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

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "bcachefs.h"
   3 #include "alloc_background.h"
   4 #include "alloc_foreground.h"
   5 #include "btree_cache.h"
   6 #include "btree_io.h"
   7 #include "btree_update.h"
   8 #include "btree_update_interior.h"
   9 #include "btree_gc.h"
  10 #include "buckets.h"
  11 #include "clock.h"
  12 #include "debug.h"
  13 #include "ec.h"
  14 #include "error.h"
  15 #include "recovery.h"
  16
  17 #include <linux/kthread.h>
  18 #include <linux/math64.h>
  19 #include <linux/random.h>
  20 #include <linux/rculist.h>
  21 #include <linux/rcupdate.h>
  22 #include <linux/sched/task.h>
  23 #include <linux/sort.h>
  24 #include <trace/events/bcachefs.h>
  25
  26 static const char * const bch2_alloc_field_names[] = {
  27 #define x(name, bytes) #name,
  28         BCH_ALLOC_FIELDS()
  29 #undef x
  30         NULL
  31 };
  32
  33 static void bch2_recalc_oldest_io(struct bch_fs *, struct bch_dev *, int);
  34
  35 /* Ratelimiting/PD controllers */
  36
  37 static void pd_controllers_update(struct work_struct *work)
  38 {
  39         struct bch_fs *c = container_of(to_delayed_work(work),
  40                                            struct bch_fs,
  41                                            pd_controllers_update);
  42         struct bch_dev *ca;
  43         unsigned i;
  44
  45         for_each_member_device(ca, c, i) {
  46                 struct bch_dev_usage stats = bch2_dev_usage_read(c, ca);
  47
  48                 u64 free = bucket_to_sector(ca,
  49                                 __dev_buckets_free(ca, stats)) << 9;
  50                 /*
  51                  * Bytes of internal fragmentation, which can be
  52                  * reclaimed by copy GC
  53                  */
  54                 s64 fragmented = (bucket_to_sector(ca,
  55                                         stats.buckets[BCH_DATA_USER] +
  56                                         stats.buckets[BCH_DATA_CACHED]) -
  57                                   (stats.sectors[BCH_DATA_USER] +
  58                                    stats.sectors[BCH_DATA_CACHED])) << 9;
  59
  60                 fragmented = max(0LL, fragmented);
  61
  62                 bch2_pd_controller_update(&ca->copygc_pd,
  63                                          free, fragmented, -1);
  64         }
  65
  66         schedule_delayed_work(&c->pd_controllers_update,
  67                               c->pd_controllers_update_seconds * HZ);
  68 }
  69
  70 /* Persistent alloc info: */
  71
  72 static inline u64 get_alloc_field(const struct bch_alloc *a,
  73                                   const void **p, unsigned field)
  74 {
  75         unsigned bytes = BCH_ALLOC_FIELD_BYTES[field];
  76         u64 v;
  77
  78         if (!(a->fields & (1 << field)))
  79                 return 0;
  80
  81         switch (bytes) {
  82         case 1:
  83                 v = *((const u8 *) *p);
  84                 break;
  85         case 2:
  86                 v = le16_to_cpup(*p);
  87                 break;
  88         case 4:
  89                 v = le32_to_cpup(*p);
  90                 break;
  91         case 8:
  92                 v = le64_to_cpup(*p);
  93                 break;
  94         default:
  95                 BUG();
  96         }
  97
  98         *p += bytes;
  99         return v;
 100 }
 101
 102 static inline void put_alloc_field(struct bkey_i_alloc *a, void **p,
 103                                    unsigned field, u64 v)
 104 {
 105         unsigned bytes = BCH_ALLOC_FIELD_BYTES[field];
 106
 107         if (!v)
 108                 return;
 109
 110         a->v.fields |= 1 << field;
 111
 112         switch (bytes) {
 113         case 1:
 114                 *((u8 *) *p) = v;
 115                 break;
 116         case 2:
 117                 *((__le16 *) *p) = cpu_to_le16(v);
 118                 break;
 119         case 4:
 120                 *((__le32 *) *p) = cpu_to_le32(v);
 121                 break;
 122         case 8:
 123                 *((__le64 *) *p) = cpu_to_le64(v);
 124                 break;
 125         default:
 126                 BUG();
 127         }
 128
 129         *p += bytes;
 130 }
 131
 132 struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 133 {
 134         struct bkey_alloc_unpacked ret = { .gen = 0 };
 135
 136         if (k.k->type == KEY_TYPE_alloc) {
 137                 const struct bch_alloc *a = bkey_s_c_to_alloc(k).v;
 138                 const void *d = a->data;
 139                 unsigned idx = 0;
 140
 141                 ret.gen = a->gen;
 142
 143 #define x(_name, _bits) ret._name = get_alloc_field(a, &d, idx++);
 144                 BCH_ALLOC_FIELDS()
 145 #undef  x
 146         }
 147         return ret;
 148 }
 149
 150 void bch2_alloc_pack(struct bkey_i_alloc *dst,
 151                      const struct bkey_alloc_unpacked src)
 152 {
 153         unsigned idx = 0;
 154         void *d = dst->v.data;
 155         unsigned bytes;
 156
 157         dst->v.fields   = 0;
 158         dst->v.gen      = src.gen;
 159
 160 #define x(_name, _bits) put_alloc_field(dst, &d, idx++, src._name);
 161         BCH_ALLOC_FIELDS()
 162 #undef  x
 163
 164         bytes = (void *) d - (void *) &dst->v;
 165         set_bkey_val_bytes(&dst->k, bytes);
 166         memset_u64s_tail(&dst->v, 0, bytes);
 167 }
 168
 169 static unsigned bch_alloc_val_u64s(const struct bch_alloc *a)
 170 {
 171         unsigned i, bytes = offsetof(struct bch_alloc, data);
 172
 173         for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_FIELD_BYTES); i++)
 174                 if (a->fields & (1 << i))
 175                         bytes += BCH_ALLOC_FIELD_BYTES[i];
 176
 177         return DIV_ROUND_UP(bytes, sizeof(u64));
 178 }
 179
 180 const char *bch2_alloc_invalid(const struct bch_fs *c, struct bkey_s_c k)
 181 {
 182         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 183
 184         if (k.k->p.inode >= c->sb.nr_devices ||
 185             !c->devs[k.k->p.inode])
 186                 return "invalid device";
 187
 188         /* allow for unknown fields */
 189         if (bkey_val_u64s(a.k) < bch_alloc_val_u64s(a.v))
 190                 return "incorrect value size";
 191
 192         return NULL;
 193 }
 194
 195 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c,
 196                         struct bkey_s_c k)
 197 {
 198         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 199         const void *d = a.v->data;
 200         unsigned i;
 201
 202         pr_buf(out, "gen %u", a.v->gen);
 203
 204         for (i = 0; i < BCH_ALLOC_FIELD_NR; i++)
 205                 if (a.v->fields & (1 << i))
 206                         pr_buf(out, " %s %llu",
 207                                bch2_alloc_field_names[i],
 208                                get_alloc_field(a.v, &d, i));
 209 }
 210
 211 static int bch2_alloc_read_fn(struct bch_fs *c, enum btree_id id,
 212                               unsigned level, struct bkey_s_c k)
 213 {
 214         if (!level)
 215                 bch2_mark_key(c, k, 0, 0, NULL, 0,
 216                               BTREE_TRIGGER_ALLOC_READ|
 217                               BTREE_TRIGGER_NOATOMIC);
 218
 219         return 0;
 220 }
 221
 222 int bch2_alloc_read(struct bch_fs *c, struct journal_keys *journal_keys)
 223 {
 224         struct bch_dev *ca;
 225         unsigned i;
 226         int ret = 0;
 227
