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

   1 /*
   2  * Primary bucket allocation code
   3  *
   4  * Copyright 2012 Google, Inc.
   5  *
   6  * Allocation in bcache is done in terms of buckets:
   7  *
   8  * Each bucket has associated an 8 bit gen; this gen corresponds to the gen in
   9  * btree pointers - they must match for the pointer to be considered valid.
  10  *
  11  * Thus (assuming a bucket has no dirty data or metadata in it) we can reuse a
  12  * bucket simply by incrementing its gen.
  13  *
  14  * The gens (along with the priorities; it's really the gens are important but
  15  * the code is named as if it's the priorities) are written in an arbitrary list
  16  * of buckets on disk, with a pointer to them in the journal header.
  17  *
  18  * When we invalidate a bucket, we have to write its new gen to disk and wait
  19  * for that write to complete before we use it - otherwise after a crash we
  20  * could have pointers that appeared to be good but pointed to data that had
  21  * been overwritten.
  22  *
  23  * Since the gens and priorities are all stored contiguously on disk, we can
  24  * batch this up: We fill up the free_inc list with freshly invalidated buckets,
  25  * call prio_write(), and when prio_write() finishes we pull buckets off the
  26  * free_inc list and optionally discard them.
  27  *
  28  * free_inc isn't the only freelist - if it was, we'd often have to sleep while
  29  * priorities and gens were being written before we could allocate. c->free is a
  30  * smaller freelist, and buckets on that list are always ready to be used.
  31  *
  32  * If we've got discards enabled, that happens when a bucket moves from the
  33  * free_inc list to the free list.
  34  *
  35  * It's important to ensure that gens don't wrap around - with respect to
  36  * either the oldest gen in the btree or the gen on disk. This is quite
  37  * difficult to do in practice, but we explicitly guard against it anyways - if
  38  * a bucket is in danger of wrapping around we simply skip invalidating it that
  39  * time around, and we garbage collect or rewrite the priorities sooner than we
  40  * would have otherwise.
  41  *
  42  * bch2_bucket_alloc() allocates a single bucket from a specific device.
  43  *
  44  * bch2_bucket_alloc_set() allocates one or more buckets from different devices
  45  * in a given filesystem.
  46  *
  47  * invalidate_buckets() drives all the processes described above. It's called
  48  * from bch2_bucket_alloc() and a few other places that need to make sure free
  49  * buckets are ready.
  50  *
  51  * invalidate_buckets_(lru|fifo)() find buckets that are available to be
  52  * invalidated, and then invalidate them and stick them on the free_inc list -
  53  * in either lru or fifo order.
  54  */
  55
  56 #include "bcachefs.h"
  57 #include "alloc.h"
  58 #include "btree_update.h"
  59 #include "btree_gc.h"
  60 #include "buckets.h"
  61 #include "checksum.h"
  62 #include "clock.h"
  63 #include "debug.h"
  64 #include "error.h"
  65 #include "extents.h"
  66 #include "io.h"
  67 #include "journal.h"
  68 #include "super-io.h"
  69
  70 #include <linux/blkdev.h>
  71 #include <linux/kthread.h>
  72 #include <linux/math64.h>
  73 #include <linux/random.h>
  74 #include <linux/rculist.h>
  75 #include <linux/rcupdate.h>
  76 #include <linux/sched/task.h>
  77 #include <linux/sort.h>
  78 #include <trace/events/bcachefs.h>
  79
  80 static void bch2_recalc_min_prio(struct bch_fs *, struct bch_dev *, int);
  81
  82 /* Ratelimiting/PD controllers */
  83
  84 static void pd_controllers_update(struct work_struct *work)
  85 {
  86         struct bch_fs *c = container_of(to_delayed_work(work),
  87                                            struct bch_fs,
  88                                            pd_controllers_update);
  89         struct bch_dev *ca;
  90         unsigned i, iter;
  91
  92         /* All units are in bytes */
  93         u64 faster_tiers_size   = 0;
  94         u64 faster_tiers_dirty  = 0;
  95
  96         u64 copygc_can_free     = 0;
  97
  98         rcu_read_lock();
  99         for (i = 0; i < ARRAY_SIZE(c->tiers); i++) {
 100                 bch2_pd_controller_update(&c->tiers[i].pd,
 101                                 div_u64(faster_tiers_size *
 102                                         c->tiering_percent, 100),
 103                                 faster_tiers_dirty,
 104                                 -1);
 105
 106                 for_each_member_device_rcu(ca, c, iter, &c->tiers[i].devs) {
 107                         struct bch_dev_usage stats = bch2_dev_usage_read(c, ca);
 108
 109                         u64 size = bucket_to_sector(ca, ca->mi.nbuckets -
 110                                         ca->mi.first_bucket) << 9;
 111                         u64 dirty = bucket_to_sector(ca,
 112                                         stats.buckets[BCH_DATA_USER]) << 9;
 113                         u64 free = bucket_to_sector(ca,
 114                                         __dev_buckets_free(ca, stats)) << 9;
 115                         /*
 116                          * Bytes of internal fragmentation, which can be
 117                          * reclaimed by copy GC
 118                          */
 119                         s64 fragmented = (bucket_to_sector(ca,
 120                                                 stats.buckets[BCH_DATA_USER] +
 121                                                 stats.buckets[BCH_DATA_CACHED]) -
 122                                           (stats.sectors[BCH_DATA_USER] +
 123                                            stats.sectors[BCH_DATA_CACHED])) << 9;
 124
 125                         fragmented = max(0LL, fragmented);
 126
 127                         bch2_pd_controller_update(&ca->copygc_pd,
 128                                                  free, fragmented, -1);
 129
 130                         faster_tiers_size               += size;
 131                         faster_tiers_dirty              += dirty;
 132
 133                         copygc_can_free                 += fragmented;
 134                 }
 135         }
 136
 137         rcu_read_unlock();
 138
 139         /*
 140          * Throttle foreground writes if tier 0 is running out of free buckets,
 141          * and either tiering or copygc can free up space.
 142          *
 143          * Target will be small if there isn't any work to do - we don't want to
 144          * throttle foreground writes if we currently have all the free space
 145          * we're ever going to have.
 146          *
 147          * Otherwise, if there's work to do, try to keep 20% of tier0 available
 148          * for foreground writes.
