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

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright 2012 Google, Inc.
   4  *
   5  * Foreground allocator code: allocate buckets from freelist, and allocate in
   6  * sector granularity from writepoints.
   7  *
   8  * bch2_bucket_alloc() allocates a single bucket from a specific device.
   9  *
  10  * bch2_bucket_alloc_set() allocates one or more buckets from different devices
  11  * in a given filesystem.
  12  */
  13
  14 #include "bcachefs.h"
  15 #include "alloc_background.h"
  16 #include "alloc_foreground.h"
  17 #include "btree_gc.h"
  18 #include "buckets.h"
  19 #include "clock.h"
  20 #include "debug.h"
  21 #include "disk_groups.h"
  22 #include "ec.h"
  23 #include "io.h"
  24
  25 #include <linux/math64.h>
  26 #include <linux/rculist.h>
  27 #include <linux/rcupdate.h>
  28 #include <trace/events/bcachefs.h>
  29
  30 /*
  31  * Open buckets represent a bucket that's currently being allocated from.  They
  32  * serve two purposes:
  33  *
  34  *  - They track buckets that have been partially allocated, allowing for
  35  *    sub-bucket sized allocations - they're used by the sector allocator below
  36  *
  37  *  - They provide a reference to the buckets they own that mark and sweep GC
  38  *    can find, until the new allocation has a pointer to it inserted into the
  39  *    btree
  40  *
  41  * When allocating some space with the sector allocator, the allocation comes
  42  * with a reference to an open bucket - the caller is required to put that
  43  * reference _after_ doing the index update that makes its allocation reachable.
  44  */
  45
  46 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
  47 {
  48         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
  49
  50         if (ob->ec) {
  51                 bch2_ec_bucket_written(c, ob);
  52                 return;
  53         }
  54
  55         percpu_down_read(&c->mark_lock);
  56         spin_lock(&ob->lock);
  57
  58         bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr), false);
  59         ob->valid = false;
  60         ob->type = 0;
  61
  62         spin_unlock(&ob->lock);
  63         percpu_up_read(&c->mark_lock);
  64
  65         spin_lock(&c->freelist_lock);
  66         ob->freelist = c->open_buckets_freelist;
  67         c->open_buckets_freelist = ob - c->open_buckets;
  68
  69         c->open_buckets_nr_free++;
  70         ca->nr_open_buckets--;
  71         spin_unlock(&c->freelist_lock);
  72
  73         closure_wake_up(&c->open_buckets_wait);
  74 }
  75
  76 void bch2_open_bucket_write_error(struct bch_fs *c,
  77                                   struct open_buckets *obs,
  78                                   unsigned dev)
  79 {
  80         struct open_bucket *ob;
  81         unsigned i;
  82
  83         open_bucket_for_each(c, obs, ob, i)
  84                 if (ob->ptr.dev == dev &&
  85                     ob->ec)
  86                         bch2_ec_bucket_cancel(c, ob);
  87 }
  88
  89 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
  90 {
  91         struct open_bucket *ob;
  92
  93         BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
  94
  95         ob = c->open_buckets + c->open_buckets_freelist;
  96         c->open_buckets_freelist = ob->freelist;
  97         atomic_set(&ob->pin, 1);
  98         ob->type = 0;
  99
 100         c->open_buckets_nr_free--;
 101         return ob;
 102 }
 103
 104 static void open_bucket_free_unused(struct bch_fs *c,
 105                                     struct write_point *wp,
 106                                     struct open_bucket *ob)
 107 {
 108         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 109         bool may_realloc = wp->type == BCH_DATA_user;
 110
 111         BUG_ON(ca->open_buckets_partial_nr >
 112                ARRAY_SIZE(ca->open_buckets_partial));
 113
 114         if (ca->open_buckets_partial_nr <
 115             ARRAY_SIZE(ca->open_buckets_partial) &&
 116             may_realloc) {
 117                 spin_lock(&c->freelist_lock);
 118                 ob->on_partial_list = true;
 119                 ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
 120                         ob - c->open_buckets;
 121                 spin_unlock(&c->freelist_lock);
 122
 123                 closure_wake_up(&c->open_buckets_wait);