 228         ret = bch2_btree_and_journal_walk(c, journal_keys, BTREE_ID_ALLOC,
 229                                           NULL, bch2_alloc_read_fn);
 230         if (ret) {
 231                 bch_err(c, "error reading alloc info: %i", ret);
 232                 return ret;
 233         }
 234
 235         percpu_down_write(&c->mark_lock);
 236         bch2_dev_usage_from_buckets(c);
 237         percpu_up_write(&c->mark_lock);
 238
 239         mutex_lock(&c->bucket_clock[READ].lock);
 240         for_each_member_device(ca, c, i) {
 241                 down_read(&ca->bucket_lock);
 242                 bch2_recalc_oldest_io(c, ca, READ);
 243                 up_read(&ca->bucket_lock);
 244         }
 245         mutex_unlock(&c->bucket_clock[READ].lock);
 246
 247         mutex_lock(&c->bucket_clock[WRITE].lock);
 248         for_each_member_device(ca, c, i) {
 249                 down_read(&ca->bucket_lock);
 250                 bch2_recalc_oldest_io(c, ca, WRITE);
 251                 up_read(&ca->bucket_lock);
 252         }
 253         mutex_unlock(&c->bucket_clock[WRITE].lock);
 254
 255         return 0;
 256 }
 257
 258 enum alloc_write_ret {
 259         ALLOC_WROTE,
 260         ALLOC_NOWROTE,
 261         ALLOC_END,
 262 };
 263
 264 static int bch2_alloc_write_key(struct btree_trans *trans,
 265                                 struct btree_iter *iter,
 266                                 unsigned flags)
 267 {
 268         struct bch_fs *c = trans->c;
 269         struct bkey_s_c k;
 270         struct bch_dev *ca;
 271         struct bucket_array *ba;
 272         struct bucket *g;
 273         struct bucket_mark m;
 274         struct bkey_alloc_unpacked old_u, new_u;
 275         __BKEY_PADDED(k, 8) alloc_key; /* hack: */
 276         struct bkey_i_alloc *a;
 277         int ret;
 278 retry:
 279         k = bch2_btree_iter_peek_slot(iter);
 280         ret = bkey_err(k);
 281         if (ret)
 282                 goto err;
 283
 284         old_u = bch2_alloc_unpack(k);
 285
 286         if (iter->pos.inode >= c->sb.nr_devices ||
 287             !c->devs[iter->pos.inode])
 288                 return ALLOC_END;
 289
 290         percpu_down_read(&c->mark_lock);
 291         ca      = bch_dev_bkey_exists(c, iter->pos.inode);
 292         ba      = bucket_array(ca);
 293
 294         if (iter->pos.offset >= ba->nbuckets) {
 295                 percpu_up_read(&c->mark_lock);
 296                 return ALLOC_END;
 297         }
 298
 299         g       = &ba->b[iter->pos.offset];
 300         m       = READ_ONCE(g->mark);
 301         new_u   = alloc_mem_to_key(g, m);
 302         percpu_up_read(&c->mark_lock);
 303
 304         if (!bkey_alloc_unpacked_cmp(old_u, new_u))
 305                 return ALLOC_NOWROTE;
 306
 307         a = bkey_alloc_init(&alloc_key.k);
 308         a->k.p = iter->pos;
 309         bch2_alloc_pack(a, new_u);
 310
 311         bch2_trans_update(trans, iter, &a->k_i,
 312                           BTREE_TRIGGER_NORUN);
 313         ret = bch2_trans_commit(trans, NULL, NULL,
 314                                 BTREE_INSERT_NOFAIL|
 315                                 BTREE_INSERT_USE_RESERVE|
 316                                 flags);
 317 err:
 318         if (ret == -EINTR)
 319                 goto retry;
 320         return ret;
 321 }
 322
 323 int bch2_alloc_write(struct bch_fs *c, unsigned flags, bool *wrote)
 324 {
 325         struct btree_trans trans;
 326         struct btree_iter *iter;
 327         struct bch_dev *ca;
 328         unsigned i;
 329         int ret = 0;
 330
 331         BUG_ON(BKEY_ALLOC_VAL_U64s_MAX > 8);
 332
 333         bch2_trans_init(&trans, c, 0, 0);
 334
 335         iter = bch2_trans_get_iter(&trans, BTREE_ID_ALLOC, POS_MIN,
 336                                    BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 337
 338         for_each_rw_member(ca, c, i) {
 339                 unsigned first_bucket;
 340
 341                 percpu_down_read(&c->mark_lock);
 342                 first_bucket = bucket_array(ca)->first_bucket;
 343                 percpu_up_read(&c->mark_lock);
 344
 345                 bch2_btree_iter_set_pos(iter, POS(i, first_bucket));
 346
 347                 while (1) {
 348                         ret = bch2_alloc_write_key(&trans, iter, flags);
 349                         if (ret < 0 || ret == ALLOC_END)
 350                                 break;
 351                         if (ret == ALLOC_WROTE)
 352                                 *wrote = true;
 353                         bch2_btree_iter_next_slot(iter);
 354                 }
 355
 356                 if (ret < 0) {
 357                         percpu_ref_put(&ca->io_ref);
 358                         break;
 359                 }
 360         }
 361
 362         bch2_trans_exit(&trans);
 363
 364         return ret < 0 ? ret : 0;
 365 }
 366
 367 int bch2_alloc_replay_key(struct bch_fs *c, struct bkey_i *k)
 368 {
 369         struct btree_trans trans;
 370         struct btree_iter *iter;
 371         int ret;
 372
 373         bch2_trans_init(&trans, c, 0, 0);
 374
 375         iter = bch2_trans_get_iter(&trans, BTREE_ID_ALLOC, k->k.p,
 376                                    BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 377
 378         ret = bch2_alloc_write_key(&trans, iter,
 379                                    BTREE_INSERT_NOFAIL|
 380                                    BTREE_INSERT_LAZY_RW|
 381                                    BTREE_INSERT_JOURNAL_REPLAY);
 382         bch2_trans_exit(&trans);
 383         return ret < 0 ? ret : 0;
 384 }
 385
 386 /* Bucket IO clocks: */
 387
 388 static void bch2_recalc_oldest_io(struct bch_fs *c, struct bch_dev *ca, int rw)
 389 {
 390         struct bucket_clock *clock = &c->bucket_clock[rw];
 391         struct bucket_array *buckets = bucket_array(ca);
 392         struct bucket *g;
 393         u16 max_last_io = 0;
 394         unsigned i;
 395
 396         lockdep_assert_held(&c->bucket_clock[rw].lock);
 397
 398         /* Recalculate max_last_io for this device: */
 399         for_each_bucket(g, buckets)
 400                 max_last_io = max(max_last_io, bucket_last_io(c, g, rw));
 401
 402         ca->max_last_bucket_io[rw] = max_last_io;
 403
 404         /* Recalculate global max_last_io: */
 405         max_last_io = 0;
 406
 407         for_each_member_device(ca, c, i)
 408                 max_last_io = max(max_last_io, ca->max_last_bucket_io[rw]);
 409
 410         clock->max_last_io = max_last_io;
 411 }
 412
 413 static void bch2_rescale_bucket_io_times(struct bch_fs *c, int rw)
 414 {
 415         struct bucket_clock *clock = &c->bucket_clock[rw];
 416         struct bucket_array *buckets;
 417         struct bch_dev *ca;
 418         struct bucket *g;
 419         unsigned i;
 420
 421         trace_rescale_prios(c);
 422
 423         for_each_member_device(ca, c, i) {
 424                 down_read(&ca->bucket_lock);
 425                 buckets = bucket_array(ca);
 426
 427                 for_each_bucket(g, buckets)
 428                         g->io_time[rw] = clock->hand -
 429                         bucket_last_io(c, g, rw) / 2;
 430
 431                 bch2_recalc_oldest_io(c, ca, rw);
 432
 433                 up_read(&ca->bucket_lock);
 434         }
 435 }
 436
 437 static inline u64 bucket_clock_freq(u64 capacity)
 438 {
 439         return max(capacity >> 10, 2028ULL);
 440 }
 441
 442 static void bch2_inc_clock_hand(struct io_timer *timer)
 443 {