 149          */
 150         if (c->fastest_tier)
 151                 copygc_can_free = U64_MAX;
 152
 153         schedule_delayed_work(&c->pd_controllers_update,
 154                               c->pd_controllers_update_seconds * HZ);
 155 }
 156
 157 /* Persistent alloc info: */
 158
 159 static unsigned bch_alloc_val_u64s(const struct bch_alloc *a)
 160 {
 161         unsigned bytes = offsetof(struct bch_alloc, data);
 162
 163         if (a->fields & (1 << BCH_ALLOC_FIELD_READ_TIME))
 164                 bytes += 2;
 165         if (a->fields & (1 << BCH_ALLOC_FIELD_WRITE_TIME))
 166                 bytes += 2;
 167
 168         return DIV_ROUND_UP(bytes, sizeof(u64));
 169 }
 170
 171 static const char *bch2_alloc_invalid(const struct bch_fs *c,
 172                                       struct bkey_s_c 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         switch (k.k->type) {
 179         case BCH_ALLOC: {
 180                 struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 181
 182                 if (bch_alloc_val_u64s(a.v) != bkey_val_u64s(a.k))
 183                         return "incorrect value size";
 184                 break;
 185         }
 186         default:
 187                 return "invalid type";
 188         }
 189
 190         return NULL;
 191 }
 192
 193 static void bch2_alloc_to_text(struct bch_fs *c, char *buf,
 194                                size_t size, struct bkey_s_c k)
 195 {
 196         buf[0] = '\0';
 197
 198         switch (k.k->type) {
 199         case BCH_ALLOC:
 200                 break;
 201         }
 202 }
 203
 204 const struct bkey_ops bch2_bkey_alloc_ops = {
 205         .key_invalid    = bch2_alloc_invalid,
 206         .val_to_text    = bch2_alloc_to_text,
 207 };
 208
 209 static inline unsigned get_alloc_field(const u8 **p, unsigned bytes)
 210 {
 211         unsigned v;
 212
 213         switch (bytes) {
 214         case 1:
 215                 v = **p;
 216                 break;
 217         case 2:
 218                 v = le16_to_cpup((void *) *p);
 219                 break;
 220         case 4:
 221                 v = le32_to_cpup((void *) *p);
 222                 break;
 223         default:
 224                 BUG();
 225         }
 226
 227         *p += bytes;
 228         return v;
 229 }
 230
 231 static inline void put_alloc_field(u8 **p, unsigned bytes, unsigned v)
 232 {
 233         switch (bytes) {
 234         case 1:
 235                 **p = v;
 236                 break;
 237         case 2:
 238                 *((__le16 *) *p) = cpu_to_le16(v);
 239                 break;
 240         case 4:
 241                 *((__le32 *) *p) = cpu_to_le32(v);
 242                 break;
 243         default:
 244                 BUG();
 245         }
 246
 247         *p += bytes;
 248 }
 249
 250 static void bch2_alloc_read_key(struct bch_fs *c, struct bkey_s_c k)
 251 {
 252         struct bch_dev *ca;
 253         struct bkey_s_c_alloc a;
 254         struct bucket_mark new;
 255         struct bucket *g;
 256         const u8 *d;
 257
 258         if (k.k->type != BCH_ALLOC)
 259                 return;
 260
 261         a = bkey_s_c_to_alloc(k);
 262         ca = bch_dev_bkey_exists(c, a.k->p.inode);
 263
 264         if (a.k->p.offset >= ca->mi.nbuckets)
 265                 return;
 266
 267         lg_local_lock(&c->usage_lock);
 268
 269         g = bucket(ca, a.k->p.offset);
 270         bucket_cmpxchg(g, new, ({
 271                 new.gen = a.v->gen;
 272                 new.gen_valid = 1;
 273         }));
 274
 275         d = a.v->data;
 276         if (a.v->fields & (1 << BCH_ALLOC_FIELD_READ_TIME))
 277                 g->prio[READ] = get_alloc_field(&d, 2);
 278         if (a.v->fields & (1 << BCH_ALLOC_FIELD_WRITE_TIME))
 279                 g->prio[WRITE] = get_alloc_field(&d, 2);
 280
 281         lg_local_unlock(&c->usage_lock);
 282 }
 283
 284 int bch2_alloc_read(struct bch_fs *c, struct list_head *journal_replay_list)
 285 {
 286         struct journal_replay *r;
 287         struct btree_iter iter;
 288         struct bkey_s_c k;
 289         struct bch_dev *ca;
 290         unsigned i;
 291         int ret;
 292
 293         if (!c->btree_roots[BTREE_ID_ALLOC].b)
 294                 return 0;
 295
 296         for_each_btree_key(&iter, c, BTREE_ID_ALLOC, POS_MIN, 0, k) {
 297                 bch2_alloc_read_key(c, k);
 298                 bch2_btree_iter_cond_resched(&iter);
 299         }
 300
 301         ret = bch2_btree_iter_unlock(&iter);
 302         if (ret)
 303                 return ret;
 304
 305         list_for_each_entry(r, journal_replay_list, list) {
 306                 struct bkey_i *k, *n;
 307                 struct jset_entry *entry;
 308
 309                 for_each_jset_key(k, n, entry, &r->j)
 310                         if (entry->btree_id == BTREE_ID_ALLOC)
 311                                 bch2_alloc_read_key(c, bkey_i_to_s_c(k));
 312         }
 313
 314         mutex_lock(&c->prio_clock[READ].lock);
 315         for_each_member_device(ca, c, i) {
 316                 down_read(&ca->bucket_lock);
 317                 bch2_recalc_min_prio(c, ca, READ);
 318                 up_read(&ca->bucket_lock);
 319         }
 320         mutex_unlock(&c->prio_clock[READ].lock);
 321
 322         mutex_lock(&c->prio_clock[WRITE].lock);
 323         for_each_member_device(ca, c, i) {
 324                 down_read(&ca->bucket_lock);
 325                 bch2_recalc_min_prio(c, ca, WRITE);
 326                 up_read(&ca->bucket_lock);
 327         }
 328         mutex_unlock(&c->prio_clock[WRITE].lock);
 329
 330         return 0;
 331 }
 332
 333 static int __bch2_alloc_write_key(struct bch_fs *c, struct bch_dev *ca,
 334                                   size_t b, struct btree_iter *iter,
 335                                   u64 *journal_seq)
 336 {
 337         struct bucket_mark m;
 338         __BKEY_PADDED(k, DIV_ROUND_UP(sizeof(struct bch_alloc), 8)) alloc_key;
 339         struct bucket *g;
 340         struct bkey_i_alloc *a;
 341         u8 *d;
 342         int ret;
 343
 344         bch2_btree_iter_set_pos(iter, POS(ca->dev_idx, b));
 345
 346         do {
 347                 ret = bch2_btree_iter_traverse(iter);
 348                 if (ret)
 349                         break;
 350
 351                 lg_local_lock(&c->usage_lock);
 352                 g = bucket(ca, b);
 353
 354                 /* read mark under btree node lock: */
 355                 m = READ_ONCE(g->mark);
 356                 a = bkey_alloc_init(&alloc_key.k);
 357                 a->k.p          = iter->pos;
 358                 a->v.fields     = 0;
 359                 a->v.gen        = m.gen;
 360                 set_bkey_val_u64s(&a->k, bch_alloc_val_u64s(&a->v));
 361
 362                 d = a->v.data;
 363                 if (a->v.fields & (1 << BCH_ALLOC_FIELD_READ_TIME))
 364                         put_alloc_field(&d, 2, g->prio[READ]);
 365                 if (a->v.fields & (1 << BCH_ALLOC_FIELD_WRITE_TIME))
 366                         put_alloc_field(&d, 2, g->prio[WRITE]);
 367                 lg_local_unlock(&c->usage_lock);
 368
 369                 ret = bch2_btree_insert_at(c, NULL, NULL, journal_seq,
 370                                            BTREE_INSERT_ATOMIC|
 371                                            BTREE_INSERT_NOFAIL|
 372                                            BTREE_INSERT_USE_RESERVE|
 373                                            BTREE_INSERT_USE_ALLOC_RESERVE|
 374                                            BTREE_INSERT_NOWAIT,
 375                                            BTREE_INSERT_ENTRY(iter, &a->k_i));
 376                 bch2_btree_iter_cond_resched(iter);
 377         } while (ret == -EINTR);
 378
 379         return ret;
 380 }
 381
 382 int bch2_alloc_replay_key(struct bch_fs *c, struct bpos pos)
 383 {
 384         struct bch_dev *ca;
 385         struct btree_iter iter;
 386         int ret;
 387
 388         if (pos.inode >= c->sb.nr_devices || !c->devs[pos.inode])
 389                 return 0;
 390
 391         ca = bch_dev_bkey_exists(c, pos.inode);
 392
 393         if (pos.offset >= ca->mi.nbuckets)
 394                 return 0;
 395
 396         bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, POS_MIN,
 397                              BTREE_ITER_INTENT);
 398
 399         ret = __bch2_alloc_write_key(c, ca, pos.offset, &iter, NULL);
 400         bch2_btree_iter_unlock(&iter);
 401         return ret;
 402 }
 403
 404 static int bch2_alloc_write(struct bch_fs *c, struct bch_dev *ca, u64 *journal_seq)
 405 {
 406         struct btree_iter iter;
 407         unsigned long bucket;
 408         int ret = 0;
 409
 410         bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, POS_MIN,
 411                              BTREE_ITER_INTENT);
 412
 413         down_read(&ca->bucket_lock);
 414         for_each_set_bit(bucket, ca->buckets_dirty, ca->mi.nbuckets) {
 415                 ret = __bch2_alloc_write_key(c, ca, bucket, &iter, journal_seq);
 416                 if (ret)
 417                         break;
 418
 419                 clear_bit(bucket, ca->buckets_dirty);
 420         }
 421         up_read(&ca->bucket_lock);
 422
 423         bch2_btree_iter_unlock(&iter);
 424         return ret;
 425 }
 426
 427 /* Bucket IO clocks: */
 428
 429 static void bch2_recalc_min_prio(struct bch_fs *c, struct bch_dev *ca, int rw)
 430 {
 431         struct prio_clock *clock = &c->prio_clock[rw];
 432         struct bucket_array *buckets = bucket_array(ca);
 433         struct bucket *g;
 434         u16 max_delta = 1;
 435         unsigned i;
 436
 437         lockdep_assert_held(&c->prio_clock[rw].lock);
 438
 439         /* Determine min prio for this particular device */
 440         for_each_bucket(g, buckets)
 441                 max_delta = max(max_delta, (u16) (clock->hand - g->prio[rw]));
 442
 443         ca->min_prio[rw] = clock->hand - max_delta;
 444
 445         /*
 446          * This may possibly increase the min prio for the whole device, check
 447          * that as well.