 124                 closure_wake_up(&c->freelist_wait);
 125         } else {
 126                 bch2_open_bucket_put(c, ob);
 127         }
 128 }
 129
 130 static void verify_not_stale(struct bch_fs *c, const struct open_buckets *obs)
 131 {
 132 #ifdef CONFIG_BCACHEFS_DEBUG
 133         struct open_bucket *ob;
 134         unsigned i;
 135
 136         open_bucket_for_each(c, obs, ob, i) {
 137                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 138
 139                 BUG_ON(ptr_stale(ca, &ob->ptr));
 140         }
 141 #endif
 142 }
 143
 144 /* _only_ for allocating the journal on a new device: */
 145 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
 146 {
 147         struct bucket_array *buckets;
 148         ssize_t b;
 149
 150         rcu_read_lock();
 151         buckets = bucket_array(ca);
 152
 153         for (b = buckets->first_bucket; b < buckets->nbuckets; b++)
 154                 if (is_available_bucket(buckets->b[b].mark) &&
 155                     (!ca->buckets_nouse || !test_bit(b, ca->buckets_nouse)) &&
 156                     !buckets->b[b].mark.owned_by_allocator)
 157                         goto success;
 158         b = -1;
 159 success:
 160         rcu_read_unlock();
 161         return b;
 162 }
 163
 164 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
 165 {
 166         switch (reserve) {
 167         case RESERVE_BTREE:
 168         case RESERVE_BTREE_MOVINGGC:
 169                 return 0;
 170         case RESERVE_MOVINGGC:
 171                 return OPEN_BUCKETS_COUNT / 4;
 172         default:
 173                 return OPEN_BUCKETS_COUNT / 2;
 174         }
 175 }
 176
 177 /**
 178  * bch_bucket_alloc - allocate a single bucket from a specific device
 179  *
 180  * Returns index of bucket on success, 0 on failure
 181  * */
 182 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 183                                       enum alloc_reserve reserve,
 184                                       bool may_alloc_partial,
 185                                       struct closure *cl)
 186 {
 187         struct open_bucket *ob;
 188         long b = 0;
 189
 190         spin_lock(&c->freelist_lock);
 191
 192         if (may_alloc_partial) {
 193                 int i;
 194
 195                 for (i = ca->open_buckets_partial_nr - 1; i >= 0; --i) {
 196                         ob = c->open_buckets + ca->open_buckets_partial[i];
 197
 198                         if (reserve <= ob->alloc_reserve) {
 199                                 array_remove_item(ca->open_buckets_partial,
 200                                                   ca->open_buckets_partial_nr,
 201                                                   i);
 202                                 ob->on_partial_list = false;
 203                                 ob->alloc_reserve = reserve;
 204                                 spin_unlock(&c->freelist_lock);
 205                                 return ob;
 206                         }
 207                 }
 208         }
 209
 210         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
 211                 if (cl)
 212                         closure_wait(&c->open_buckets_wait, cl);
 213
 214                 if (!c->blocked_allocate_open_bucket)
 215                         c->blocked_allocate_open_bucket = local_clock();
 216
 217                 spin_unlock(&c->freelist_lock);
 218                 trace_open_bucket_alloc_fail(ca, reserve);
 219                 return ERR_PTR(-OPEN_BUCKETS_EMPTY);
 220         }
 221
 222         if (likely(fifo_pop(&ca->free[RESERVE_NONE], b)))
 223                 goto out;
 224
 225         switch (reserve) {
 226         case RESERVE_BTREE_MOVINGGC:
 227         case RESERVE_MOVINGGC:
 228                 if (fifo_pop(&ca->free[RESERVE_MOVINGGC], b))
 229                         goto out;
 230                 break;
 231         default:
 232                 break;
 233         }
 234
 235         if (cl)
 236                 closure_wait(&c->freelist_wait, cl);
 237
 238         if (!c->blocked_allocate)
 239                 c->blocked_allocate = local_clock();
 240
 241         spin_unlock(&c->freelist_lock);
 242
 243         trace_bucket_alloc_fail(ca, reserve);
 244         return ERR_PTR(-FREELIST_EMPTY);
 245 out:
 246         verify_not_on_freelist(c, ca, b);
 247
 248         ob = bch2_open_bucket_alloc(c);
 249
