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 static int __bch2_alloc_write_key(struct bch_fs *c, struct bch_dev *ca,
 313                                   size_t b, struct btree_iter *iter,
 314                                   u64 *journal_seq, unsigned flags)
 315 {
 316 #if 0
 317         __BKEY_PADDED(k, BKEY_ALLOC_VAL_U64s_MAX) alloc_key;
 318 #else
 319         /* hack: */
 320         __BKEY_PADDED(k, 8) alloc_key;
 321 #endif
 322         struct bkey_i_alloc *a = bkey_alloc_init(&alloc_key.k);
 323         struct bucket *g;
 324         struct bucket_mark m, new;
 325         int ret;
 326
 327         BUG_ON(BKEY_ALLOC_VAL_U64s_MAX > 8);
 328
 329         a->k.p = POS(ca->dev_idx, b);
 330
 331         bch2_btree_iter_set_pos(iter, a->k.p);
 332
 333         ret = bch2_btree_iter_traverse(iter);
 334         if (ret)
 335                 return ret;
 336
 337         percpu_down_read_preempt_disable(&c->mark_lock);
 338         g = bucket(ca, b);
 339         m = READ_ONCE(g->mark);
 340
 341         if (!m.dirty) {
 342                 percpu_up_read_preempt_enable(&c->mark_lock);
 343                 return 0;
 344         }
 345
 346         __alloc_write_key(a, g, m);
 347         percpu_up_read_preempt_enable(&c->mark_lock);
 348
 349         bch2_btree_iter_cond_resched(iter);
 350
 351         ret = bch2_btree_insert_at(c, NULL, journal_seq,
 352                                    BTREE_INSERT_NOCHECK_RW|
 353                                    BTREE_INSERT_NOFAIL|
 354                                    BTREE_INSERT_USE_RESERVE|
 355                                    BTREE_INSERT_USE_ALLOC_RESERVE|
 356                                    BTREE_INSERT_NOMARK|
 357                                    flags,
 358                                    BTREE_INSERT_ENTRY(iter, &a->k_i));
 359         if (ret)
 360                 return ret;
 361
 362         new = m;
 363         new.dirty = false;
 364         atomic64_cmpxchg(&g->_mark.v, m.v.counter, new.v.counter);
 365
 366         if (ca->buckets_written)
 367                 set_bit(b, ca->buckets_written);
 368
 369         return 0;
 370 }
 371
 372 int bch2_alloc_replay_key(struct bch_fs *c, struct bkey_i *k)
 373 {
 374         struct bch_dev *ca;
 375         struct btree_iter iter;
 376         int ret;
 377
 378         if (k->k.p.inode >= c->sb.nr_devices ||
 379             !c->devs[k->k.p.inode])
 380                 return 0;
 381
 382         ca = bch_dev_bkey_exists(c, k->k.p.inode);
 383
 384         if (k->k.p.offset >= ca->mi.nbuckets)
 385                 return 0;
 386
 387         bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, k->k.p,
 388                              BTREE_ITER_INTENT);
 389
 390         ret = bch2_btree_iter_traverse(&iter);
 391         if (ret)
 392                 goto err;
 393
 394         /* check buckets_written with btree node locked: */
 395
 396         ret = test_bit(k->k.p.offset, ca->buckets_written)
 397                 ? 0
 398                 : bch2_btree_insert_at(c, NULL, NULL,
 399                                        BTREE_INSERT_NOFAIL|
 400                                        BTREE_INSERT_JOURNAL_REPLAY|
 401                                        BTREE_INSERT_NOMARK,
 402                                        BTREE_INSERT_ENTRY(&iter, k));
 403 err:
 404         bch2_btree_iter_unlock(&iter);
 405         return ret;
 406 }
 407
 408 int bch2_alloc_write(struct bch_fs *c, bool nowait, bool *wrote)
 409 {
 410         struct bch_dev *ca;
 411         unsigned i;
 412         int ret = 0;
 413
 414         *wrote = false;
 415
 416         for_each_rw_member(ca, c, i) {
 417                 struct btree_iter iter;
 418                 struct bucket_array *buckets;
 419                 size_t b;
 420
 421                 bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, POS_MIN,
 422                                      BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 423
 424                 down_read(&ca->bucket_lock);
 425                 buckets = bucket_array(ca);
 426
 427                 for (b = buckets->first_bucket;
 428                      b < buckets->nbuckets;
 429                      b++) {
 430                         if (!buckets->b[b].mark.dirty)
 431                                 continue;
 432
 433                         ret = __bch2_alloc_write_key(c, ca, b, &iter, NULL,
 434                                                      nowait
 435                                                      ? BTREE_INSERT_NOWAIT
 436                                                      : 0);
 437                         if (ret)
 438                                 break;
 439
 440                         *wrote = true;
 441                 }
 442                 up_read(&ca->bucket_lock);
 443                 bch2_btree_iter_unlock(&iter);
 444
 445                 if (ret) {
 446                         percpu_ref_put(&ca->io_ref);
 447                         break;
 448                 }
 449         }
 450
 451         return ret;
 452 }
 453
 454 /* Bucket IO clocks: */
 455
 456 static void bch2_recalc_oldest_io(struct bch_fs *c, struct bch_dev *ca, int rw)
 457 {
 458         struct bucket_clock *clock = &c->bucket_clock[rw];
 459         struct bucket_array *buckets = bucket_array(ca);
 460         struct bucket *g;
 461         u16 max_last_io = 0;
 462         unsigned i;
 463
 464         lockdep_assert_held(&c->bucket_clock[rw].lock);
 465
 466         /* Recalculate max_last_io for this device: */
 467         for_each_bucket(g, buckets)
 468                 max_last_io = max(max_last_io, bucket_last_io(c, g, rw));
 469
 470         ca->max_last_bucket_io[rw] = max_last_io;
 471
 472         /* Recalculate global max_last_io: */
 473         max_last_io = 0;
 474
 475         for_each_member_device(ca, c, i)
 476                 max_last_io = max(max_last_io, ca->max_last_bucket_io[rw]);
 477
 478         clock->max_last_io = max_last_io;
