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_cache.h"
  59 #include "btree_io.h"
  60 #include "btree_update.h"
  61 #include "btree_update_interior.h"
  62 #include "btree_gc.h"
  63 #include "buckets.h"
  64 #include "checksum.h"
  65 #include "clock.h"
  66 #include "debug.h"
  67 #include "disk_groups.h"
  68 #include "error.h"
  69 #include "extents.h"
  70 #include "io.h"
  71 #include "journal.h"
  72 #include "journal_io.h"
  73 #include "super-io.h"
  74
  75 #include <linux/blkdev.h>
  76 #include <linux/kthread.h>
  77 #include <linux/math64.h>
  78 #include <linux/random.h>
  79 #include <linux/rculist.h>
  80 #include <linux/rcupdate.h>
  81 #include <linux/sched/task.h>
  82 #include <linux/sort.h>
  83 #include <trace/events/bcachefs.h>
  84
  85 static void bch2_recalc_oldest_io(struct bch_fs *, struct bch_dev *, int);
  86
  87 /* Ratelimiting/PD controllers */
  88
  89 static void pd_controllers_update(struct work_struct *work)
  90 {
  91         struct bch_fs *c = container_of(to_delayed_work(work),
  92                                            struct bch_fs,
  93                                            pd_controllers_update);
  94         struct bch_dev *ca;
  95         unsigned i;
  96
  97         for_each_member_device(ca, c, i) {
  98                 struct bch_dev_usage stats = bch2_dev_usage_read(c, ca);
  99
 100                 u64 free = bucket_to_sector(ca,
 101                                 __dev_buckets_free(ca, stats)) << 9;
 102                 /*
 103                  * Bytes of internal fragmentation, which can be
 104                  * reclaimed by copy GC
 105                  */
 106                 s64 fragmented = (bucket_to_sector(ca,
 107                                         stats.buckets[BCH_DATA_USER] +
 108                                         stats.buckets[BCH_DATA_CACHED]) -
 109                                   (stats.sectors[BCH_DATA_USER] +
 110                                    stats.sectors[BCH_DATA_CACHED])) << 9;
 111
 112                 fragmented = max(0LL, fragmented);
 113
 114                 bch2_pd_controller_update(&ca->copygc_pd,
 115                                          free, fragmented, -1);
 116         }
 117
 118         schedule_delayed_work(&c->pd_controllers_update,
 119                               c->pd_controllers_update_seconds * HZ);
 120 }
 121
 122 /* Persistent alloc info: */
 123
 124 static unsigned bch_alloc_val_u64s(const struct bch_alloc *a)
 125 {
 126         unsigned bytes = offsetof(struct bch_alloc, data);
 127
 128         if (a->fields & (1 << BCH_ALLOC_FIELD_READ_TIME))
 129                 bytes += 2;
 130         if (a->fields & (1 << BCH_ALLOC_FIELD_WRITE_TIME))
 131                 bytes += 2;
 132
 133         return DIV_ROUND_UP(bytes, sizeof(u64));
 134 }
 135
 136 const char *bch2_alloc_invalid(const struct bch_fs *c, struct bkey_s_c k)
 137 {
 138         if (k.k->p.inode >= c->sb.nr_devices ||
 139             !c->devs[k.k->p.inode])
 140                 return "invalid device";
 141
 142         switch (k.k->type) {
 143         case BCH_ALLOC: {
 144                 struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 145
 146                 if (bch_alloc_val_u64s(a.v) != bkey_val_u64s(a.k))
 147                         return "incorrect value size";
 148                 break;
 149         }
 150         default:
 151                 return "invalid type";
 152         }
 153
 154         return NULL;
 155 }
 156
 157 void bch2_alloc_to_text(struct bch_fs *c, char *buf,
 158                         size_t size, struct bkey_s_c k)
 159 {
 160         buf[0] = '\0';
 161
 162         switch (k.k->type) {
 163         case BCH_ALLOC:
 164                 break;
 165         }
 166 }
 167
 168 static inline unsigned get_alloc_field(const u8 **p, unsigned bytes)
 169 {
 170         unsigned v;
 171
 172         switch (bytes) {
 173         case 1:
 174                 v = **p;
 175                 break;
 176         case 2:
 177                 v = le16_to_cpup((void *) *p);
 178                 break;
 179         case 4:
 180                 v = le32_to_cpup((void *) *p);
 181                 break;
 182         default:
 183                 BUG();
 184         }
 185
 186         *p += bytes;
 187         return v;
 188 }
 189
 190 static inline void put_alloc_field(u8 **p, unsigned bytes, unsigned v)
 191 {
 192         switch (bytes) {
 193         case 1:
 194                 **p = v;
 195                 break;
 196         case 2:
 197                 *((__le16 *) *p) = cpu_to_le16(v);
 198                 break;
 199         case 4:
 200                 *((__le32 *) *p) = cpu_to_le32(v);
 201                 break;
 202         default:
 203                 BUG();
 204         }
 205
 206         *p += bytes;
 207 }
 208
 209 static void bch2_alloc_read_key(struct bch_fs *c, struct bkey_s_c k)
 210 {
 211         struct bch_dev *ca;
 212         struct bkey_s_c_alloc a;
 213         struct bucket_mark new;
 214         struct bucket *g;
 215         const u8 *d;
 216
 217         if (k.k->type != BCH_ALLOC)
 218                 return;
 219
 220         a = bkey_s_c_to_alloc(k);
 221         ca = bch_dev_bkey_exists(c, a.k->p.inode);
 222
 223         if (a.k->p.offset >= ca->mi.nbuckets)
 224                 return;
 225
 226         lg_local_lock(&c->usage_lock);
 227
 228         g = bucket(ca, a.k->p.offset);
 229         bucket_cmpxchg(g, new, ({
 230                 new.gen = a.v->gen;
 231                 new.gen_valid = 1;
 232         }));
 233
 234         d = a.v->data;
 235         if (a.v->fields & (1 << BCH_ALLOC_FIELD_READ_TIME))
 236                 g->io_time[READ] = get_alloc_field(&d, 2);
 237         if (a.v->fields & (1 << BCH_ALLOC_FIELD_WRITE_TIME))
 238                 g->io_time[WRITE] = get_alloc_field(&d, 2);
 239
 240         lg_local_unlock(&c->usage_lock);
 241 }
 242
 243 int bch2_alloc_read(struct bch_fs *c, struct list_head *journal_replay_list)
 244 {
 245         struct journal_replay *r;
 246         struct btree_iter iter;
 247         struct bkey_s_c k;
 248         struct bch_dev *ca;
 249         unsigned i;
 250         int ret;
 251
 252         for_each_btree_key(&iter, c, BTREE_ID_ALLOC, POS_MIN, 0, k) {
 253                 bch2_alloc_read_key(c, k);
 254                 bch2_btree_iter_cond_resched(&iter);
 255         }
 256
 257         ret = bch2_btree_iter_unlock(&iter);
 258         if (ret)
 259                 return ret;
 260
 261         list_for_each_entry(r, journal_replay_list, list) {
 262                 struct bkey_i *k, *n;
 263                 struct jset_entry *entry;
 264
 265                 for_each_jset_key(k, n, entry, &r->j)
 266                         if (entry->btree_id == BTREE_ID_ALLOC)
 267                                 bch2_alloc_read_key(c, bkey_i_to_s_c(k));
 268         }
 269
 270         mutex_lock(&c->bucket_clock[READ].lock);
 271         for_each_member_device(ca, c, i) {
 272                 down_read(&ca->bucket_lock);
 273                 bch2_recalc_oldest_io(c, ca, READ);
 274                 up_read(&ca->bucket_lock);
 275         }
 276         mutex_unlock(&c->bucket_clock[READ].lock);
 277
 278         mutex_lock(&c->bucket_clock[WRITE].lock);
 279         for_each_member_device(ca, c, i) {
 280                 down_read(&ca->bucket_lock);
 281                 bch2_recalc_oldest_io(c, ca, WRITE);
 282                 up_read(&ca->bucket_lock);
 283         }
 284         mutex_unlock(&c->bucket_clock[WRITE].lock);
 285
 286         return 0;
 287 }
 288
 289 static int __bch2_alloc_write_key(struct bch_fs *c, struct bch_dev *ca,
 290                                   size_t b, struct btree_iter *iter,
 291                                   u64 *journal_seq)
 292 {
 293         struct bucket_mark m;
 294         __BKEY_PADDED(k, DIV_ROUND_UP(sizeof(struct bch_alloc), 8)) alloc_key;
 295         struct bucket *g;
 296         struct bkey_i_alloc *a;
 297         u8 *d;
 298         int ret;
 299
 300         bch2_btree_iter_set_pos(iter, POS(ca->dev_idx, b));
 301
 302         do {
 303                 ret = btree_iter_err(bch2_btree_iter_peek_slot(iter));
 304                 if (ret)
 305                         break;
 306
 307                 lg_local_lock(&c->usage_lock);
 308                 g = bucket(ca, b);
 309
 310                 /* read mark under btree node lock: */
 311                 m = READ_ONCE(g->mark);
 312                 a = bkey_alloc_init(&alloc_key.k);
 313                 a->k.p          = iter->pos;
 314                 a->v.fields     = 0;
 315                 a->v.gen        = m.gen;
 316                 set_bkey_val_u64s(&a->k, bch_alloc_val_u64s(&a->v));
 317
 318                 d = a->v.data;
 319                 if (a->v.fields & (1 << BCH_ALLOC_FIELD_READ_TIME))
 320                         put_alloc_field(&d, 2, g->io_time[READ]);
 321                 if (a->v.fields & (1 << BCH_ALLOC_FIELD_WRITE_TIME))
 322                         put_alloc_field(&d, 2, g->io_time[WRITE]);
 323                 lg_local_unlock(&c->usage_lock);
 324
 325                 ret = bch2_btree_insert_at(c, NULL, NULL, journal_seq,
 326                                            BTREE_INSERT_ATOMIC|
 327                                            BTREE_INSERT_NOFAIL|
 328                                            BTREE_INSERT_USE_RESERVE|
 329                                            BTREE_INSERT_USE_ALLOC_RESERVE|
 330                                            BTREE_INSERT_NOWAIT,
 331                                            BTREE_INSERT_ENTRY(iter, &a->k_i));
 332                 bch2_btree_iter_cond_resched(iter);
 333         } while (ret == -EINTR);
 334
 335         return ret;
 336 }
 337
 338 int bch2_alloc_replay_key(struct bch_fs *c, struct bpos pos)
