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 "ec.h"
  13 #include "error.h"
  14 #include "journal_io.h"
  15
  16 #include <linux/kthread.h>
  17 #include <linux/math64.h>
  18 #include <linux/random.h>
  19 #include <linux/rculist.h>
  20 #include <linux/rcupdate.h>
  21 #include <linux/sched/task.h>
  22 #include <linux/sort.h>
  23 #include <trace/events/bcachefs.h>
  24
  25 static const char * const bch2_alloc_field_names[] = {
  26 #define x(name, bytes) #name,
  27         BCH_ALLOC_FIELDS()
  28 #undef x
  29         NULL
  30 };
  31
  32 static void bch2_recalc_oldest_io(struct bch_fs *, struct bch_dev *, int);
  33
  34 /* Ratelimiting/PD controllers */
  35
  36 static void pd_controllers_update(struct work_struct *work)
  37 {
  38         struct bch_fs *c = container_of(to_delayed_work(work),
  39                                            struct bch_fs,
  40                                            pd_controllers_update);
  41         struct bch_dev *ca;
  42         unsigned i;
  43
  44         for_each_member_device(ca, c, i) {
  45                 struct bch_dev_usage stats = bch2_dev_usage_read(c, ca);
  46
  47                 u64 free = bucket_to_sector(ca,
  48                                 __dev_buckets_free(ca, stats)) << 9;
  49                 /*
  50                  * Bytes of internal fragmentation, which can be
  51                  * reclaimed by copy GC
  52                  */
  53                 s64 fragmented = (bucket_to_sector(ca,
  54                                         stats.buckets[BCH_DATA_USER] +
  55                                         stats.buckets[BCH_DATA_CACHED]) -
  56                                   (stats.sectors[BCH_DATA_USER] +
  57                                    stats.sectors[BCH_DATA_CACHED])) << 9;
  58
  59                 fragmented = max(0LL, fragmented);
  60
  61                 bch2_pd_controller_update(&ca->copygc_pd,
  62                                          free, fragmented, -1);
  63         }
  64
  65         schedule_delayed_work(&c->pd_controllers_update,
  66                               c->pd_controllers_update_seconds * HZ);
  67 }
  68
  69 /* Persistent alloc info: */
  70
  71 static inline u64 get_alloc_field(const struct bch_alloc *a,
  72                                   const void **p, unsigned field)
  73 {
  74         unsigned bytes = BCH_ALLOC_FIELD_BYTES[field];
  75         u64 v;
  76
  77         if (!(a->fields & (1 << field)))
  78                 return 0;
  79
  80         switch (bytes) {
  81         case 1:
  82                 v = *((const u8 *) *p);
  83                 break;
  84         case 2:
  85                 v = le16_to_cpup(*p);
  86                 break;
  87         case 4:
  88                 v = le32_to_cpup(*p);
  89                 break;
  90         case 8:
  91                 v = le64_to_cpup(*p);
  92                 break;
  93         default:
  94                 BUG();
  95         }
  96
  97         *p += bytes;
  98         return v;
  99 }
 100
 101 static inline void put_alloc_field(struct bkey_i_alloc *a, void **p,
 102                                    unsigned field, u64 v)
 103 {
 104         unsigned bytes = BCH_ALLOC_FIELD_BYTES[field];
 105
 106         if (!v)
 107                 return;
 108
 109         a->v.fields |= 1 << field;
 110
 111         switch (bytes) {
 112         case 1:
 113                 *((u8 *) *p) = v;
 114                 break;
 115         case 2:
 116                 *((__le16 *) *p) = cpu_to_le16(v);
 117                 break;
 118         case 4:
 119                 *((__le32 *) *p) = cpu_to_le32(v);
 120                 break;
 121         case 8:
 122                 *((__le64 *) *p) = cpu_to_le64(v);
 123                 break;
 124         default:
 125                 BUG();
 126         }
 127
 128         *p += bytes;
 129 }
 130
 131 struct bkey_alloc_unpacked bch2_alloc_unpack(const struct bch_alloc *a)
 132 {
 133         struct bkey_alloc_unpacked ret = { .gen = a->gen };
 134         const void *d = a->data;
 135         unsigned idx = 0;
 136
 137 #define x(_name, _bits) ret._name = get_alloc_field(a, &d, idx++);
 138         BCH_ALLOC_FIELDS()
 139 #undef  x
 140         return ret;
 141 }
 142
 143 static void bch2_alloc_pack(struct bkey_i_alloc *dst,
 144                             const struct bkey_alloc_unpacked src)
 145 {
 146         unsigned idx = 0;
 147         void *d = dst->v.data;
 148
 149         dst->v.fields   = 0;
 150         dst->v.gen      = src.gen;
 151
 152 #define x(_name, _bits) put_alloc_field(dst, &d, idx++, src._name);
 153         BCH_ALLOC_FIELDS()
 154 #undef  x
 155
 156         set_bkey_val_bytes(&dst->k, (void *) d - (void *) &dst->v);
 157 }
 158
 159 static unsigned bch_alloc_val_u64s(const struct bch_alloc *a)
 160 {
 161         unsigned i, bytes = offsetof(struct bch_alloc, data);
 162
 163         for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_FIELD_BYTES); i++)
 164                 if (a->fields & (1 << i))
 165                         bytes += BCH_ALLOC_FIELD_BYTES[i];
 166
 167         return DIV_ROUND_UP(bytes, sizeof(u64));
 168 }
 169
 170 const char *bch2_alloc_invalid(const struct bch_fs *c, struct bkey_s_c k)
 171 {
 172         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 173
 174         if (k.k->p.inode >= c->sb.nr_devices ||
 175             !c->devs[k.k->p.inode])
 176                 return "invalid device";
 177
 178         /* allow for unknown fields */
 179         if (bkey_val_u64s(a.k) < bch_alloc_val_u64s(a.v))
 180                 return "incorrect value size";
 181
 182         return NULL;
 183 }
 184
 185 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c,
 186                         struct bkey_s_c k)
 187 {
 188         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 189         const void *d = a.v->data;
 190         unsigned i;
 191
 192         pr_buf(out, "gen %u", a.v->gen);
 193
 194         for (i = 0; i < BCH_ALLOC_FIELD_NR; i++)
 195                 if (a.v->fields & (1 << i))
 196                         pr_buf(out, " %s %llu",
 197                                bch2_alloc_field_names[i],
 198                                get_alloc_field(a.v, &d, i));
 199 }
 200
 201 static void __alloc_read_key(struct bucket *g, const struct bch_alloc *a)
 202 {
 203         const void *d = a->data;
 204         unsigned idx = 0, data_type, dirty_sectors, cached_sectors;
 205         struct bucket_mark m;
 206
 207         g->io_time[READ]        = get_alloc_field(a, &d, idx++);
 208         g->io_time[WRITE]       = get_alloc_field(a, &d, idx++);
 209         data_type               = get_alloc_field(a, &d, idx++);
 210         dirty_sectors           = get_alloc_field(a, &d, idx++);
 211         cached_sectors          = get_alloc_field(a, &d, idx++);
 212         g->oldest_gen           = get_alloc_field(a, &d, idx++);
 213
 214         bucket_cmpxchg(g, m, ({
 215                 m.gen                   = a->gen;
 216                 m.data_type             = data_type;
 217                 m.dirty_sectors         = dirty_sectors;
 218                 m.cached_sectors        = cached_sectors;
 219         }));
 220
 221         g->gen_valid            = 1;
 222 }
 223
 224 static void __alloc_write_key(struct bkey_i_alloc *a, struct bucket *g,