 479 }
 480
 481 static void bch2_rescale_bucket_io_times(struct bch_fs *c, int rw)
 482 {
 483         struct bucket_clock *clock = &c->bucket_clock[rw];
 484         struct bucket_array *buckets;
 485         struct bch_dev *ca;
 486         struct bucket *g;
 487         unsigned i;
 488
 489         trace_rescale_prios(c);
 490
 491         for_each_member_device(ca, c, i) {
 492                 down_read(&ca->bucket_lock);
 493                 buckets = bucket_array(ca);
 494
 495                 for_each_bucket(g, buckets)
 496                         g->io_time[rw] = clock->hand -
 497                         bucket_last_io(c, g, rw) / 2;
 498
 499                 bch2_recalc_oldest_io(c, ca, rw);
 500
 501                 up_read(&ca->bucket_lock);
 502         }
 503 }
 504
 505 static inline u64 bucket_clock_freq(u64 capacity)
 506 {
 507         return max(capacity >> 10, 2028ULL);
 508 }
 509
 510 static void bch2_inc_clock_hand(struct io_timer *timer)
 511 {
 512         struct bucket_clock *clock = container_of(timer,
 513                                                 struct bucket_clock, rescale);
 514         struct bch_fs *c = container_of(clock,
 515                                         struct bch_fs, bucket_clock[clock->rw]);
 516         struct bch_dev *ca;
 517         u64 capacity;
 518         unsigned i;
 519
 520         mutex_lock(&clock->lock);
 521
 522         /* if clock cannot be advanced more, rescale prio */
 523         if (clock->max_last_io >= U16_MAX - 2)
 524                 bch2_rescale_bucket_io_times(c, clock->rw);
 525
 526         BUG_ON(clock->max_last_io >= U16_MAX - 2);
 527
 528         for_each_member_device(ca, c, i)
 529                 ca->max_last_bucket_io[clock->rw]++;
 530         clock->max_last_io++;
 531         clock->hand++;
 532
 533         mutex_unlock(&clock->lock);
 534
 535         capacity = READ_ONCE(c->capacity);
 536
 537         if (!capacity)
 538                 return;
 539
 540         /*
 541          * we only increment when 0.1% of the filesystem capacity has been read
 542          * or written too, this determines if it's time
 543          *
 544          * XXX: we shouldn't really be going off of the capacity of devices in
 545          * RW mode (that will be 0 when we're RO, yet we can still service
 546          * reads)
 547          */
 548         timer->expire += bucket_clock_freq(capacity);
 549
 550         bch2_io_timer_add(&c->io_clock[clock->rw], timer);
 551 }
 552
 553 static void bch2_bucket_clock_init(struct bch_fs *c, int rw)
 554 {
 555         struct bucket_clock *clock = &c->bucket_clock[rw];
 556
 557         clock->hand             = 1;
 558         clock->rw               = rw;
 559         clock->rescale.fn       = bch2_inc_clock_hand;
 560         clock->rescale.expire   = bucket_clock_freq(c->capacity);
 561         mutex_init(&clock->lock);
 562 }
 563
 564 /* Background allocator thread: */
 565
 566 /*
 567  * Scans for buckets to be invalidated, invalidates them, rewrites prios/gens
 568  * (marking them as invalidated on disk), then optionally issues discard
 569  * commands to the newly free buckets, then puts them on the various freelists.
 570  */
 571
 572 #define BUCKET_GC_GEN_MAX       96U
 573
 574 /**
 575  * wait_buckets_available - wait on reclaimable buckets
 576  *
 577  * If there aren't enough available buckets to fill up free_inc, wait until
 578  * there are.
 579  */
 580 static int wait_buckets_available(struct bch_fs *c, struct bch_dev *ca)
 581 {
 582         unsigned long gc_count = c->gc_count;
 583         int ret = 0;
 584
 585         while (1) {
 586                 set_current_state(TASK_INTERRUPTIBLE);
 587                 if (kthread_should_stop()) {
 588                         ret = 1;
 589                         break;
 590                 }
 591
 592                 if (gc_count != c->gc_count)
 593                         ca->inc_gen_really_needs_gc = 0;
 594
 595                 if ((ssize_t) (dev_buckets_available(c, ca) -
 596                                ca->inc_gen_really_needs_gc) >=
 597                     (ssize_t) fifo_free(&ca->free_inc))
 598                         break;
 599
 600                 up_read(&c->gc_lock);
 601                 schedule();
 602                 try_to_freeze();
 603                 down_read(&c->gc_lock);
 604         }
 605
 606         __set_current_state(TASK_RUNNING);
 607         return ret;
 608 }
 609
 610 static bool bch2_can_invalidate_bucket(struct bch_dev *ca,
 611                                        size_t bucket,
 612                                        struct bucket_mark mark)
 613 {
 614         u8 gc_gen;
 615
 616         if (!is_available_bucket(mark))
 617                 return false;
 618
 619         if (ca->buckets_nouse &&
 620             test_bit(bucket, ca->buckets_nouse))
 621                 return false;
 622
 623         gc_gen = bucket_gc_gen(ca, bucket);
 624
 625         if (gc_gen >= BUCKET_GC_GEN_MAX / 2)
 626                 ca->inc_gen_needs_gc++;
 627
 628         if (gc_gen >= BUCKET_GC_GEN_MAX)
 629                 ca->inc_gen_really_needs_gc++;
 630
 631         return gc_gen < BUCKET_GC_GEN_MAX;
 632 }
 633
 634 /*
 635  * Determines what order we're going to reuse buckets, smallest bucket_key()
 636  * first.
 637  *
 638  *
 639  * - We take into account the read prio of the bucket, which gives us an
 640  *   indication of how hot the data is -- we scale the prio so that the prio
 641  *   farthest from the clock is worth 1/8th of the closest.
 642  *
 643  * - The number of sectors of cached data in the bucket, which gives us an
 644  *   indication of the cost in cache misses this eviction will cause.
 645  *
 646  * - If hotness * sectors used compares equal, we pick the bucket with the
 647  *   smallest bucket_gc_gen() - since incrementing the same bucket's generation
 648  *   number repeatedly forces us to run mark and sweep gc to avoid generation
 649  *   number wraparound.