 339 {
 340         struct bch_dev *ca;
 341         struct btree_iter iter;
 342         int ret;
 343
 344         if (pos.inode >= c->sb.nr_devices || !c->devs[pos.inode])
 345                 return 0;
 346
 347         ca = bch_dev_bkey_exists(c, pos.inode);
 348
 349         if (pos.offset >= ca->mi.nbuckets)
 350                 return 0;
 351
 352         bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, POS_MIN,
 353                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 354
 355         ret = __bch2_alloc_write_key(c, ca, pos.offset, &iter, NULL);
 356         bch2_btree_iter_unlock(&iter);
 357         return ret;
 358 }
 359
 360 int bch2_alloc_write(struct bch_fs *c)
 361 {
 362         struct bch_dev *ca;
 363         unsigned i;
 364         int ret = 0;
 365
 366         for_each_rw_member(ca, c, i) {
 367                 struct btree_iter iter;
 368                 unsigned long bucket;
 369
 370                 bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, POS_MIN,
 371                                      BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 372
 373                 down_read(&ca->bucket_lock);
 374                 for_each_set_bit(bucket, ca->buckets_dirty, ca->mi.nbuckets) {
 375                         ret = __bch2_alloc_write_key(c, ca, bucket, &iter, NULL);
 376                         if (ret)
 377                                 break;
 378
 379                         clear_bit(bucket, ca->buckets_dirty);
 380                 }
 381                 up_read(&ca->bucket_lock);
 382                 bch2_btree_iter_unlock(&iter);
 383
 384                 if (ret) {
 385                         percpu_ref_put(&ca->io_ref);
 386                         break;
 387                 }
 388         }
 389
 390         return ret;
 391 }
 392
 393 /* Bucket IO clocks: */
 394
 395 static void bch2_recalc_oldest_io(struct bch_fs *c, struct bch_dev *ca, int rw)
 396 {
 397         struct bucket_clock *clock = &c->bucket_clock[rw];
 398         struct bucket_array *buckets = bucket_array(ca);
 399         struct bucket *g;
 400         u16 max_last_io = 0;
 401         unsigned i;
 402
 403         lockdep_assert_held(&c->bucket_clock[rw].lock);
 404
 405         /* Recalculate max_last_io for this device: */
 406         for_each_bucket(g, buckets)
 407                 max_last_io = max(max_last_io, bucket_last_io(c, g, rw));
 408
 409         ca->max_last_bucket_io[rw] = max_last_io;
 410
 411         /* Recalculate global max_last_io: */
 412         max_last_io = 0;
 413
 414         for_each_member_device(ca, c, i)
 415                 max_last_io = max(max_last_io, ca->max_last_bucket_io[rw]);
 416
 417         clock->max_last_io = max_last_io;
 418 }
 419
 420 static void bch2_rescale_bucket_io_times(struct bch_fs *c, int rw)
 421 {
 422         struct bucket_clock *clock = &c->bucket_clock[rw];
 423         struct bucket_array *buckets;
 424         struct bch_dev *ca;
 425         struct bucket *g;
 426         unsigned i;
 427
 428         trace_rescale_prios(c);
 429
 430         for_each_member_device(ca, c, i) {
 431                 down_read(&ca->bucket_lock);
 432                 buckets = bucket_array(ca);
 433
 434                 for_each_bucket(g, buckets)
 435                         g->io_time[rw] = clock->hand -
 436                         bucket_last_io(c, g, rw) / 2;
 437
 438                 bch2_recalc_oldest_io(c, ca, rw);
 439
 440                 up_read(&ca->bucket_lock);
 441         }
 442 }
 443
 444 static void bch2_inc_clock_hand(struct io_timer *timer)
 445 {
 446         struct bucket_clock *clock = container_of(timer,
 447                                                 struct bucket_clock, rescale);
 448         struct bch_fs *c = container_of(clock,
 449                                         struct bch_fs, bucket_clock[clock->rw]);
 450         struct bch_dev *ca;
 451         u64 capacity;
 452         unsigned i;
 453
 454         mutex_lock(&clock->lock);
 455
 456         /* if clock cannot be advanced more, rescale prio */
 457         if (clock->max_last_io >= U16_MAX - 2)
 458                 bch2_rescale_bucket_io_times(c, clock->rw);
 459
 460         BUG_ON(clock->max_last_io >= U16_MAX - 2);
 461
 462         for_each_member_device(ca, c, i)
 463                 ca->max_last_bucket_io[clock->rw]++;
 464         clock->max_last_io++;
 465         clock->hand++;
 466
 467         mutex_unlock(&clock->lock);
 468
 469         capacity = READ_ONCE(c->capacity);
 470
 471         if (!capacity)
 472                 return;
 473
 474         /*
 475          * we only increment when 0.1% of the filesystem capacity has been read
 476          * or written too, this determines if it's time
 477          *
 478          * XXX: we shouldn't really be going off of the capacity of devices in
 479          * RW mode (that will be 0 when we're RO, yet we can still service
 480          * reads)
 481          */
 482         timer->expire += capacity >> 10;
 483
 484         bch2_io_timer_add(&c->io_clock[clock->rw], timer);
 485 }
 486
 487 static void bch2_bucket_clock_init(struct bch_fs *c, int rw)
 488 {
 489         struct bucket_clock *clock = &c->bucket_clock[rw];
 490
 491         clock->hand             = 1;
 492         clock->rw               = rw;
 493         clock->rescale.fn       = bch2_inc_clock_hand;
 494         clock->rescale.expire   = c->capacity >> 10;
 495         mutex_init(&clock->lock);
 496 }
 497
 498 /* Background allocator thread: */
 499
 500 /*
 501  * Scans for buckets to be invalidated, invalidates them, rewrites prios/gens
 502  * (marking them as invalidated on disk), then optionally issues discard
 503  * commands to the newly free buckets, then puts them on the various freelists.
 504  */
 505
 506 static void verify_not_on_freelist(struct bch_fs *c, struct bch_dev *ca,
 507                                    size_t bucket)
 508 {
 509         if (expensive_debug_checks(c) &&
 510             test_bit(BCH_FS_ALLOCATOR_STARTED, &c->flags)) {
 511                 size_t iter;
 512                 long i;
 513                 unsigned j;
 514
 515                 for (j = 0; j < RESERVE_NR; j++)
 516                         fifo_for_each_entry(i, &ca->free[j], iter)
 517                                 BUG_ON(i == bucket);
 518                 fifo_for_each_entry(i, &ca->free_inc, iter)
 519                         BUG_ON(i == bucket);
 520         }
 521 }
 522
 523 #define BUCKET_GC_GEN_MAX       96U
 524
 525 /**
 526  * wait_buckets_available - wait on reclaimable buckets
 527  *
 528  * If there aren't enough available buckets to fill up free_inc, wait until
 529  * there are.
 530  */
 531 static int wait_buckets_available(struct bch_fs *c, struct bch_dev *ca)
 532 {
 533         unsigned long gc_count = c->gc_count;
 534         int ret = 0;
 535
 536         while (1) {
 537                 set_current_state(TASK_INTERRUPTIBLE);
 538                 if (kthread_should_stop()) {
 539                         ret = 1;
 540                         break;
 541                 }
 542
 543                 if (gc_count != c->gc_count)
 544                         ca->inc_gen_really_needs_gc = 0;
 545
 546                 if ((ssize_t) (dev_buckets_available(c, ca) -
 547                                ca->inc_gen_really_needs_gc) >=
 548                     (ssize_t) fifo_free(&ca->free_inc))
 549                         break;
 550
 551                 up_read(&c->gc_lock);
 552                 schedule();
 553                 try_to_freeze();
 554                 down_read(&c->gc_lock);
 555         }
 556
 557         __set_current_state(TASK_RUNNING);
 558         return ret;
 559 }
 560
 561 static bool bch2_can_invalidate_bucket(struct bch_dev *ca,
 562                                        size_t bucket,
 563                                        struct bucket_mark mark)
 564 {
 565         u8 gc_gen;
 566
 567         if (!is_available_bucket(mark))
 568                 return false;
 569
 570         gc_gen = bucket_gc_gen(ca, bucket);
 571
 572         if (gc_gen >= BUCKET_GC_GEN_MAX / 2)
 573                 ca->inc_gen_needs_gc++;
 574
 575         if (gc_gen >= BUCKET_GC_GEN_MAX)
 576                 ca->inc_gen_really_needs_gc++;
 577
 578         return gc_gen < BUCKET_GC_GEN_MAX;
 579 }
 580
 581 static void bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca,
 582                                        size_t bucket)
 583 {
 584         struct bucket_mark m;
 585
 586         spin_lock(&c->freelist_lock);
 587         if (!bch2_invalidate_bucket(c, ca, bucket, &m)) {
 588                 spin_unlock(&c->freelist_lock);
 589                 return;
 590         }
 591
 592         verify_not_on_freelist(c, ca, bucket);
 593         BUG_ON(!fifo_push(&ca->free_inc, bucket));
 594         spin_unlock(&c->freelist_lock);
 595
 596         /* gc lock held: */
 597         bucket_io_clock_reset(c, ca, bucket, READ);
 598         bucket_io_clock_reset(c, ca, bucket, WRITE);
 599
 600         if (m.cached_sectors) {
 601                 ca->allocator_invalidating_data = true;
 602         } else if (m.journal_seq_valid) {
 603                 u64 journal_seq = atomic64_read(&c->journal.seq);
 604                 u64 bucket_seq  = journal_seq;
 605
 606                 bucket_seq &= ~((u64) U16_MAX);
 607                 bucket_seq |= m.journal_seq;
 608
 609                 if (bucket_seq > journal_seq)
 610                         bucket_seq -= 1 << 16;
 611
 612                 ca->allocator_journal_seq_flush =
 613                         max(ca->allocator_journal_seq_flush, bucket_seq);
 614         }
 615 }
 616
 617 /*
 618  * Determines what order we're going to reuse buckets, smallest bucket_key()
 619  * first.