 250         spin_lock(&ob->lock);
 251
 252         ob->valid       = true;
 253         ob->sectors_free = ca->mi.bucket_size;
 254         ob->alloc_reserve = reserve;
 255         ob->ptr         = (struct bch_extent_ptr) {
 256                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
 257                 .gen    = bucket(ca, b)->mark.gen,
 258                 .offset = bucket_to_sector(ca, b),
 259                 .dev    = ca->dev_idx,
 260         };
 261
 262         spin_unlock(&ob->lock);
 263
 264         if (c->blocked_allocate_open_bucket) {
 265                 bch2_time_stats_update(
 266                         &c->times[BCH_TIME_blocked_allocate_open_bucket],
 267                         c->blocked_allocate_open_bucket);
 268                 c->blocked_allocate_open_bucket = 0;
 269         }
 270
 271         if (c->blocked_allocate) {
 272                 bch2_time_stats_update(
 273                         &c->times[BCH_TIME_blocked_allocate],
 274                         c->blocked_allocate);
 275                 c->blocked_allocate = 0;
 276         }
 277
 278         ca->nr_open_buckets++;
 279         spin_unlock(&c->freelist_lock);
 280
 281         bch2_wake_allocator(ca);
 282
 283         trace_bucket_alloc(ca, reserve);
 284         return ob;
 285 }
 286
 287 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 288                             unsigned l, unsigned r)
 289 {
 290         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 291                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 292 }
 293
 294 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 295
 296 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 297                                           struct dev_stripe_state *stripe,
 298                                           struct bch_devs_mask *devs)
 299 {
 300         struct dev_alloc_list ret = { .nr = 0 };
 301         unsigned i;
 302
 303         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 304                 ret.devs[ret.nr++] = i;
 305
 306         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 307         return ret;
 308 }
 309
 310 void bch2_dev_stripe_increment(struct bch_dev *ca,
 311                                struct dev_stripe_state *stripe)
 312 {
 313         u64 *v = stripe->next_alloc + ca->dev_idx;
 314         u64 free_space = dev_buckets_available(ca);
 315         u64 free_space_inv = free_space
 316                 ? div64_u64(1ULL << 48, free_space)
 317                 : 1ULL << 48;
 318         u64 scale = *v / 4;
 319
 320         if (*v + free_space_inv >= *v)
 321                 *v += free_space_inv;
 322         else
 323                 *v = U64_MAX;
 324
 325         for (v = stripe->next_alloc;
 326              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 327                 *v = *v < scale ? 0 : *v - scale;
 328 }
 329
 330 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 331 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 332
 333 static void add_new_bucket(struct bch_fs *c,
 334                            struct open_buckets *ptrs,
 335                            struct bch_devs_mask *devs_may_alloc,
 336                            unsigned *nr_effective,
 337                            bool *have_cache,
 338                            unsigned flags,
 339                            struct open_bucket *ob)
 340 {
 341         unsigned durability =
 342                 bch_dev_bkey_exists(c, ob->ptr.dev)->mi.durability;
 343
 344         __clear_bit(ob->ptr.dev, devs_may_alloc->d);
 345         *nr_effective   += (flags & BUCKET_ALLOC_USE_DURABILITY)
 346                 ? durability : 1;
 347         *have_cache     |= !durability;
 348
 349         ob_push(c, ptrs, ob);
 350 }
 351
 352 int bch2_bucket_alloc_set(struct bch_fs *c,
 353                       struct open_buckets *ptrs,
 354                       struct dev_stripe_state *stripe,
 355                       struct bch_devs_mask *devs_may_alloc,
 356                       unsigned nr_replicas,
 357                       unsigned *nr_effective,
 358                       bool *have_cache,
 359                       enum alloc_reserve reserve,
 360                       unsigned flags,
 361                       struct closure *cl)
 362 {
 363         struct dev_alloc_list devs_sorted =