 650  */
 651
 652 static unsigned long bucket_sort_key(struct bch_fs *c, struct bch_dev *ca,
 653                                      size_t b, struct bucket_mark m)
 654 {
 655         unsigned last_io = bucket_last_io(c, bucket(ca, b), READ);
 656         unsigned max_last_io = ca->max_last_bucket_io[READ];
 657
 658         /*
 659          * Time since last read, scaled to [0, 8) where larger value indicates
 660          * more recently read data:
 661          */
 662         unsigned long hotness = (max_last_io - last_io) * 7 / max_last_io;
 663
 664         /* How much we want to keep the data in this bucket: */
 665         unsigned long data_wantness =
 666                 (hotness + 1) * bucket_sectors_used(m);
 667
 668         unsigned long needs_journal_commit =
 669                 bucket_needs_journal_commit(m, c->journal.last_seq_ondisk);
 670
 671         return  (data_wantness << 9) |
 672                 (needs_journal_commit << 8) |
 673                 (bucket_gc_gen(ca, b) / 16);
 674 }
 675
 676 static inline int bucket_alloc_cmp(alloc_heap *h,
 677                                    struct alloc_heap_entry l,
 678                                    struct alloc_heap_entry r)
 679 {
 680         return (l.key > r.key) - (l.key < r.key) ?:
 681                 (l.nr < r.nr)  - (l.nr  > r.nr) ?:
 682                 (l.bucket > r.bucket) - (l.bucket < r.bucket);
 683 }
 684
 685 static inline int bucket_idx_cmp(const void *_l, const void *_r)
 686 {
 687         const struct alloc_heap_entry *l = _l, *r = _r;
 688
 689         return (l->bucket > r->bucket) - (l->bucket < r->bucket);
 690 }
 691
 692 static void find_reclaimable_buckets_lru(struct bch_fs *c, struct bch_dev *ca)
 693 {
 694         struct bucket_array *buckets;
 695         struct alloc_heap_entry e = { 0 };
 696         size_t b, i, nr = 0;
 697
 698         ca->alloc_heap.used = 0;
 699
 700         mutex_lock(&c->bucket_clock[READ].lock);
 701         down_read(&ca->bucket_lock);
 702
 703         buckets = bucket_array(ca);
 704
 705         bch2_recalc_oldest_io(c, ca, READ);
 706
 707         /*
 708          * Find buckets with lowest read priority, by building a maxheap sorted
 709          * by read priority and repeatedly replacing the maximum element until
 710          * all buckets have been visited.
 711          */
 712         for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++) {
 713                 struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
 714                 unsigned long key = bucket_sort_key(c, ca, b, m);
 715
 716                 if (!bch2_can_invalidate_bucket(ca, b, m))
 717                         continue;
 718
 719                 if (e.nr && e.bucket + e.nr == b && e.key == key) {
 720                         e.nr++;
 721                 } else {
 722                         if (e.nr)
 723                                 heap_add_or_replace(&ca->alloc_heap, e,
 724                                         -bucket_alloc_cmp, NULL);
 725
 726                         e = (struct alloc_heap_entry) {
 727                                 .bucket = b,
 728                                 .nr     = 1,
 729                                 .key    = key,
 730                         };
 731                 }
 732
 733                 cond_resched();
 734         }
 735
 736         if (e.nr)
 737                 heap_add_or_replace(&ca->alloc_heap, e,
 738                                 -bucket_alloc_cmp, NULL);
 739
 740         for (i = 0; i < ca->alloc_heap.used; i++)
 741                 nr += ca->alloc_heap.data[i].nr;
 742
 743         while (nr - ca->alloc_heap.data[0].nr >= ALLOC_SCAN_BATCH(ca)) {
 744                 nr -= ca->alloc_heap.data[0].nr;
 745                 heap_pop(&ca->alloc_heap, e, -bucket_alloc_cmp, NULL);
 746         }
 747
 748         up_read(&ca->bucket_lock);
 749         mutex_unlock(&c->bucket_clock[READ].lock);
 750 }
 751
 752 static void find_reclaimable_buckets_fifo(struct bch_fs *c, struct bch_dev *ca)
 753 {
 754         struct bucket_array *buckets = bucket_array(ca);
 755         struct bucket_mark m;
 756         size_t b, start;
 757
 758         if (ca->fifo_last_bucket <  ca->mi.first_bucket ||
 759             ca->fifo_last_bucket >= ca->mi.nbuckets)
 760                 ca->fifo_last_bucket = ca->mi.first_bucket;
 761
 762         start = ca->fifo_last_bucket;
 763
 764         do {
 765                 ca->fifo_last_bucket++;
 766                 if (ca->fifo_last_bucket == ca->mi.nbuckets)
 767                         ca->fifo_last_bucket = ca->mi.first_bucket;
 768
 769                 b = ca->fifo_last_bucket;
 770                 m = READ_ONCE(buckets->b[b].mark);
 771
 772                 if (bch2_can_invalidate_bucket(ca, b, m)) {
 773                         struct alloc_heap_entry e = { .bucket = b, .nr = 1, };
 774
 775                         heap_add(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
 776                         if (heap_full(&ca->alloc_heap))
 777                                 break;
 778                 }
 779
 780                 cond_resched();
 781         } while (ca->fifo_last_bucket != start);
 782 }
 783
 784 static void find_reclaimable_buckets_random(struct bch_fs *c, struct bch_dev *ca)
 785 {
 786         struct bucket_array *buckets = bucket_array(ca);
 787         struct bucket_mark m;
 788         size_t checked, i;
 789
 790         for (checked = 0;
 791              checked < ca->mi.nbuckets / 2;
 792              checked++) {
 793                 size_t b = bch2_rand_range(ca->mi.nbuckets -
 794                                            ca->mi.first_bucket) +
 795                         ca->mi.first_bucket;
 796
 797                 m = READ_ONCE(buckets->b[b].mark);
 798
 799                 if (bch2_can_invalidate_bucket(ca, b, m)) {
 800                         struct alloc_heap_entry e = { .bucket = b, .nr = 1, };
 801
 802                         heap_add(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