 444         struct bucket_clock *clock = container_of(timer,
 445                                                 struct bucket_clock, rescale);
 446         struct bch_fs *c = container_of(clock,
 447                                         struct bch_fs, bucket_clock[clock->rw]);
 448         struct bch_dev *ca;
 449         u64 capacity;
 450         unsigned i;
 451
 452         mutex_lock(&clock->lock);
 453
 454         /* if clock cannot be advanced more, rescale prio */
 455         if (clock->max_last_io >= U16_MAX - 2)
 456                 bch2_rescale_bucket_io_times(c, clock->rw);
 457
 458         BUG_ON(clock->max_last_io >= U16_MAX - 2);
 459
 460         for_each_member_device(ca, c, i)
 461                 ca->max_last_bucket_io[clock->rw]++;
 462         clock->max_last_io++;
 463         clock->hand++;
 464
 465         mutex_unlock(&clock->lock);
 466
 467         capacity = READ_ONCE(c->capacity);
 468
 469         if (!capacity)
 470                 return;
 471
 472         /*
 473          * we only increment when 0.1% of the filesystem capacity has been read
 474          * or written too, this determines if it's time
 475          *
 476          * XXX: we shouldn't really be going off of the capacity of devices in
 477          * RW mode (that will be 0 when we're RO, yet we can still service
 478          * reads)
 479          */
 480         timer->expire += bucket_clock_freq(capacity);
 481
 482         bch2_io_timer_add(&c->io_clock[clock->rw], timer);
 483 }
 484
 485 static void bch2_bucket_clock_init(struct bch_fs *c, int rw)
 486 {
 487         struct bucket_clock *clock = &c->bucket_clock[rw];
 488
 489         clock->hand             = 1;
 490         clock->rw               = rw;
 491         clock->rescale.fn       = bch2_inc_clock_hand;
 492         clock->rescale.expire   = bucket_clock_freq(c->capacity);
 493         mutex_init(&clock->lock);
 494 }
 495
 496 /* Background allocator thread: */
 497
 498 /*
 499  * Scans for buckets to be invalidated, invalidates them, rewrites prios/gens
 500  * (marking them as invalidated on disk), then optionally issues discard
 501  * commands to the newly free buckets, then puts them on the various freelists.
 502  */
 503
 504 #define BUCKET_GC_GEN_MAX       96U
 505
 506 /**
 507  * wait_buckets_available - wait on reclaimable buckets
 508  *
 509  * If there aren't enough available buckets to fill up free_inc, wait until
 510  * there are.
 511  */
 512 static int wait_buckets_available(struct bch_fs *c, struct bch_dev *ca)
 513 {
 514         unsigned long gc_count = c->gc_count;
 515         int ret = 0;
 516
 517         ca->allocator_state = ALLOCATOR_BLOCKED;
 518         closure_wake_up(&c->freelist_wait);
 519
 520         while (1) {
 521                 set_current_state(TASK_INTERRUPTIBLE);
 522                 if (kthread_should_stop()) {
 523                         ret = 1;
 524                         break;
 525                 }
 526
 527                 if (gc_count != c->gc_count)
 528                         ca->inc_gen_really_needs_gc = 0;
 529
 530                 if ((ssize_t) (dev_buckets_available(c, ca) -
 531                                ca->inc_gen_really_needs_gc) >=
 532                     (ssize_t) fifo_free(&ca->free_inc))
 533                         break;
 534
 535                 up_read(&c->gc_lock);
 536                 schedule();
 537                 try_to_freeze();
 538                 down_read(&c->gc_lock);
 539         }
 540
 541         __set_current_state(TASK_RUNNING);
 542         ca->allocator_state = ALLOCATOR_RUNNING;
 543         closure_wake_up(&c->freelist_wait);
 544
 545         return ret;
 546 }
 547
 548 static bool bch2_can_invalidate_bucket(struct bch_dev *ca,
 549                                        size_t bucket,
 550                                        struct bucket_mark mark)
 551 {
 552         u8 gc_gen;
 553
 554         if (!is_available_bucket(mark))
 555                 return false;
 556
 557         if (ca->buckets_nouse &&
 558             test_bit(bucket, ca->buckets_nouse))
 559                 return false;
 560
 561         gc_gen = bucket_gc_gen(ca, bucket);
 562
 563         if (gc_gen >= BUCKET_GC_GEN_MAX / 2)
 564                 ca->inc_gen_needs_gc++;
 565
 566         if (gc_gen >= BUCKET_GC_GEN_MAX)
 567                 ca->inc_gen_really_needs_gc++;
 568
 569         return gc_gen < BUCKET_GC_GEN_MAX;
 570 }
 571
 572 /*
 573  * Determines what order we're going to reuse buckets, smallest bucket_key()
 574  * first.
 575  *
 576  *
 577  * - We take into account the read prio of the bucket, which gives us an
 578  *   indication of how hot the data is -- we scale the prio so that the prio
 579  *   farthest from the clock is worth 1/8th of the closest.
 580  *
 581  * - The number of sectors of cached data in the bucket, which gives us an
 582  *   indication of the cost in cache misses this eviction will cause.
 583  *
 584  * - If hotness * sectors used compares equal, we pick the bucket with the
 585  *   smallest bucket_gc_gen() - since incrementing the same bucket's generation
 586  *   number repeatedly forces us to run mark and sweep gc to avoid generation
 587  *   number wraparound.
 588  */
 589
 590 static unsigned long bucket_sort_key(struct bch_fs *c, struct bch_dev *ca,
 591                                      size_t b, struct bucket_mark m)
 592 {
 593         unsigned last_io = bucket_last_io(c, bucket(ca, b), READ);
 594         unsigned max_last_io = ca->max_last_bucket_io[READ];
 595
 596         /*
 597          * Time since last read, scaled to [0, 8) where larger value indicates
 598          * more recently read data:
 599          */
 600         unsigned long hotness = (max_last_io - last_io) * 7 / max_last_io;
 601
 602         /* How much we want to keep the data in this bucket: */
 603         unsigned long data_wantness =
 604                 (hotness + 1) * bucket_sectors_used(m);
 605
 606         unsigned long needs_journal_commit =
 607                 bucket_needs_journal_commit(m, c->journal.last_seq_ondisk);
 608
 609         return  (data_wantness << 9) |
 610                 (needs_journal_commit << 8) |
 611                 (bucket_gc_gen(ca, b) / 16);
 612 }
 613
 614 static inline int bucket_alloc_cmp(alloc_heap *h,
 615                                    struct alloc_heap_entry l,
 616                                    struct alloc_heap_entry r)
 617 {
 618         return  cmp_int(l.key, r.key) ?:
 619                 cmp_int(r.nr, l.nr) ?:
 620                 cmp_int(l.bucket, r.bucket);
 621 }
 622
 623 static inline int bucket_idx_cmp(const void *_l, const void *_r)
 624 {
 625         const struct alloc_heap_entry *l = _l, *r = _r;
 626
 627         return cmp_int(l->bucket, r->bucket);
 628 }
 629
 630 static void find_reclaimable_buckets_lru(struct bch_fs *c, struct bch_dev *ca)
 631 {
 632         struct bucket_array *buckets;
 633         struct alloc_heap_entry e = { 0 };
 634         size_t b, i, nr = 0;
 635
 636         ca->alloc_heap.used = 0;
 637
 638         mutex_lock(&c->bucket_clock[READ].lock);
 639         down_read(&ca->bucket_lock);
 640
 641         buckets = bucket_array(ca);
 642
 643         bch2_recalc_oldest_io(c, ca, READ);
 644
 645         /*
 646          * Find buckets with lowest read priority, by building a maxheap sorted
 647          * by read priority and repeatedly replacing the maximum element until
 648          * all buckets have been visited.