 620  *
 621  *
 622  * - We take into account the read prio of the bucket, which gives us an
 623  *   indication of how hot the data is -- we scale the prio so that the prio
 624  *   farthest from the clock is worth 1/8th of the closest.
 625  *
 626  * - The number of sectors of cached data in the bucket, which gives us an
 627  *   indication of the cost in cache misses this eviction will cause.
 628  *
 629  * - If hotness * sectors used compares equal, we pick the bucket with the
 630  *   smallest bucket_gc_gen() - since incrementing the same bucket's generation
 631  *   number repeatedly forces us to run mark and sweep gc to avoid generation
 632  *   number wraparound.
 633  */
 634
 635 static unsigned long bucket_sort_key(struct bch_fs *c, struct bch_dev *ca,
 636                                      size_t b, struct bucket_mark m)
 637 {
 638         unsigned last_io = bucket_last_io(c, bucket(ca, b), READ);
 639         unsigned max_last_io = ca->max_last_bucket_io[READ];
 640
 641         /*
 642          * Time since last read, scaled to [0, 8) where larger value indicates
 643          * more recently read data:
 644          */
 645         unsigned long hotness = (max_last_io - last_io) * 7 / max_last_io;
 646
 647         /* How much we want to keep the data in this bucket: */
 648         unsigned long data_wantness =
 649                 (hotness + 1) * bucket_sectors_used(m);
 650
 651         unsigned long needs_journal_commit =
 652                 bucket_needs_journal_commit(m, c->journal.last_seq_ondisk);
 653
 654         return  (data_wantness << 9) |
 655                 (needs_journal_commit << 8) |
 656                 bucket_gc_gen(ca, b);
 657 }
 658
 659 static inline int bucket_alloc_cmp(alloc_heap *h,
 660                                    struct alloc_heap_entry l,
 661                                    struct alloc_heap_entry r)
 662 {
 663         return (l.key > r.key) - (l.key < r.key) ?:
 664                 (l.nr < r.nr)  - (l.nr  > r.nr) ?:
 665                 (l.bucket > r.bucket) - (l.bucket < r.bucket);
 666 }
 667
 668 static void find_reclaimable_buckets_lru(struct bch_fs *c, struct bch_dev *ca)
 669 {
 670         struct bucket_array *buckets;
 671         struct alloc_heap_entry e = { 0 };
 672         size_t b;
 673
 674         ca->alloc_heap.used = 0;
 675
 676         mutex_lock(&c->bucket_clock[READ].lock);
 677         down_read(&ca->bucket_lock);
 678
 679         buckets = bucket_array(ca);
 680
 681         bch2_recalc_oldest_io(c, ca, READ);
 682
 683         /*
 684          * Find buckets with lowest read priority, by building a maxheap sorted
 685          * by read priority and repeatedly replacing the maximum element until
 686          * all buckets have been visited.
 687          */
 688         for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++) {
 689                 struct bucket_mark m = READ_ONCE(buckets->b[b].mark);
 690                 unsigned long key = bucket_sort_key(c, ca, b, m);
 691
 692                 if (!bch2_can_invalidate_bucket(ca, b, m))
 693                         continue;
 694
 695                 if (e.nr && e.bucket + e.nr == b && e.key == key) {
 696                         e.nr++;
 697                 } else {
 698                         if (e.nr)
 699                                 heap_add_or_replace(&ca->alloc_heap, e, -bucket_alloc_cmp);
 700
 701                         e = (struct alloc_heap_entry) {
 702                                 .bucket = b,
 703                                 .nr     = 1,
 704                                 .key    = key,
 705                         };
 706                 }
 707
 708                 cond_resched();
 709         }
 710
 711         if (e.nr)
 712                 heap_add_or_replace(&ca->alloc_heap, e, -bucket_alloc_cmp);
 713
 714         up_read(&ca->bucket_lock);
 715         mutex_unlock(&c->bucket_clock[READ].lock);
 716
 717         heap_resort(&ca->alloc_heap, bucket_alloc_cmp);
 718
 719         while (heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp)) {
 720                 for (b = e.bucket;
 721                      b < e.bucket + e.nr;
 722                      b++) {
 723                         if (fifo_full(&ca->free_inc))
 724                                 return;
 725
 726                         bch2_invalidate_one_bucket(c, ca, b);
 727                 }
 728         }
 729 }
 730
 731 static void find_reclaimable_buckets_fifo(struct bch_fs *c, struct bch_dev *ca)
 732 {
 733         struct bucket_array *buckets = bucket_array(ca);
 734         struct bucket_mark m;
 735         size_t b, checked;
 736
 737         for (checked = 0;
 738              checked < ca->mi.nbuckets && !fifo_full(&ca->free_inc);
 739              checked++) {
 740                 if (ca->fifo_last_bucket <  ca->mi.first_bucket ||
 741                     ca->fifo_last_bucket >= ca->mi.nbuckets)
 742                         ca->fifo_last_bucket = ca->mi.first_bucket;
 743
 744                 b = ca->fifo_last_bucket++;
 745
 746                 m = READ_ONCE(buckets->b[b].mark);
 747
 748                 if (bch2_can_invalidate_bucket(ca, b, m))
 749                         bch2_invalidate_one_bucket(c, ca, b);
 750
 751                 cond_resched();
 752         }
 753 }
 754
 755 static void find_reclaimable_buckets_random(struct bch_fs *c, struct bch_dev *ca)
 756 {
 757         struct bucket_array *buckets = bucket_array(ca);
 758         struct bucket_mark m;
 759         size_t checked;
 760
 761         for (checked = 0;
 762              checked < ca->mi.nbuckets / 2 && !fifo_full(&ca->free_inc);
 763              checked++) {
 764                 size_t b = bch2_rand_range(ca->mi.nbuckets -
 765                                            ca->mi.first_bucket) +
 766                         ca->mi.first_bucket;
 767
 768                 m = READ_ONCE(buckets->b[b].mark);
 769
 770                 if (bch2_can_invalidate_bucket(ca, b, m))
 771                         bch2_invalidate_one_bucket(c, ca, b);
 772
 773                 cond_resched();
 774         }
 775 }
 776
 777 static void find_reclaimable_buckets(struct bch_fs *c, struct bch_dev *ca)
 778 {
 779         ca->inc_gen_needs_gc                    = 0;
 780         ca->inc_gen_really_needs_gc             = 0;
 781
 782         switch (ca->mi.replacement) {
 783         case CACHE_REPLACEMENT_LRU:
 784                 find_reclaimable_buckets_lru(c, ca);
 785                 break;
 786         case CACHE_REPLACEMENT_FIFO:
 787                 find_reclaimable_buckets_fifo(c, ca);
 788                 break;
 789         case CACHE_REPLACEMENT_RANDOM:
 790                 find_reclaimable_buckets_random(c, ca);
 791                 break;
 792         }
 793 }
 794
 795 static int size_t_cmp(const void *_l, const void *_r)
 796 {
 797         const size_t *l = _l, *r = _r;
 798
 799         return (*l > *r) - (*l < *r);
 800 }
 801
 802 static void sort_free_inc(struct bch_fs *c, struct bch_dev *ca)
 803 {
 804         BUG_ON(ca->free_inc.front);
 805
 806         spin_lock(&c->freelist_lock);
 807         sort(ca->free_inc.data,
 808              ca->free_inc.back,
 809              sizeof(ca->free_inc.data[0]),
 810              size_t_cmp, NULL);
 811         spin_unlock(&c->freelist_lock);
 812 }
 813
 814 static int bch2_invalidate_free_inc(struct bch_fs *c, struct bch_dev *ca,
 815                                     u64 *journal_seq, size_t nr)
 816 {
 817         struct btree_iter iter;
 818         int ret = 0;
 819
 820         bch2_btree_iter_init(&iter, c, BTREE_ID_ALLOC, POS(ca->dev_idx, 0),
 821                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 822
 823         /*
 824          * XXX: if ca->nr_invalidated != 0, just return if we'd block doing the
 825          * btree update or journal_res_get
 826          */
 827         while (ca->nr_invalidated < min(nr, fifo_used(&ca->free_inc))) {
 828                 size_t b = fifo_idx_entry(&ca->free_inc, ca->nr_invalidated);
 829
 830                 ret = __bch2_alloc_write_key(c, ca, b, &iter, journal_seq);
 831                 if (ret)
 832                         break;
 833
 834                 ca->nr_invalidated++;
 835         }
 836
 837         bch2_btree_iter_unlock(&iter);
 838         return ret;
 839 }
 840
 841 static bool __push_invalidated_bucket(struct bch_fs *c, struct bch_dev *ca, size_t bucket)
 842 {
 843         unsigned i;
 844
 845         /*
 846          * Don't remove from free_inc until after it's added to
 847          * freelist, so gc can find it:
 848          */
 849         spin_lock(&c->freelist_lock);
 850         for (i = 0; i < RESERVE_NR; i++)
 851                 if (fifo_push(&ca->free[i], bucket)) {
 852                         fifo_pop(&ca->free_inc, bucket);
 853                         --ca->nr_invalidated;
 854                         closure_wake_up(&c->freelist_wait);