 225                               struct bucket_mark m)
 226 {
 227         unsigned idx = 0;
 228         void *d = a->v.data;
 229
 230         a->v.fields     = 0;
 231         a->v.gen        = m.gen;
 232
 233         d = a->v.data;
 234         put_alloc_field(a, &d, idx++, g->io_time[READ]);
 235         put_alloc_field(a, &d, idx++, g->io_time[WRITE]);
 236         put_alloc_field(a, &d, idx++, m.data_type);
 237         put_alloc_field(a, &d, idx++, m.dirty_sectors);
 238         put_alloc_field(a, &d, idx++, m.cached_sectors);
 239         put_alloc_field(a, &d, idx++, g->oldest_gen);
 240
 241         set_bkey_val_bytes(&a->k, (void *) d - (void *) &a->v);
 242 }
 243
 244 static void bch2_alloc_read_key(struct bch_fs *c, struct bkey_s_c k)
 245 {
 246         struct bch_dev *ca;
 247         struct bkey_s_c_alloc a;
 248
 249         if (k.k->type != KEY_TYPE_alloc)
 250                 return;
 251
 252         a = bkey_s_c_to_alloc(k);
 253         ca = bch_dev_bkey_exists(c, a.k->p.inode);
 254
 255         if (a.k->p.offset >= ca->mi.nbuckets)
 256                 return;
 257
 258         percpu_down_read_preempt_disable(&c->mark_lock);
 259         __alloc_read_key(bucket(ca, a.k->p.offset), a.v);
 260         percpu_up_read_preempt_enable(&c->mark_lock);
 261 }
 262
 263 int bch2_alloc_read(struct bch_fs *c, struct list_head *journal_replay_list)
 264 {
 265         struct journal_replay *r;
 266         struct btree_iter iter;
 267         struct bkey_s_c k;
 268         struct bch_dev *ca;
 269         unsigned i;
 270         int ret;
 271
 272         for_each_btree_key(&iter, c, BTREE_ID_ALLOC, POS_MIN, 0, k) {
 273                 bch2_alloc_read_key(c, k);
 274                 bch2_btree_iter_cond_resched(&iter);
 275         }
 276
 277         ret = bch2_btree_iter_unlock(&iter);
 278         if (ret)
 279                 return ret;
 280
 281         list_for_each_entry(r, journal_replay_list, list) {
 282                 struct bkey_i *k, *n;
 283                 struct jset_entry *entry;
 284
 285                 for_each_jset_key(k, n, entry, &r->j)
 286                         if (entry->btree_id == BTREE_ID_ALLOC)
 287                                 bch2_alloc_read_key(c, bkey_i_to_s_c(k));
 288         }
 289
 290         for_each_member_device(ca, c, i)
 291                 bch2_dev_usage_from_buckets(c, ca);
 292
 293         mutex_lock(&c->bucket_clock[READ].lock);
 294         for_each_member_device(ca, c, i) {
 295                 down_read(&ca->bucket_lock);
 296                 bch2_recalc_oldest_io(c, ca, READ);
 297                 up_read(&ca->bucket_lock);
 298         }
 299         mutex_unlock(&c->bucket_clock[READ].lock);
 300
 301         mutex_lock(&c->bucket_clock[WRITE].lock);
 302         for_each_member_device(ca, c, i) {
 303                 down_read(&ca->bucket_lock);
 304                 bch2_recalc_oldest_io(c, ca, WRITE);
 305                 up_read(&ca->bucket_lock);
 306         }
 307         mutex_unlock(&c->bucket_clock[WRITE].lock);
 308
 309         return 0;
 310 }
 311
 312 int bch2_alloc_replay_key(struct bch_fs *c, struct bkey_i *k)
 313 {
 314         struct btree_trans trans;
 315         struct btree_iter *iter;
 316         struct bch_dev *ca;
 317         int ret;
 318
 319         if (k->k.p.inode >= c->sb.nr_devices ||
 320             !c->devs[k->k.p.inode])
 321                 return 0;
 322
 323         ca = bch_dev_bkey_exists(c, k->k.p.inode);
 324
 325         if (k->k.p.offset >= ca->mi.nbuckets)
 326                 return 0;
 327
 328         bch2_trans_init(&trans, c);
 329
 330         iter = bch2_trans_get_iter(&trans, BTREE_ID_ALLOC, k->k.p,
 331                                    BTREE_ITER_INTENT);
 332
 333         ret = bch2_btree_iter_traverse(iter);
 334         if (ret)
 335                 goto err;
 336
 337         /* check buckets_written with btree node locked: */
 338         if (test_bit(k->k.p.offset, ca->buckets_written)) {
 339                 ret = 0;
 340                 goto err;
 341         }
 342
 343         bch2_trans_update(&trans, BTREE_INSERT_ENTRY(iter, k));
 344
 345         ret = bch2_trans_commit(&trans, NULL, NULL,
 346                                 BTREE_INSERT_NOFAIL|
 347                                 BTREE_INSERT_LAZY_RW|
 348                                 BTREE_INSERT_JOURNAL_REPLAY|
 349                                 BTREE_INSERT_NOMARK);
 350 err:
 351         bch2_trans_exit(&trans);
 352         return ret;
 353 }
 354
 355 static int __bch2_alloc_write_key(struct btree_trans *trans, struct bch_dev *ca,
 356                                   size_t b, struct btree_iter *iter,
 357                                   u64 *journal_seq, unsigned flags)
 358 {
 359         struct bch_fs *c = trans->c;
 360 #if 0
 361         __BKEY_PADDED(k, BKEY_ALLOC_VAL_U64s_MAX) alloc_key;
 362 #else
 363         /* hack: */
 364         __BKEY_PADDED(k, 8) alloc_key;
 365 #endif
 366         struct bkey_i_alloc *a = bkey_alloc_init(&alloc_key.k);
 367         struct bucket *g;
 368         struct bucket_mark m, new;
 369         int ret;
 370
 371         BUG_ON(BKEY_ALLOC_VAL_U64s_MAX > 8);
 372
 373         a->k.p = POS(ca->dev_idx, b);
 374
 375         bch2_btree_iter_set_pos(iter, a->k.p);
 376
 377         ret = bch2_btree_iter_traverse(iter);
 378         if (ret)
 379                 return ret;
 380
 381         percpu_down_read_preempt_disable(&c->mark_lock);
 382         g = bucket(ca, b);
 383         m = READ_ONCE(g->mark);
 384
 385         if (!m.dirty) {
 386                 percpu_up_read_preempt_enable(&c->mark_lock);
 387                 return 0;
 388         }
 389
 390         __alloc_write_key(a, g, m);
 391         percpu_up_read_preempt_enable(&c->mark_lock);
 392
 393         bch2_btree_iter_cond_resched(iter);
 394
 395         bch2_trans_update(trans, BTREE_INSERT_ENTRY(iter, &a->k_i));
 396
 397         ret = bch2_trans_commit(trans, NULL, journal_seq,
 398                                    BTREE_INSERT_NOCHECK_RW|
 399                                    BTREE_INSERT_NOFAIL|
 400                                    BTREE_INSERT_USE_RESERVE|
 401                                    BTREE_INSERT_USE_ALLOC_RESERVE|
 402                                    BTREE_INSERT_NOMARK|
 403                                    flags);
 404         if (ret)
 405                 return ret;
 406
 407         new = m;
 408         new.dirty = false;
 409         atomic64_cmpxchg(&g->_mark.v, m.v.counter, new.v.counter);
 410
 411         if (ca->buckets_written)
 412                 set_bit(b, ca->buckets_written);
 413
 414         return 0;
 415 }
 416
 417 int bch2_alloc_write(struct bch_fs *c, bool nowait, bool *wrote)
 418 {
 419         struct bch_dev *ca;
 420         unsigned i;
 421         int ret = 0;
 422
 423         *wrote = false;
 424
 425         for_each_rw_member(ca, c, i) {
 426                 struct btree_trans trans;
 427                 struct btree_iter *iter;
 428                 struct bucket_array *buckets;
 429                 size_t b;
 430
 431                 bch2_trans_init(&trans, c);
 432
 433                 iter = bch2_trans_get_iter(&trans, BTREE_ID_ALLOC, POS_MIN,
 434                                            BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 435