 448          */
 449         max_delta = 1;
 450
 451         for_each_member_device(ca, c, i)
 452                 max_delta = max(max_delta,
 453                                 (u16) (clock->hand - ca->min_prio[rw]));
 454
 455         clock->min_prio = clock->hand - max_delta;
 456 }
 457
 458 static void bch2_rescale_prios(struct bch_fs *c, int rw)
 459 {
 460         struct prio_clock *clock = &c->prio_clock[rw];
 461         struct bucket_array *buckets;
 462         struct bch_dev *ca;
 463         struct bucket *g;
 464         unsigned i;
 465
 466         trace_rescale_prios(c);
 467
 468         for_each_member_device(ca, c, i) {
 469                 down_read(&ca->bucket_lock);
 470                 buckets = bucket_array(ca);
 471
 472                 for_each_bucket(g, buckets)
 473                         g->prio[rw] = clock->hand -
 474                         (clock->hand - g->prio[rw]) / 2;
 475
 476                 bch2_recalc_min_prio(c, ca, rw);
 477
 478                 up_read(&ca->bucket_lock);
 479         }
 480 }
 481
 482 static void bch2_inc_clock_hand(struct io_timer *timer)
 483 {
 484         struct prio_clock *clock = container_of(timer,
 485                                                 struct prio_clock, rescale);
 486         struct bch_fs *c = container_of(clock,
 487                                         struct bch_fs, prio_clock[clock->rw]);
 488         u64 capacity;
 489
 490         mutex_lock(&clock->lock);
 491
 492         clock->hand++;
 493
 494         /* if clock cannot be advanced more, rescale prio */
 495         if (clock->hand == (u16) (clock->min_prio - 1))
 496                 bch2_rescale_prios(c, clock->rw);
 497
 498         mutex_unlock(&clock->lock);
 499
 500         capacity = READ_ONCE(c->capacity);
 501
 502         if (!capacity)
 503                 return;
 504
 505         /*
 506          * we only increment when 0.1% of the filesystem capacity has been read
 507          * or written too, this determines if it's time
 508          *
 509          * XXX: we shouldn't really be going off of the capacity of devices in
 510          * RW mode (that will be 0 when we're RO, yet we can still service
 511          * reads)
 512          */
 513         timer->expire += capacity >> 10;
 514
 515         bch2_io_timer_add(&c->io_clock[clock->rw], timer);
 516 }
 517
 518 static void bch2_prio_timer_init(struct bch_fs *c, int rw)
 519 {
 520         struct prio_clock *clock = &c->prio_clock[rw];
 521
 522         clock->hand             = 1;
 523         clock->rw               = rw;
 524         clock->rescale.fn       = bch2_inc_clock_hand;
 525         clock->rescale.expire   = c->capacity >> 10;
 526         mutex_init(&clock->lock);
 527 }
 528
 529 /* Background allocator thread: */
 530
 531 /*
 532  * Scans for buckets to be invalidated, invalidates them, rewrites prios/gens
 533  * (marking them as invalidated on disk), then optionally issues discard
 534  * commands to the newly free buckets, then puts them on the various freelists.
 535  */
 536
 537 static void verify_not_on_freelist(struct bch_fs *c, struct bch_dev *ca,
 538                                    size_t bucket)
 539 {
 540         if (expensive_debug_checks(c)) {
 541                 size_t iter;
 542                 long i;
 543                 unsigned j;
 544
 545                 for (j = 0; j < RESERVE_NR; j++)
 546                         fifo_for_each_entry(i, &ca->free[j], iter)
 547                                 BUG_ON(i == bucket);
 548                 fifo_for_each_entry(i, &ca->free_inc, iter)
 549                         BUG_ON(i == bucket);
 550         }
 551 }
 552
 553 #define BUCKET_GC_GEN_MAX       96U
 554
 555 /**
 556  * wait_buckets_available - wait on reclaimable buckets
 557  *
 558  * If there aren't enough available buckets to fill up free_inc, wait until
 559  * there are.
 560  */
 561 static int wait_buckets_available(struct bch_fs *c, struct bch_dev *ca)
 562 {
 563         unsigned long gc_count = c->gc_count;
 564         int ret = 0;
 565
 566         while (1) {
 567                 set_current_state(TASK_INTERRUPTIBLE);
 568                 if (kthread_should_stop()) {
 569                         ret = -1;
 570                         break;
 571                 }
 572
 573                 if (gc_count != c->gc_count)
 574                         ca->inc_gen_really_needs_gc = 0;
 575
 576                 if ((ssize_t) (dev_buckets_available(c, ca) -
 577                                ca->inc_gen_really_needs_gc) >=
 578                     (ssize_t) fifo_free(&ca->free_inc))
 579                         break;
 580
 581                 up_read(&c->gc_lock);
 582                 schedule();
 583                 try_to_freeze();
 584                 down_read(&c->gc_lock);
 585         }
 586
 587         __set_current_state(TASK_RUNNING);
 588         return ret;
 589 }
 590
 591 static bool bch2_can_invalidate_bucket(struct bch_dev *ca,
 592                                        size_t bucket,
 593                                        struct bucket_mark mark)
 594 {
 595         u8 gc_gen;
 596
 597         if (!is_available_bucket(mark))
 598                 return false;
 599
 600         gc_gen = bucket_gc_gen(ca, bucket);
 601
 602         if (gc_gen >= BUCKET_GC_GEN_MAX / 2)
 603                 ca->inc_gen_needs_gc++;
 604
 605         if (gc_gen >= BUCKET_GC_GEN_MAX)
 606                 ca->inc_gen_really_needs_gc++;
 607
 608         return gc_gen < BUCKET_GC_GEN_MAX;
 609 }
 610
 611 static void bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca,
 612                                        size_t bucket)
 613 {
 614         struct bucket_mark m;
 615
 616         spin_lock(&c->freelist_lock);
 617         if (!bch2_invalidate_bucket(c, ca, bucket, &m)) {
 618                 spin_unlock(&c->freelist_lock);
 619                 return;
 620         }
 621
 622         verify_not_on_freelist(c, ca, bucket);
 623         BUG_ON(!fifo_push(&ca->free_inc, bucket));
 624         spin_unlock(&c->freelist_lock);
 625
 626         /* gc lock held: */
 627         bucket_io_clock_reset(c, ca, bucket, READ);
 628         bucket_io_clock_reset(c, ca, bucket, WRITE);
 629
 630         if (m.cached_sectors) {
 631                 ca->allocator_invalidating_data = true;
 632         } else if (m.journal_seq_valid) {
 633                 u64 journal_seq = atomic64_read(&c->journal.seq);
 634                 u64 bucket_seq  = journal_seq;
 635
 636                 bucket_seq &= ~((u64) U16_MAX);
 637                 bucket_seq |= m.journal_seq;
 638
 639                 if (bucket_seq > journal_seq)
 640                         bucket_seq -= 1 << 16;
 641
 642                 ca->allocator_journal_seq_flush =
 643                         max(ca->allocator_journal_seq_flush, bucket_seq);
 644         }
 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         /*
 669          * Time since last read, scaled to [0, 8) where larger value indicates
 670          * more recently read data:
 671          */
 672         unsigned long hotness =
 673                 (bucket(ca, b)->prio[READ]      - ca->min_prio[READ]) * 7 /
 674                 (c->prio_clock[READ].hand       - ca->min_prio[READ]);
 675
 676         /* How much we want to keep the data in this bucket: */
 677         unsigned long data_wantness =
 678                 (hotness + 1) * bucket_sectors_used(m);
 679
 680         unsigned long needs_journal_commit =
 681                 bucket_needs_journal_commit(m, c->journal.last_seq_ondisk);
 682
 683         return  (data_wantness << 9) |
 684                 (needs_journal_commit << 8) |
 685                 bucket_gc_gen(ca, b);
 686 }
 687
 688 static inline int bucket_alloc_cmp(alloc_heap *h,
 689                                    struct alloc_heap_entry l,
 690                                    struct alloc_heap_entry r)
 691 {
 692         return (l.key > r.key) - (l.key < r.key);
 693 }
 694
 695 static void invalidate_buckets_lru(struct bch_fs *c, struct bch_dev *ca)
 696 {
 697         struct bucket_array *buckets;
 698         struct alloc_heap_entry e;
 699         size_t b;
 700
 701         ca->alloc_heap.used = 0;
 702
 703         mutex_lock(&c->prio_clock[READ].lock);
 704         down_read(&ca->bucket_lock);
 705
 706         buckets = bucket_array(ca);
 707
 708         bch2_recalc_min_prio(c, ca, READ);
 709
 710         /*
 711          * Find buckets with lowest read priority, by building a maxheap sorted
 712          * by read priority and repeatedly replacing the maximum element until
 713          * all buckets have been visited.