 855                         spin_unlock(&c->freelist_lock);
 856                         return true;
 857                 }
 858         spin_unlock(&c->freelist_lock);
 859
 860         return false;
 861 }
 862
 863 static int push_invalidated_bucket(struct bch_fs *c, struct bch_dev *ca, size_t bucket)
 864 {
 865         int ret = 0;
 866
 867         while (1) {
 868                 set_current_state(TASK_INTERRUPTIBLE);
 869
 870                 if (__push_invalidated_bucket(c, ca, bucket))
 871                         break;
 872
 873                 if ((current->flags & PF_KTHREAD) &&
 874                     kthread_should_stop()) {
 875                         ret = 1;
 876                         break;
 877                 }
 878
 879                 schedule();
 880                 try_to_freeze();
 881         }
 882
 883         __set_current_state(TASK_RUNNING);
 884         return ret;
 885 }
 886
 887 /*
 888  * Given an invalidated, ready to use bucket: issue a discard to it if enabled,
 889  * then add it to the freelist, waiting until there's room if necessary:
 890  */
 891 static int discard_invalidated_buckets(struct bch_fs *c, struct bch_dev *ca)
 892 {
 893         while (ca->nr_invalidated) {
 894                 size_t bucket = fifo_peek(&ca->free_inc);
 895
 896                 BUG_ON(fifo_empty(&ca->free_inc) || !ca->nr_invalidated);
 897
 898                 if (ca->mi.discard &&
 899                     blk_queue_discard(bdev_get_queue(ca->disk_sb.bdev)))
 900                         blkdev_issue_discard(ca->disk_sb.bdev,
 901                                              bucket_to_sector(ca, bucket),
 902                                              ca->mi.bucket_size, GFP_NOIO, 0);
 903
 904                 if (push_invalidated_bucket(c, ca, bucket))
 905                         return 1;
 906         }
 907
 908         return 0;
 909 }
 910
 911 /**
 912  * bch_allocator_thread - move buckets from free_inc to reserves
 913  *
 914  * The free_inc FIFO is populated by find_reclaimable_buckets(), and
 915  * the reserves are depleted by bucket allocation. When we run out
 916  * of free_inc, try to invalidate some buckets and write out
 917  * prios and gens.
 918  */
 919 static int bch2_allocator_thread(void *arg)
 920 {
 921         struct bch_dev *ca = arg;
 922         struct bch_fs *c = ca->fs;
 923         u64 journal_seq;
 924         int ret;
 925
 926         set_freezable();
 927
 928         while (1) {
 929                 while (1) {
 930                         cond_resched();
 931
 932                         pr_debug("discarding %zu invalidated buckets",
 933                                  ca->nr_invalidated);
 934
 935                         ret = discard_invalidated_buckets(c, ca);
 936                         if (ret)
 937                                 goto stop;
 938
 939                         if (fifo_empty(&ca->free_inc))
 940                                 break;
 941
 942                         pr_debug("invalidating %zu buckets",
 943                                  fifo_used(&ca->free_inc));
 944
 945                         journal_seq = 0;
 946                         ret = bch2_invalidate_free_inc(c, ca, &journal_seq, SIZE_MAX);
 947                         if (ret) {
 948                                 bch_err(ca, "error invalidating buckets: %i", ret);
 949                                 goto stop;
 950                         }
 951
 952                         if (!ca->nr_invalidated) {
 953                                 bch_err(ca, "allocator thread unable to make forward progress!");
 954                                 goto stop;
 955                         }
 956
 957                         if (ca->allocator_invalidating_data)
 958                                 ret = bch2_journal_flush_seq(&c->journal, journal_seq);
 959                         else if (ca->allocator_journal_seq_flush)
 960                                 ret = bch2_journal_flush_seq(&c->journal,
 961                                                        ca->allocator_journal_seq_flush);
 962
 963                         /*
 964                          * journal error - buckets haven't actually been
 965                          * invalidated, can't discard them:
 966                          */
 967                         if (ret) {
 968                                 bch_err(ca, "journal error: %i", ret);
 969                                 goto stop;
 970                         }
 971                 }
 972
 973                 pr_debug("free_inc now empty");
 974
 975                 /* Reset front/back so we can easily sort fifo entries later: */
 976                 ca->free_inc.front = ca->free_inc.back  = 0;
 977                 ca->allocator_journal_seq_flush         = 0;
 978                 ca->allocator_invalidating_data         = false;
 979
 980                 down_read(&c->gc_lock);
 981                 while (1) {
 982                         size_t prev = fifo_used(&ca->free_inc);
 983
 984                         if (test_bit(BCH_FS_GC_FAILURE, &c->flags)) {
 985                                 up_read(&c->gc_lock);
 986                                 bch_err(ca, "gc failure");
 987                                 goto stop;
 988                         }
 989
 990                         /*
 991                          * Find some buckets that we can invalidate, either
 992                          * they're completely unused, or only contain clean data
 993                          * that's been written back to the backing device or
 994                          * another cache tier
 995                          */
 996
 997                         pr_debug("scanning for reclaimable buckets");
 998
 999                         find_reclaimable_buckets(c, ca);
1000
1001                         pr_debug("found %zu buckets (free_inc %zu/%zu)",
1002                                  fifo_used(&ca->free_inc) - prev,
1003                                  fifo_used(&ca->free_inc), ca->free_inc.size);
1004
1005                         trace_alloc_batch(ca, fifo_used(&ca->free_inc),
1006                                           ca->free_inc.size);
1007
1008                         if ((ca->inc_gen_needs_gc >= ca->free_inc.size ||
1009                              (!fifo_full(&ca->free_inc) &&
1010                               ca->inc_gen_really_needs_gc >=
1011                               fifo_free(&ca->free_inc))) &&
1012                             c->gc_thread) {
1013                                 atomic_inc(&c->kick_gc);
1014                                 wake_up_process(c->gc_thread);
1015                         }
1016
1017                         if (fifo_full(&ca->free_inc))
1018                                 break;
1019
1020                         if (!fifo_empty(&ca->free_inc) &&
1021                             !fifo_full(&ca->free[RESERVE_MOVINGGC]))
1022                                 break;
1023
1024                         /*
1025                          * copygc may be waiting until either its reserve fills
1026                          * up, or we can't make forward progress:
1027                          */
1028                         ca->allocator_blocked = true;
1029                         closure_wake_up(&c->freelist_wait);
1030
1031                         ret = wait_buckets_available(c, ca);
1032                         if (ret) {
1033                                 up_read(&c->gc_lock);
1034                                 goto stop;
1035                         }
1036                 }
1037
1038                 ca->allocator_blocked = false;
1039                 up_read(&c->gc_lock);
1040
1041                 pr_debug("free_inc now %zu/%zu",
1042                          fifo_used(&ca->free_inc),
1043                          ca->free_inc.size);
1044
1045                 sort_free_inc(c, ca);
1046
1047                 /*
1048                  * free_inc is now full of newly-invalidated buckets: next,
1049                  * write out the new bucket gens:
1050                  */
1051         }
1052
1053 stop:
1054         pr_debug("alloc thread stopping (ret %i)", ret);
1055         return 0;
1056 }
1057
1058 /* Allocation */
1059
1060 /*
1061  * Open buckets represent a bucket that's currently being allocated from.  They
1062  * serve two purposes:
1063  *
1064  *  - They track buckets that have been partially allocated, allowing for
1065  *    sub-bucket sized allocations - they're used by the sector allocator below
1066  *
1067  *  - They provide a reference to the buckets they own that mark and sweep GC
1068  *    can find, until the new allocation has a pointer to it inserted into the
1069  *    btree
1070  *
1071  * When allocating some space with the sector allocator, the allocation comes
1072  * with a reference to an open bucket - the caller is required to put that
1073  * reference _after_ doing the index update that makes its allocation reachable.