 803                         if (heap_full(&ca->alloc_heap))
 804                                 break;
 805                 }
 806
 807                 cond_resched();
 808         }
 809
 810         sort(ca->alloc_heap.data,
 811              ca->alloc_heap.used,
 812              sizeof(ca->alloc_heap.data[0]),
 813              bucket_idx_cmp, NULL);
 814
 815         /* remove duplicates: */
 816         for (i = 0; i + 1 < ca->alloc_heap.used; i++)
 817                 if (ca->alloc_heap.data[i].bucket ==
 818                     ca->alloc_heap.data[i + 1].bucket)
 819                         ca->alloc_heap.data[i].nr = 0;
 820 }
 821
 822 static size_t find_reclaimable_buckets(struct bch_fs *c, struct bch_dev *ca)
 823 {
 824         size_t i, nr = 0;
 825
 826         ca->inc_gen_needs_gc                    = 0;
 827
 828         switch (ca->mi.replacement) {
 829         case CACHE_REPLACEMENT_LRU:
 830                 find_reclaimable_buckets_lru(c, ca);
 831                 break;
 832         case CACHE_REPLACEMENT_FIFO:
 833                 find_reclaimable_buckets_fifo(c, ca);
 834                 break;
 835         case CACHE_REPLACEMENT_RANDOM:
 836                 find_reclaimable_buckets_random(c, ca);
 837                 break;
 838         }
 839
 840         heap_resort(&ca->alloc_heap, bucket_alloc_cmp, NULL);
 841
 842         for (i = 0; i < ca->alloc_heap.used; i++)
 843                 nr += ca->alloc_heap.data[i].nr;
 844
 845         return nr;
 846 }
 847
 848 static inline long next_alloc_bucket(struct bch_dev *ca)
 849 {
 850         struct alloc_heap_entry e, *top = ca->alloc_heap.data;
 851
 852         while (ca->alloc_heap.used) {
 853                 if (top->nr) {
 854                         size_t b = top->bucket;
 855
 856                         top->bucket++;
 857                         top->nr--;
 858                         return b;
 859                 }
 860
 861                 heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
 862         }
 863
 864         return -1;
 865 }
 866
 867 /*
 868  * returns sequence number of most recent journal entry that updated this
 869  * bucket:
 870  */
 871 static u64 bucket_journal_seq(struct bch_fs *c, struct bucket_mark m)
 872 {
 873         if (m.journal_seq_valid) {
 874                 u64 journal_seq = atomic64_read(&c->journal.seq);
 875                 u64 bucket_seq  = journal_seq;
 876
 877                 bucket_seq &= ~((u64) U16_MAX);
 878                 bucket_seq |= m.journal_seq;
 879
 880                 if (bucket_seq > journal_seq)
 881                         bucket_seq -= 1 << 16;
 882
 883                 return bucket_seq;
 884         } else {
 885                 return 0;
 886         }
 887 }
 888
 889 static int bch2_invalidate_one_bucket2(struct bch_fs *c, struct bch_dev *ca,
 890                                        struct btree_iter *iter,
 891                                        u64 *journal_seq, unsigned flags)
 892 {
 893 #if 0
 894         __BKEY_PADDED(k, BKEY_ALLOC_VAL_U64s_MAX) alloc_key;
 895 #else
 896         /* hack: */
 897         __BKEY_PADDED(k, 8) alloc_key;
 898 #endif
 899         struct bkey_i_alloc *a;
 900         struct bkey_alloc_unpacked u;
 901         struct bucket_mark m;
 902         struct bkey_s_c k;
 903         bool invalidating_cached_data;
 904         size_t b;
 905         int ret;
 906
 907         BUG_ON(!ca->alloc_heap.used ||
 908                !ca->alloc_heap.data[0].nr);
 909         b = ca->alloc_heap.data[0].bucket;
 910
 911         /* first, put on free_inc and mark as owned by allocator: */
 912         percpu_down_read_preempt_disable(&c->mark_lock);
 913         spin_lock(&c->freelist_lock);
 914
 915         verify_not_on_freelist(c, ca, b);
 916
 917         BUG_ON(!fifo_push(&ca->free_inc, b));
 918
 919         bch2_mark_alloc_bucket(c, ca, b, true, gc_pos_alloc(c, NULL), 0);
 920         m = bucket(ca, b)->mark;
 921
 922         spin_unlock(&c->freelist_lock);
 923         percpu_up_read_preempt_enable(&c->mark_lock);
 924
 925         bch2_btree_iter_cond_resched(iter);
 926
 927         BUG_ON(BKEY_ALLOC_VAL_U64s_MAX > 8);
 928
 929         bch2_btree_iter_set_pos(iter, POS(ca->dev_idx, b));
 930 retry:
 931         k = bch2_btree_iter_peek_slot(iter);
 932         ret = btree_iter_err(k);
 933         if (ret)
 934                 return ret;
 935
 936         if (k.k && k.k->type == KEY_TYPE_alloc)
 937                 u = bch2_alloc_unpack(bkey_s_c_to_alloc(k).v);
 938         else
 939                 memset(&u, 0, sizeof(u));
 940
 941         invalidating_cached_data = u.cached_sectors != 0;
 942
 943         //BUG_ON(u.dirty_sectors);
 944         u.data_type     = 0;
 945         u.dirty_sectors = 0;
 946         u.cached_sectors = 0;
 947         u.read_time     = c->bucket_clock[READ].hand;
 948         u.write_time    = c->bucket_clock[WRITE].hand;
 949         u.gen++;
 950
 951         a = bkey_alloc_init(&alloc_key.k);
 952         a->k.p = iter->pos;
 953         bch2_alloc_pack(a, u);
 954
 955         ret = bch2_btree_insert_at(c, NULL,
 956                         invalidating_cached_data ? journal_seq : NULL,
 957                         BTREE_INSERT_ATOMIC|
 958                         BTREE_INSERT_NOCHECK_RW|
 959                         BTREE_INSERT_NOFAIL|
 960                         BTREE_INSERT_USE_RESERVE|
 961                         BTREE_INSERT_USE_ALLOC_RESERVE|
 962                         flags,
 963                         BTREE_INSERT_ENTRY(iter, &a->k_i));
 964         if (ret == -EINTR)
 965                 goto retry;
 966
 967         if (!ret) {
 968                 /* remove from alloc_heap: */
 969                 struct alloc_heap_entry e, *top = ca->alloc_heap.data;
 970
 971                 top->bucket++;
 972                 top->nr--;
 973
 974                 if (!top->nr)
 975                         heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
 976
 977                 /*
