 for_each_btree_key(&iter, c, BTREE_ID_ALLOC, POS(ca->dev_idx, 0), 0, k) {
1264                         if (k.k->type != BCH_ALLOC)
1265                                 continue;
1266
1267                         bu = k.k->p.offset;
1268                         m = READ_ONCE(bucket(ca, bu)->mark);
1269
1270                         if (!is_available_bucket(m) || m.cached_sectors)
1271                                 continue;
1272
1273                         percpu_down_read_preempt_disable(&c->usage_lock);
1274                         bch2_mark_alloc_bucket(c, ca, bu, true,
1275                                         gc_pos_alloc(c, NULL),
1276                                         BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
1277                                         BCH_BUCKET_MARK_GC_LOCK_HELD);
1278                         percpu_up_read_preempt_enable(&c->usage_lock);
1279
1280                         fifo_push(&ca->free_inc, bu);
1281
1282                         if (fifo_full(&ca->free_inc))
1283                                 break;
1284                 }
1285                 bch2_btree_iter_unlock(&iter);
1286         }
1287
1288         /* did we find enough buckets? */
1289         for_each_rw_member(ca, c, dev_iter)
1290                 if (fifo_used(&ca->free_inc) < ca->free[RESERVE_BTREE].size) {
1291                         percpu_ref_put(&ca->io_ref);
1292                         goto not_enough;
1293                 }
1294
1295         return 0;
1296 not_enough:
1297         pr_debug("did not find enough empty buckets; issuing discards");
1298
1299         /* clear out free_inc, we'll be using it again below: */
1300         for_each_rw_member(ca, c, dev_iter)
1301                 discard_invalidated_buckets(c, ca);
1302
1303         pr_debug("scanning for reclaimable buckets");
1304
1305         for_each_rw_member(ca, c, dev_iter) {
1306                 find_reclaimable_buckets(c, ca);
1307
1308                 while (!fifo_full(&ca->free[RESERVE_BTREE]) &&
1309                        (bu = next_alloc_bucket(ca)) >= 0) {
1310                         invalidating_data |=
1311                                 bch2_invalidate_one_bucket(c, ca, bu, &journal_seq);
1312
1313                         fifo_push(&ca->free[RESERVE_BTREE], bu);
1314                         set_bit(bu, ca->buckets_dirty);
1315                 }
1316         }
1317
1318         pr_debug("done scanning for reclaimable buckets");
1319
1320         /*
1321          * We're moving buckets to freelists _before_ they've been marked as
1322          * invalidated on disk - we have to so that we can allocate new btree
1323          * nodes to mark them as invalidated on disk.
1324          *
1325          * However, we can't _write_ to any of these buckets yet - they might
1326          * have cached data in them, which is live until they're marked as
1327          * invalidated on disk:
1328          */
1329         if (invalidating_data) {
1330                 BUG();
1331                 pr_info("holding writes");
1332                 pr_debug("invalidating existing data");
1333                 set_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags);
1334         } else {
1335                 pr_debug("issuing discards");
1336                 allocator_start_issue_discards(c);
1337         }
1338
1339         /*
1340          * XXX: it's possible for this to deadlock waiting on journal reclaim,
1341          * since we're holding btree writes. What then?
1342          */
1343         ret = bch2_alloc_write(c);
1344         if (ret)
1345                 return ret;
1346
1347         if (invalidating_data) {
1348                 pr_debug("flushing journal");
1349
1350                 ret = bch2_journal_flush_seq(&c->journal, journal_seq);
1351                 if (ret)
1352                         return ret;
1353
1354                 pr_debug("issuing discards");
1355                 allocator_start_issue_discards(c);
1356         }
1357
1358         set_bit(BCH_FS_ALLOCATOR_STARTED, &c->flags);
1359
1360         /* now flush dirty btree nodes: */
1361         if (invalidating_data)
1362                 flush_held_btree_writes(c);
1363
1364         return 0;
1365 }
1366
1367 int bch2_fs_allocator_start(struct bch_fs *c)
1368 {
1369         struct bch_dev *ca;
1370         unsigned i;
1371         int ret;
1372
1373         down_read(&c->gc_lock);
1374         ret = __bch2_fs_allocator_start(c);
1375         up_read(&c->gc_lock);
1376
1377         if (ret)
1378                 return ret;
1379
1380         for_each_rw_member(ca, c, i) {
1381                 ret = bch2_dev_allocator_start(ca);
1382                 if (ret) {
1383                         percpu_ref_put(&ca->io_ref);
1384                         return ret;
1385                 }
1386         }
1387
1388         return bch2_alloc_write(c);
1389 }
1390
1391 void bch2_fs_allocator_init(struct bch_fs *c)
1392 {
1393         struct open_bucket *ob;
1394         struct write_point *wp;
1395
1396         mutex_init(&c->write_points_hash_lock);
1397         spin_lock_init(&c->freelist_lock);
1398         bch2_bucket_clock_init(c, READ);
1399         bch2_bucket_clock_init(c, WRITE);
1400
1401         /* open bucket 0 is a sentinal NULL: */
1402         spin_lock_init(&c->open_buckets[0].lock);
1403
1404         for (ob = c->open_buckets + 1;
1405              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1406                 spin_lock_init(&ob->lock);
1407                 c->open_buckets_nr_free++;
1408
1409                 ob->freelist = c->open_buckets_freelist;
1410                 c->open_buckets_freelist = ob - c->open_buckets;
1411         }
1412
1413         writepoint_init(&c->btree_write_point, BCH_DATA_BTREE);
1414         writepoint_init(&c->rebalance_write_point, BCH_DATA_USER);
1415
1416         for (wp = c->write_points;
1417              wp < c->write_points + ARRAY_SIZE(c->write_points); wp++) {
1418                 writepoint_init(wp, BCH_DATA_USER);
1419
1420                 wp->last_used   = sched_clock();
1421                 wp->write_point = (unsigned long) wp;
1422                 hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
1423         }
1424
1425         c->pd_controllers_update_seconds = 5;
1426         INIT_DELAYED_WORK(&c->pd_controllers_update, pd_controllers_update);
1427 }