 436                 down_read(&ca->bucket_lock);
 437                 buckets = bucket_array(ca);
 438
 439                 for (b = buckets->first_bucket;
 440                      b < buckets->nbuckets;
 441                      b++) {
 442                         if (!buckets->b[b].mark.dirty)
 443                                 continue;
 444
 445                         ret = __bch2_alloc_write_key(&trans, ca, b, iter, NULL,
 446                                                      nowait
 447                                                      ? BTREE_INSERT_NOWAIT
 448                                                      : 0);
 449                         if (ret)
 450                                 break;
 451
 452                         *wrote = true;
 453                 }
 454                 up_read(&ca->bucket_lock);
 455
 456                 bch2_trans_exit(&trans);
 457
 458                 if (ret) {
 459                         percpu_ref_put(&ca->io_ref);
 460                         break;
 461                 }
 462         }
 463
 464         return ret;
 465 }
 466
 467 /* Bucket IO clocks: */
 468
 469 static void bch2_recalc_oldest_io(struct bch_fs *c, struct bch_dev *ca, int rw)
 470 {
 471         struct bucket_clock *clock = &c->bucket_clock[rw];
 472         struct bucket_array *buckets = bucket_array(ca);
 473         struct bucket *g;
 474         u16 max_last_io = 0;
 475         unsigned i;
 476
 477         lockdep_assert_held(&c->bucket_clock[rw].lock);
 478
 479         /* Recalculate max_last_io for this device: */
 480         for_each_bucket(g, buckets)
 481                 max_last_io = max(max_last_io, bucket_last_io(c, g, rw));
 482
 483         ca->max_last_bucket_io[rw] = max_last_io;
 484
 485         /* Recalculate global max_last_io: */
 486         max_last_io = 0;
 487
 488         for_each_member_device(ca, c, i)
 489                 max_last_io = max(max_last_io, ca->max_last_bucket_io[rw]);
 490
 491         clock->max_last_io = max_last_io;
 492 }
 493
 494 static void bch2_rescale_bucket_io_times(struct bch_fs *c, int rw)
 495 {
 496         struct bucket_clock *clock = &c->bucket_clock[rw];
 497         struct bucket_array *buckets;
 498         struct bch_dev *ca;
 499         struct bucket *g;
 500         unsigned i;
 501
 502         trace_rescale_prios(c);
 503
 504         for_each_member_device(ca, c, i) {
 505                 down_read(&ca->bucket_lock);
 506                 buckets = bucket_array(ca);
 507
 508                 for_each_bucket(g, buckets)
 509                         g->io_time[rw] = clock->hand -
 510                         bucket_last_io(c, g, rw) / 2;
 511
 512                 bch2_recalc_oldest_io(c, ca, rw);
 513
 514                 up_read(&ca->bucket_lock);
 515         }
 516 }
 517
 518 static inline u64 bucket_clock_freq(u64 capacity)
 519 {
 520         return max(capacity >> 10, 2028ULL);
 521 }
 522
 523 static void bch2_inc_clock_hand(struct io_timer *timer)
 524 {
 525         struct bucket_clock *clock = container_of(timer,
 526                                                 struct bucket_clock, rescale);
 527         struct bch_fs *c = container_of(clock,
 528                                         struct bch_fs, bucket_clock[clock->rw]);
 529         struct bch_dev *ca;
 530         u64 capacity;
 531         unsigned i;
 532
 533         mutex_lock(&clock->lock);
 534
 535         /* if clock cannot be advanced more, rescale prio */
 536         if (clock->max_last_io >= U16_MAX - 2)
 537                 bch2_rescale_bucket_io_times(c, clock->rw);
 538
 539         BUG_ON(clock->max_last_io >= U16_MAX - 2);
 540
 541         for_each_member_device(ca, c, i)
 542                 ca->max_last_bucket_io[clock->rw]++;
 543         clock->max_last_io++;
 544         clock->hand++;
 545
 546         mutex_unlock(&clock->lock);
 547
 548         capacity = READ_ONCE(c->capacity);
 549
 550         if (!capacity)
 551                 return;
 552
 553         /*
 554          * we only increment when 0.1% of the filesystem capacity has been read
 555          * or written too, this determines if it's time
 556          *
 557          * XXX: we shouldn't really be going off of the capacity of devices in
 558          * RW mode (that will be 0 when we're RO, yet we can still service
 559          * reads)
 560          */
 561         timer->expire += bucket_clock_freq(capacity);
 562
 563         bch2_io_timer_add(&c->io_clock[clock->rw], timer);
 564 }
 565
 566 static void bch2_bucket_clock_init(struct bch_fs *c, int rw)
 567 {
 568         struct bucket_clock *clock = &c->bucket_clock[rw];
 569
 570         clock->hand             = 1;
 571         clock->rw               = rw;
 572         clock->rescale.fn       = bch2_inc_clock_hand;
 573         clock->rescale.expire   = bucket_clock_freq(c->capacity);
 574         mutex_init(&clock->lock);
 575 }
 576
 577 /* Background allocator thread: */
 578
 579 /*
 580  * Scans for buckets to be invalidated, invalidates them, rewrites prios/gens
 581  * (marking them as invalidated on disk), then optionally issues discard
 582  * commands to the newly free buckets, then puts them on the various freelists.
 583  */
 584
 585 #define BUCKET_GC_GEN_MAX       96U
 586
 587 /**
 588  * wait_buckets_available - wait on reclaimable buckets
 589  *
 590  * If there aren't enough available buckets to fill up free_inc, wait until
 591  * there are.
 592  */
 593 static int wait_buckets_available(struct bch_fs *c, struct bch_dev *ca)
 594 {
 595         unsigned long gc_count = c->gc_count;
 596         int ret = 0;
 597
 598         while (1) {
 599                 set_current_state(TASK_INTERRUPTIBLE);
 600                 if (kthread_should_stop()) {
 601                         ret = 1;
 602                         break;
 603                 }
 604
 605                 if (gc_count != c->gc_count)
 606                         ca->inc_gen_really_needs_gc = 0;
 607
 608                 if ((ssize_t) (dev_buckets_available(c, ca) -
 609                                ca->inc_gen_really_needs_gc) >=
 610                     (ssize_t) fifo_free(&ca->free_inc))
 611                         break;
 612
 613                 up_read(&c->gc_lock);
 614                 schedule();
 615                 try_to_freeze();
 616                 down_read(&c->gc_lock);
 617         }
 618
 619         __set_current_state(TASK_RUNNING);
 620         return ret;
 621 }
 622
 623 static bool bch2_can_invalidate_bucket(struct bch_dev *ca,
 624                                        size_t bucket,
 625                                        struct bucket_mark mark)
 626 {
 627         u8 gc_gen;
 628
 629         if (!is_available_bucket(mark))
 630                 return false;
 631
 632         if (ca->buckets_nouse &&
 633             test_bit(bucket, ca->buckets_nouse))
 634                 return false;
 635
 636         gc_gen = bucket_gc_gen(ca, bucket);
 637
 638         if (gc_gen >= BUCKET_GC_GEN_MAX / 2)
 639                 ca->inc_gen_needs_gc++;
 640
 641         if (gc_gen >= BUCKET_GC_GEN_MAX)
 642                 ca->inc_gen_really_needs_gc++;
 643
 644         return gc_gen < BUCKET_GC_GEN_MAX;
 645 }
 646
 647 /*
 648  * Determines what order we're going to reuse buckets, smallest bucket_key()
 649  * first.