 978                  * Make sure we flush the last journal entry that updated this
 979                  * bucket (i.e. deleting the last reference) before writing to
 980                  * this bucket again:
 981                  */
 982                 *journal_seq = max(*journal_seq, bucket_journal_seq(c, m));
 983         } else {
 984                 size_t b2;
 985
 986                 /* remove from free_inc: */
 987                 percpu_down_read_preempt_disable(&c->mark_lock);
 988                 spin_lock(&c->freelist_lock);
 989
 990                 bch2_mark_alloc_bucket(c, ca, b, false,
 991                                        gc_pos_alloc(c, NULL), 0);
 992
 993                 BUG_ON(!fifo_pop_back(&ca->free_inc, b2));
 994                 BUG_ON(b != b2);
 995
 996                 spin_unlock(&c->freelist_lock);
 997                 percpu_up_read_preempt_enable(&c->mark_lock);
 998         }
 999
1000         return ret;
1001 }
1002
1003 static bool bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca,
1004                                        size_t bucket, u64 *flush_seq)
1005 {
1006         struct bucket_mark m;
1007
1008         percpu_down_read_preempt_disable(&c->mark_lock);
1009         spin_lock(&c->freelist_lock);
1010
1011         bch2_invalidate_bucket(c, ca, bucket, &m);
1012
1013         verify_not_on_freelist(c, ca, bucket);
1014         BUG_ON(!fifo_push(&ca->free_inc, bucket));
1015
1016         spin_unlock(&c->freelist_lock);
1017
1018         bucket_io_clock_reset(c, ca, bucket, READ);
1019         bucket_io_clock_reset(c, ca, bucket, WRITE);
1020
1021         percpu_up_read_preempt_enable(&c->mark_lock);
1022
1023         *flush_seq = max(*flush_seq, bucket_journal_seq(c, m));
1024
1025         return m.cached_sectors != 0;
1026 }
1027
1028 /*
1029  * Pull buckets off ca->alloc_heap, invalidate them, move them to ca->free_inc:
1030  */
1031 static int bch2_invalidate_buckets(struct bch_fs *c, struct bch_dev *ca)
1032 {
1033         struct btree_iter iter;
1034         u64 journal_seq = 0;
1035         int ret = 0;
1036
1037         bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, POS(ca->dev_idx, 0),
1038                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
1039
1040         /* Only use nowait if we've already invalidated at least one bucket: */
1041         while (!ret &&
1042                !fifo_full(&ca->free_inc) &&
1043                ca->alloc_heap.used)
1044                 ret = bch2_invalidate_one_bucket2(c, ca, &iter, &journal_seq,
1045                                 BTREE_INSERT_GC_LOCK_HELD|
1046                                 (!fifo_empty(&ca->free_inc)
1047                                  ? BTREE_INSERT_NOWAIT : 0));
1048
1049         bch2_btree_iter_unlock(&iter);
1050
1051         /* If we used NOWAIT, don't return the error: */
1052         if (!fifo_empty(&ca->free_inc))
1053                 ret = 0;
1054         if (ret) {
1055                 bch_err(ca, "error invalidating buckets: %i", ret);
1056                 return ret;
1057         }
1058
1059         if (journal_seq)
1060                 ret = bch2_journal_flush_seq(&c->journal, journal_seq);
1061         if (ret) {
1062                 bch_err(ca, "journal error: %i", ret);
1063                 return ret;
1064         }
1065
1066         return 0;
1067 }
1068
1069 static int push_invalidated_bucket(struct bch_fs *c, struct bch_dev *ca, size_t bucket)
1070 {
1071         unsigned i;
1072         int ret = 0;
1073
1074         while (1) {
1075                 set_current_state(TASK_INTERRUPTIBLE);
1076
1077                 spin_lock(&c->freelist_lock);
1078                 for (i = 0; i < RESERVE_NR; i++)
1079                         if (fifo_push(&ca->free[i], bucket)) {
1080                                 fifo_pop(&ca->free_inc, bucket);
1081
1082                                 closure_wake_up(&c->freelist_wait);
1083                                 ca->allocator_blocked_full = false;
1084
1085                                 spin_unlock(&c->freelist_lock);
1086                                 goto out;
1087                         }
1088
1089                 if (!ca->allocator_blocked_full) {
1090                         ca->allocator_blocked_full = true;
1091                         closure_wake_up(&c->freelist_wait);
1092                 }
1093
1094                 spin_unlock(&c->freelist_lock);
1095
1096                 if ((current->flags & PF_KTHREAD) &&
1097                     kthread_should_stop()) {
1098                         ret = 1;
1099                         break;
1100                 }
1101
1102                 schedule();
1103                 try_to_freeze();
1104         }
1105 out:
1106         __set_current_state(TASK_RUNNING);
1107         return ret;
1108 }
1109
1110 /*
1111  * Pulls buckets off free_inc, discards them (if enabled), then adds them to
1112  * freelists, waiting until there's room if necessary:
1113  */
1114 static int discard_invalidated_buckets(struct bch_fs *c, struct bch_dev *ca)
1115 {
1116         while (!fifo_empty(&ca->free_inc)) {
1117                 size_t bucket = fifo_peek(&ca->free_inc);
1118
1119                 if (ca->mi.discard &&
1120                     blk_queue_discard(bdev_get_queue(ca->disk_sb.bdev)))
1121                         blkdev_issue_discard(ca->disk_sb.bdev,
1122                                              bucket_to_sector(ca, bucket),
1123                                              ca->mi.bucket_size, GFP_NOIO, 0);
1124
1125                 if (push_invalidated_bucket(c, ca, bucket))
1126                         return 1;
1127         }
1128
1129         return 0;
1130 }
1131
1132 /**
1133  * bch_allocator_thread - move buckets from free_inc to reserves
1134  *
1135  * The free_inc FIFO is populated by find_reclaimable_buckets(), and
1136  * the reserves are depleted by bucket allocation. When we run out
1137  * of free_inc, try to invalidate some buckets and write out
1138  * prios and gens.