 714          */
 715         for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++) {
 716                 struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
 717
 718                 if (!bch2_can_invalidate_bucket(ca, b, m))
 719                         continue;
 720
 721                 e = (struct alloc_heap_entry) {
 722                         .bucket = b,
 723                         .key    = bucket_sort_key(c, ca, b, m)
 724                 };
 725
 726                 heap_add_or_replace(&ca->alloc_heap, e, -bucket_alloc_cmp);
 727         }
 728
 729         up_read(&ca->bucket_lock);
 730         mutex_unlock(&c->prio_clock[READ].lock);
 731
 732         heap_resort(&ca->alloc_heap, bucket_alloc_cmp);
 733
 734         /*
 735          * If we run out of buckets to invalidate, bch2_allocator_thread() will
 736          * kick stuff and retry us
 737          */
 738         while (!fifo_full(&ca->free_inc) &&
 739                heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp))
 740                 bch2_invalidate_one_bucket(c, ca, e.bucket);
 741 }
 742
 743 static void invalidate_buckets_fifo(struct bch_fs *c, struct bch_dev *ca)
 744 {
 745         struct bucket_array *buckets = bucket_array(ca);
 746         struct bucket_mark m;
 747         size_t b, checked;
 748
 749         for (checked = 0;
 750              checked < ca->mi.nbuckets && !fifo_full(&ca->free_inc);
 751              checked++) {
 752                 if (ca->fifo_last_bucket <  ca->mi.first_bucket ||
 753                     ca->fifo_last_bucket >= ca->mi.nbuckets)
 754                         ca->fifo_last_bucket = ca->mi.first_bucket;
 755
 756                 b = ca->fifo_last_bucket++;
 757
 758                 m = READ_ONCE(buckets->b[b].mark);
 759
 760                 if (bch2_can_invalidate_bucket(ca, b, m))
 761                         bch2_invalidate_one_bucket(c, ca, b);
 762         }
 763 }
 764
 765 static void invalidate_buckets_random(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 checked;
 770
 771         for (checked = 0;
 772              checked < ca->mi.nbuckets / 2 && !fifo_full(&ca->free_inc);
 773              checked++) {
 774                 size_t b = bch2_rand_range(ca->mi.nbuckets -
 775                                            ca->mi.first_bucket) +
 776                         ca->mi.first_bucket;
 777
 778                 m = READ_ONCE(buckets->b[b].mark);
 779
 780                 if (bch2_can_invalidate_bucket(ca, b, m))
 781                         bch2_invalidate_one_bucket(c, ca, b);
 782         }
 783 }
 784
 785 static void invalidate_buckets(struct bch_fs *c, struct bch_dev *ca)
 786 {
 787         ca->inc_gen_needs_gc                    = 0;
 788         ca->inc_gen_really_needs_gc             = 0;
 789
 790         switch (ca->mi.replacement) {
 791                 case CACHE_REPLACEMENT_LRU:
 792                         invalidate_buckets_lru(c, ca);
 793                         break;
 794                 case CACHE_REPLACEMENT_FIFO:
 795                         invalidate_buckets_fifo(c, ca);
 796                         break;
 797                 case CACHE_REPLACEMENT_RANDOM:
 798                         invalidate_buckets_random(c, ca);
 799                         break;
 800         }
 801 }
 802
 803 static int size_t_cmp(const void *_l, const void *_r)
 804 {
 805         const size_t *l = _l, *r = _r;
 806
 807         return (*l > *r) - (*l < *r);
 808 }
 809
 810 static int bch2_invalidate_free_inc(struct bch_fs *c, struct bch_dev *ca,
 811                                     u64 *journal_seq)
 812 {
 813         struct btree_iter iter;
 814         unsigned nr_invalidated = 0;
 815         size_t b, i;
 816         int ret = 0;
 817
 818         bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, POS(ca->dev_idx, 0),
 819                              BTREE_ITER_INTENT);
 820
 821         fifo_for_each_entry(b, &ca->free_inc, i) {
 822                 ret = __bch2_alloc_write_key(c, ca, b, &iter, journal_seq);
 823                 if (ret)
 824                         break;
 825
 826                 nr_invalidated++;
 827         }
 828
 829         bch2_btree_iter_unlock(&iter);
 830         return nr_invalidated ?: ret;
 831 }
 832
 833 /*
 834  * Given an invalidated, ready to use bucket: issue a discard to it if enabled,
 835  * then add it to the freelist, waiting until there's room if necessary:
 836  */
 837 static void discard_invalidated_bucket(struct bch_fs *c, struct bch_dev *ca,
 838                                        long bucket)
 839 {
 840         if (ca->mi.discard &&
 841             blk_queue_discard(bdev_get_queue(ca->disk_sb.bdev)))
 842                 blkdev_issue_discard(ca->disk_sb.bdev,
 843                                      bucket_to_sector(ca, bucket),
 844                                      ca->mi.bucket_size, GFP_NOIO, 0);
 845
 846         while (1) {
 847                 bool pushed = false;
 848                 unsigned i;
 849
 850                 set_current_state(TASK_INTERRUPTIBLE);
 851
 852                 /*
 853                  * Don't remove from free_inc until after it's added to
 854                  * freelist, so gc can find it:
 855                  */
 856                 spin_lock(&c->freelist_lock);
 857                 for (i = 0; i < RESERVE_NR; i++)
 858                         if (fifo_push(&ca->free[i], bucket)) {
 859                                 fifo_pop(&ca->free_inc, bucket);
 860                                 closure_wake_up(&c->freelist_wait);
 861                                 pushed = true;
 862                                 break;
 863                         }
 864                 spin_unlock(&c->freelist_lock);
 865
 866                 if (pushed)
 867                         break;
 868
 869                 if (kthread_should_stop())
 870                         break;
 871
 872                 schedule();
 873                 try_to_freeze();
 874         }
 875
 876         __set_current_state(TASK_RUNNING);
 877 }
 878
 879 /**
 880  * bch_allocator_thread - move buckets from free_inc to reserves
 881  *
 882  * The free_inc FIFO is populated by invalidate_buckets(), and
 883  * the reserves are depleted by bucket allocation. When we run out
 884  * of free_inc, try to invalidate some buckets and write out
 885  * prios and gens.
 886  */
 887 static int bch2_allocator_thread(void *arg)
 888 {
 889         struct bch_dev *ca = arg;
 890         struct bch_fs *c = ca->fs;
 891         u64 journal_seq;
 892         size_t bucket;
 893         int ret;
 894
 895         set_freezable();
 896
 897         while (1) {
 898                 while (1) {
 899                         while (ca->nr_invalidated) {
 900                                 BUG_ON(fifo_empty(&ca->free_inc));
 901
 902                                 bucket = fifo_peek(&ca->free_inc);
 903                                 discard_invalidated_bucket(c, ca, bucket);
 904                                 if (kthread_should_stop())
 905                                         return 0;
 906                                 --ca->nr_invalidated;
 907                         }
 908
 909                         if (fifo_empty(&ca->free_inc))
 910                                 break;
 911
 912                         journal_seq = 0;
 913                         ret = bch2_invalidate_free_inc(c, ca, &journal_seq);
 914                         if (ret < 0)
 915                                 return 0;
 916
 917                         ca->nr_invalidated = ret;
 918
 919                         if (ca->nr_invalidated == fifo_used(&ca->free_inc)) {
 920                                 ca->alloc_thread_started = true;
 921                                 bch2_alloc_write(c, ca, &journal_seq);
 922                         }
 923
 924                         if (ca->allocator_invalidating_data)
 925                                 bch2_journal_flush_seq(&c->journal, journal_seq);
 926                         else if (ca->allocator_journal_seq_flush)
 927                                 bch2_journal_flush_seq(&c->journal,
 928                                                        ca->allocator_journal_seq_flush);
 929                 }
 930
 931                 /* Reset front/back so we can easily sort fifo entries later: */
 932                 ca->free_inc.front = ca->free_inc.back  = 0;
 933                 ca->allocator_journal_seq_flush         = 0;
 934                 ca->allocator_invalidating_data         = false;
 935
 936                 down_read(&c->gc_lock);
 937                 if (test_bit(BCH_FS_GC_FAILURE, &c->flags)) {
 938                         up_read(&c->gc_lock);
 939                         return 0;
 940                 }
 941
 942                 while (1) {
 943                         /*
 944                          * Find some buckets that we can invalidate, either
 945                          * they're completely unused, or only contain clean data
 946                          * that's been written back to the backing device or
 947                          * another cache tier
 948                          */
 949
 950                         invalidate_buckets(c, ca);
 951                         trace_alloc_batch(ca, fifo_used(&ca->free_inc),
 952                                           ca->free_inc.size);
 953
 954                         if ((ca->inc_gen_needs_gc >= ca->free_inc.size ||
 955                              (!fifo_full(&ca->free_inc) &&
 956                               ca->inc_gen_really_needs_gc >=
 957                               fifo_free(&ca->free_inc))) &&
 958                             c->gc_thread) {
 959                                 atomic_inc(&c->kick_gc);
 960                                 wake_up_process(c->gc_thread);
 961                         }
 962
 963                         if (fifo_full(&ca->free_inc))
 964                                 break;
 965
 966                         if (wait_buckets_available(c, ca)) {
 967                                 up_read(&c->gc_lock);
 968                                 return 0;
 969                         }
 970                 }
 971                 up_read(&c->gc_lock);
 972
 973                 BUG_ON(ca->free_inc.front);
 974
 975                 spin_lock(&c->freelist_lock);
 976                 sort(ca->free_inc.data,
 977                      ca->free_inc.back,
 978                      sizeof(ca->free_inc.data[0]),
 979                      size_t_cmp, NULL);
 980                 spin_unlock(&c->freelist_lock);
 981
 982                 /*
 983                  * free_inc is now full of newly-invalidated buckets: next,
 984                  * write out the new bucket gens:
 985                  */
 986         }
 987 }
 988
 989 /* Allocation */
 990
 991 /*
 992  * Open buckets represent a bucket that's currently being allocated from.  They
 993  * serve two purposes:
 994  *
 995  *  - They track buckets that have been partially allocated, allowing for
 996  *    sub-bucket sized allocations - they're used by the sector allocator below
 997  *
 998  *  - They provide a reference to the buckets they own that mark and sweep GC
 999  *    can find, until the new allocation has a pointer to it inserted into the
1000  *    btree
1001  *
1002  * When allocating some space with the sector allocator, the allocation comes
1003  * with a reference to an open bucket - the caller is required to put that
1004  * reference _after_ doing the index update that makes its allocation reachable.