 650  *
 651  *
 652  * - We take into account the read prio of the bucket, which gives us an
 653  *   indication of how hot the data is -- we scale the prio so that the prio
 654  *   farthest from the clock is worth 1/8th of the closest.
 655  *
 656  * - The number of sectors of cached data in the bucket, which gives us an
 657  *   indication of the cost in cache misses this eviction will cause.
 658  *
 659  * - If hotness * sectors used compares equal, we pick the bucket with the
 660  *   smallest bucket_gc_gen() - since incrementing the same bucket's generation
 661  *   number repeatedly forces us to run mark and sweep gc to avoid generation
 662  *   number wraparound.
 663  */
 664
 665 static unsigned long bucket_sort_key(struct bch_fs *c, struct bch_dev *ca,
 666                                      size_t b, struct bucket_mark m)
 667 {
 668         unsigned last_io = bucket_last_io(c, bucket(ca, b), READ);
 669         unsigned max_last_io = ca->max_last_bucket_io[READ];
 670
 671         /*
 672          * Time since last read, scaled to [0, 8) where larger value indicates
 673          * more recently read data:
 674          */
 675         unsigned long hotness = (max_last_io - last_io) * 7 / max_last_io;
 676
 677         /* How much we want to keep the data in this bucket: */
 678         unsigned long data_wantness =
 679                 (hotness + 1) * bucket_sectors_used(m);
 680
 681         unsigned long needs_journal_commit =
 682                 bucket_needs_journal_commit(m, c->journal.last_seq_ondisk);
 683
 684         return  (data_wantness << 9) |
 685                 (needs_journal_commit << 8) |
 686                 (bucket_gc_gen(ca, b) / 16);
 687 }
 688
 689 static inline int bucket_alloc_cmp(alloc_heap *h,
 690                                    struct alloc_heap_entry l,
 691                                    struct alloc_heap_entry r)
 692 {
 693         return (l.key > r.key) - (l.key < r.key) ?:
 694                 (l.nr < r.nr)  - (l.nr  > r.nr) ?:
 695                 (l.bucket > r.bucket) - (l.bucket < r.bucket);
 696 }
 697
 698 static inline int bucket_idx_cmp(const void *_l, const void *_r)
 699 {
 700         const struct alloc_heap_entry *l = _l, *r = _r;
 701
 702         return (l->bucket > r->bucket) - (l->bucket < r->bucket);
 703 }
 704
 705 static void find_reclaimable_buckets_lru(struct bch_fs *c, struct bch_dev *ca)
 706 {
 707         struct bucket_array *buckets;
 708         struct alloc_heap_entry e = { 0 };
 709         size_t b, i, nr = 0;
 710
 711         ca->alloc_heap.used = 0;
 712
 713         mutex_lock(&c->bucket_clock[READ].lock);
 714         down_read(&ca->bucket_lock);
 715
 716         buckets = bucket_array(ca);
 717
 718         bch2_recalc_oldest_io(c, ca, READ);
 719
 720         /*
 721          * Find buckets with lowest read priority, by building a maxheap sorted
 722          * by read priority and repeatedly replacing the maximum element until
 723          * all buckets have been visited.
 724          */
 725         for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++) {
 726                 struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
 727                 unsigned long key = bucket_sort_key(c, ca, b, m);
 728
 729                 if (!bch2_can_invalidate_bucket(ca, b, m))
 730                         continue;
 731
 732                 if (e.nr && e.bucket + e.nr == b && e.key == key) {
 733                         e.nr++;
 734                 } else {
 735                         if (e.nr)
 736                                 heap_add_or_replace(&ca->alloc_heap, e,
 737                                         -bucket_alloc_cmp, NULL);
 738
 739                         e = (struct alloc_heap_entry) {
 740                                 .bucket = b,
 741                                 .nr     = 1,
 742                                 .key    = key,
 743                         };
 744                 }
 745
 746                 cond_resched();
 747         }
 748
 749         if (e.nr)
 750                 heap_add_or_replace(&ca->alloc_heap, e,
 751                                 -bucket_alloc_cmp, NULL);
 752
 753         for (i = 0; i < ca->alloc_heap.used; i++)
 754                 nr += ca->alloc_heap.data[i].nr;
 755
 756         while (nr - ca->alloc_heap.data[0].nr >= ALLOC_SCAN_BATCH(ca)) {
 757                 nr -= ca->alloc_heap.data[0].nr;
 758                 heap_pop(&ca->alloc_heap, e, -bucket_alloc_cmp, NULL);
 759         }
 760
 761         up_read(&ca->bucket_lock);
 762         mutex_unlock(&c->bucket_clock[READ].lock);
 763 }
 764
 765 static void find_reclaimable_buckets_fifo(struct bch_fs *c, struct bch_dev *ca)
 766 {
 767         struct bucket_array *buckets = bucket_array(ca);
 768         struct bucket_mark m;
 769         size_t b, start;
 770
 771         if (ca->fifo_last_bucket <  ca->mi.first_bucket ||
 772             ca->fifo_last_bucket >= ca->mi.nbuckets)
 773                 ca->fifo_last_bucket = ca->mi.first_bucket;
 774
 775         start = ca->fifo_last_bucket;
 776
 777         do {
 778                 ca->fifo_last_bucket++;
 779                 if (ca->fifo_last_bucket == ca->mi.nbuckets)
 780                         ca->fifo_last_bucket = ca->mi.first_bucket;
 781
 782                 b = ca->fifo_last_bucket;
 783                 m = READ_ONCE(buckets->b[b].mark);
 784
 785                 if (bch2_can_invalidate_bucket(ca, b, m)) {
 786                         struct alloc_heap_entry e = { .bucket = b, .nr = 1, };
 787
 788                         heap_add(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
 789                         if (heap_full(&ca->alloc_heap))
 790                                 break;
 791                 }
 792
 793                 cond_resched();
 794         } while (ca->fifo_last_bucket != start);
 795 }
 796
 797 static void find_reclaimable_buckets_random(struct bch_fs *c, struct bch_dev *ca)
 798 {
 799         struct bucket_array *buckets = bucket_array(ca);
 800         struct bucket_mark m;
 801         size_t checked, i;
 802
 803         for (checked = 0;
 804              checked < ca->mi.nbuckets / 2;
 805              checked++) {
 806                 size_t b = bch2_rand_range(ca->mi.nbuckets -
 807                                            ca->mi.first_bucket) +
 808                         ca->mi.first_bucket;
 809
 810                 m = READ_ONCE(buckets->b[b].mark);
 811
 812                 if (bch2_can_invalidate_bucket(ca, b, m)) {
 813                         struct alloc_heap_entry e = { .bucket = b, .nr = 1, };
 814
 815                         heap_add(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
 816                         if (heap_full(&ca->alloc_heap))
 817                                 break;
 818                 }
 819
 820                 cond_resched();
 821         }
 822
 823         sort(ca->alloc_heap.data,
 824              ca->alloc_heap.used,
 825              sizeof(ca->alloc_heap.data[0]),
 826              bucket_idx_cmp, NULL);
 827
 828         /* remove duplicates: */
 829         for (i = 0; i + 1 < ca->alloc_heap.used; i++)
 830                 if (ca->alloc_heap.data[i].bucket ==
 831                     ca->alloc_heap.data[i + 1].bucket)
 832                         ca->alloc_heap.data[i].nr = 0;
 833 }
 834
 835 static size_t find_reclaimable_buckets(struct bch_fs *c, struct bch_dev *ca)
 836 {
 837         size_t i, nr = 0;
 838
 839         ca->inc_gen_needs_gc                    = 0;
 840
 841         switch (ca->mi.replacement) {
 842         case CACHE_REPLACEMENT_LRU:
 843                 find_reclaimable_buckets_lru(c, ca);
 844                 break;
 845         case CACHE_REPLACEMENT_FIFO:
 846                 find_reclaimable_buckets_fifo(c, ca);
 847                 break;
 848         case CACHE_REPLACEMENT_RANDOM:
 849                 find_reclaimable_buckets_random(c, ca);
 850                 break;
 851         }
 852
 853         heap_resort(&ca->alloc_heap, bucket_alloc_cmp, NULL);
 854
 855         for (i = 0; i < ca->alloc_heap.used; i++)
 856                 nr += ca->alloc_heap.data[i].nr;