1074  */
1075
1076 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
1077 {
1078         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
1079
1080         spin_lock(&ob->lock);
1081         bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr),
1082                                false, gc_pos_alloc(c, ob), 0);
1083         ob->valid = false;
1084         spin_unlock(&ob->lock);
1085
1086         spin_lock(&c->freelist_lock);
1087         ob->freelist = c->open_buckets_freelist;
1088         c->open_buckets_freelist = ob - c->open_buckets;
1089         c->open_buckets_nr_free++;
1090         spin_unlock(&c->freelist_lock);
1091
1092         closure_wake_up(&c->open_buckets_wait);
1093 }
1094
1095 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
1096 {
1097         struct open_bucket *ob;
1098
1099         BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
1100
1101         ob = c->open_buckets + c->open_buckets_freelist;
1102         c->open_buckets_freelist = ob->freelist;
1103         atomic_set(&ob->pin, 1);
1104
1105         c->open_buckets_nr_free--;
1106         return ob;
1107 }
1108
1109 /* _only_ for allocating the journal on a new device: */
1110 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
1111 {
1112         struct bucket_array *buckets;
1113         ssize_t b;
1114
1115         rcu_read_lock();
1116         buckets = bucket_array(ca);
1117
1118         for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++)
1119                 if (is_available_bucket(buckets->b[b].mark))
1120                         goto success;
1121         b = -1;
1122 success:
1123         rcu_read_unlock();
1124         return b;
1125 }
1126
1127 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
1128 {
1129         switch (reserve) {
1130         case RESERVE_ALLOC:
1131                 return 0;
1132         case RESERVE_BTREE:
1133                 return BTREE_NODE_RESERVE / 2;
1134         default:
1135                 return BTREE_NODE_RESERVE;
1136         }
1137 }
1138
1139 /**
1140  * bch_bucket_alloc - allocate a single bucket from a specific device
1141  *
1142  * Returns index of bucket on success, 0 on failure
1143  * */
1144 int bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
1145                       enum alloc_reserve reserve,
1146                       bool may_alloc_partial,
1147                       struct closure *cl)
1148 {
1149         struct bucket_array *buckets;
1150         struct open_bucket *ob;
1151         long bucket;
1152
1153         spin_lock(&c->freelist_lock);
1154         if (may_alloc_partial &&
1155             ca->open_buckets_partial_nr) {
1156                 int ret = ca->open_buckets_partial[--ca->open_buckets_partial_nr];
1157                 c->open_buckets[ret].on_partial_list = false;
1158                 spin_unlock(&c->freelist_lock);
1159                 return ret;
1160         }
1161
1162         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
1163                 if (cl)
1164                         closure_wait(&c->open_buckets_wait, cl);
1165                 spin_unlock(&c->freelist_lock);
1166                 trace_open_bucket_alloc_fail(ca, reserve);
1167                 return OPEN_BUCKETS_EMPTY;
1168         }
1169
1170         if (likely(fifo_pop(&ca->free[RESERVE_NONE], bucket)))
1171                 goto out;
1172
1173         switch (reserve) {
1174         case RESERVE_ALLOC:
1175                 if (fifo_pop(&ca->free[RESERVE_BTREE], bucket))
1176                         goto out;
1177                 break;
1178         case RESERVE_BTREE:
1179                 if (fifo_used(&ca->free[RESERVE_BTREE]) * 2 >=
1180                     ca->free[RESERVE_BTREE].size &&
1181                     fifo_pop(&ca->free[RESERVE_BTREE], bucket))
1182                         goto out;
1183                 break;
1184         case RESERVE_MOVINGGC:
1185                 if (fifo_pop(&ca->free[RESERVE_MOVINGGC], bucket))
1186                         goto out;
1187                 break;
1188         default:
1189                 break;
1190         }
1191
1192         if (cl)
1193                 closure_wait(&c->freelist_wait, cl);
1194
1195         spin_unlock(&c->freelist_lock);
1196
1197         trace_bucket_alloc_fail(ca, reserve);
1198         return FREELIST_EMPTY;
1199 out:
1200         verify_not_on_freelist(c, ca, bucket);
1201
1202         ob = bch2_open_bucket_alloc(c);
1203
1204         spin_lock(&ob->lock);
1205         lg_local_lock(&c->usage_lock);
1206         buckets = bucket_array(ca);
1207
1208         ob->valid       = true;
1209         ob->sectors_free = ca->mi.bucket_size;
1210         ob->ptr         = (struct bch_extent_ptr) {
1211                 .gen    = buckets->b[bucket].mark.gen,
1212                 .offset = bucket_to_sector(ca, bucket),
1213                 .dev    = ca->dev_idx,
1214         };
1215
1216         bucket_io_clock_reset(c, ca, bucket, READ);
1217         bucket_io_clock_reset(c, ca, bucket, WRITE);
1218
1219         lg_local_unlock(&c->usage_lock);
1220         spin_unlock(&ob->lock);
1221
1222         spin_unlock(&c->freelist_lock);
1223
1224         bch2_wake_allocator(ca);
1225
1226         trace_bucket_alloc(ca, reserve);
1227         return ob - c->open_buckets;
1228 }
1229
1230 static int __dev_alloc_cmp(struct write_point *wp,
1231                            unsigned l, unsigned r)
1232 {
1233         return ((wp->next_alloc[l] > wp->next_alloc[r]) -
1234                 (wp->next_alloc[l] < wp->next_alloc[r]));
1235 }
1236
1237 #define dev_alloc_cmp(l, r) __dev_alloc_cmp(wp, l, r)
1238
1239 struct dev_alloc_list bch2_wp_alloc_list(struct bch_fs *c,
1240                                          struct write_point *wp,
1241                                          struct bch_devs_mask *devs)
1242 {
1243         struct dev_alloc_list ret = { .nr = 0 };
1244         struct bch_dev *ca;
1245         unsigned i;
1246
1247         for_each_member_device_rcu(ca, c, i, devs)
1248                 ret.devs[ret.nr++] = i;
1249
1250         bubble_sort(ret.devs, ret.nr, dev_alloc_cmp);
1251         return ret;
1252 }
1253
1254 void bch2_wp_rescale(struct bch_fs *c, struct bch_dev *ca,
1255                      struct write_point *wp)
1256 {
1257         u64 *v = wp->next_alloc + ca->dev_idx;
1258         u64 free_space = dev_buckets_free(c, ca);
1259         u64 free_space_inv = free_space
1260                 ? div64_u64(1ULL << 48, free_space)
1261                 : 1ULL << 48;
1262         u64 scale = *v / 4;
1263
1264         if (*v + free_space_inv >= *v)
1265                 *v += free_space_inv;
1266         else
1267                 *v = U64_MAX;
1268
1269         for (v = wp->next_alloc;
1270              v < wp->next_alloc + ARRAY_SIZE(wp->next_alloc); v++)
1271                 *v = *v < scale ? 0 : *v - scale;
1272 }
1273
1274 static enum bucket_alloc_ret bch2_bucket_alloc_set(struct bch_fs *c,
1275                                         struct write_point *wp,
1276                                         unsigned nr_replicas,
1277                                         enum alloc_reserve reserve,
1278                                         struct bch_devs_mask *devs,
1279                                         struct closure *cl)
1280 {
1281         enum bucket_alloc_ret ret = NO_DEVICES;
1282         struct dev_alloc_list devs_sorted;
1283         struct bch_dev *ca;
1284         unsigned i, nr_ptrs_effective = 0;
1285         bool have_cache_dev = false;
1286
1287         BUG_ON(nr_replicas > ARRAY_SIZE(wp->ptrs));
1288
1289         for (i = wp->first_ptr; i < wp->nr_ptrs; i++) {
1290                 ca = bch_dev_bkey_exists(c, wp->ptrs[i]->ptr.dev);
1291
1292                 nr_ptrs_effective += ca->mi.durability;
1293                 have_cache_dev |= !ca->mi.durability;
1294         }
1295
1296         if (nr_ptrs_effective >= nr_replicas)
1297                 return ALLOC_SUCCESS;
1298
1299         rcu_read_lock();
1300         devs_sorted = bch2_wp_alloc_list(c, wp, devs);
1301
1302         for (i = 0; i < devs_sorted.nr; i++) {
1303                 int ob;
1304
1305                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
1306                 if (!ca)
1307                         continue;
1308
1309                 if (!ca->mi.durability &&
1310                     (have_cache_dev ||
1311                      wp->type != BCH_DATA_USER))
1312                         continue;
1313
1314                 ob = bch2_bucket_alloc(c, ca, reserve,
1315                                        wp->type == BCH_DATA_USER, cl);
1316                 if (ob < 0) {
1317                         ret = ob;
1318                         if (ret == OPEN_BUCKETS_EMPTY)
1319                                 break;
1320                         continue;
1321                 }
1322
1323                 BUG_ON(ob <= 0 || ob > U8_MAX);
1324                 BUG_ON(wp->nr_ptrs >= ARRAY_SIZE(wp->ptrs));
1325
1326                 wp->ptrs[wp->nr_ptrs++] = c->open_buckets + ob;
1327
1328                 bch2_wp_rescale(c, ca, wp);
1329
1330                 nr_ptrs_effective += ca->mi.durability;
1331                 have_cache_dev |= !ca->mi.durability;
1332
1333                 __clear_bit(ca->dev_idx, devs->d);
1334
1335                 if (nr_ptrs_effective >= nr_replicas) {
1336                         ret = ALLOC_SUCCESS;
1337                         break;
1338                 }
1339         }
1340         rcu_read_unlock();
1341
1342         EBUG_ON(reserve == RESERVE_MOVINGGC &&
1343                 ret != ALLOC_SUCCESS &&
1344                 ret != OPEN_BUCKETS_EMPTY);
1345
1346         switch (ret) {
1347         case ALLOC_SUCCESS:
1348                 return 0;
1349         case NO_DEVICES:
1350                 return -EROFS;
1351         case FREELIST_EMPTY:
1352         case OPEN_BUCKETS_EMPTY:
1353                 return cl ? -EAGAIN : -ENOSPC;
1354         default:
1355                 BUG();
1356         }
1357 }
1358
1359 /* Sector allocator */
1360
1361 static void writepoint_drop_ptr(struct bch_fs *c,
1362                                 struct write_point *wp,
1363                                 unsigned i)
1364 {
1365         struct open_bucket *ob = wp->ptrs[i];
1366         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
1367