 649          */
 650         for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++) {
 651                 struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
 652                 unsigned long key = bucket_sort_key(c, ca, b, m);
 653
 654                 if (!bch2_can_invalidate_bucket(ca, b, m))
 655                         continue;
 656
 657                 if (e.nr && e.bucket + e.nr == b && e.key == key) {
 658                         e.nr++;
 659                 } else {
 660                         if (e.nr)
 661                                 heap_add_or_replace(&ca->alloc_heap, e,
 662                                         -bucket_alloc_cmp, NULL);
 663
 664                         e = (struct alloc_heap_entry) {
 665                                 .bucket = b,
 666                                 .nr     = 1,
 667                                 .key    = key,
 668                         };
 669                 }
 670
 671                 cond_resched();
 672         }
 673
 674         if (e.nr)
 675                 heap_add_or_replace(&ca->alloc_heap, e,
 676                                 -bucket_alloc_cmp, NULL);
 677
 678         for (i = 0; i < ca->alloc_heap.used; i++)
 679                 nr += ca->alloc_heap.data[i].nr;
 680
 681         while (nr - ca->alloc_heap.data[0].nr >= ALLOC_SCAN_BATCH(ca)) {
 682                 nr -= ca->alloc_heap.data[0].nr;
 683                 heap_pop(&ca->alloc_heap, e, -bucket_alloc_cmp, NULL);
 684         }
 685
 686         up_read(&ca->bucket_lock);
 687         mutex_unlock(&c->bucket_clock[READ].lock);
 688 }
 689
 690 static void find_reclaimable_buckets_fifo(struct bch_fs *c, struct bch_dev *ca)
 691 {
 692         struct bucket_array *buckets = bucket_array(ca);
 693         struct bucket_mark m;
 694         size_t b, start;
 695
 696         if (ca->fifo_last_bucket <  ca->mi.first_bucket ||
 697             ca->fifo_last_bucket >= ca->mi.nbuckets)
 698                 ca->fifo_last_bucket = ca->mi.first_bucket;
 699
 700         start = ca->fifo_last_bucket;
 701
 702         do {
 703                 ca->fifo_last_bucket++;
 704                 if (ca->fifo_last_bucket == ca->mi.nbuckets)
 705                         ca->fifo_last_bucket = ca->mi.first_bucket;
 706
 707                 b = ca->fifo_last_bucket;
 708                 m = READ_ONCE(buckets->b[b].mark);
 709
 710                 if (bch2_can_invalidate_bucket(ca, b, m)) {
 711                         struct alloc_heap_entry e = { .bucket = b, .nr = 1, };
 712
 713                         heap_add(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
 714                         if (heap_full(&ca->alloc_heap))
 715                                 break;
 716                 }
 717
 718                 cond_resched();
 719         } while (ca->fifo_last_bucket != start);
 720 }
 721
 722 static void find_reclaimable_buckets_random(struct bch_fs *c, struct bch_dev *ca)
 723 {
 724         struct bucket_array *buckets = bucket_array(ca);
 725         struct bucket_mark m;
 726         size_t checked, i;
 727
 728         for (checked = 0;
 729              checked < ca->mi.nbuckets / 2;
 730              checked++) {
 731                 size_t b = bch2_rand_range(ca->mi.nbuckets -
 732                                            ca->mi.first_bucket) +
 733                         ca->mi.first_bucket;
 734
 735                 m = READ_ONCE(buckets->b[b].mark);
 736
 737                 if (bch2_can_invalidate_bucket(ca, b, m)) {
 738                         struct alloc_heap_entry e = { .bucket = b, .nr = 1, };
 739
 740                         heap_add(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
 741                         if (heap_full(&ca->alloc_heap))
 742                                 break;
 743                 }
 744
 745                 cond_resched();
 746         }
 747
 748         sort(ca->alloc_heap.data,
 749              ca->alloc_heap.used,
 750              sizeof(ca->alloc_heap.data[0]),
 751              bucket_idx_cmp, NULL);
 752
 753         /* remove duplicates: */
 754         for (i = 0; i + 1 < ca->alloc_heap.used; i++)
 755                 if (ca->alloc_heap.data[i].bucket ==
 756                     ca->alloc_heap.data[i + 1].bucket)
 757                         ca->alloc_heap.data[i].nr = 0;
 758 }
 759
 760 static size_t find_reclaimable_buckets(struct bch_fs *c, struct bch_dev *ca)
 761 {
 762         size_t i, nr = 0;
 763
 764         ca->inc_gen_needs_gc                    = 0;
 765
 766         switch (ca->mi.replacement) {
 767         case CACHE_REPLACEMENT_LRU:
 768                 find_reclaimable_buckets_lru(c, ca);
 769                 break;
 770         case CACHE_REPLACEMENT_FIFO:
 771                 find_reclaimable_buckets_fifo(c, ca);
 772                 break;
 773         case CACHE_REPLACEMENT_RANDOM:
 774                 find_reclaimable_buckets_random(c, ca);
 775                 break;
 776         }
 777
 778         heap_resort(&ca->alloc_heap, bucket_alloc_cmp, NULL);
 779
 780         for (i = 0; i < ca->alloc_heap.used; i++)
 781                 nr += ca->alloc_heap.data[i].nr;
 782
 783         return nr;
 784 }
 785
 786 static inline long next_alloc_bucket(struct bch_dev *ca)
 787 {
 788         struct alloc_heap_entry e, *top = ca->alloc_heap.data;
 789
 790         while (ca->alloc_heap.used) {
 791                 if (top->nr) {
 792                         size_t b = top->bucket;
 793
 794                         top->bucket++;
 795                         top->nr--;
 796                         return b;
 797                 }
 798
 799                 heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
 800         }
 801
 802         return -1;
 803 }
 804
 805 /*
 806  * returns sequence number of most recent journal entry that updated this
 807  * bucket:
 808  */
 809 static u64 bucket_journal_seq(struct bch_fs *c, struct bucket_mark m)
 810 {
 811         if (m.journal_seq_valid) {
 812                 u64 journal_seq = atomic64_read(&c->journal.seq);
 813                 u64 bucket_seq  = journal_seq;
 814
 815                 bucket_seq &= ~((u64) U16_MAX);
 816                 bucket_seq |= m.journal_seq;
 817
 818                 if (bucket_seq > journal_seq)
 819                         bucket_seq -= 1 << 16;
 820
 821                 return bucket_seq;
 822         } else {
 823                 return 0;
 824         }
 825 }
 826
 827 static int bch2_invalidate_one_bucket2(struct btree_trans *trans,
 828                                        struct bch_dev *ca,
 829                                        struct btree_iter *iter,
 830                                        u64 *journal_seq, unsigned flags)
 831 {
 832 #if 0
 833         __BKEY_PADDED(k, BKEY_ALLOC_VAL_U64s_MAX) alloc_key;
 834 #else
 835         /* hack: */
 836         __BKEY_PADDED(k, 8) alloc_key;
 837 #endif
 838         struct bch_fs *c = trans->c;
 839         struct bkey_i_alloc *a;
 840         struct bkey_alloc_unpacked u;
 841         struct bucket *g;
 842         struct bucket_mark m;
 843         struct bkey_s_c k;
 844         bool invalidating_cached_data;
 845         size_t b;
 846         int ret = 0;
 847
 848         BUG_ON(!ca->alloc_heap.used ||
 849                !ca->alloc_heap.data[0].nr);