 857
 858         return nr;
 859 }
 860
 861 static inline long next_alloc_bucket(struct bch_dev *ca)
 862 {
 863         struct alloc_heap_entry e, *top = ca->alloc_heap.data;
 864
 865         while (ca->alloc_heap.used) {
 866                 if (top->nr) {
 867                         size_t b = top->bucket;
 868
 869                         top->bucket++;
 870                         top->nr--;
 871                         return b;
 872                 }
 873
 874                 heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
 875         }
 876
 877         return -1;
 878 }
 879
 880 /*
 881  * returns sequence number of most recent journal entry that updated this
 882  * bucket:
 883  */
 884 static u64 bucket_journal_seq(struct bch_fs *c, struct bucket_mark m)
 885 {
 886         if (m.journal_seq_valid) {
 887                 u64 journal_seq = atomic64_read(&c->journal.seq);
 888                 u64 bucket_seq  = journal_seq;
 889
 890                 bucket_seq &= ~((u64) U16_MAX);
 891                 bucket_seq |= m.journal_seq;
 892
 893                 if (bucket_seq > journal_seq)
 894                         bucket_seq -= 1 << 16;
 895
 896                 return bucket_seq;
 897         } else {
 898                 return 0;
 899         }
 900 }
 901
 902 static int bch2_invalidate_one_bucket2(struct btree_trans *trans,
 903                                        struct bch_dev *ca,
 904                                        struct btree_iter *iter,
 905                                        u64 *journal_seq, unsigned flags)
 906 {
 907 #if 0
 908         __BKEY_PADDED(k, BKEY_ALLOC_VAL_U64s_MAX) alloc_key;
 909 #else
 910         /* hack: */
 911         __BKEY_PADDED(k, 8) alloc_key;
 912 #endif
 913         struct bch_fs *c = trans->c;
 914         struct bkey_i_alloc *a;
 915         struct bkey_alloc_unpacked u;
 916         struct bucket_mark m;
 917         struct bkey_s_c k;
 918         bool invalidating_cached_data;
 919         size_t b;
 920         int ret;
 921
 922         BUG_ON(!ca->alloc_heap.used ||
 923                !ca->alloc_heap.data[0].nr);
 924         b = ca->alloc_heap.data[0].bucket;
 925
 926         /* first, put on free_inc and mark as owned by allocator: */
 927         percpu_down_read_preempt_disable(&c->mark_lock);
 928         spin_lock(&c->freelist_lock);
 929
 930         verify_not_on_freelist(c, ca, b);
 931
 932         BUG_ON(!fifo_push(&ca->free_inc, b));
 933
 934         bch2_mark_alloc_bucket(c, ca, b, true, gc_pos_alloc(c, NULL), 0);
 935         m = bucket(ca, b)->mark;
 936
 937         spin_unlock(&c->freelist_lock);
 938         percpu_up_read_preempt_enable(&c->mark_lock);
 939
 940         bch2_btree_iter_cond_resched(iter);
 941
 942         BUG_ON(BKEY_ALLOC_VAL_U64s_MAX > 8);
 943
 944         bch2_btree_iter_set_pos(iter, POS(ca->dev_idx, b));
 945 retry:
 946         k = bch2_btree_iter_peek_slot(iter);
 947         ret = btree_iter_err(k);
 948         if (ret)
 949                 return ret;
 950
 951         if (k.k && k.k->type == KEY_TYPE_alloc)
 952                 u = bch2_alloc_unpack(bkey_s_c_to_alloc(k).v);
 953         else
 954                 memset(&u, 0, sizeof(u));
 955
 956         invalidating_cached_data = m.cached_sectors != 0;
 957
 958         //BUG_ON(u.dirty_sectors);
 959         u.data_type     = 0;
 960         u.dirty_sectors = 0;
 961         u.cached_sectors = 0;
 962         u.read_time     = c->bucket_clock[READ].hand;
 963         u.write_time    = c->bucket_clock[WRITE].hand;
 964
 965         /*
 966          * The allocator has to start before journal replay is finished - thus,
 967          * we have to trust the in memory bucket @m, not the version in the
 968          * btree:
 969          */
 970         u.gen           = m.gen + 1;
 971
 972         a = bkey_alloc_init(&alloc_key.k);
 973         a->k.p = iter->pos;
 974         bch2_alloc_pack(a, u);
 975
 976         bch2_trans_update(trans, BTREE_INSERT_ENTRY(iter, &a->k_i));
 977
 978         /*
 979          * XXX:
 980          * when using deferred btree updates, we have journal reclaim doing
 981          * btree updates and thus requiring the allocator to make forward
 982          * progress, and here the allocator is requiring space in the journal -
 983          * so we need a journal pre-reservation:
 984          */
 985         ret = bch2_trans_commit(trans, NULL,
 986                                 invalidating_cached_data ? journal_seq : NULL,
 987                                 BTREE_INSERT_ATOMIC|
 988                                 BTREE_INSERT_NOUNLOCK|
 989                                 BTREE_INSERT_NOCHECK_RW|
 990                                 BTREE_INSERT_NOFAIL|
 991                                 BTREE_INSERT_USE_RESERVE|
 992                                 BTREE_INSERT_USE_ALLOC_RESERVE|
 993                                 flags);
 994         if (ret == -EINTR)
 995                 goto retry;
 996
 997         if (!ret) {
 998                 /* remove from alloc_heap: */
 999                 struct alloc_heap_entry e, *top = ca->alloc_heap.data;
1000
1001                 top->bucket++;
1002                 top->nr--;
1003
1004                 if (!top->nr)
1005                         heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
1006
1007                 /* with btree still locked: */
1008                 if (ca->buckets_written)
1009                         set_bit(b, ca->buckets_written);
1010
1011                 /*
1012                  * Make sure we flush the last journal entry that updated this
1013                  * bucket (i.e. deleting the last reference) before writing to
1014                  * this bucket again:
1015                  */
1016                 *journal_seq = max(*journal_seq, bucket_journal_seq(c, m));
1017         } else {
1018                 size_t b2;
1019
1020                 /* remove from free_inc: */
1021                 percpu_down_read_preempt_disable(&c->mark_lock);
1022                 spin_lock(&c->freelist_lock);
1023
1024                 bch2_mark_alloc_bucket(c, ca, b, false,
1025                                        gc_pos_alloc(c, NULL), 0);
1026
1027                 BUG_ON(!fifo_pop_back(&ca->free_inc, b2));
1028                 BUG_ON(b != b2);
1029
1030                 spin_unlock(&c->freelist_lock);
1031                 percpu_up_read_preempt_enable(&c->mark_lock);
1032         }
1033
1034         return ret;
1035 }
1036
1037 static bool bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca,
1038                                        size_t bucket, u64 *flush_seq)
1039 {
1040         struct bucket_mark m;
1041
1042         percpu_down_read_preempt_disable(&c->mark_lock);
1043         spin_lock(&c->freelist_lock);
1044
1045         bch2_invalidate_bucket(c, ca, bucket, &m);
1046
1047         verify_not_on_freelist(c, ca, bucket);
1048         BUG_ON(!fifo_push(&ca->free_inc, bucket));
1049
1050         spin_unlock(&c->freelist_lock);
1051
1052         bucket_io_clock_reset(c, ca, bucket, READ);
1053         bucket_io_clock_reset(c, ca, bucket, WRITE);
1054
1055         percpu_up_read_preempt_enable(&c->mark_lock);
1056
1057         *flush_seq = max(*flush_seq, bucket_journal_seq(c, m));
1058
1059         return m.cached_sectors != 0;
1060 }
1061
1062 /*
1063  * Pull buckets off ca->alloc_heap, invalidate them, move them to ca->free_inc:
1064  */
1065 static int bch2_invalidate_buckets(struct bch_fs *c, struct bch_dev *ca)
1066 {
1067         struct btree_trans trans;
1068         struct btree_iter *iter;
1069         u64 journal_seq = 0;
1070         int ret = 0;
1071
1072         bch2_trans_init(&trans, c);
1073
1074         iter = bch2_trans_get_iter(&trans, BTREE_ID_ALLOC,
1075                                    POS(ca->dev_idx, 0),
1076                                    BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
1077
1078         /* Only use nowait if we've already invalidated at least one bucket: */
1079         while (!ret &&
1080                !fifo_full(&ca->free_inc) &&
1081                ca->alloc_heap.used)