1005  */
1006
1007 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
1008 {
1009         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
1010
1011         spin_lock(&ob->lock);
1012         bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr),
1013                                false, gc_pos_alloc(c, ob), 0);
1014         ob->valid = false;
1015         spin_unlock(&ob->lock);
1016
1017         spin_lock(&c->freelist_lock);
1018         ob->freelist = c->open_buckets_freelist;
1019         c->open_buckets_freelist = ob - c->open_buckets;
1020         c->open_buckets_nr_free++;
1021         spin_unlock(&c->freelist_lock);
1022
1023         closure_wake_up(&c->open_buckets_wait);
1024 }
1025
1026 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
1027 {
1028         struct open_bucket *ob;
1029
1030         BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
1031
1032         ob = c->open_buckets + c->open_buckets_freelist;
1033         c->open_buckets_freelist = ob->freelist;
1034         atomic_set(&ob->pin, 1);
1035
1036         c->open_buckets_nr_free--;
1037         return ob;
1038 }
1039
1040 /*
1041  * XXX: allocation on startup is still sketchy. There is insufficient
1042  * synchronization for bch2_bucket_alloc_startup() to work correctly after
1043  * bch2_alloc_write() has been called, and we aren't currently doing anything
1044  * to guarantee that this won't happen.
1045  *
1046  * Even aside from that, it's really difficult to avoid situations where on
1047  * startup we write out a pointer to a freshly allocated bucket before the
1048  * corresponding gen - when we're still digging ourself out of the "i need to
1049  * allocate to write bucket gens, but i need to write bucket gens to allocate"
1050  * hole.
1051  *
1052  * Fortunately, bch2_btree_mark_key_initial() will detect and repair this
1053  * easily enough...
1054  */
1055 static long bch2_bucket_alloc_startup(struct bch_fs *c, struct bch_dev *ca)
1056 {
1057         struct bucket_array *buckets;
1058         ssize_t b;
1059
1060         if (!down_read_trylock(&c->gc_lock))
1061                 return -1;
1062
1063         if (test_bit(BCH_FS_GC_FAILURE, &c->flags)) {
1064                 up_read(&c->gc_lock);
1065                 return -1;
1066         }
1067
1068         spin_unlock(&c->freelist_lock);
1069
1070         down_read(&ca->bucket_lock);
1071         buckets = bucket_array(ca);
1072
1073         spin_lock(&c->freelist_lock);
1074
1075         for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++)
1076                 if (is_startup_available_bucket(buckets->b[b].mark) &&
1077                     bch2_mark_alloc_bucket_startup(c, ca, b)) {
1078                         set_bit(b, ca->buckets_dirty);
1079                         goto success;
1080                 }
1081         b = -1;
1082 success:
1083         up_read(&ca->bucket_lock);
1084         up_read(&c->gc_lock);
1085         return b;
1086 }
1087
1088 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
1089 {
1090         switch (reserve) {
1091         case RESERVE_ALLOC:
1092                 return 0;
1093         case RESERVE_BTREE:
1094                 return BTREE_NODE_RESERVE / 2;
1095         default:
1096                 return BTREE_NODE_RESERVE;
1097         }
1098 }
1099
1100 /**
1101  * bch_bucket_alloc - allocate a single bucket from a specific device
1102  *
1103  * Returns index of bucket on success, 0 on failure
1104  * */
1105 int bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
1106                       enum alloc_reserve reserve,
1107                       bool may_alloc_partial,
1108                       struct closure *cl)
1109 {
1110         struct bucket_array *buckets;
1111         struct open_bucket *ob;
1112         long bucket;
1113
1114         spin_lock(&c->freelist_lock);
1115         if (may_alloc_partial &&
1116             ca->open_buckets_partial_nr) {
1117                 int ret = ca->open_buckets_partial[--ca->open_buckets_partial_nr];
1118                 c->open_buckets[ret].on_partial_list = false;
1119                 spin_unlock(&c->freelist_lock);
1120                 return ret;
1121         }
1122
1123         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
1124                 if (cl)
1125                         closure_wait(&c->open_buckets_wait, cl);
1126                 spin_unlock(&c->freelist_lock);
1127                 trace_open_bucket_alloc_fail(ca, reserve);
1128                 return OPEN_BUCKETS_EMPTY;
1129         }
1130
1131         if (likely(fifo_pop(&ca->free[RESERVE_NONE], bucket)))
1132                 goto out;
1133
1134         switch (reserve) {
1135         case RESERVE_ALLOC:
1136                 if (fifo_pop(&ca->free[RESERVE_BTREE], bucket))
1137                         goto out;
1138                 break;
1139         case RESERVE_BTREE:
1140                 if (fifo_used(&ca->free[RESERVE_BTREE]) * 2 >=
1141                     ca->free[RESERVE_BTREE].size &&
1142                     fifo_pop(&ca->free[RESERVE_BTREE], bucket))
1143                         goto out;
1144                 break;
1145         case RESERVE_MOVINGGC:
1146                 if (fifo_pop(&ca->free[RESERVE_MOVINGGC], bucket))
1147                         goto out;
1148                 break;
1149         default:
1150                 break;
1151         }
1152
1153         if (unlikely(!ca->alloc_thread_started) &&
1154             (reserve == RESERVE_ALLOC) &&
1155             (bucket = bch2_bucket_alloc_startup(c, ca)) >= 0)
1156                 goto out;
1157
1158         spin_unlock(&c->freelist_lock);
1159
1160         trace_bucket_alloc_fail(ca, reserve);
1161         return FREELIST_EMPTY;
1162 out:
1163         verify_not_on_freelist(c, ca, bucket);
1164
1165         ob = bch2_open_bucket_alloc(c);
1166
1167         spin_lock(&ob->lock);
1168         lg_local_lock(&c->usage_lock);
1169         buckets = bucket_array(ca);
1170
1171         ob->valid       = true;
1172         ob->sectors_free = ca->mi.bucket_size;
1173         ob->ptr         = (struct bch_extent_ptr) {
1174                 .gen    = buckets->b[bucket].mark.gen,
1175                 .offset = bucket_to_sector(ca, bucket),
1176                 .dev    = ca->dev_idx,
1177         };
1178
1179         bucket_io_clock_reset(c, ca, bucket, READ);
1180         bucket_io_clock_reset(c, ca, bucket, WRITE);
1181
1182         lg_local_unlock(&c->usage_lock);
1183         spin_unlock(&ob->lock);
1184
1185         spin_unlock(&c->freelist_lock);
1186
1187         bch2_wake_allocator(ca);
1188
1189         trace_bucket_alloc(ca, reserve);
1190         return ob - c->open_buckets;
1191 }
1192
1193 struct dev_alloc_list bch2_wp_alloc_list(struct bch_fs *c,
1194                                          struct write_point *wp,
1195                                          struct bch_devs_mask *devs)
1196 {
1197         struct dev_alloc_list ret = { .nr = 0 };
1198         struct bch_dev *ca, *ca2;
1199         unsigned i, j;
1200
1201         for_each_member_device_rcu(ca, c, i, devs) {
1202                 for (j = 0; j < ret.nr; j++) {
1203                         unsigned idx = ret.devs[j];
1204
1205                         ca2 = rcu_dereference(c->devs[idx]);
1206                         if (!ca2)
1207                                 break;
1208
1209                         if (ca->mi.tier < ca2->mi.tier)
1210                                 break;
1211
1212                         if (ca->mi.tier == ca2->mi.tier &&
1213                             wp->next_alloc[i] < wp->next_alloc[idx])
1214                                 break;
1215                 }
1216
1217                 array_insert_item(ret.devs, ret.nr, j, i);
1218         }
1219
1220         return ret;
1221 }
1222
1223 void bch2_wp_rescale(struct bch_fs *c, struct bch_dev *ca,
1224                      struct write_point *wp)
1225 {
1226         unsigned i;
1227
1228         for (i = 0; i < ARRAY_SIZE(wp->next_alloc); i++)
1229                 wp->next_alloc[i] >>= 1;
1230 }
1231
1232 static enum bucket_alloc_ret __bch2_bucket_alloc_set(struct bch_fs *c,
1233                                         struct write_point *wp,