1139  */
1140 static int bch2_allocator_thread(void *arg)
1141 {
1142         struct bch_dev *ca = arg;
1143         struct bch_fs *c = ca->fs;
1144         size_t nr;
1145         int ret;
1146
1147         set_freezable();
1148
1149         while (1) {
1150                 cond_resched();
1151
1152                 pr_debug("discarding %zu invalidated buckets",
1153                          fifo_used(&ca->free_inc));
1154
1155                 ret = discard_invalidated_buckets(c, ca);
1156                 if (ret)
1157                         goto stop;
1158
1159                 down_read(&c->gc_lock);
1160
1161                 ret = bch2_invalidate_buckets(c, ca);
1162                 if (ret) {
1163                         up_read(&c->gc_lock);
1164                         goto stop;
1165                 }
1166
1167                 if (!fifo_empty(&ca->free_inc)) {
1168                         up_read(&c->gc_lock);
1169                         continue;
1170                 }
1171
1172                 pr_debug("free_inc now empty");
1173
1174                 do {
1175                         /*
1176                          * Find some buckets that we can invalidate, either
1177                          * they're completely unused, or only contain clean data
1178                          * that's been written back to the backing device or
1179                          * another cache tier
1180                          */
1181
1182                         pr_debug("scanning for reclaimable buckets");
1183
1184                         nr = find_reclaimable_buckets(c, ca);
1185
1186                         pr_debug("found %zu buckets", nr);
1187
1188                         trace_alloc_batch(ca, nr, ca->alloc_heap.size);
1189
1190                         if ((ca->inc_gen_needs_gc >= ALLOC_SCAN_BATCH(ca) ||
1191                              ca->inc_gen_really_needs_gc) &&
1192                             c->gc_thread) {
1193                                 atomic_inc(&c->kick_gc);
1194                                 wake_up_process(c->gc_thread);
1195                         }
1196
1197                         /*
1198                          * If we found any buckets, we have to invalidate them
1199                          * before we scan for more - but if we didn't find very
1200                          * many we may want to wait on more buckets being
1201                          * available so we don't spin:
1202                          */
1203                         if (!nr ||
1204                             (nr < ALLOC_SCAN_BATCH(ca) &&
1205                              !fifo_full(&ca->free[RESERVE_MOVINGGC]))) {
1206                                 ca->allocator_blocked = true;
1207                                 closure_wake_up(&c->freelist_wait);
1208
1209                                 ret = wait_buckets_available(c, ca);
1210                                 if (ret) {
1211                                         up_read(&c->gc_lock);
1212                                         goto stop;
1213                                 }
1214                         }
1215                 } while (!nr);
1216
1217                 ca->allocator_blocked = false;
1218                 up_read(&c->gc_lock);
1219
1220                 pr_debug("%zu buckets to invalidate", nr);
1221
1222                 /*
1223                  * alloc_heap is now full of newly-invalidated buckets: next,
1224                  * write out the new bucket gens:
1225                  */
1226         }
1227
1228 stop:
1229         pr_debug("alloc thread stopping (ret %i)", ret);
1230         return 0;
1231 }
1232
1233 /* Startup/shutdown (ro/rw): */
1234
1235 void bch2_recalc_capacity(struct bch_fs *c)
1236 {
1237         struct bch_dev *ca;
1238         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
1239         unsigned bucket_size_max = 0;
1240         unsigned long ra_pages = 0;
1241         unsigned i, j;
1242
1243         lockdep_assert_held(&c->state_lock);
1244
1245         for_each_online_member(ca, c, i) {
1246                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_bdi;
1247
1248                 ra_pages += bdi->ra_pages;
1249         }
1250
1251         bch2_set_ra_pages(c, ra_pages);
1252
1253         for_each_rw_member(ca, c, i) {
1254                 u64 dev_reserve = 0;
1255
1256                 /*
1257                  * We need to reserve buckets (from the number
1258                  * of currently available buckets) against
1259                  * foreground writes so that mainly copygc can
1260                  * make forward progress.
1261                  *
1262                  * We need enough to refill the various reserves
1263                  * from scratch - copygc will use its entire
1264                  * reserve all at once, then run against when
1265                  * its reserve is refilled (from the formerly
1266                  * available buckets).
1267                  *
1268                  * This reserve is just used when considering if
1269                  * allocations for foreground writes must wait -
1270                  * not -ENOSPC calculations.
1271                  */
1272                 for (j = 0; j < RESERVE_NONE; j++)
1273                         dev_reserve += ca->free[j].size;
1274
1275                 dev_reserve += 1;       /* btree write point */
1276                 dev_reserve += 1;       /* copygc write point */
1277                 dev_reserve += 1;       /* rebalance write point */
1278
1279                 dev_reserve *= ca->mi.bucket_size;
1280
1281                 ca->copygc_threshold = dev_reserve;
1282
1283                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
1284                                              ca->mi.first_bucket);
1285
1286                 reserved_sectors += dev_reserve * 2;
1287
1288                 bucket_size_max = max_t(unsigned, bucket_size_max,
1289                                         ca->mi.bucket_size);
1290         }
1291
1292         gc_reserve = c->opts.gc_reserve_bytes
1293                 ? c->opts.gc_reserve_bytes >> 9
1294                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
1295
1296         reserved_sectors = max(gc_reserve, reserved_sectors);
1297
1298         reserved_sectors = min(reserved_sectors, capacity);
1299
1300         c->capacity = capacity - reserved_sectors;
1301
1302         c->bucket_size_max = bucket_size_max;
1303
1304         if (c->capacity) {
1305                 bch2_io_timer_add(&c->io_clock[READ],
1306                                  &c->bucket_clock[READ].rescale);
1307                 bch2_io_timer_add(&c->io_clock[WRITE],
1308                                  &c->bucket_clock[WRITE].rescale);
1309         } else {
1310                 bch2_io_timer_del(&c->io_clock[READ],
1311                                  &c->bucket_clock[READ].rescale);
1312                 bch2_io_timer_del(&c->io_clock[WRITE],
1313                                  &c->bucket_clock[WRITE].rescale);
1314         }