 850         b = ca->alloc_heap.data[0].bucket;
 851
 852         /* first, put on free_inc and mark as owned by allocator: */
 853         percpu_down_read(&c->mark_lock);
 854         spin_lock(&c->freelist_lock);
 855
 856         verify_not_on_freelist(c, ca, b);
 857
 858         BUG_ON(!fifo_push(&ca->free_inc, b));
 859
 860         g = bucket(ca, b);
 861         m = READ_ONCE(g->mark);
 862
 863         bch2_mark_alloc_bucket(c, ca, b, true, gc_pos_alloc(c, NULL), 0);
 864
 865         spin_unlock(&c->freelist_lock);
 866         percpu_up_read(&c->mark_lock);
 867
 868         invalidating_cached_data = m.cached_sectors != 0;
 869         if (!invalidating_cached_data)
 870                 goto out;
 871
 872         /*
 873          * If the read-only path is trying to shut down, we can't be generating
 874          * new btree updates:
 875          */
 876         if (test_bit(BCH_FS_ALLOCATOR_STOPPING, &c->flags)) {
 877                 ret = 1;
 878                 goto out;
 879         }
 880
 881         BUG_ON(BKEY_ALLOC_VAL_U64s_MAX > 8);
 882
 883         bch2_btree_iter_set_pos(iter, POS(ca->dev_idx, b));
 884 retry:
 885         k = bch2_btree_iter_peek_slot(iter);
 886         ret = bkey_err(k);
 887         if (ret)
 888                 return ret;
 889
 890         /*
 891          * The allocator has to start before journal replay is finished - thus,
 892          * we have to trust the in memory bucket @m, not the version in the
 893          * btree:
 894          */
 895         percpu_down_read(&c->mark_lock);
 896         g = bucket(ca, b);
 897         m = READ_ONCE(g->mark);
 898         u = alloc_mem_to_key(g, m);
 899         percpu_up_read(&c->mark_lock);
 900
 901         invalidating_cached_data = m.cached_sectors != 0;
 902
 903         u.gen++;
 904         u.data_type     = 0;
 905         u.dirty_sectors = 0;
 906         u.cached_sectors = 0;
 907         u.read_time     = c->bucket_clock[READ].hand;
 908         u.write_time    = c->bucket_clock[WRITE].hand;
 909
 910         a = bkey_alloc_init(&alloc_key.k);
 911         a->k.p = iter->pos;
 912         bch2_alloc_pack(a, u);
 913
 914         bch2_trans_update(trans, iter, &a->k_i,
 915                           BTREE_TRIGGER_BUCKET_INVALIDATE);
 916
 917         /*
 918          * XXX:
 919          * when using deferred btree updates, we have journal reclaim doing
 920          * btree updates and thus requiring the allocator to make forward
 921          * progress, and here the allocator is requiring space in the journal -
 922          * so we need a journal pre-reservation:
 923          */
 924         ret = bch2_trans_commit(trans, NULL,
 925                                 invalidating_cached_data ? journal_seq : NULL,
 926                                 BTREE_INSERT_NOUNLOCK|
 927                                 BTREE_INSERT_NOCHECK_RW|
 928                                 BTREE_INSERT_NOFAIL|
 929                                 BTREE_INSERT_USE_RESERVE|
 930                                 BTREE_INSERT_USE_ALLOC_RESERVE|
 931                                 flags);
 932         if (ret == -EINTR)
 933                 goto retry;
 934 out:
 935         if (!ret) {
 936                 /* remove from alloc_heap: */
 937                 struct alloc_heap_entry e, *top = ca->alloc_heap.data;
 938
 939                 top->bucket++;
 940                 top->nr--;
 941
 942                 if (!top->nr)
 943                         heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
 944
 945                 /*
 946                  * Make sure we flush the last journal entry that updated this
 947                  * bucket (i.e. deleting the last reference) before writing to
 948                  * this bucket again:
 949                  */
 950                 *journal_seq = max(*journal_seq, bucket_journal_seq(c, m));
 951         } else {
 952                 size_t b2;
 953
 954                 /* remove from free_inc: */
 955                 percpu_down_read(&c->mark_lock);
 956                 spin_lock(&c->freelist_lock);
 957
 958                 bch2_mark_alloc_bucket(c, ca, b, false,
 959                                        gc_pos_alloc(c, NULL), 0);
 960
 961                 BUG_ON(!fifo_pop_back(&ca->free_inc, b2));
 962                 BUG_ON(b != b2);
 963
 964                 spin_unlock(&c->freelist_lock);
 965                 percpu_up_read(&c->mark_lock);
 966         }
 967
 968         return ret < 0 ? ret : 0;
 969 }
 970
 971 static bool bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca,
 972                                        size_t bucket, u64 *flush_seq)
 973 {
 974         struct bucket_mark m;
 975
 976         percpu_down_read(&c->mark_lock);
 977         spin_lock(&c->freelist_lock);
 978
 979         bch2_invalidate_bucket(c, ca, bucket, &m);
 980
 981         verify_not_on_freelist(c, ca, bucket);
 982         BUG_ON(!fifo_push(&ca->free_inc, bucket));
 983
 984         spin_unlock(&c->freelist_lock);
 985
 986         bucket_io_clock_reset(c, ca, bucket, READ);
 987         bucket_io_clock_reset(c, ca, bucket, WRITE);
 988
 989         percpu_up_read(&c->mark_lock);
 990
 991         *flush_seq = max(*flush_seq, bucket_journal_seq(c, m));
 992
 993         return m.cached_sectors != 0;
 994 }
 995
 996 /*
 997  * Pull buckets off ca->alloc_heap, invalidate them, move them to ca->free_inc:
 998  */
 999 static int bch2_invalidate_buckets(struct bch_fs *c, struct bch_dev *ca)
1000 {
1001         struct btree_trans trans;
1002         struct btree_iter *iter;
1003         u64 journal_seq = 0;
1004         int ret = 0;
1005
1006         bch2_trans_init(&trans, c, 0, 0);
1007
1008         iter = bch2_trans_get_iter(&trans, BTREE_ID_ALLOC,
1009                                    POS(ca->dev_idx, 0),
1010                                    BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
1011
1012         /* Only use nowait if we've already invalidated at least one bucket: */
1013         while (!ret &&
1014                !fifo_full(&ca->free_inc) &&
1015                ca->alloc_heap.used)
1016                 ret = bch2_invalidate_one_bucket2(&trans, ca, iter, &journal_seq,
1017                                 BTREE_INSERT_GC_LOCK_HELD|
1018                                 (!fifo_empty(&ca->free_inc)
1019                                  ? BTREE_INSERT_NOWAIT : 0));
1020
1021         bch2_trans_exit(&trans);
1022
1023         /* If we used NOWAIT, don't return the error: */
1024         if (!fifo_empty(&ca->free_inc))
1025                 ret = 0;
1026         if (ret) {
1027                 bch_err(ca, "error invalidating buckets: %i", ret);
1028                 return ret;
1029         }
1030
1031         if (journal_seq)
1032                 ret = bch2_journal_flush_seq(&c->journal, journal_seq);
1033         if (ret) {
1034                 bch_err(ca, "journal error: %i", ret);
1035                 return ret;
1036         }
1037
1038         return 0;
1039 }
1040
1041 static int push_invalidated_bucket(struct bch_fs *c, struct bch_dev *ca, size_t bucket)
1042 {
1043         unsigned i;
1044         int ret = 0;
1045
1046         while (1) {
1047                 set_current_state(TASK_INTERRUPTIBLE);
1048
1049                 spin_lock(&c->freelist_lock);
1050                 for (i = 0; i < RESERVE_NR; i++) {