 364                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 365         struct bch_dev *ca;
 366         int ret = -INSUFFICIENT_DEVICES;
 367         unsigned i;
 368
 369         BUG_ON(*nr_effective >= nr_replicas);
 370
 371         for (i = 0; i < devs_sorted.nr; i++) {
 372                 struct open_bucket *ob;
 373
 374                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
 375                 if (!ca)
 376                         continue;
 377
 378                 if (!ca->mi.durability && *have_cache)
 379                         continue;
 380
 381                 ob = bch2_bucket_alloc(c, ca, reserve,
 382                                 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
 383                 if (IS_ERR(ob)) {
 384                         ret = PTR_ERR(ob);
 385
 386                         if (cl)
 387                                 return ret;
 388                         continue;
 389                 }
 390
 391                 add_new_bucket(c, ptrs, devs_may_alloc,
 392                                nr_effective, have_cache, flags, ob);
 393
 394                 bch2_dev_stripe_increment(ca, stripe);
 395
 396                 if (*nr_effective >= nr_replicas)
 397                         return 0;
 398         }
 399
 400         return ret;
 401 }
 402
 403 /* Allocate from stripes: */
 404
 405 /*
 406  * if we can't allocate a new stripe because there are already too many
 407  * partially filled stripes, force allocating from an existing stripe even when
 408  * it's to a device we don't want:
 409  */
 410
 411 static int bucket_alloc_from_stripe(struct bch_fs *c,
 412                          struct open_buckets *ptrs,
 413                          struct write_point *wp,
 414                          struct bch_devs_mask *devs_may_alloc,
 415                          u16 target,
 416                          unsigned erasure_code,
 417                          unsigned nr_replicas,
 418                          unsigned *nr_effective,
 419                          bool *have_cache,
 420                          unsigned flags,
 421                          struct closure *cl)
 422 {
 423         struct dev_alloc_list devs_sorted;
 424         struct ec_stripe_head *h;
 425         struct open_bucket *ob;
 426         struct bch_dev *ca;
 427         unsigned i, ec_idx;
 428
 429         if (!erasure_code)
 430                 return 0;
 431
 432         if (nr_replicas < 2)
 433                 return 0;
 434
 435         if (ec_open_bucket(c, ptrs))
 436                 return 0;
 437
 438         h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1,
 439                                     wp == &c->copygc_write_point,
 440                                     cl);
 441         if (IS_ERR(h))
 442                 return -PTR_ERR(h);
 443         if (!h)
 444                 return 0;
 445
 446         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 447
 448         for (i = 0; i < devs_sorted.nr; i++)
 449                 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 450                         if (!h->s->blocks[ec_idx])
 451                                 continue;
 452
 453                         ob = c->open_buckets + h->s->blocks[ec_idx];
 454                         if (ob->ptr.dev == devs_sorted.devs[i] &&
 455                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 456                                 goto got_bucket;
 457                 }
 458         goto out_put_head;
 459 got_bucket:
 460         ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 461
 462         ob->ec_idx      = ec_idx;
 463         ob->ec          = h->s;
 464
 465         add_new_bucket(c, ptrs, devs_may_alloc,
 466                        nr_effective, have_cache, flags, ob);
 467         atomic_inc(&h->s->pin);
 468 out_put_head:
 469         bch2_ec_stripe_head_put(c, h);
 470         return 0;
 471 }
 472
 473 /* Sector allocator */
 474
 475 static void get_buckets_from_writepoint(struct bch_fs *c,
 476                                         struct open_buckets *ptrs,
 477                                         struct write_point *wp,
 478                                         struct bch_devs_mask *devs_may_alloc,
 479                                         unsigned nr_replicas,
 480                                         unsigned *nr_effective,
 481                                         bool *have_cache,