1082                 ret = bch2_invalidate_one_bucket2(&trans, ca, iter, &journal_seq,
1083                                 BTREE_INSERT_GC_LOCK_HELD|
1084                                 (!fifo_empty(&ca->free_inc)
1085                                  ? BTREE_INSERT_NOWAIT : 0));
1086
1087         bch2_trans_exit(&trans);
1088
1089         /* If we used NOWAIT, don't return the error: */
1090         if (!fifo_empty(&ca->free_inc))
1091                 ret = 0;
1092         if (ret) {
1093                 bch_err(ca, "error invalidating buckets: %i", ret);
1094                 return ret;
1095         }
1096
1097         if (journal_seq)
1098                 ret = bch2_journal_flush_seq(&c->journal, journal_seq);
1099         if (ret) {
1100                 bch_err(ca, "journal error: %i", ret);
1101                 return ret;
1102         }
1103
1104         return 0;
1105 }
1106
1107 static int push_invalidated_bucket(struct bch_fs *c, struct bch_dev *ca, size_t bucket)
1108 {
1109         unsigned i;
1110         int ret = 0;
1111
1112         while (1) {
1113                 set_current_state(TASK_INTERRUPTIBLE);
1114
1115                 spin_lock(&c->freelist_lock);
1116                 for (i = 0; i < RESERVE_NR; i++)
1117                         if (fifo_push(&ca->free[i], bucket)) {
1118                                 fifo_pop(&ca->free_inc, bucket);
1119
1120                                 closure_wake_up(&c->freelist_wait);
1121                                 ca->allocator_blocked_full = false;
1122
1123                                 spin_unlock(&c->freelist_lock);
1124                                 goto out;
1125                         }
1126
1127                 if (!ca->allocator_blocked_full) {
1128                         ca->allocator_blocked_full = true;
1129                         closure_wake_up(&c->freelist_wait);
1130                 }
1131
1132                 spin_unlock(&c->freelist_lock);
1133
1134                 if ((current->flags & PF_KTHREAD) &&
1135                     kthread_should_stop()) {
1136                         ret = 1;
1137                         break;
1138                 }
1139
1140                 schedule();
1141                 try_to_freeze();
1142         }
1143 out:
1144         __set_current_state(TASK_RUNNING);
1145         return ret;
1146 }
1147
1148 /*
1149  * Pulls buckets off free_inc, discards them (if enabled), then adds them to
1150  * freelists, waiting until there's room if necessary:
1151  */
1152 static int discard_invalidated_buckets(struct bch_fs *c, struct bch_dev *ca)
1153 {
1154         while (!fifo_empty(&ca->free_inc)) {
1155                 size_t bucket = fifo_peek(&ca->free_inc);
1156
1157                 if (ca->mi.discard &&
1158                     blk_queue_discard(bdev_get_queue(ca->disk_sb.bdev)))
1159                         blkdev_issue_discard(ca->disk_sb.bdev,
1160                                              bucket_to_sector(ca, bucket),
1161                                              ca->mi.bucket_size, GFP_NOIO, 0);
1162
1163                 if (push_invalidated_bucket(c, ca, bucket))
1164                         return 1;
1165         }
1166
1167         return 0;
1168 }
1169
1170 /**
1171  * bch_allocator_thread - move buckets from free_inc to reserves
1172  *
1173  * The free_inc FIFO is populated by find_reclaimable_buckets(), and
1174  * the reserves are depleted by bucket allocation. When we run out
1175  * of free_inc, try to invalidate some buckets and write out
1176  * prios and gens.
1177  */
1178 static int bch2_allocator_thread(void *arg)
1179 {
1180         struct bch_dev *ca = arg;
1181         struct bch_fs *c = ca->fs;
1182         size_t nr;
1183         int ret;
1184
1185         set_freezable();
1186
1187         while (1) {
1188                 cond_resched();
1189
1190                 pr_debug("discarding %zu invalidated buckets",
1191                          fifo_used(&ca->free_inc));
1192
1193                 ret = discard_invalidated_buckets(c, ca);
1194                 if (ret)
1195                         goto stop;
1196
1197                 down_read(&c->gc_lock);
1198
1199                 ret = bch2_invalidate_buckets(c, ca);
1200                 if (ret) {
1201                         up_read(&c->gc_lock);
1202                         goto stop;
1203                 }
1204
1205                 if (!fifo_empty(&ca->free_inc)) {
1206                         up_read(&c->gc_lock);
1207                         continue;
1208                 }
1209
1210                 pr_debug("free_inc now empty");
1211
1212                 do {
1213                         /*
1214                          * Find some buckets that we can invalidate, either
1215                          * they're completely unused, or only contain clean data
1216                          * that's been written back to the backing device or
1217                          * another cache tier
1218                          */
1219
1220                         pr_debug("scanning for reclaimable buckets");
1221
1222                         nr = find_reclaimable_buckets(c, ca);
1223
1224                         pr_debug("found %zu buckets", nr);
1225
1226                         trace_alloc_batch(ca, nr, ca->alloc_heap.size);
1227
1228                         if ((ca->inc_gen_needs_gc >= ALLOC_SCAN_BATCH(ca) ||
1229                              ca->inc_gen_really_needs_gc) &&
1230                             c->gc_thread) {
1231                                 atomic_inc(&c->kick_gc);
1232                                 wake_up_process(c->gc_thread);
1233                         }
1234
1235                         /*
1236                          * If we found any buckets, we have to invalidate them
1237                          * before we scan for more - but if we didn't find very
1238                          * many we may want to wait on more buckets being
1239                          * available so we don't spin:
1240                          */
1241                         if (!nr ||
1242                             (nr < ALLOC_SCAN_BATCH(ca) &&
1243                              !fifo_full(&ca->free[RESERVE_MOVINGGC]))) {
1244                                 ca->allocator_blocked = true;
1245                                 closure_wake_up(&c->freelist_wait);
1246
1247                                 ret = wait_buckets_available(c, ca);
1248                                 if (ret) {
1249                                         up_read(&c->gc_lock);
1250                                         goto stop;
1251                                 }
1252                         }
1253                 } while (!nr);
1254
1255                 ca->allocator_blocked = false;
1256                 up_read(&c->gc_lock);
1257
1258                 pr_debug("%zu buckets to invalidate", nr);
1259
1260                 /*
1261                  * alloc_heap is now full of newly-invalidated buckets: next,
1262                  * write out the new bucket gens:
1263                  */
1264         }
1265
1266 stop:
1267         pr_debug("alloc thread stopping (ret %i)", ret);
1268         return 0;
1269 }
1270
1271 /* Startup/shutdown (ro/rw): */
1272
1273 void bch2_recalc_capacity(struct bch_fs *c)
1274 {
1275         struct bch_dev *ca;
1276         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
1277         unsigned bucket_size_max = 0;
1278         unsigned long ra_pages = 0;
1279         unsigned i, j;
1280
1281         lockdep_assert_held(&c->state_lock);
1282
1283         for_each_online_member(ca, c, i) {
1284                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_bdi;
1285
1286                 ra_pages += bdi->ra_pages;
1287         }
1288
1289         bch2_set_ra_pages(c, ra_pages);
1290
1291         for_each_rw_member(ca, c, i) {
1292                 u64 dev_reserve = 0;
1293
1294                 /*
1295                  * We need to reserve buckets (from the number
1296                  * of currently available buckets) against
1297                  * foreground writes so that mainly copygc can
1298                  * make forward progress.