1234                                         unsigned nr_replicas,
1235                                         enum alloc_reserve reserve,
1236                                         struct bch_devs_mask *devs,
1237                                         struct closure *cl)
1238 {
1239         enum bucket_alloc_ret ret = NO_DEVICES;
1240         struct dev_alloc_list devs_sorted;
1241         u64 buckets_free;
1242         unsigned i;
1243
1244         BUG_ON(nr_replicas > ARRAY_SIZE(wp->ptrs));
1245
1246         if (wp->nr_ptrs >= nr_replicas)
1247                 return ALLOC_SUCCESS;
1248
1249         rcu_read_lock();
1250         devs_sorted = bch2_wp_alloc_list(c, wp, devs);
1251
1252         for (i = 0; i < devs_sorted.nr; i++) {
1253                 struct bch_dev *ca =
1254                         rcu_dereference(c->devs[devs_sorted.devs[i]]);
1255                 int ob;
1256
1257                 if (!ca)
1258                         continue;
1259
1260                 ob = bch2_bucket_alloc(c, ca, reserve,
1261                                        wp->type == BCH_DATA_USER, cl);
1262                 if (ob < 0) {
1263                         ret = ob;
1264                         if (ret == OPEN_BUCKETS_EMPTY)
1265                                 break;
1266                         continue;
1267                 }
1268
1269                 BUG_ON(ob <= 0 || ob > U8_MAX);
1270                 BUG_ON(wp->nr_ptrs >= ARRAY_SIZE(wp->ptrs));
1271                 wp->ptrs[wp->nr_ptrs++] = c->open_buckets + ob;
1272
1273                 buckets_free = U64_MAX, dev_buckets_free(c, ca);
1274                 if (buckets_free)
1275                         wp->next_alloc[ca->dev_idx] +=
1276                                 div64_u64(U64_MAX, buckets_free *
1277                                           ca->mi.bucket_size);
1278                 else
1279                         wp->next_alloc[ca->dev_idx] = U64_MAX;
1280                 bch2_wp_rescale(c, ca, wp);
1281
1282                 __clear_bit(ca->dev_idx, devs->d);
1283
1284                 if (wp->nr_ptrs == nr_replicas) {
1285                         ret = ALLOC_SUCCESS;
1286                         break;
1287                 }
1288         }
1289
1290         EBUG_ON(reserve == RESERVE_MOVINGGC &&
1291                 ret != ALLOC_SUCCESS &&
1292                 ret != OPEN_BUCKETS_EMPTY);
1293         rcu_read_unlock();
1294         return ret;
1295 }
1296
1297 static int bch2_bucket_alloc_set(struct bch_fs *c, struct write_point *wp,
1298                                 unsigned nr_replicas,
1299                                 enum alloc_reserve reserve,
1300                                 struct bch_devs_mask *devs,
1301                                 struct closure *cl)
1302 {
1303         bool waiting = false;
1304
1305         while (1) {
1306                 switch (__bch2_bucket_alloc_set(c, wp, nr_replicas,
1307                                                 reserve, devs, cl)) {
1308                 case ALLOC_SUCCESS:
1309                         if (waiting)
1310                                 closure_wake_up(&c->freelist_wait);
1311
1312                         return 0;
1313
1314                 case NO_DEVICES:
1315                         if (waiting)
1316                                 closure_wake_up(&c->freelist_wait);
1317                         return -EROFS;
1318
1319                 case FREELIST_EMPTY:
1320                         if (!cl)
1321                                 return -ENOSPC;
1322
1323                         if (waiting)
1324                                 return -EAGAIN;
1325
1326                         /* Retry allocation after adding ourself to waitlist: */
1327                         closure_wait(&c->freelist_wait, cl);
1328                         waiting = true;
1329                         break;
1330                 case OPEN_BUCKETS_EMPTY:
1331                         return cl ? -EAGAIN : -ENOSPC;
1332                 default:
1333                         BUG();
1334                 }
1335         }
1336 }
1337
1338 /* Sector allocator */
1339
1340 static void writepoint_drop_ptrs(struct bch_fs *c,
1341                                  struct write_point *wp,
1342                                  struct bch_devs_mask *devs,
1343                                  unsigned nr_ptrs_dislike)
1344 {
1345         int i;
1346
1347         if (!nr_ptrs_dislike)
1348                 return;
1349
1350         for (i = wp->nr_ptrs - 1; i >= 0; --i) {
1351                 struct open_bucket *ob = wp->ptrs[i];
1352                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
1353
1354                 if (nr_ptrs_dislike && !test_bit(ob->ptr.dev, devs->d)) {
1355                         BUG_ON(ca->open_buckets_partial_nr >=
1356                                ARRAY_SIZE(ca->open_buckets_partial));
1357
1358                         spin_lock(&c->freelist_lock);
1359                         ob->on_partial_list = true;
1360                         ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
1361                                 ob - c->open_buckets;
1362                         spin_unlock(&c->freelist_lock);
1363
1364                         closure_wake_up(&c->open_buckets_wait);
1365                         closure_wake_up(&c->freelist_wait);
1366
1367                         array_remove_item(wp->ptrs, wp->nr_ptrs, i);
1368                         --nr_ptrs_dislike;
1369                 }
1370         }
1371 }
1372
1373 static void verify_not_stale(struct bch_fs *c, const struct write_point *wp)
1374 {
1375 #ifdef CONFIG_BCACHEFS_DEBUG
1376         struct open_bucket *ob;
1377         unsigned i;
1378
1379         writepoint_for_each_ptr(wp, ob, i) {
1380                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
1381
1382                 BUG_ON(ptr_stale(ca, &ob->ptr));
1383         }
1384 #endif
1385 }
1386
1387 static int open_bucket_add_buckets(struct bch_fs *c,
1388                                    struct bch_devs_mask *_devs,
1389                                    struct write_point *wp,
1390                                    struct bch_devs_list *devs_have,
1391                                    unsigned nr_replicas,
1392                                    enum alloc_reserve reserve,
1393                                    struct closure *cl)
1394 {
1395         struct bch_devs_mask devs = c->rw_devs[wp->type];
1396         struct open_bucket *ob;
1397         unsigned i;
1398
1399         if (wp->nr_ptrs >= nr_replicas)
1400                 return 0;
1401
1402         /* Don't allocate from devices we already have pointers to: */
1403         for (i = 0; i < devs_have->nr; i++)
1404                 __clear_bit(devs_have->devs[i], devs.d);
1405
1406         writepoint_for_each_ptr(wp, ob, i)
1407                 __clear_bit(ob->ptr.dev, devs.d);
1408
1409         if (_devs)
1410                 bitmap_and(devs.d, devs.d, _devs->d, BCH_SB_MEMBERS_MAX);
1411
1412         return bch2_bucket_alloc_set(c, wp, nr_replicas, reserve, &devs, cl);
1413 }
1414
1415 static struct write_point *__writepoint_find(struct hlist_head *head,
1416                                              unsigned long write_point)
1417 {
1418         struct write_point *wp;
1419
1420         hlist_for_each_entry_rcu(wp, head, node)
1421                 if (wp->write_point == write_point)
1422                         return wp;
1423
1424         return NULL;
1425 }
1426
1427 static struct hlist_head *writepoint_hash(struct bch_fs *c,
1428                                           unsigned long write_point)
1429 {
1430         unsigned hash =
1431                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
1432
1433         return &c->write_points_hash[hash];
1434 }
1435
1436 static struct write_point *writepoint_find(struct bch_fs *c,
1437                                            unsigned long write_point)
1438 {
1439         struct write_point *wp, *oldest;
1440         struct hlist_head *head;
1441
1442         if (!(write_point & 1UL)) {
1443                 wp = (struct write_point *) write_point;