1368         BUG_ON(ca->open_buckets_partial_nr >=
1369                ARRAY_SIZE(ca->open_buckets_partial));
1370
1371         if (wp->type == BCH_DATA_USER) {
1372                 spin_lock(&c->freelist_lock);
1373                 ob->on_partial_list = true;
1374                 ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
1375                         ob - c->open_buckets;
1376                 spin_unlock(&c->freelist_lock);
1377
1378                 closure_wake_up(&c->open_buckets_wait);
1379                 closure_wake_up(&c->freelist_wait);
1380         } else {
1381                 bch2_open_bucket_put(c, ob);
1382         }
1383
1384         array_remove_item(wp->ptrs, wp->nr_ptrs, i);
1385
1386         if (i < wp->first_ptr)
1387                 wp->first_ptr--;
1388 }
1389
1390 static void writepoint_drop_ptrs(struct bch_fs *c,
1391                                  struct write_point *wp,
1392                                  u16 target, bool in_target)
1393 {
1394         int i;
1395
1396         for (i = wp->first_ptr - 1; i >= 0; --i) {
1397                 struct bch_dev *ca = bch_dev_bkey_exists(c, wp->ptrs[i]->ptr.dev);
1398
1399                 if (dev_in_target(ca, target) == in_target)
1400                         writepoint_drop_ptr(c, wp, i);
1401         }
1402 }
1403
1404 static void verify_not_stale(struct bch_fs *c, const struct write_point *wp)
1405 {
1406 #ifdef CONFIG_BCACHEFS_DEBUG
1407         struct open_bucket *ob;
1408         unsigned i;
1409
1410         writepoint_for_each_ptr_all(wp, ob, i) {
1411                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
1412
1413                 BUG_ON(ptr_stale(ca, &ob->ptr));
1414         }
1415 #endif
1416 }
1417
1418 static int open_bucket_add_buckets(struct bch_fs *c,
1419                                    u16 target,
1420                                    struct write_point *wp,
1421                                    struct bch_devs_list *devs_have,
1422                                    unsigned nr_replicas,
1423                                    enum alloc_reserve reserve,
1424                                    struct closure *cl)
1425 {
1426         struct bch_devs_mask devs = c->rw_devs[wp->type];
1427         struct open_bucket *ob;
1428         unsigned i;
1429
1430         /* Don't allocate from devices we already have pointers to: */
1431         for (i = 0; i < devs_have->nr; i++)
1432                 __clear_bit(devs_have->devs[i], devs.d);
1433
1434         writepoint_for_each_ptr_all(wp, ob, i)
1435                 __clear_bit(ob->ptr.dev, devs.d);
1436
1437         if (target) {
1438                 const struct bch_devs_mask *t;
1439
1440                 rcu_read_lock();
1441                 t = bch2_target_to_mask(c, target);
1442                 if (t)
1443                         bitmap_and(devs.d, devs.d, t->d, BCH_SB_MEMBERS_MAX);
1444                 rcu_read_unlock();
1445         }
1446
1447         return bch2_bucket_alloc_set(c, wp, nr_replicas, reserve, &devs, cl);
1448 }
1449
1450 static struct write_point *__writepoint_find(struct hlist_head *head,
1451                                              unsigned long write_point)
1452 {
1453         struct write_point *wp;
1454
1455         hlist_for_each_entry_rcu(wp, head, node)
1456                 if (wp->write_point == write_point)
1457                         return wp;
1458
1459         return NULL;
1460 }
1461
1462 static struct hlist_head *writepoint_hash(struct bch_fs *c,
1463                                           unsigned long write_point)
1464 {
1465         unsigned hash =
1466                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
1467
1468         return &c->write_points_hash[hash];
1469 }
1470
1471 static struct write_point *writepoint_find(struct bch_fs *c,
1472                                            unsigned long write_point)
1473 {
1474         struct write_point *wp, *oldest;
1475         struct hlist_head *head;
1476
1477         if (!(write_point & 1UL)) {
1478                 wp = (struct write_point *) write_point;
1479                 mutex_lock(&wp->lock);
1480                 return wp;
1481         }
1482
1483         head = writepoint_hash(c, write_point);
1484 restart_find:
1485         wp = __writepoint_find(head, write_point);
1486         if (wp) {
1487 lock_wp:
1488                 mutex_lock(&wp->lock);
1489                 if (wp->write_point == write_point)
1490                         goto out;
1491                 mutex_unlock(&wp->lock);
1492                 goto restart_find;
1493         }
1494
1495         oldest = NULL;
1496         for (wp = c->write_points;
1497              wp < c->write_points + ARRAY_SIZE(c->write_points);
1498              wp++)
1499                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1500                         oldest = wp;
1501
1502         mutex_lock(&oldest->lock);
1503         mutex_lock(&c->write_points_hash_lock);
1504         wp = __writepoint_find(head, write_point);
1505         if (wp && wp != oldest) {
1506                 mutex_unlock(&c->write_points_hash_lock);
1507                 mutex_unlock(&oldest->lock);
1508                 goto lock_wp;
1509         }
1510
1511         wp = oldest;
1512         hlist_del_rcu(&wp->node);
1513         wp->write_point = write_point;
1514         hlist_add_head_rcu(&wp->node, head);
1515         mutex_unlock(&c->write_points_hash_lock);
1516 out:
1517         wp->last_used = sched_clock();
1518         return wp;
1519 }
1520
1521 /*
1522  * Get us an open_bucket we can allocate from, return with it locked:
1523  */
1524 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
1525                                 unsigned target,
1526                                 struct write_point_specifier write_point,
1527                                 struct bch_devs_list *devs_have,
1528                                 unsigned nr_replicas,
1529                                 unsigned nr_replicas_required,
1530                                 enum alloc_reserve reserve,
1531                                 unsigned flags,
1532                                 struct closure *cl)
1533 {
1534         struct write_point *wp;
1535         struct open_bucket *ob;
1536         struct bch_dev *ca;
1537         unsigned nr_ptrs_have, nr_ptrs_effective;
1538         int ret, i, cache_idx = -1;
1539
1540         BUG_ON(!nr_replicas || !nr_replicas_required);
1541
1542         wp = writepoint_find(c, write_point.v);
1543
1544         wp->first_ptr = 0;
1545
1546         /* does writepoint have ptrs we can't use? */
1547         writepoint_for_each_ptr(wp, ob, i)
1548                 if (bch2_dev_list_has_dev(*devs_have, ob->ptr.dev)) {
1549                         swap(wp->ptrs[i], wp->ptrs[wp->first_ptr]);
1550                         wp->first_ptr++;
1551                 }
1552
1553         nr_ptrs_have = wp->first_ptr;
1554
1555         /* does writepoint have ptrs we don't want to use? */
1556         if (target)
1557                 writepoint_for_each_ptr(wp, ob, i)
1558                         if (!dev_idx_in_target(c, ob->ptr.dev, target)) {
1559                                 swap(wp->ptrs[i], wp->ptrs[wp->first_ptr]);
1560                                 wp->first_ptr++;
1561                         }
1562
1563         if (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS) {
1564                 ret = open_bucket_add_buckets(c, target, wp, devs_have,
1565                                               nr_replicas, reserve, cl);
1566         } else {
1567                 ret = open_bucket_add_buckets(c, target, wp, devs_have,
1568                                               nr_replicas, reserve, NULL);
1569                 if (!ret)
1570                         goto alloc_done;
1571
1572                 wp->first_ptr = nr_ptrs_have;
1573
1574                 ret = open_bucket_add_buckets(c, 0, wp, devs_have,
1575                                               nr_replicas, reserve, cl);
1576         }
1577
1578         if (ret && ret != -EROFS)
1579                 goto err;
1580 alloc_done:
1581         /* check for more than one cache: */
1582         for (i = wp->nr_ptrs - 1; i >= wp->first_ptr; --i) {
1583                 ca = bch_dev_bkey_exists(c, wp->ptrs[i]->ptr.dev);
1584
1585                 if (ca->mi.durability)
1586                         continue;
1587
1588                 /*
1589                  * if we ended up with more than one cache device, prefer the
1590                  * one in the target we want:
1591                  */
1592                 if (cache_idx >= 0) {
1593                         if (!dev_in_target(ca, target)) {
1594                                 writepoint_drop_ptr(c, wp, i);
1595                         } else {
1596                                 writepoint_drop_ptr(c, wp, cache_idx);
1597                                 cache_idx = i;
1598                         }
1599                 } else {
1600                         cache_idx = i;
1601                 }
1602         }
1603
1604         /* we might have more effective replicas than required: */
1605         nr_ptrs_effective = 0;
1606         writepoint_for_each_ptr(wp, ob, i) {
1607                 ca = bch_dev_bkey_exists(c, ob->ptr.dev);
1608                 nr_ptrs_effective += ca->mi.durability;
1609         }
1610
1611         if (ret == -EROFS &&
1612             nr_ptrs_effective >= nr_replicas_required)
1613                 ret = 0;
1614
1615         if (ret)
1616                 goto err;
1617
1618         if (nr_ptrs_effective > nr_replicas) {
1619                 writepoint_for_each_ptr(wp, ob, i) {
1620                         ca = bch_dev_bkey_exists(c, ob->ptr.dev);
1621
1622                         if (ca->mi.durability &&
1623                             ca->mi.durability <= nr_ptrs_effective - nr_replicas &&
1624                             !dev_idx_in_target(c, ob->ptr.dev, target)) {
1625                                 swap(wp->ptrs[i], wp->ptrs[wp->first_ptr]);
1626                                 wp->first_ptr++;
1627                                 nr_ptrs_effective -= ca->mi.durability;
1628                         }
1629                 }
1630         }
1631
1632         if (nr_ptrs_effective > nr_replicas) {
1633                 writepoint_for_each_ptr(wp, ob, i) {