1299                  *
1300                  * We need enough to refill the various reserves
1301                  * from scratch - copygc will use its entire
1302                  * reserve all at once, then run against when
1303                  * its reserve is refilled (from the formerly
1304                  * available buckets).
1305                  *
1306                  * This reserve is just used when considering if
1307                  * allocations for foreground writes must wait -
1308                  * not -ENOSPC calculations.
1309                  */
1310                 for (j = 0; j < RESERVE_NONE; j++)
1311                         dev_reserve += ca->free[j].size;
1312
1313                 dev_reserve += 1;       /* btree write point */
1314                 dev_reserve += 1;       /* copygc write point */
1315                 dev_reserve += 1;       /* rebalance write point */
1316
1317                 dev_reserve *= ca->mi.bucket_size;
1318
1319                 ca->copygc_threshold = dev_reserve;
1320
1321                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
1322                                              ca->mi.first_bucket);
1323
1324                 reserved_sectors += dev_reserve * 2;
1325
1326                 bucket_size_max = max_t(unsigned, bucket_size_max,
1327                                         ca->mi.bucket_size);
1328         }
1329
1330         gc_reserve = c->opts.gc_reserve_bytes
1331                 ? c->opts.gc_reserve_bytes >> 9
1332                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
1333
1334         reserved_sectors = max(gc_reserve, reserved_sectors);
1335
1336         reserved_sectors = min(reserved_sectors, capacity);
1337
1338         c->capacity = capacity - reserved_sectors;
1339
1340         c->bucket_size_max = bucket_size_max;
1341
1342         if (c->capacity) {
1343                 bch2_io_timer_add(&c->io_clock[READ],
1344                                  &c->bucket_clock[READ].rescale);
1345                 bch2_io_timer_add(&c->io_clock[WRITE],
1346                                  &c->bucket_clock[WRITE].rescale);
1347         } else {
1348                 bch2_io_timer_del(&c->io_clock[READ],
1349                                  &c->bucket_clock[READ].rescale);
1350                 bch2_io_timer_del(&c->io_clock[WRITE],
1351                                  &c->bucket_clock[WRITE].rescale);
1352         }
1353
1354         /* Wake up case someone was waiting for buckets */
1355         closure_wake_up(&c->freelist_wait);
1356 }
1357
1358 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
1359 {
1360         struct open_bucket *ob;
1361         bool ret = false;
1362
1363         for (ob = c->open_buckets;
1364              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1365              ob++) {
1366                 spin_lock(&ob->lock);
1367                 if (ob->valid && !ob->on_partial_list &&
1368                     ob->ptr.dev == ca->dev_idx)
1369                         ret = true;
1370                 spin_unlock(&ob->lock);
1371         }
1372
1373         return ret;
1374 }
1375
1376 /* device goes ro: */
1377 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
1378 {
1379         unsigned i;
1380
1381         BUG_ON(ca->alloc_thread);
1382
1383         /* First, remove device from allocation groups: */
1384
1385         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1386                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
1387
1388         /*
1389          * Capacity is calculated based off of devices in allocation groups:
1390          */
1391         bch2_recalc_capacity(c);
1392
1393         /* Next, close write points that point to this device... */
1394         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1395                 bch2_writepoint_stop(c, ca, &c->write_points[i]);
1396
1397         bch2_writepoint_stop(c, ca, &ca->copygc_write_point);
1398         bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
1399         bch2_writepoint_stop(c, ca, &c->btree_write_point);
1400
1401         mutex_lock(&c->btree_reserve_cache_lock);
1402         while (c->btree_reserve_cache_nr) {
1403                 struct btree_alloc *a =
1404                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1405
1406                 bch2_open_buckets_put(c, &a->ob);
1407         }
1408         mutex_unlock(&c->btree_reserve_cache_lock);
1409
1410         while (1) {
1411                 struct open_bucket *ob;
1412
1413                 spin_lock(&c->freelist_lock);
1414                 if (!ca->open_buckets_partial_nr) {
1415                         spin_unlock(&c->freelist_lock);
1416                         break;
1417                 }
1418                 ob = c->open_buckets +
1419                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
1420                 ob->on_partial_list = false;
1421                 spin_unlock(&c->freelist_lock);
1422
1423                 bch2_open_bucket_put(c, ob);
1424         }
1425
1426         bch2_ec_stop_dev(c, ca);
1427
1428         /*
1429          * Wake up threads that were blocked on allocation, so they can notice
1430          * the device can no longer be removed and the capacity has changed:
1431          */
1432         closure_wake_up(&c->freelist_wait);
1433
1434         /*
1435          * journal_res_get() can block waiting for free space in the journal -
1436          * it needs to notice there may not be devices to allocate from anymore:
1437          */
1438         wake_up(&c->journal.wait);
1439
1440         /* Now wait for any in flight writes: */
1441
1442         closure_wait_event(&c->open_buckets_wait,
1443                            !bch2_dev_has_open_write_point(c, ca));
1444 }
1445
1446 /* device goes rw: */
1447 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
1448 {
1449         unsigned i;
1450
1451         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1452                 if (ca->mi.data_allowed & (1 << i))
1453                         set_bit(ca->dev_idx, c->rw_devs[i].d);
1454 }
1455
1456 void bch2_dev_allocator_quiesce(struct bch_fs *c, struct bch_dev *ca)
1457 {
1458         if (ca->alloc_thread)
1459                 closure_wait_event(&c->freelist_wait, ca->allocator_blocked_full);
1460 }
1461
1462 /* stop allocator thread: */
1463 void bch2_dev_allocator_stop(struct bch_dev *ca)
1464 {
1465         struct task_struct *p;
1466
1467         p = rcu_dereference_protected(ca->alloc_thread, 1);
1468         ca->alloc_thread = NULL;
1469
1470         /*
1471          * We need an rcu barrier between setting ca->alloc_thread = NULL and
1472          * the thread shutting down to avoid bch2_wake_allocator() racing:
1473          *
1474          * XXX: it would be better to have the rcu barrier be asynchronous
1475          * instead of blocking us here
1476          */
1477         synchronize_rcu();
1478
1479         if (p) {
1480                 kthread_stop(p);
1481                 put_task_struct(p);
1482         }
1483 }
1484
1485 /* start allocator thread: */
1486 int bch2_dev_allocator_start(struct bch_dev *ca)
1487 {
1488         struct task_struct *p;
1489
1490         /*
1491          * allocator thread already started?