1315
1316         /* Wake up case someone was waiting for buckets */
1317         closure_wake_up(&c->freelist_wait);
1318 }
1319
1320 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
1321 {
1322         struct open_bucket *ob;
1323         bool ret = false;
1324
1325         for (ob = c->open_buckets;
1326              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1327              ob++) {
1328                 spin_lock(&ob->lock);
1329                 if (ob->valid && !ob->on_partial_list &&
1330                     ob->ptr.dev == ca->dev_idx)
1331                         ret = true;
1332                 spin_unlock(&ob->lock);
1333         }
1334
1335         return ret;
1336 }
1337
1338 /* device goes ro: */
1339 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
1340 {
1341         unsigned i;
1342
1343         BUG_ON(ca->alloc_thread);
1344
1345         /* First, remove device from allocation groups: */
1346
1347         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1348                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
1349
1350         /*
1351          * Capacity is calculated based off of devices in allocation groups:
1352          */
1353         bch2_recalc_capacity(c);
1354
1355         /* Next, close write points that point to this device... */
1356         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1357                 bch2_writepoint_stop(c, ca, &c->write_points[i]);
1358
1359         bch2_writepoint_stop(c, ca, &ca->copygc_write_point);
1360         bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
1361         bch2_writepoint_stop(c, ca, &c->btree_write_point);
1362
1363         mutex_lock(&c->btree_reserve_cache_lock);
1364         while (c->btree_reserve_cache_nr) {
1365                 struct btree_alloc *a =
1366                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1367
1368                 bch2_open_buckets_put(c, &a->ob);
1369         }
1370         mutex_unlock(&c->btree_reserve_cache_lock);
1371
1372         while (1) {
1373                 struct open_bucket *ob;
1374
1375                 spin_lock(&c->freelist_lock);
1376                 if (!ca->open_buckets_partial_nr) {
1377                         spin_unlock(&c->freelist_lock);
1378                         break;
1379                 }
1380                 ob = c->open_buckets +
1381                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
1382                 ob->on_partial_list = false;
1383                 spin_unlock(&c->freelist_lock);
1384
1385                 bch2_open_bucket_put(c, ob);
1386         }
1387
1388         bch2_ec_stop_dev(c, ca);
1389
1390         /*
1391          * Wake up threads that were blocked on allocation, so they can notice
1392          * the device can no longer be removed and the capacity has changed:
1393          */
1394         closure_wake_up(&c->freelist_wait);
1395
1396         /*
1397          * journal_res_get() can block waiting for free space in the journal -
1398          * it needs to notice there may not be devices to allocate from anymore:
1399          */
1400         wake_up(&c->journal.wait);
1401
1402         /* Now wait for any in flight writes: */
1403
1404         closure_wait_event(&c->open_buckets_wait,
1405                            !bch2_dev_has_open_write_point(c, ca));
1406 }
1407
1408 /* device goes rw: */
1409 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
1410 {
1411         unsigned i;
1412
1413         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1414                 if (ca->mi.data_allowed & (1 << i))
1415                         set_bit(ca->dev_idx, c->rw_devs[i].d);
1416 }
1417
1418 void bch2_dev_allocator_quiesce(struct bch_fs *c, struct bch_dev *ca)
1419 {
1420         if (ca->alloc_thread)
1421                 closure_wait_event(&c->freelist_wait, ca->allocator_blocked_full);
1422 }
1423
1424 /* stop allocator thread: */
1425 void bch2_dev_allocator_stop(struct bch_dev *ca)
1426 {
1427         struct task_struct *p;
1428
1429         p = rcu_dereference_protected(ca->alloc_thread, 1);
1430         ca->alloc_thread = NULL;
1431
1432         /*
1433          * We need an rcu barrier between setting ca->alloc_thread = NULL and
1434          * the thread shutting down to avoid bch2_wake_allocator() racing:
1435          *
1436          * XXX: it would be better to have the rcu barrier be asynchronous
1437          * instead of blocking us here
1438          */
1439         synchronize_rcu();
1440
1441         if (p) {
1442                 kthread_stop(p);
1443                 put_task_struct(p);
1444         }
1445 }
1446
1447 /* start allocator thread: */
1448 int bch2_dev_allocator_start(struct bch_dev *ca)
1449 {
1450         struct task_struct *p;
1451
1452         /*
1453          * allocator thread already started?
1454          */
1455         if (ca->alloc_thread)
1456                 return 0;
1457
1458         p = kthread_create(bch2_allocator_thread, ca,
1459                            "bch_alloc[%s]", ca->name);
1460         if (IS_ERR(p))
1461                 return PTR_ERR(p);
1462
1463         get_task_struct(p);
1464         rcu_assign_pointer(ca->alloc_thread, p);
1465         wake_up_process(p);
1466         return 0;
1467 }
1468
1469 static bool flush_held_btree_writes(struct bch_fs *c)
1470 {
1471         struct bucket_table *tbl;
1472         struct rhash_head *pos;
1473         struct btree *b;
1474         bool nodes_unwritten;
1475         size_t i;
1476 again:
1477         cond_resched();
1478         nodes_unwritten = false;
1479
1480         rcu_read_lock();
1481         for_each_cached_btree(b, c, tbl, i, pos)
1482                 if (btree_node_need_write(b)) {
1483                         if (btree_node_may_write(b)) {
1484                                 rcu_read_unlock();
1485                                 btree_node_lock_type(c, b, SIX_LOCK_read);
1486                                 bch2_btree_node_write(c, b, SIX_LOCK_read);
1487                                 six_unlock_read(&b->lock);
1488                                 goto again;
1489                         } else {
1490                                 nodes_unwritten = true;
1491                         }
1492                 }
1493         rcu_read_unlock();
1494
1495         if (c->btree_roots_dirty) {
1496                 bch2_journal_meta(&c->journal);
1497                 goto again;
1498         }
1499
1500         return !nodes_unwritten &&
1501                 !bch2_btree_interior_updates_nr_pending(c);
1502 }
1503
1504 static void allocator_start_issue_discards(struct bch_fs *c)
1505 {
1506         struct bch_dev *ca;
1507         unsigned dev_iter;
1508         size_t bu;
1509
1510         for_each_rw_member(ca, c, dev_iter)
1511                 while (fifo_pop(&ca->free_inc, bu))
1512                         blkdev_issue_discard(ca->disk_sb.bdev,
1513                                              bucket_to_sector(ca, bu),
1514                                              ca->mi.bucket_size, GFP_NOIO, 0);