1051
1052                         /*
1053                          * Don't strand buckets on the copygc freelist until
1054                          * after recovery is finished:
1055                          */
1056                         if (!test_bit(BCH_FS_STARTED, &c->flags) &&
1057                             i == RESERVE_MOVINGGC)
1058                                 continue;
1059
1060                         if (fifo_push(&ca->free[i], bucket)) {
1061                                 fifo_pop(&ca->free_inc, bucket);
1062
1063                                 closure_wake_up(&c->freelist_wait);
1064                                 ca->allocator_state = ALLOCATOR_RUNNING;
1065
1066                                 spin_unlock(&c->freelist_lock);
1067                                 goto out;
1068                         }
1069                 }
1070
1071                 if (ca->allocator_state != ALLOCATOR_BLOCKED_FULL) {
1072                         ca->allocator_state = ALLOCATOR_BLOCKED_FULL;
1073                         closure_wake_up(&c->freelist_wait);
1074                 }
1075
1076                 spin_unlock(&c->freelist_lock);
1077
1078                 if ((current->flags & PF_KTHREAD) &&
1079                     kthread_should_stop()) {
1080                         ret = 1;
1081                         break;
1082                 }
1083
1084                 schedule();
1085                 try_to_freeze();
1086         }
1087 out:
1088         __set_current_state(TASK_RUNNING);
1089         return ret;
1090 }
1091
1092 /*
1093  * Pulls buckets off free_inc, discards them (if enabled), then adds them to
1094  * freelists, waiting until there's room if necessary:
1095  */
1096 static int discard_invalidated_buckets(struct bch_fs *c, struct bch_dev *ca)
1097 {
1098         while (!fifo_empty(&ca->free_inc)) {
1099                 size_t bucket = fifo_peek(&ca->free_inc);
1100
1101                 if (ca->mi.discard &&
1102                     blk_queue_discard(bdev_get_queue(ca->disk_sb.bdev)))
1103                         blkdev_issue_discard(ca->disk_sb.bdev,
1104                                              bucket_to_sector(ca, bucket),
1105                                              ca->mi.bucket_size, GFP_NOIO, 0);
1106
1107                 if (push_invalidated_bucket(c, ca, bucket))
1108                         return 1;
1109         }
1110
1111         return 0;
1112 }
1113
1114 /**
1115  * bch_allocator_thread - move buckets from free_inc to reserves
1116  *
1117  * The free_inc FIFO is populated by find_reclaimable_buckets(), and
1118  * the reserves are depleted by bucket allocation. When we run out
1119  * of free_inc, try to invalidate some buckets and write out
1120  * prios and gens.
1121  */
1122 static int bch2_allocator_thread(void *arg)
1123 {
1124         struct bch_dev *ca = arg;
1125         struct bch_fs *c = ca->fs;
1126         size_t nr;
1127         int ret;
1128
1129         set_freezable();
1130         ca->allocator_state = ALLOCATOR_RUNNING;
1131
1132         while (1) {
1133                 cond_resched();
1134
1135                 pr_debug("discarding %zu invalidated buckets",
1136                          fifo_used(&ca->free_inc));
1137
1138                 ret = discard_invalidated_buckets(c, ca);
1139                 if (ret)
1140                         goto stop;
1141
1142                 down_read(&c->gc_lock);
1143
1144                 ret = bch2_invalidate_buckets(c, ca);
1145                 if (ret) {
1146                         up_read(&c->gc_lock);
1147                         goto stop;
1148                 }
1149
1150                 if (!fifo_empty(&ca->free_inc)) {
1151                         up_read(&c->gc_lock);
1152                         continue;
1153                 }
1154
1155                 pr_debug("free_inc now empty");
1156
1157                 do {
1158                         /*
1159                          * Find some buckets that we can invalidate, either
1160                          * they're completely unused, or only contain clean data
1161                          * that's been written back to the backing device or
1162                          * another cache tier
1163                          */
1164
1165                         pr_debug("scanning for reclaimable buckets");
1166
1167                         nr = find_reclaimable_buckets(c, ca);
1168
1169                         pr_debug("found %zu buckets", nr);
1170
1171                         trace_alloc_batch(ca, nr, ca->alloc_heap.size);
1172
1173                         if ((ca->inc_gen_needs_gc >= ALLOC_SCAN_BATCH(ca) ||
1174                              ca->inc_gen_really_needs_gc) &&
1175                             c->gc_thread) {
1176                                 atomic_inc(&c->kick_gc);
1177                                 wake_up_process(c->gc_thread);
1178                         }
1179
1180                         /*
1181                          * If we found any buckets, we have to invalidate them
1182                          * before we scan for more - but if we didn't find very
1183                          * many we may want to wait on more buckets being
1184                          * available so we don't spin:
1185                          */
1186                         if (!nr ||
1187                             (nr < ALLOC_SCAN_BATCH(ca) &&
1188                              !fifo_empty(&ca->free[RESERVE_NONE]))) {
1189                                 ret = wait_buckets_available(c, ca);
1190                                 if (ret) {
1191                                         up_read(&c->gc_lock);
1192                                         goto stop;
1193                                 }
1194                         }
1195                 } while (!nr);
1196
1197                 up_read(&c->gc_lock);
1198
1199                 pr_debug("%zu buckets to invalidate", nr);
1200
1201                 /*
1202                  * alloc_heap is now full of newly-invalidated buckets: next,
1203                  * write out the new bucket gens:
1204                  */
1205         }
1206
1207 stop:
1208         pr_debug("alloc thread stopping (ret %i)", ret);
1209         ca->allocator_state = ALLOCATOR_STOPPED;
1210         closure_wake_up(&c->freelist_wait);
1211         return 0;
1212 }
1213
1214 /* Startup/shutdown (ro/rw): */
1215
1216 void bch2_recalc_capacity(struct bch_fs *c)
1217 {
1218         struct bch_dev *ca;
1219         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
1220         unsigned bucket_size_max = 0;
1221         unsigned long ra_pages = 0;
1222         unsigned i, j;
1223
1224         lockdep_assert_held(&c->state_lock);
1225
1226         for_each_online_member(ca, c, i) {
1227                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_bdi;
1228
1229                 ra_pages += bdi->ra_pages;
1230         }
1231
1232         bch2_set_ra_pages(c, ra_pages);
1233
1234         for_each_rw_member(ca, c, i) {
1235                 u64 dev_reserve = 0;
1236
1237                 /*
1238                  * We need to reserve buckets (from the number
1239                  * of currently available buckets) against
1240                  * foreground writes so that mainly copygc can
1241                  * make forward progress.