1492          */
1493         if (ca->alloc_thread)
1494                 return 0;
1495
1496         p = kthread_create(bch2_allocator_thread, ca,
1497                            "bch_alloc[%s]", ca->name);
1498         if (IS_ERR(p))
1499                 return PTR_ERR(p);
1500
1501         get_task_struct(p);
1502         rcu_assign_pointer(ca->alloc_thread, p);
1503         wake_up_process(p);
1504         return 0;
1505 }
1506
1507 static bool flush_held_btree_writes(struct bch_fs *c)
1508 {
1509         struct bucket_table *tbl;
1510         struct rhash_head *pos;
1511         struct btree *b;
1512         bool nodes_unwritten;
1513         size_t i;
1514 again:
1515         cond_resched();
1516         nodes_unwritten = false;
1517
1518         rcu_read_lock();
1519         for_each_cached_btree(b, c, tbl, i, pos)
1520                 if (btree_node_need_write(b)) {
1521                         if (btree_node_may_write(b)) {
1522                                 rcu_read_unlock();
1523                                 btree_node_lock_type(c, b, SIX_LOCK_read);
1524                                 bch2_btree_node_write(c, b, SIX_LOCK_read);
1525                                 six_unlock_read(&b->lock);
1526                                 goto again;
1527                         } else {
1528                                 nodes_unwritten = true;
1529                         }
1530                 }
1531         rcu_read_unlock();
1532
1533         if (c->btree_roots_dirty) {
1534                 bch2_journal_meta(&c->journal);
1535                 goto again;
1536         }
1537
1538         return !nodes_unwritten &&
1539                 !bch2_btree_interior_updates_nr_pending(c);
1540 }
1541
1542 static void allocator_start_issue_discards(struct bch_fs *c)
1543 {
1544         struct bch_dev *ca;
1545         unsigned dev_iter;
1546         size_t bu;
1547
1548         for_each_rw_member(ca, c, dev_iter)
1549                 while (fifo_pop(&ca->free_inc, bu))
1550                         blkdev_issue_discard(ca->disk_sb.bdev,
1551                                              bucket_to_sector(ca, bu),
1552                                              ca->mi.bucket_size, GFP_NOIO, 0);
1553 }
1554
1555 static int resize_free_inc(struct bch_dev *ca)
1556 {
1557         alloc_fifo free_inc;
1558
1559         if (!fifo_full(&ca->free_inc))
1560                 return 0;
1561
1562         if (!init_fifo(&free_inc,
1563                        ca->free_inc.size * 2,
1564                        GFP_KERNEL))
1565                 return -ENOMEM;
1566
1567         fifo_move(&free_inc, &ca->free_inc);
1568         swap(free_inc, ca->free_inc);
1569         free_fifo(&free_inc);
1570         return 0;
1571 }
1572
1573 static bool bch2_fs_allocator_start_fast(struct bch_fs *c)
1574 {
1575         struct bch_dev *ca;
1576         unsigned dev_iter;
1577         bool ret = true;
1578
1579         if (test_alloc_startup(c))
1580                 return false;
1581
1582         down_read(&c->gc_lock);
1583
1584         /* Scan for buckets that are already invalidated: */
1585         for_each_rw_member(ca, c, dev_iter) {
1586                 struct bucket_array *buckets;
1587                 struct bucket_mark m;
1588                 long bu;
1589
1590                 down_read(&ca->bucket_lock);
1591                 buckets = bucket_array(ca);
1592
1593                 for (bu = buckets->first_bucket;
1594                      bu < buckets->nbuckets; bu++) {
1595                         m = READ_ONCE(buckets->b[bu].mark);
1596
1597                         if (!buckets->b[bu].gen_valid ||
1598                             !is_available_bucket(m) ||
1599                             m.cached_sectors ||
1600                             (ca->buckets_nouse &&
1601                              test_bit(bu, ca->buckets_nouse)))
1602                                 continue;
1603
1604                         percpu_down_read_preempt_disable(&c->mark_lock);
1605                         bch2_mark_alloc_bucket(c, ca, bu, true,
1606                                         gc_pos_alloc(c, NULL), 0);
1607                         percpu_up_read_preempt_enable(&c->mark_lock);
1608
1609                         fifo_push(&ca->free_inc, bu);
1610
1611                         discard_invalidated_buckets(c, ca);
1612
1613                         if (fifo_full(&ca->free[RESERVE_BTREE]))
1614                                 break;
1615                 }
1616                 up_read(&ca->bucket_lock);
1617         }
1618
1619         up_read(&c->gc_lock);
1620
1621         /* did we find enough buckets? */
1622         for_each_rw_member(ca, c, dev_iter)
1623                 if (!fifo_full(&ca->free[RESERVE_BTREE]))
1624                         ret = false;
1625
1626         return ret;
1627 }
1628
1629 int bch2_fs_allocator_start(struct bch_fs *c)
1630 {
1631         struct bch_dev *ca;
1632         unsigned dev_iter;
1633         u64 journal_seq = 0;
1634         bool wrote;
1635         long bu;
1636         int ret = 0;
1637
1638         if (!test_alloc_startup(c) &&
1639             bch2_fs_allocator_start_fast(c))
1640                 return 0;
1641
1642         pr_debug("not enough empty buckets; scanning for reclaimable buckets");
1643
1644         /*
1645          * We're moving buckets to freelists _before_ they've been marked as
1646          * invalidated on disk - we have to so that we can allocate new btree
1647          * nodes to mark them as invalidated on disk.
1648          *
1649          * However, we can't _write_ to any of these buckets yet - they might
1650          * have cached data in them, which is live until they're marked as
1651          * invalidated on disk:
1652          */
1653         set_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags);
1654
1655         down_read(&c->gc_lock);
1656         do {
1657                 wrote = false;
1658
1659                 for_each_rw_member(ca, c, dev_iter) {
1660                         find_reclaimable_buckets(c, ca);
1661
1662                         while (!fifo_full(&ca->free[RESERVE_BTREE]) &&
1663                                (bu = next_alloc_bucket(ca)) >= 0) {
1664                                 ret = resize_free_inc(ca);
1665                                 if (ret) {
1666                                         percpu_ref_put(&ca->io_ref);
1667                                         up_read(&c->gc_lock);
1668                                         goto err;
1669                                 }
1670
1671                                 bch2_invalidate_one_bucket(c, ca, bu,
1672                                                            &journal_seq);
1673
1674                                 fifo_push(&ca->free[RESERVE_BTREE], bu);
1675                         }
1676                 }
1677
1678                 pr_debug("done scanning for reclaimable buckets");
1679
1680                 /*
1681                  * XXX: it's possible for this to deadlock waiting on journal reclaim,
1682                  * since we're holding btree writes. What then?
1683                  */
1684                 ret = bch2_alloc_write(c, true, &wrote);
1685
1686                 /*
1687                  * If bch2_alloc_write() did anything, it may have used some
1688                  * buckets, and we need the RESERVE_BTREE freelist full - so we
1689                  * need to loop and scan again.
1690                  * And if it errored, it may have been because there weren't
1691                  * enough buckets, so just scan and loop again as long as it
1692                  * made some progress:
1693                  */
1694         } while (wrote);
1695         up_read(&c->gc_lock);
1696
1697         if (ret)
1698                 goto err;
1699
1700         pr_debug("flushing journal");
1701
1702         ret = bch2_journal_flush(&c->journal);
1703         if (ret)
1704                 goto err;
1705
1706         pr_debug("issuing discards");
1707         allocator_start_issue_discards(c);
1708 err:
1709         clear_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags);
1710         closure_wait_event(&c->btree_interior_update_wait,
1711                            flush_held_btree_writes(c));
1712
1713         return ret;
1714 }
1715
1716 void bch2_fs_allocator_background_init(struct bch_fs *c)
1717 {
1718         spin_lock_init(&c->freelist_lock);
1719         bch2_bucket_clock_init(c, READ);
1720         bch2_bucket_clock_init(c, WRITE);
1721
1722         c->pd_controllers_update_seconds = 5;
1723         INIT_DELAYED_WORK(&c->pd_controllers_update, pd_controllers_update);
1724 }