1515 }
1516
1517 static int resize_free_inc(struct bch_dev *ca)
1518 {
1519         alloc_fifo free_inc;
1520
1521         if (!fifo_full(&ca->free_inc))
1522                 return 0;
1523
1524         if (!init_fifo(&free_inc,
1525                        ca->free_inc.size * 2,
1526                        GFP_KERNEL))
1527                 return -ENOMEM;
1528
1529         fifo_move(&free_inc, &ca->free_inc);
1530         swap(free_inc, ca->free_inc);
1531         free_fifo(&free_inc);
1532         return 0;
1533 }
1534
1535 static bool bch2_fs_allocator_start_fast(struct bch_fs *c)
1536 {
1537         struct bch_dev *ca;
1538         unsigned dev_iter;
1539         bool ret = true;
1540
1541         if (test_alloc_startup(c))
1542                 return false;
1543
1544         down_read(&c->gc_lock);
1545
1546         /* Scan for buckets that are already invalidated: */
1547         for_each_rw_member(ca, c, dev_iter) {
1548                 struct bucket_array *buckets;
1549                 struct bucket_mark m;
1550                 long bu;
1551
1552                 down_read(&ca->bucket_lock);
1553                 buckets = bucket_array(ca);
1554
1555                 for (bu = buckets->first_bucket;
1556                      bu < buckets->nbuckets; bu++) {
1557                         m = READ_ONCE(buckets->b[bu].mark);
1558
1559                         if (!buckets->b[bu].gen_valid ||
1560                             !is_available_bucket(m) ||
1561                             m.cached_sectors ||
1562                             (ca->buckets_nouse &&
1563                              test_bit(bu, ca->buckets_nouse)))
1564                                 continue;
1565
1566                         percpu_down_read_preempt_disable(&c->mark_lock);
1567                         bch2_mark_alloc_bucket(c, ca, bu, true,
1568                                         gc_pos_alloc(c, NULL), 0);
1569                         percpu_up_read_preempt_enable(&c->mark_lock);
1570
1571                         fifo_push(&ca->free_inc, bu);
1572
1573                         discard_invalidated_buckets(c, ca);
1574
1575                         if (fifo_full(&ca->free[RESERVE_BTREE]))
1576                                 break;
1577                 }
1578                 up_read(&ca->bucket_lock);
1579         }
1580
1581         up_read(&c->gc_lock);
1582
1583         /* did we find enough buckets? */
1584         for_each_rw_member(ca, c, dev_iter)
1585                 if (!fifo_full(&ca->free[RESERVE_BTREE]))
1586                         ret = false;
1587
1588         return ret;
1589 }
1590
1591 static int __bch2_fs_allocator_start(struct bch_fs *c)
1592 {
1593         struct bch_dev *ca;
1594         unsigned dev_iter;
1595         u64 journal_seq = 0;
1596         bool wrote;
1597         long bu;
1598         int ret = 0;
1599
1600         pr_debug("not enough empty buckets; scanning for reclaimable buckets");
1601
1602         /*
1603          * We're moving buckets to freelists _before_ they've been marked as
1604          * invalidated on disk - we have to so that we can allocate new btree
1605          * nodes to mark them as invalidated on disk.
1606          *
1607          * However, we can't _write_ to any of these buckets yet - they might
1608          * have cached data in them, which is live until they're marked as
1609          * invalidated on disk:
1610          */
1611         set_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags);
1612
1613         down_read(&c->gc_lock);
1614         do {
1615                 wrote = false;
1616
1617                 for_each_rw_member(ca, c, dev_iter) {
1618                         find_reclaimable_buckets(c, ca);
1619
1620                         while (!fifo_full(&ca->free[RESERVE_BTREE]) &&
1621                                (bu = next_alloc_bucket(ca)) >= 0) {
1622                                 ret = resize_free_inc(ca);
1623                                 if (ret) {
1624                                         percpu_ref_put(&ca->io_ref);
1625                                         up_read(&c->gc_lock);
1626                                         goto err;
1627                                 }
1628
1629                                 bch2_invalidate_one_bucket(c, ca, bu,
1630                                                            &journal_seq);
1631
1632                                 fifo_push(&ca->free[RESERVE_BTREE], bu);
1633                         }
1634                 }
1635
1636                 pr_debug("done scanning for reclaimable buckets");
1637
1638                 /*
1639                  * XXX: it's possible for this to deadlock waiting on journal reclaim,
1640                  * since we're holding btree writes. What then?
1641                  */
1642                 ret = bch2_alloc_write(c, true, &wrote);
1643
1644                 /*
1645                  * If bch2_alloc_write() did anything, it may have used some
1646                  * buckets, and we need the RESERVE_BTREE freelist full - so we
1647                  * need to loop and scan again.
1648                  * And if it errored, it may have been because there weren't
1649                  * enough buckets, so just scan and loop again as long as it
1650                  * made some progress:
1651                  */
1652         } while (wrote);
1653         up_read(&c->gc_lock);
1654
1655         if (ret)
1656                 goto err;
1657
1658         pr_debug("flushing journal");
1659
1660         ret = bch2_journal_flush(&c->journal);
1661         if (ret)
1662                 goto err;
1663
1664         pr_debug("issuing discards");
1665         allocator_start_issue_discards(c);
1666 err:
1667         clear_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags);
1668         closure_wait_event(&c->btree_interior_update_wait,
1669                            flush_held_btree_writes(c));
1670
1671         return ret;
1672 }
1673
1674 int bch2_fs_allocator_start(struct bch_fs *c)
1675 {
1676         struct bch_dev *ca;
1677         unsigned i;
1678         int ret;
1679
1680         ret = bch2_fs_allocator_start_fast(c) ? 0 :
1681                 __bch2_fs_allocator_start(c);
1682         if (ret)
1683                 return ret;
1684
1685         set_bit(BCH_FS_ALLOCATOR_STARTED, &c->flags);
1686
1687         for_each_rw_member(ca, c, i) {
1688                 ret = bch2_dev_allocator_start(ca);
1689                 if (ret) {
1690                         percpu_ref_put(&ca->io_ref);
1691                         return ret;
1692                 }
1693         }
1694
1695         set_bit(BCH_FS_ALLOCATOR_RUNNING, &c->flags);
1696         return 0;
1697 }
1698
1699 void bch2_fs_allocator_background_init(struct bch_fs *c)
1700 {
1701         spin_lock_init(&c->freelist_lock);
1702         bch2_bucket_clock_init(c, READ);
1703         bch2_bucket_clock_init(c, WRITE);
1704
1705         c->pd_controllers_update_seconds = 5;
1706         INIT_DELAYED_WORK(&c->pd_controllers_update, pd_controllers_update);
1707 }