1444                 mutex_lock(&wp->lock);
1445                 return wp;
1446         }
1447
1448         head = writepoint_hash(c, write_point);
1449 restart_find:
1450         wp = __writepoint_find(head, write_point);
1451         if (wp) {
1452 lock_wp:
1453                 mutex_lock(&wp->lock);
1454                 if (wp->write_point == write_point)
1455                         goto out;
1456                 mutex_unlock(&wp->lock);
1457                 goto restart_find;
1458         }
1459
1460         oldest = NULL;
1461         for (wp = c->write_points;
1462              wp < c->write_points + ARRAY_SIZE(c->write_points);
1463              wp++)
1464                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1465                         oldest = wp;
1466
1467         mutex_lock(&oldest->lock);
1468         mutex_lock(&c->write_points_hash_lock);
1469         wp = __writepoint_find(head, write_point);
1470         if (wp && wp != oldest) {
1471                 mutex_unlock(&c->write_points_hash_lock);
1472                 mutex_unlock(&oldest->lock);
1473                 goto lock_wp;
1474         }
1475
1476         wp = oldest;
1477         hlist_del_rcu(&wp->node);
1478         wp->write_point = write_point;
1479         hlist_add_head_rcu(&wp->node, head);
1480         mutex_unlock(&c->write_points_hash_lock);
1481 out:
1482         wp->last_used = sched_clock();
1483         return wp;
1484 }
1485
1486 /*
1487  * Get us an open_bucket we can allocate from, return with it locked:
1488  */
1489 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
1490                                 struct bch_devs_mask *devs,
1491                                 struct write_point_specifier write_point,
1492                                 struct bch_devs_list *devs_have,
1493                                 unsigned nr_replicas,
1494                                 unsigned nr_replicas_required,
1495                                 enum alloc_reserve reserve,
1496                                 unsigned flags,
1497                                 struct closure *cl)
1498 {
1499         struct write_point *wp;
1500         struct open_bucket *ob;
1501         unsigned i, nr_ptrs_dislike = 0, nr_ptrs_have = 0;
1502         int ret;
1503
1504         BUG_ON(!nr_replicas || !nr_replicas_required);
1505
1506         wp = writepoint_find(c, write_point.v);
1507
1508         /* does ob have ptrs we don't need? */
1509         writepoint_for_each_ptr(wp, ob, i)
1510                 if (bch2_dev_list_has_dev(*devs_have, ob->ptr.dev))
1511                         nr_ptrs_have++;
1512                 else if (devs && !test_bit(ob->ptr.dev, devs->d))
1513                         nr_ptrs_dislike++;
1514
1515         ret = open_bucket_add_buckets(c, devs, wp, devs_have,
1516                                 nr_replicas + nr_ptrs_have + nr_ptrs_dislike,
1517                                 reserve, cl);
1518         if (ret && ret != -EROFS)
1519                 goto err;
1520
1521         if (wp->nr_ptrs <
1522             nr_ptrs_have + nr_ptrs_dislike + nr_replicas_required) {
1523                 ret = -EROFS;
1524                 goto err;
1525         }
1526
1527         if ((int) wp->nr_ptrs - nr_ptrs_dislike < nr_replicas)
1528                 nr_ptrs_dislike = clamp_t(int, wp->nr_ptrs - nr_replicas,
1529                                           0, nr_ptrs_dislike);
1530
1531         /* Remove pointers we don't want to use: */
1532         writepoint_drop_ptrs(c, wp, devs, nr_ptrs_dislike);
1533
1534         /*
1535          * Move pointers to devices we already have to end of open bucket
1536          * pointer list - note that removing pointers we don't want to use might
1537          * have changed nr_ptrs_have:
1538          */
1539         if (nr_ptrs_have) {
1540                 i = nr_ptrs_have = 0;
1541                 while (i < wp->nr_ptrs - nr_ptrs_have)
1542                         if (bch2_dev_list_has_dev(*devs_have, wp->ptrs[i]->ptr.dev)) {
1543                                 nr_ptrs_have++;
1544                                 swap(wp->ptrs[i], wp->ptrs[wp->nr_ptrs - nr_ptrs_have]);
1545                         } else {
1546                                 i++;
1547                         }
1548         }
1549
1550         wp->nr_ptrs_can_use =
1551                 min_t(unsigned, nr_replicas, wp->nr_ptrs - nr_ptrs_have);
1552
1553         BUG_ON(wp->nr_ptrs_can_use < nr_replicas_required ||
1554                wp->nr_ptrs_can_use > wp->nr_ptrs);
1555
1556         wp->sectors_free = UINT_MAX;
1557
1558         for (i = 0; i < wp->nr_ptrs_can_use; i++)
1559                 wp->sectors_free = min(wp->sectors_free,
1560                                        wp->ptrs[i]->sectors_free);
1561
1562         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1563
1564         verify_not_stale(c, wp);
1565
1566         return wp;
1567 err:
1568         mutex_unlock(&wp->lock);
1569         return ERR_PTR(ret);
1570 }
1571
1572 /*
1573  * Append pointers to the space we just allocated to @k, and mark @sectors space
1574  * as allocated out of @ob
1575  */
1576 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1577                                     struct bkey_i_extent *e, unsigned sectors)
1578 {
1579         unsigned i;
1580
1581         BUG_ON(sectors > wp->sectors_free);
1582         wp->sectors_free -= sectors;
1583
1584         for (i = 0; i < wp->nr_ptrs_can_use; i++) {
1585                 struct open_bucket *ob = wp->ptrs[i];
1586                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
1587                 struct bch_extent_ptr tmp = ob->ptr;
1588
1589                 EBUG_ON(bch2_extent_has_device(extent_i_to_s_c(e), ob->ptr.dev));
1590
1591                 tmp.cached = bkey_extent_is_cached(&e->k);
1592                 tmp.offset += ca->mi.bucket_size - ob->sectors_free;
1593                 extent_ptr_append(e, tmp);
1594
1595                 BUG_ON(sectors > ob->sectors_free);
1596                 ob->sectors_free -= sectors;
1597         }
1598 }
1599
1600 /*
1601  * Append pointers to the space we just allocated to @k, and mark @sectors space
1602  * as allocated out of @ob
1603  */
1604 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1605 {
1606         int i;
1607
1608         for (i = wp->nr_ptrs - 1; i >= 0; --i) {
1609                 struct open_bucket *ob = wp->ptrs[i];
1610
1611                 if (!ob->sectors_free) {
1612                         array_remove_item(wp->ptrs, wp->nr_ptrs, i);
1613                         bch2_open_bucket_put(c, ob);
1614                 }
1615         }
1616
1617         mutex_unlock(&wp->lock);
1618 }
1619
1620 /* Startup/shutdown (ro/rw): */
1621
1622 void bch2_recalc_capacity(struct bch_fs *c)
1623 {
1624         struct bch_tier *fastest_tier = NULL, *slowest_tier = NULL, *tier;
1625         struct bch_dev *ca;
1626         u64 total_capacity, capacity = 0, reserved_sectors = 0;
1627         unsigned long ra_pages = 0;
1628         unsigned i, j;
1629
1630         lockdep_assert_held(&c->state_lock);
1631
1632         for_each_online_member(ca, c, i) {
1633                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_bdi;
1634
1635                 ra_pages += bdi->ra_pages;
1636         }
1637
1638         bch2_set_ra_pages(c, ra_pages);
1639
1640         /* Find fastest, slowest tiers with devices: */
1641
1642         for (tier = c->tiers;
1643              tier < c->tiers + ARRAY_SIZE(c->tiers); tier++) {
1644                 if (!dev_mask_nr(&tier->devs))
1645                         continue;
1646                 if (!fastest_tier)
1647                         fastest_tier = tier;
1648                 slowest_tier = tier;
1649         }
1650
1651         c->fastest_tier = fastest_tier != slowest_tier ? fastest_tier : NULL;
1652         c->fastest_devs = fastest_tier != slowest_tier ? &fastest_tier->devs : NULL;
1653
1654         if (!fastest_tier)
1655                 goto set_capacity;
1656
1657         /*
1658          * Capacity of the filesystem is the capacity of all the devices in the
1659          * slowest (highest) tier - we don't include lower tier devices.