 482                                         unsigned flags,
 483                                         bool need_ec)
 484 {
 485         struct open_buckets ptrs_skip = { .nr = 0 };
 486         struct open_bucket *ob;
 487         unsigned i;
 488
 489         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 490                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 491
 492                 if (*nr_effective < nr_replicas &&
 493                     test_bit(ob->ptr.dev, devs_may_alloc->d) &&
 494                     (ca->mi.durability ||
 495                      (wp->type == BCH_DATA_user && !*have_cache)) &&
 496                     (ob->ec || !need_ec)) {
 497                         add_new_bucket(c, ptrs, devs_may_alloc,
 498                                        nr_effective, have_cache,
 499                                        flags, ob);
 500                 } else {
 501                         ob_push(c, &ptrs_skip, ob);
 502                 }
 503         }
 504         wp->ptrs = ptrs_skip;
 505 }
 506
 507 static int open_bucket_add_buckets(struct bch_fs *c,
 508                         struct open_buckets *ptrs,
 509                         struct write_point *wp,
 510                         struct bch_devs_list *devs_have,
 511                         u16 target,
 512                         unsigned erasure_code,
 513                         unsigned nr_replicas,
 514                         unsigned *nr_effective,
 515                         bool *have_cache,
 516                         enum alloc_reserve reserve,
 517                         unsigned flags,
 518                         struct closure *_cl)
 519 {
 520         struct bch_devs_mask devs;
 521         struct open_bucket *ob;
 522         struct closure *cl = NULL;
 523         int ret;
 524         unsigned i;
 525
 526         rcu_read_lock();
 527         devs = target_rw_devs(c, wp->type, target);
 528         rcu_read_unlock();
 529
 530         /* Don't allocate from devices we already have pointers to: */
 531         for (i = 0; i < devs_have->nr; i++)
 532                 __clear_bit(devs_have->devs[i], devs.d);
 533
 534         open_bucket_for_each(c, ptrs, ob, i)
 535                 __clear_bit(ob->ptr.dev, devs.d);
 536
 537         if (erasure_code) {
 538                 if (!ec_open_bucket(c, ptrs)) {
 539                         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 540                                                     nr_replicas, nr_effective,
 541                                                     have_cache, flags, true);
 542                         if (*nr_effective >= nr_replicas)
 543                                 return 0;
 544                 }
 545
 546                 if (!ec_open_bucket(c, ptrs)) {
 547                         ret = bucket_alloc_from_stripe(c, ptrs, wp, &devs,
 548                                                  target, erasure_code,
 549                                                  nr_replicas, nr_effective,
 550                                                  have_cache, flags, _cl);
 551                         if (ret == -FREELIST_EMPTY ||
 552                             ret == -OPEN_BUCKETS_EMPTY)
 553                                 return ret;
 554                         if (*nr_effective >= nr_replicas)
 555                                 return 0;
 556                 }
 557         }
 558
 559         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 560                                     nr_replicas, nr_effective,
 561                                     have_cache, flags, false);
 562         if (*nr_effective >= nr_replicas)
 563                 return 0;
 564
 565         percpu_down_read(&c->mark_lock);
 566         rcu_read_lock();
 567
 568 retry_blocking:
 569         /*
 570          * Try nonblocking first, so that if one device is full we'll try from
 571          * other devices:
 572          */
 573         ret = bch2_bucket_alloc_set(c, ptrs, &wp->stripe, &devs,
 574                                 nr_replicas, nr_effective, have_cache,
 575                                 reserve, flags, cl);
 576         if (ret && ret != -INSUFFICIENT_DEVICES && !cl && _cl) {
 577                 cl = _cl;
 578                 goto retry_blocking;
 579         }
 580
 581         rcu_read_unlock();