1242                  *
1243                  * We need enough to refill the various reserves
1244                  * from scratch - copygc will use its entire
1245                  * reserve all at once, then run against when
1246                  * its reserve is refilled (from the formerly
1247                  * available buckets).
1248                  *
1249                  * This reserve is just used when considering if
1250                  * allocations for foreground writes must wait -
1251                  * not -ENOSPC calculations.
1252                  */
1253                 for (j = 0; j < RESERVE_NONE; j++)
1254                         dev_reserve += ca->free[j].size;
1255
1256                 dev_reserve += 1;       /* btree write point */
1257                 dev_reserve += 1;       /* copygc write point */
1258                 dev_reserve += 1;       /* rebalance write point */
1259
1260                 dev_reserve *= ca->mi.bucket_size;
1261
1262                 ca->copygc_threshold = dev_reserve;
1263
1264                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
1265                                              ca->mi.first_bucket);
1266
1267                 reserved_sectors += dev_reserve * 2;
1268
1269                 bucket_size_max = max_t(unsigned, bucket_size_max,
1270                                         ca->mi.bucket_size);
1271         }
1272
1273         gc_reserve = c->opts.gc_reserve_bytes
1274                 ? c->opts.gc_reserve_bytes >> 9
1275                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
1276
1277         reserved_sectors = max(gc_reserve, reserved_sectors);
1278
1279         reserved_sectors = min(reserved_sectors, capacity);
1280
1281         c->capacity = capacity - reserved_sectors;
1282
1283         c->bucket_size_max = bucket_size_max;
1284
1285         if (c->capacity) {
1286                 bch2_io_timer_add(&c->io_clock[READ],
1287                                  &c->bucket_clock[READ].rescale);
1288                 bch2_io_timer_add(&c->io_clock[WRITE],
1289                                  &c->bucket_clock[WRITE].rescale);
1290         } else {
1291                 bch2_io_timer_del(&c->io_clock[READ],
1292                                  &c->bucket_clock[READ].rescale);
1293                 bch2_io_timer_del(&c->io_clock[WRITE],
1294                                  &c->bucket_clock[WRITE].rescale);
1295         }
1296
1297         /* Wake up case someone was waiting for buckets */
1298         closure_wake_up(&c->freelist_wait);
1299 }
1300
1301 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
1302 {
1303         struct open_bucket *ob;
1304         bool ret = false;
1305
1306         for (ob = c->open_buckets;
1307              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1308              ob++) {
1309                 spin_lock(&ob->lock);
1310                 if (ob->valid && !ob->on_partial_list &&
1311                     ob->ptr.dev == ca->dev_idx)
1312                         ret = true;
1313                 spin_unlock(&ob->lock);
1314         }
1315
1316         return ret;
1317 }
1318
1319 /* device goes ro: */
1320 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
1321 {
1322         unsigned i;
1323
1324         BUG_ON(ca->alloc_thread);
1325
1326         /* First, remove device from allocation groups: */
1327
1328         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1329                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
1330
1331         /*
1332          * Capacity is calculated based off of devices in allocation groups:
1333          */
1334         bch2_recalc_capacity(c);
1335
1336         /* Next, close write points that point to this device... */
1337         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1338                 bch2_writepoint_stop(c, ca, &c->write_points[i]);
1339
1340         bch2_writepoint_stop(c, ca, &ca->copygc_write_point);
1341         bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
1342         bch2_writepoint_stop(c, ca, &c->btree_write_point);
1343
1344         mutex_lock(&c->btree_reserve_cache_lock);
1345         while (c->btree_reserve_cache_nr) {
1346                 struct btree_alloc *a =
1347                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1348
1349                 bch2_open_buckets_put(c, &a->ob);
1350         }
1351         mutex_unlock(&c->btree_reserve_cache_lock);
1352
1353         while (1) {
1354                 struct open_bucket *ob;
1355
1356                 spin_lock(&c->freelist_lock);
1357                 if (!ca->open_buckets_partial_nr) {
1358                         spin_unlock(&c->freelist_lock);
1359                         break;
1360                 }
1361                 ob = c->open_buckets +
1362                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
1363                 ob->on_partial_list = false;
1364                 spin_unlock(&c->freelist_lock);
1365
1366                 bch2_open_bucket_put(c, ob);
1367         }
1368
1369         bch2_ec_stop_dev(c, ca);
1370
1371         /*
1372          * Wake up threads that were blocked on allocation, so they can notice
1373          * the device can no longer be removed and the capacity has changed:
1374          */
1375         closure_wake_up(&c->freelist_wait);
1376
1377         /*
1378          * journal_res_get() can block waiting for free space in the journal -
1379          * it needs to notice there may not be devices to allocate from anymore:
1380          */
1381         wake_up(&c->journal.wait);
1382
1383         /* Now wait for any in flight writes: */
1384
1385         closure_wait_event(&c->open_buckets_wait,
1386                            !bch2_dev_has_open_write_point(c, ca));
1387 }
1388
1389 /* device goes rw: */
1390 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
1391 {
1392         unsigned i;
1393
1394         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1395                 if (ca->mi.data_allowed & (1 << i))
1396                         set_bit(ca->dev_idx, c->rw_devs[i].d);
1397 }
1398
1399 void bch2_dev_allocator_quiesce(struct bch_fs *c, struct bch_dev *ca)
1400 {
1401         if (ca->alloc_thread)
1402                 closure_wait_event(&c->freelist_wait,
1403                                    ca->allocator_state != ALLOCATOR_RUNNING);
1404 }
1405
1406 /* stop allocator thread: */
1407 void bch2_dev_allocator_stop(struct bch_dev *ca)
1408 {
1409         struct task_struct *p;
1410
1411         p = rcu_dereference_protected(ca->alloc_thread, 1);
1412         ca->alloc_thread = NULL;
1413
1414         /*
1415          * We need an rcu barrier between setting ca->alloc_thread = NULL and
1416          * the thread shutting down to avoid bch2_wake_allocator() racing:
1417          *
1418          * XXX: it would be better to have the rcu barrier be asynchronous
1419          * instead of blocking us here
1420          */
1421         synchronize_rcu();
1422
1423         if (p) {
1424                 kthread_stop(p);
1425                 put_task_struct(p);
1426         }
1427 }
1428
1429 /* start allocator thread: */
1430 int bch2_dev_allocator_start(struct bch_dev *ca)
1431 {
1432         struct task_struct *p;
1433
1434         /*
1435          * allocator thread already started?