1660          */
1661         for_each_member_device_rcu(ca, c, i, &slowest_tier->devs) {
1662                 size_t reserve = 0;
1663
1664                 /*
1665                  * We need to reserve buckets (from the number
1666                  * of currently available buckets) against
1667                  * foreground writes so that mainly copygc can
1668                  * make forward progress.
1669                  *
1670                  * We need enough to refill the various reserves
1671                  * from scratch - copygc will use its entire
1672                  * reserve all at once, then run against when
1673                  * its reserve is refilled (from the formerly
1674                  * available buckets).
1675                  *
1676                  * This reserve is just used when considering if
1677                  * allocations for foreground writes must wait -
1678                  * not -ENOSPC calculations.
1679                  */
1680                 for (j = 0; j < RESERVE_NONE; j++)
1681                         reserve += ca->free[j].size;
1682
1683                 reserve += ca->free_inc.size;
1684
1685                 reserve += ARRAY_SIZE(c->write_points);
1686
1687                 if (ca->mi.tier)
1688                         reserve += 1;   /* tiering write point */
1689                 reserve += 1;           /* btree write point */
1690
1691                 reserved_sectors += bucket_to_sector(ca, reserve);
1692
1693                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
1694                                              ca->mi.first_bucket);
1695         }
1696 set_capacity:
1697         total_capacity = capacity;
1698
1699         capacity *= (100 - c->opts.gc_reserve_percent);
1700         capacity = div64_u64(capacity, 100);
1701
1702         BUG_ON(reserved_sectors > total_capacity);
1703
1704         capacity = min(capacity, total_capacity - reserved_sectors);
1705
1706         c->capacity = capacity;
1707
1708         if (c->capacity) {
1709                 bch2_io_timer_add(&c->io_clock[READ],
1710                                  &c->prio_clock[READ].rescale);
1711                 bch2_io_timer_add(&c->io_clock[WRITE],
1712                                  &c->prio_clock[WRITE].rescale);
1713         } else {
1714                 bch2_io_timer_del(&c->io_clock[READ],
1715                                  &c->prio_clock[READ].rescale);
1716                 bch2_io_timer_del(&c->io_clock[WRITE],
1717                                  &c->prio_clock[WRITE].rescale);
1718         }
1719
1720         /* Wake up case someone was waiting for buckets */
1721         closure_wake_up(&c->freelist_wait);
1722 }
1723
1724 static void bch2_stop_write_point(struct bch_fs *c, struct bch_dev *ca,
1725                                   struct write_point *wp)
1726 {
1727         struct bch_devs_mask not_self;
1728
1729         bitmap_complement(not_self.d, ca->self.d, BCH_SB_MEMBERS_MAX);
1730
1731         mutex_lock(&wp->lock);
1732         writepoint_drop_ptrs(c, wp, &not_self, wp->nr_ptrs);
1733         mutex_unlock(&wp->lock);
1734 }
1735
1736 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
1737 {
1738         struct open_bucket *ob;
1739         bool ret = false;
1740
1741         for (ob = c->open_buckets;
1742              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1743              ob++) {
1744                 spin_lock(&ob->lock);
1745                 if (ob->valid && !ob->on_partial_list &&
1746                     ob->ptr.dev == ca->dev_idx)
1747                         ret = true;
1748                 spin_unlock(&ob->lock);
1749         }
1750
1751         return ret;
1752 }
1753
1754 /* device goes ro: */
1755 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
1756 {
1757         unsigned i;
1758
1759         BUG_ON(ca->alloc_thread);
1760
1761         /* First, remove device from allocation groups: */
1762
1763         clear_bit(ca->dev_idx, c->tiers[ca->mi.tier].devs.d);
1764         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1765                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
1766
1767         /*
1768          * Capacity is calculated based off of devices in allocation groups:
1769          */
1770         bch2_recalc_capacity(c);
1771
1772         /* Next, close write points that point to this device... */
1773         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1774                 bch2_stop_write_point(c, ca, &c->write_points[i]);
1775
1776         bch2_stop_write_point(c, ca, &ca->copygc_write_point);
1777         bch2_stop_write_point(c, ca, &c->tiers[ca->mi.tier].wp);
1778         bch2_stop_write_point(c, ca, &c->btree_write_point);
1779
1780         mutex_lock(&c->btree_reserve_cache_lock);
1781         while (c->btree_reserve_cache_nr) {
1782                 struct btree_alloc *a =
1783                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1784
1785                 bch2_open_bucket_put_refs(c, &a->ob.nr, a->ob.refs);
1786         }
1787         mutex_unlock(&c->btree_reserve_cache_lock);
1788
1789         /*
1790          * Wake up threads that were blocked on allocation, so they can notice
1791          * the device can no longer be removed and the capacity has changed:
1792          */
1793         closure_wake_up(&c->freelist_wait);
1794
1795         /*
1796          * journal_res_get() can block waiting for free space in the journal -
1797          * it needs to notice there may not be devices to allocate from anymore:
1798          */
1799         wake_up(&c->journal.wait);
1800
1801         /* Now wait for any in flight writes: */
1802
1803         closure_wait_event(&c->open_buckets_wait,
1804                            !bch2_dev_has_open_write_point(c, ca));
1805 }
1806
1807 /* device goes rw: */
1808 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
1809 {
1810         unsigned i;
1811
1812         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1813                 if (ca->mi.data_allowed & (1 << i))
1814                         set_bit(ca->dev_idx, c->rw_devs[i].d);
1815         set_bit(ca->dev_idx, c->tiers[ca->mi.tier].devs.d);
1816 }
1817
1818 /* stop allocator thread: */
1819 void bch2_dev_allocator_stop(struct bch_dev *ca)
1820 {
1821         struct task_struct *p = ca->alloc_thread;
1822
1823         ca->alloc_thread = NULL;
1824
1825         /*
1826          * We need an rcu barrier between setting ca->alloc_thread = NULL and
1827          * the thread shutting down to avoid bch2_wake_allocator() racing:
1828          *
1829          * XXX: it would be better to have the rcu barrier be asynchronous
1830          * instead of blocking us here
1831          */
1832         synchronize_rcu();
1833
1834         if (p) {
1835                 kthread_stop(p);
1836                 put_task_struct(p);
1837         }
1838 }
1839
1840 /* start allocator thread: */
1841 int bch2_dev_allocator_start(struct bch_dev *ca)
1842 {
1843         struct task_struct *p;
1844
1845         /*
1846          * allocator thread already started?
1847          */
1848         if (ca->alloc_thread)
1849                 return 0;
1850
1851         p = kthread_create(bch2_allocator_thread, ca, "bcache_allocator");
1852         if (IS_ERR(p))
1853                 return PTR_ERR(p);
1854
1855         get_task_struct(p);
1856         ca->alloc_thread = p;
1857         wake_up_process(p);
1858         return 0;
1859 }
1860
1861 void bch2_fs_allocator_init(struct bch_fs *c)
1862 {
1863         struct open_bucket *ob;
1864         struct write_point *wp;
1865         unsigned i;
1866
1867         mutex_init(&c->write_points_hash_lock);
1868         spin_lock_init(&c->freelist_lock);
1869         bch2_prio_timer_init(c, READ);
1870         bch2_prio_timer_init(c, WRITE);
1871
1872         /* open bucket 0 is a sentinal NULL: */
1873         spin_lock_init(&c->open_buckets[0].lock);
1874
1875         for (ob = c->open_buckets + 1;
1876              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1877                 spin_lock_init(&ob->lock);
1878                 c->open_buckets_nr_free++;
1879
1880                 ob->freelist = c->open_buckets_freelist;
1881                 c->open_buckets_freelist = ob - c->open_buckets;
1882         }
1883
1884         writepoint_init(&c->btree_write_point, BCH_DATA_BTREE);
1885
1886         for (i = 0; i < ARRAY_SIZE(c->tiers); i++)
1887                 writepoint_init(&c->tiers[i].wp, BCH_DATA_USER);
1888
1889         for (wp = c->write_points;
1890              wp < c->write_points + ARRAY_SIZE(c->write_points); wp++) {
1891                 writepoint_init(wp, BCH_DATA_USER);
1892
1893                 wp->last_used   = sched_clock();
1894                 wp->write_point = (unsigned long) wp;
1895                 hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
1896         }
1897
1898         c->pd_controllers_update_seconds = 5;
1899         INIT_DELAYED_WORK(&c->pd_controllers_update, pd_controllers_update);
1900 }