1634                         ca = bch_dev_bkey_exists(c, ob->ptr.dev);
1635
1636                         if (ca->mi.durability &&
1637                             ca->mi.durability <= nr_ptrs_effective - nr_replicas) {
1638                                 swap(wp->ptrs[i], wp->ptrs[wp->first_ptr]);
1639                                 wp->first_ptr++;
1640                                 nr_ptrs_effective -= ca->mi.durability;
1641                         }
1642                 }
1643         }
1644
1645         /* Remove pointers we don't want to use: */
1646         if (target)
1647                 writepoint_drop_ptrs(c, wp, target, false);
1648
1649         BUG_ON(wp->first_ptr >= wp->nr_ptrs);
1650         BUG_ON(nr_ptrs_effective < nr_replicas_required);
1651
1652         wp->sectors_free = UINT_MAX;
1653
1654         writepoint_for_each_ptr(wp, ob, i)
1655                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1656
1657         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1658
1659         verify_not_stale(c, wp);
1660
1661         return wp;
1662 err:
1663         mutex_unlock(&wp->lock);
1664         return ERR_PTR(ret);
1665 }
1666
1667 /*
1668  * Append pointers to the space we just allocated to @k, and mark @sectors space
1669  * as allocated out of @ob
1670  */
1671 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1672                                     struct bkey_i_extent *e, unsigned sectors)
1673 {
1674         struct open_bucket *ob;
1675         unsigned i;
1676
1677         BUG_ON(sectors > wp->sectors_free);
1678         wp->sectors_free -= sectors;
1679
1680         writepoint_for_each_ptr(wp, ob, i) {
1681                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
1682                 struct bch_extent_ptr tmp = ob->ptr;
1683
1684                 EBUG_ON(bch2_extent_has_device(extent_i_to_s_c(e), ob->ptr.dev));
1685
1686                 tmp.cached = bkey_extent_is_cached(&e->k) ||
1687                         (!ca->mi.durability && wp->type == BCH_DATA_USER);
1688
1689                 tmp.offset += ca->mi.bucket_size - ob->sectors_free;
1690                 extent_ptr_append(e, tmp);
1691
1692                 BUG_ON(sectors > ob->sectors_free);
1693                 ob->sectors_free -= sectors;
1694         }
1695 }
1696
1697 /*
1698  * Append pointers to the space we just allocated to @k, and mark @sectors space
1699  * as allocated out of @ob
1700  */
1701 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1702 {
1703         int i;
1704
1705         for (i = wp->nr_ptrs - 1; i >= 0; --i) {
1706                 struct open_bucket *ob = wp->ptrs[i];
1707
1708                 if (!ob->sectors_free) {
1709                         array_remove_item(wp->ptrs, wp->nr_ptrs, i);
1710                         bch2_open_bucket_put(c, ob);
1711                 }
1712         }
1713
1714         mutex_unlock(&wp->lock);
1715 }
1716
1717 /* Startup/shutdown (ro/rw): */
1718
1719 void bch2_recalc_capacity(struct bch_fs *c)
1720 {
1721         struct bch_dev *ca;
1722         u64 total_capacity, capacity = 0, reserved_sectors = 0;
1723         unsigned long ra_pages = 0;
1724         unsigned i, j;
1725
1726         lockdep_assert_held(&c->state_lock);
1727
1728         for_each_online_member(ca, c, i) {
1729                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_bdi;
1730
1731                 ra_pages += bdi->ra_pages;
1732         }
1733
1734         bch2_set_ra_pages(c, ra_pages);
1735
1736         for_each_rw_member(ca, c, i) {
1737                 size_t reserve = 0;
1738
1739                 /*
1740                  * We need to reserve buckets (from the number
1741                  * of currently available buckets) against
1742                  * foreground writes so that mainly copygc can
1743                  * make forward progress.
1744                  *
1745                  * We need enough to refill the various reserves
1746                  * from scratch - copygc will use its entire
1747                  * reserve all at once, then run against when
1748                  * its reserve is refilled (from the formerly
1749                  * available buckets).
1750                  *
1751                  * This reserve is just used when considering if
1752                  * allocations for foreground writes must wait -
1753                  * not -ENOSPC calculations.
1754                  */
1755                 for (j = 0; j < RESERVE_NONE; j++)
1756                         reserve += ca->free[j].size;
1757
1758                 reserve += ca->free_inc.size;
1759
1760                 reserve += ARRAY_SIZE(c->write_points);
1761
1762                 reserve += 1;   /* btree write point */
1763
1764                 reserved_sectors += bucket_to_sector(ca, reserve);
1765
1766                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
1767                                              ca->mi.first_bucket);
1768         }
1769
1770         total_capacity = capacity;
1771
1772         capacity *= (100 - c->opts.gc_reserve_percent);
1773         capacity = div64_u64(capacity, 100);
1774
1775         BUG_ON(reserved_sectors > total_capacity);
1776
1777         capacity = min(capacity, total_capacity - reserved_sectors);
1778
1779         c->capacity = capacity;
1780
1781         if (c->capacity) {
1782                 bch2_io_timer_add(&c->io_clock[READ],
1783                                  &c->bucket_clock[READ].rescale);
1784                 bch2_io_timer_add(&c->io_clock[WRITE],
1785                                  &c->bucket_clock[WRITE].rescale);
1786         } else {
1787                 bch2_io_timer_del(&c->io_clock[READ],
1788                                  &c->bucket_clock[READ].rescale);
1789                 bch2_io_timer_del(&c->io_clock[WRITE],
1790                                  &c->bucket_clock[WRITE].rescale);
1791         }
1792
1793         /* Wake up case someone was waiting for buckets */
1794         closure_wake_up(&c->freelist_wait);
1795 }
1796
1797 static void bch2_stop_write_point(struct bch_fs *c, struct bch_dev *ca,
1798                                   struct write_point *wp)
1799 {
1800         struct bch_devs_mask not_self;
1801
1802         bitmap_complement(not_self.d, ca->self.d, BCH_SB_MEMBERS_MAX);
1803
1804         mutex_lock(&wp->lock);
1805         wp->first_ptr = wp->nr_ptrs;
1806         writepoint_drop_ptrs(c, wp, dev_to_target(ca->dev_idx), true);
1807         mutex_unlock(&wp->lock);
1808 }
1809
1810 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
1811 {
1812         struct open_bucket *ob;
1813         bool ret = false;
1814
1815         for (ob = c->open_buckets;
1816              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1817              ob++) {
1818                 spin_lock(&ob->lock);
1819                 if (ob->valid && !ob->on_partial_list &&
1820                     ob->ptr.dev == ca->dev_idx)
1821                         ret = true;
1822                 spin_unlock(&ob->lock);
1823         }
1824
1825         return ret;
1826 }
1827
1828 /* device goes ro: */
1829 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
1830 {
1831         unsigned i;
1832
1833         BUG_ON(ca->alloc_thread);
1834
1835         /* First, remove device from allocation groups: */
1836
1837         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1838                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
1839
1840         /*
1841          * Capacity is calculated based off of devices in allocation groups:
1842          */
1843         bch2_recalc_capacity(c);
1844
1845         /* Next, close write points that point to this device... */
1846         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1847                 bch2_stop_write_point(c, ca, &c->write_points[i]);
1848
1849         bch2_stop_write_point(c, ca, &ca->copygc_write_point);
1850         bch2_stop_write_point(c, ca, &c->rebalance_write_point);
1851         bch2_stop_write_point(c, ca, &c->btree_write_point);
1852
1853         mutex_lock(&c->btree_reserve_cache_lock);
1854         while (c->btree_reserve_cache_nr) {
1855                 struct btree_alloc *a =
1856                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1857
1858                 bch2_open_bucket_put_refs(c, &a->ob.nr, a->ob.refs);
1859         }
1860         mutex_unlock(&c->btree_reserve_cache_lock);
1861
1862         /*
1863          * Wake up threads that were blocked on allocation, so they can notice
1864          * the device can no longer be removed and the capacity has changed:
1865          */
1866         closure_wake_up(&c->freelist_wait);
1867
1868         /*
1869          * journal_res_get() can block waiting for free space in the journal -
1870          * it needs to notice there may not be devices to allocate from anymore:
1871          */
1872         wake_up(&c->journal.wait);
1873
1874         /* Now wait for any in flight writes: */
1875
1876         closure_wait_event(&c->open_buckets_wait,
1877                            !bch2_dev_has_open_write_point(c, ca));
1878 }
1879
1880 /* device goes rw: */
1881 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
1882 {
1883         unsigned i;
1884
1885         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1886                 if (ca->mi.data_allowed & (1 << i))
1887                         set_bit(ca->dev_idx, c->rw_devs[i].d);
1888 }
1889
1890 /* stop allocator thread: */
1891 void bch2_dev_allocator_stop(struct bch_dev *ca)
1892 {
1893         struct task_struct *p = ca->alloc_thread;
1894
1895         ca->alloc_thread = NULL;
1896
1897         /*
1898          * We need an rcu barrier between setting ca->alloc_thread = NULL and
1899          * the thread shutting down to avoid bch2_wake_allocator() racing:
1900          *
1901          * XXX: it would be better to have the rcu barrier be asynchronous
1902          * instead of blocking us here
1903          */
1904         synchronize_rcu();
1905
1906         if (p) {
1907                 kthread_stop(p);
1908                 put_task_struct(p);
1909         }
1910 }
1911
1912 /* start allocator thread: */
1913 int bch2_dev_allocator_start(struct bch_dev *ca)
1914 {
1915         struct task_struct *p;
1916
1917         /*
1918          * allocator thread already started?