 582         percpu_up_read(&c->mark_lock);
 583
 584         return ret;
 585 }
 586
 587 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 588                                 struct open_buckets *obs)
 589 {
 590         struct open_buckets ptrs = { .nr = 0 };
 591         struct open_bucket *ob, *ob2;
 592         unsigned i, j;
 593
 594         open_bucket_for_each(c, obs, ob, i) {
 595                 bool drop = !ca || ob->ptr.dev == ca->dev_idx;
 596
 597                 if (!drop && ob->ec) {
 598                         mutex_lock(&ob->ec->lock);
 599                         for (j = 0; j < ob->ec->new_stripe.key.v.nr_blocks; j++) {
 600                                 if (!ob->ec->blocks[j])
 601                                         continue;
 602
 603                                 ob2 = c->open_buckets + ob->ec->blocks[j];
 604                                 drop |= ob2->ptr.dev == ca->dev_idx;
 605                         }
 606                         mutex_unlock(&ob->ec->lock);
 607                 }
 608
 609                 if (drop)
 610                         bch2_open_bucket_put(c, ob);
 611                 else
 612                         ob_push(c, &ptrs, ob);
 613         }
 614
 615         *obs = ptrs;
 616 }
 617
 618 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 619                           struct write_point *wp)
 620 {
 621         mutex_lock(&wp->lock);
 622         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
 623         mutex_unlock(&wp->lock);
 624 }
 625
 626 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 627                                                  unsigned long write_point)
 628 {
 629         unsigned hash =
 630                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
 631
 632         return &c->write_points_hash[hash];
 633 }
 634
 635 static struct write_point *__writepoint_find(struct hlist_head *head,
 636                                              unsigned long write_point)
 637 {
 638         struct write_point *wp;
 639
 640         rcu_read_lock();
 641         hlist_for_each_entry_rcu(wp, head, node)
 642                 if (wp->write_point == write_point)
 643                         goto out;
 644         wp = NULL;
 645 out:
 646         rcu_read_unlock();
 647         return wp;
 648 }
 649
 650 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
 651 {
 652         u64 stranded    = c->write_points_nr * c->bucket_size_max;
 653         u64 free        = bch2_fs_usage_read_short(c).free;
 654
 655         return stranded * factor > free;
 656 }
 657
 658 static bool try_increase_writepoints(struct bch_fs *c)
 659 {
 660         struct write_point *wp;
 661
 662         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
 663             too_many_writepoints(c, 32))
 664                 return false;
 665
 666         wp = c->write_points + c->write_points_nr++;
 667         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
 668         return true;
 669 }
 670
 671 static bool try_decrease_writepoints(struct bch_fs *c,
 672                                      unsigned old_nr)
 673 {
 674         struct write_point *wp;
 675
 676         mutex_lock(&c->write_points_hash_lock);
 677         if (c->write_points_nr < old_nr) {
 678                 mutex_unlock(&c->write_points_hash_lock);
 679                 return true;
 680         }
 681
 682         if (c->write_points_nr == 1 ||
 683             !too_many_writepoints(c, 8)) {
 684                 mutex_unlock(&c->write_points_hash_lock);
 685                 return false;
 686         }
 687
 688         wp = c->write_points + --c->write_points_nr;
 689
 690         hlist_del_rcu(&wp->node);
 691         mutex_unlock(&c->write_points_hash_lock);
 692
 693         bch2_writepoint_stop(c, NULL, wp);
 694         return true;
 695 }
 696
 697 static struct write_point *writepoint_find(struct bch_fs *c,
 698                                            unsigned long write_point)
 699 {
 700         struct write_point *wp, *oldest;
 701         struct hlist_head *head;
 702
 703         if (!(write_point & 1UL)) {
 704                 wp = (struct write_point *) write_point;
 705                 mutex_lock(&wp->lock);
 706                 return wp;
 707         }
 708
 709         head = writepoint_hash(c, write_point);
 710 restart_find:
 711         wp = __writepoint_find(head, write_point);