1436          */
1437         if (ca->alloc_thread)
1438                 return 0;
1439
1440         p = kthread_create(bch2_allocator_thread, ca,
1441                            "bch_alloc[%s]", ca->name);
1442         if (IS_ERR(p))
1443                 return PTR_ERR(p);
1444
1445         get_task_struct(p);
1446         rcu_assign_pointer(ca->alloc_thread, p);
1447         wake_up_process(p);
1448         return 0;
1449 }
1450
1451 static bool flush_held_btree_writes(struct bch_fs *c)
1452 {
1453         struct bucket_table *tbl;
1454         struct rhash_head *pos;
1455         struct btree *b;
1456         bool nodes_unwritten;
1457         size_t i;
1458 again:
1459         cond_resched();
1460         nodes_unwritten = false;
1461
1462         if (bch2_journal_error(&c->journal))
1463                 return true;
1464
1465         rcu_read_lock();
1466         for_each_cached_btree(b, c, tbl, i, pos)
1467                 if (btree_node_need_write(b)) {
1468                         if (btree_node_may_write(b)) {
1469                                 rcu_read_unlock();
1470                                 btree_node_lock_type(c, b, SIX_LOCK_read);
1471                                 bch2_btree_node_write(c, b, SIX_LOCK_read);
1472                                 six_unlock_read(&b->lock);
1473                                 goto again;
1474                         } else {
1475                                 nodes_unwritten = true;
1476                         }
1477                 }
1478         rcu_read_unlock();
1479
1480         return !nodes_unwritten &&
1481                 !bch2_btree_interior_updates_nr_pending(c);
1482 }
1483
1484 static void allocator_start_issue_discards(struct bch_fs *c)
1485 {
1486         struct bch_dev *ca;
1487         unsigned dev_iter;
1488         size_t bu;
1489
1490         for_each_rw_member(ca, c, dev_iter)
1491                 while (fifo_pop(&ca->free_inc, bu))
1492                         blkdev_issue_discard(ca->disk_sb.bdev,
1493                                              bucket_to_sector(ca, bu),
1494                                              ca->mi.bucket_size, GFP_NOIO, 0);
1495 }
1496
1497 static int resize_free_inc(struct bch_dev *ca)
1498 {
1499         alloc_fifo free_inc;
1500
1501         if (!fifo_full(&ca->free_inc))
1502                 return 0;
1503
1504         if (!init_fifo(&free_inc,
1505                        ca->free_inc.size * 2,
1506                        GFP_KERNEL))
1507                 return -ENOMEM;
1508
1509         fifo_move(&free_inc, &ca->free_inc);
1510         swap(free_inc, ca->free_inc);
1511         free_fifo(&free_inc);
1512         return 0;
1513 }
1514
1515 static bool bch2_fs_allocator_start_fast(struct bch_fs *c)
1516 {
1517         struct bch_dev *ca;
1518         unsigned dev_iter;
1519         bool ret = true;
1520
1521         if (test_alloc_startup(c))
1522                 return false;
1523
1524         down_read(&c->gc_lock);
1525
1526         /* Scan for buckets that are already invalidated: */
1527         for_each_rw_member(ca, c, dev_iter) {
1528                 struct bucket_array *buckets;
1529                 struct bucket_mark m;
1530                 long bu;
1531
1532                 down_read(&ca->bucket_lock);
1533                 buckets = bucket_array(ca);
1534
1535                 for (bu = buckets->first_bucket;
1536                      bu < buckets->nbuckets; bu++) {
1537                         m = READ_ONCE(buckets->b[bu].mark);
1538
1539                         if (!buckets->b[bu].gen_valid ||
1540                             !is_available_bucket(m) ||
1541                             m.cached_sectors ||
1542                             (ca->buckets_nouse &&
1543                              test_bit(bu, ca->buckets_nouse)))
1544                                 continue;
1545
1546                         percpu_down_read(&c->mark_lock);
1547                         bch2_mark_alloc_bucket(c, ca, bu, true,
1548                                         gc_pos_alloc(c, NULL), 0);
1549                         percpu_up_read(&c->mark_lock);
1550
1551                         fifo_push(&ca->free_inc, bu);
1552
1553                         discard_invalidated_buckets(c, ca);
1554
1555                         if (fifo_full(&ca->free[RESERVE_BTREE]))
1556                                 break;
1557                 }
1558                 up_read(&ca->bucket_lock);
1559         }
1560
1561         up_read(&c->gc_lock);
1562
1563         /* did we find enough buckets? */
1564         for_each_rw_member(ca, c, dev_iter)
1565                 if (!fifo_full(&ca->free[RESERVE_BTREE]))
1566                         ret = false;
1567
1568         return ret;
1569 }
1570
1571 int bch2_fs_allocator_start(struct bch_fs *c)
1572 {
1573         struct bch_dev *ca;
1574         unsigned dev_iter;
1575         u64 journal_seq = 0;
1576         bool wrote;
1577         long bu;
1578         int ret = 0;
1579
1580         if (!test_alloc_startup(c) &&
1581             bch2_fs_allocator_start_fast(c))
1582                 return 0;
1583
1584         pr_debug("not enough empty buckets; scanning for reclaimable buckets");
1585
1586         /*
1587          * We're moving buckets to freelists _before_ they've been marked as
1588          * invalidated on disk - we have to so that we can allocate new btree
1589          * nodes to mark them as invalidated on disk.
1590          *
1591          * However, we can't _write_ to any of these buckets yet - they might
1592          * have cached data in them, which is live until they're marked as
1593          * invalidated on disk:
1594          */
1595         set_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags);
1596
1597         down_read(&c->gc_lock);
1598         do {
1599                 wrote = false;
1600
1601                 for_each_rw_member(ca, c, dev_iter) {
1602                         find_reclaimable_buckets(c, ca);
1603
1604                         while (!fifo_full(&ca->free[RESERVE_BTREE]) &&
1605                                (bu = next_alloc_bucket(ca)) >= 0) {
1606                                 ret = resize_free_inc(ca);
1607                                 if (ret) {
1608                                         percpu_ref_put(&ca->io_ref);
1609                                         up_read(&c->gc_lock);
1610                                         goto err;
1611                                 }
1612
1613                                 bch2_invalidate_one_bucket(c, ca, bu,
1614                                                            &journal_seq);
1615
1616                                 fifo_push(&ca->free[RESERVE_BTREE], bu);
1617                         }
1618                 }
1619
1620                 pr_debug("done scanning for reclaimable buckets");
1621
1622                 /*
1623                  * XXX: it's possible for this to deadlock waiting on journal reclaim,
1624                  * since we're holding btree writes. What then?
1625                  */
1626                 ret = bch2_alloc_write(c,
1627                                        BTREE_INSERT_NOCHECK_RW|
1628                                        BTREE_INSERT_USE_ALLOC_RESERVE|
1629                                        BTREE_INSERT_NOWAIT, &wrote);
1630
1631                 /*
1632                  * If bch2_alloc_write() did anything, it may have used some
1633                  * buckets, and we need the RESERVE_BTREE freelist full - so we
1634                  * need to loop and scan again.
1635                  * And if it errored, it may have been because there weren't
1636                  * enough buckets, so just scan and loop again as long as it
1637                  * made some progress:
1638                  */
1639         } while (wrote);
1640         up_read(&c->gc_lock);
1641
1642         if (ret)
1643                 goto err;
1644
1645         pr_debug("flushing journal");
1646
1647         ret = bch2_journal_flush(&c->journal);
1648         if (ret)
1649                 goto err;
1650
1651         pr_debug("issuing discards");
1652         allocator_start_issue_discards(c);
1653 err:
1654         clear_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags);
1655         closure_wait_event(&c->btree_interior_update_wait,
1656                            flush_held_btree_writes(c));
1657
1658         return ret;
1659 }
1660
1661 void bch2_fs_allocator_background_init(struct bch_fs *c)
1662 {
1663         spin_lock_init(&c->freelist_lock);
1664         bch2_bucket_clock_init(c, READ);
1665         bch2_bucket_clock_init(c, WRITE);
1666
1667         c->pd_controllers_update_seconds = 5;
1668         INIT_DELAYED_WORK(&c->pd_controllers_update, pd_controllers_update);
1669 }