1919          */
1920         if (ca->alloc_thread)
1921                 return 0;
1922
1923         p = kthread_create(bch2_allocator_thread, ca,
1924                            "bch_alloc[%s]", ca->name);
1925         if (IS_ERR(p))
1926                 return PTR_ERR(p);
1927
1928         get_task_struct(p);
1929         ca->alloc_thread = p;
1930         wake_up_process(p);
1931         return 0;
1932 }
1933
1934 static void allocator_start_issue_discards(struct bch_fs *c)
1935 {
1936         struct bch_dev *ca;
1937         unsigned dev_iter;
1938         size_t i, bu;
1939
1940         for_each_rw_member(ca, c, dev_iter) {
1941                 unsigned done = 0;
1942
1943                 fifo_for_each_entry(bu, &ca->free_inc, i) {
1944                         if (done == ca->nr_invalidated)
1945                                 break;
1946
1947                         blkdev_issue_discard(ca->disk_sb.bdev,
1948                                              bucket_to_sector(ca, bu),
1949                                              ca->mi.bucket_size, GFP_NOIO, 0);
1950                         done++;
1951                 }
1952         }
1953 }
1954
1955 static int __bch2_fs_allocator_start(struct bch_fs *c)
1956 {
1957         struct bch_dev *ca;
1958         size_t bu, i;
1959         unsigned dev_iter;
1960         u64 journal_seq = 0;
1961         bool invalidating_data = false;
1962         int ret = 0;
1963
1964         if (test_bit(BCH_FS_GC_FAILURE, &c->flags))
1965                 return -1;
1966
1967         /* Scan for buckets that are already invalidated: */
1968         for_each_rw_member(ca, c, dev_iter) {
1969                 struct btree_iter iter;
1970                 struct bucket_mark m;
1971                 struct bkey_s_c k;
1972
1973                 for_each_btree_key(&iter, c, BTREE_ID_ALLOC, POS(ca->dev_idx, 0), 0, k) {
1974                         if (k.k->type != BCH_ALLOC)
1975                                 continue;
1976
1977                         bu = k.k->p.offset;
1978                         m = READ_ONCE(bucket(ca, bu)->mark);
1979
1980                         if (!is_available_bucket(m) || m.cached_sectors)
1981                                 continue;
1982
1983                         bch2_mark_alloc_bucket(c, ca, bu, true,
1984                                         gc_pos_alloc(c, NULL),
1985                                         BCH_BUCKET_MARK_MAY_MAKE_UNAVAILABLE|
1986                                         BCH_BUCKET_MARK_GC_LOCK_HELD);
1987
1988                         fifo_push(&ca->free_inc, bu);
1989                         ca->nr_invalidated++;
1990
1991                         if (fifo_full(&ca->free_inc))
1992                                 break;
1993                 }
1994                 bch2_btree_iter_unlock(&iter);
1995         }
1996
1997         /* did we find enough buckets? */
1998         for_each_rw_member(ca, c, dev_iter)
1999                 if (fifo_used(&ca->free_inc) < ca->free[RESERVE_BTREE].size) {
2000                         percpu_ref_put(&ca->io_ref);
2001                         goto not_enough;
2002                 }
2003
2004         return 0;
2005 not_enough:
2006         pr_debug("did not find enough empty buckets; issuing discards");
2007
2008         /* clear out free_inc - find_reclaimable_buckets() assumes it's empty */
2009         for_each_rw_member(ca, c, dev_iter)
2010                 discard_invalidated_buckets(c, ca);
2011
2012         pr_debug("scanning for reclaimable buckets");
2013
2014         for_each_rw_member(ca, c, dev_iter) {
2015                 BUG_ON(!fifo_empty(&ca->free_inc));
2016                 ca->free_inc.front = ca->free_inc.back  = 0;
2017
2018                 find_reclaimable_buckets(c, ca);
2019                 sort_free_inc(c, ca);
2020
2021                 invalidating_data |= ca->allocator_invalidating_data;
2022
2023                 fifo_for_each_entry(bu, &ca->free_inc, i)
2024                         if (!fifo_push(&ca->free[RESERVE_BTREE], bu))
2025                                 break;
2026         }
2027
2028         pr_debug("done scanning for reclaimable buckets");
2029
2030         /*
2031          * We're moving buckets to freelists _before_ they've been marked as
2032          * invalidated on disk - we have to so that we can allocate new btree
2033          * nodes to mark them as invalidated on disk.
2034          *
2035          * However, we can't _write_ to any of these buckets yet - they might
2036          * have cached data in them, which is live until they're marked as
2037          * invalidated on disk:
2038          */
2039         if (invalidating_data) {
2040                 pr_debug("invalidating existing data");
2041                 set_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags);
2042         } else {
2043                 pr_debug("issuing discards");
2044                 allocator_start_issue_discards(c);
2045         }
2046
2047         /*
2048          * XXX: it's possible for this to deadlock waiting on journal reclaim,
2049          * since we're holding btree writes. What then?
2050          */
2051
2052         for_each_rw_member(ca, c, dev_iter) {
2053                 ret = bch2_invalidate_free_inc(c, ca, &journal_seq,
2054                                                ca->free[RESERVE_BTREE].size);
2055                 if (ret) {
2056                         percpu_ref_put(&ca->io_ref);
2057                         return ret;
2058                 }
2059         }
2060
2061         if (invalidating_data) {
2062                 pr_debug("flushing journal");
2063
2064                 ret = bch2_journal_flush_seq(&c->journal, journal_seq);
2065                 if (ret)
2066                         return ret;
2067
2068                 pr_debug("issuing discards");
2069                 allocator_start_issue_discards(c);
2070         }
2071
2072         for_each_rw_member(ca, c, dev_iter)
2073                 while (ca->nr_invalidated) {
2074                         BUG_ON(!fifo_pop(&ca->free_inc, bu));
2075                         ca->nr_invalidated--;
2076                 }
2077
2078         set_bit(BCH_FS_ALLOCATOR_STARTED, &c->flags);
2079
2080         /* now flush dirty btree nodes: */
2081         if (invalidating_data) {
2082                 struct bucket_table *tbl;
2083                 struct rhash_head *pos;
2084                 struct btree *b;
2085                 bool flush_updates;
2086                 size_t nr_pending_updates;
2087
2088                 clear_bit(BCH_FS_HOLD_BTREE_WRITES, &c->flags);
2089 again:
2090                 pr_debug("flushing dirty btree nodes");
2091                 cond_resched();
2092
2093                 flush_updates = false;
2094                 nr_pending_updates = bch2_btree_interior_updates_nr_pending(c);
2095
2096
2097                 rcu_read_lock();
2098                 for_each_cached_btree(b, c, tbl, i, pos)
2099                         if (btree_node_dirty(b) && (!b->written || b->level)) {
2100                                 if (btree_node_may_write(b)) {
2101                                         rcu_read_unlock();
2102                                         six_lock_read(&b->lock);
2103                                         bch2_btree_node_write(c, b, SIX_LOCK_read);
2104                                         six_unlock_read(&b->lock);
2105                                         goto again;
2106                                 } else {
2107                                         flush_updates = true;
2108                                 }
2109                         }
2110                 rcu_read_unlock();
2111
2112                 /*
2113                  * This is ugly, but it's needed to flush btree node writes
2114                  * without spinning...
2115                  */
2116                 if (flush_updates) {
2117                         closure_wait_event(&c->btree_interior_update_wait,
2118                                 bch2_btree_interior_updates_nr_pending(c) <
2119                                 nr_pending_updates);
2120                         goto again;
2121                 }
2122         }
2123
2124         return 0;
2125 }
2126
2127 int bch2_fs_allocator_start(struct bch_fs *c)
2128 {
2129         struct bch_dev *ca;
2130         unsigned i;
2131         int ret;
2132
2133         down_read(&c->gc_lock);
2134         ret = __bch2_fs_allocator_start(c);
2135         up_read(&c->gc_lock);
2136
2137         if (ret)
2138                 return ret;
2139
2140         for_each_rw_member(ca, c, i) {
2141                 ret = bch2_dev_allocator_start(ca);
2142                 if (ret) {
2143                         percpu_ref_put(&ca->io_ref);
2144                         return ret;
2145                 }
2146         }
2147
2148         return bch2_alloc_write(c);
2149 }
2150
2151 void bch2_fs_allocator_init(struct bch_fs *c)
2152 {
2153         struct open_bucket *ob;
2154         struct write_point *wp;
2155
2156         mutex_init(&c->write_points_hash_lock);
2157         spin_lock_init(&c->freelist_lock);
2158         bch2_bucket_clock_init(c, READ);
2159         bch2_bucket_clock_init(c, WRITE);
2160
2161         /* open bucket 0 is a sentinal NULL: */
2162         spin_lock_init(&c->open_buckets[0].lock);
2163
2164         for (ob = c->open_buckets + 1;
2165              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
2166                 spin_lock_init(&ob->lock);
2167                 c->open_buckets_nr_free++;
2168
2169                 ob->freelist = c->open_buckets_freelist;
2170                 c->open_buckets_freelist = ob - c->open_buckets;
2171         }
2172
2173         writepoint_init(&c->btree_write_point, BCH_DATA_BTREE);
2174         writepoint_init(&c->rebalance_write_point, BCH_DATA_USER);
2175
2176         for (wp = c->write_points;
2177              wp < c->write_points + ARRAY_SIZE(c->write_points); wp++) {
2178                 writepoint_init(wp, BCH_DATA_USER);
2179
2180                 wp->last_used   = sched_clock();
2181                 wp->write_point = (unsigned long) wp;
2182                 hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
2183         }
2184
2185         c->pd_controllers_update_seconds = 5;
2186         INIT_DELAYED_WORK(&c->pd_controllers_update, pd_controllers_update);
2187 }