 712         if (wp) {
 713 lock_wp:
 714                 mutex_lock(&wp->lock);
 715                 if (wp->write_point == write_point)
 716                         goto out;
 717                 mutex_unlock(&wp->lock);
 718                 goto restart_find;
 719         }
 720 restart_find_oldest:
 721         oldest = NULL;
 722         for (wp = c->write_points;
 723              wp < c->write_points + c->write_points_nr; wp++)
 724                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
 725                         oldest = wp;
 726
 727         mutex_lock(&oldest->lock);
 728         mutex_lock(&c->write_points_hash_lock);
 729         if (oldest >= c->write_points + c->write_points_nr ||
 730             try_increase_writepoints(c)) {
 731                 mutex_unlock(&c->write_points_hash_lock);
 732                 mutex_unlock(&oldest->lock);
 733                 goto restart_find_oldest;
 734         }
 735
 736         wp = __writepoint_find(head, write_point);
 737         if (wp && wp != oldest) {
 738                 mutex_unlock(&c->write_points_hash_lock);
 739                 mutex_unlock(&oldest->lock);
 740                 goto lock_wp;
 741         }
 742
 743         wp = oldest;
 744         hlist_del_rcu(&wp->node);
 745         wp->write_point = write_point;
 746         hlist_add_head_rcu(&wp->node, head);
 747         mutex_unlock(&c->write_points_hash_lock);
 748 out:
 749         wp->last_used = sched_clock();
 750         return wp;
 751 }
 752
 753 /*
 754  * Get us an open_bucket we can allocate from, return with it locked:
 755  */
 756 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
 757                                 unsigned target,
 758                                 unsigned erasure_code,
 759                                 struct write_point_specifier write_point,
 760                                 struct bch_devs_list *devs_have,
 761                                 unsigned nr_replicas,
 762                                 unsigned nr_replicas_required,
 763                                 enum alloc_reserve reserve,
 764                                 unsigned flags,
 765                                 struct closure *cl)
 766 {
 767         struct write_point *wp;
 768         struct open_bucket *ob;
 769         struct open_buckets ptrs;
 770         unsigned nr_effective, write_points_nr;
 771         unsigned ob_flags = 0;
 772         bool have_cache;
 773         int ret;
 774         int i;
 775
 776         if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
 777                 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
 778
 779         BUG_ON(!nr_replicas || !nr_replicas_required);
 780 retry:
 781         ptrs.nr         = 0;
 782         nr_effective    = 0;
 783         write_points_nr = c->write_points_nr;
 784         have_cache      = false;
 785
 786         wp = writepoint_find(c, write_point.v);
 787
 788         if (wp->type == BCH_DATA_user)
 789                 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
 790
 791         /* metadata may not allocate on cache devices: */
 792         if (wp->type != BCH_DATA_user)
 793                 have_cache = true;
 794
 795         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
 796                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 797                                               target, erasure_code,
 798                                               nr_replicas, &nr_effective,
 799                                               &have_cache, reserve,
 800                                               ob_flags, cl);
 801         } else {
 802                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 803                                               target, erasure_code,
 804                                               nr_replicas, &nr_effective,
 805                                               &have_cache, reserve,
 806                                               ob_flags, NULL);
 807                 if (!ret)
 808                         goto alloc_done;
 809
 810                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 811                                               0, erasure_code,
 812                                               nr_replicas, &nr_effective,
 813                                               &have_cache, reserve,
 814                                               ob_flags, cl);
 815         }
 816 alloc_done:
 817         BUG_ON(!ret && nr_effective < nr_replicas);
 818
 819         if (erasure_code && !ec_open_bucket(c, &ptrs))
 820                 pr_debug("failed to get ec bucket: ret %u", ret);
 821
 822         if (ret == -INSUFFICIENT_DEVICES &&
 823             nr_effective >= nr_replicas_required)
 824                 ret = 0;
 825
 826         if (ret)
 827                 goto err;
 828
 829         /* Free buckets we didn't use: */
 830         open_bucket_for_each(c, &wp->ptrs, ob, i)
 831                 open_bucket_free_unused(c, wp, ob);
 832
 833         wp->ptrs = ptrs;
 834
 835         wp->sectors_free = UINT_MAX;
 836
 837         open_bucket_for_each(c, &wp->ptrs, ob, i)
 838                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
 839
 840         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
 841
 842         verify_not_stale(c, &wp->ptrs);
 843
 844         return wp;
 845 err:
 846         open_bucket_for_each(c, &wp->ptrs, ob, i)
 847                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
 848                         ob_push(c, &ptrs, ob);
 849                 else
 850                         open_bucket_free_unused(c, wp, ob);
 851         wp->ptrs = ptrs;
 852
 853         mutex_unlock(&wp->lock);
 854
 855         if (ret == -FREELIST_EMPTY &&
 856             try_decrease_writepoints(c, write_points_nr))
 857                 goto retry;
 858
 859         switch (ret) {
 860         case -OPEN_BUCKETS_EMPTY:
 861         case -FREELIST_EMPTY:
 862                 return cl ? ERR_PTR(-EAGAIN) : ERR_PTR(-ENOSPC);
 863         case -INSUFFICIENT_DEVICES:
 864                 return ERR_PTR(-EROFS);
 865         default:
 866                 BUG();
 867         }
 868 }
 869
 870 /*
 871  * Append pointers to the space we just allocated to @k, and mark @sectors space
 872  * as allocated out of @ob
 873  */
 874 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
 875                                     struct bkey_i *k, unsigned sectors)
 876
 877 {
 878         struct open_bucket *ob;
 879         unsigned i;
 880
 881         BUG_ON(sectors > wp->sectors_free);
 882         wp->sectors_free -= sectors;
 883
 884         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 885                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 886                 struct bch_extent_ptr tmp = ob->ptr;
 887
 888                 tmp.cached = !ca->mi.durability &&
 889                         wp->type == BCH_DATA_user;
 890
 891                 tmp.offset += ca->mi.bucket_size - ob->sectors_free;
 892                 bch2_bkey_append_ptr(k, tmp);
 893
 894                 BUG_ON(sectors > ob->sectors_free);
 895                 ob->sectors_free -= sectors;
 896         }
 897 }
 898
 899 /*
 900  * Append pointers to the space we just allocated to @k, and mark @sectors space
 901  * as allocated out of @ob
 902  */
 903 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
 904 {
 905         struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
 906         struct open_bucket *ob;
 907         unsigned i;
 908
 909         open_bucket_for_each(c, &wp->ptrs, ob, i)
 910                 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
 911         wp->ptrs = keep;
 912
 913         mutex_unlock(&wp->lock);
 914
 915         bch2_open_buckets_put(c, &ptrs);
 916 }
 917
 918 static inline void writepoint_init(struct write_point *wp,
 919                                    enum bch_data_type type)
 920 {
 921         mutex_init(&wp->lock);
 922         wp->type = type;
 923 }
 924
 925 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
 926 {
 927         struct open_bucket *ob;
 928         struct write_point *wp;
 929
 930         mutex_init(&c->write_points_hash_lock);
 931         c->write_points_nr = ARRAY_SIZE(c->write_points);
 932
 933         /* open bucket 0 is a sentinal NULL: */
 934         spin_lock_init(&c->open_buckets[0].lock);
 935
 936         for (ob = c->open_buckets + 1;
 937              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
 938                 spin_lock_init(&ob->lock);
 939                 c->open_buckets_nr_free++;
 940
 941                 ob->freelist = c->open_buckets_freelist;
 942                 c->open_buckets_freelist = ob - c->open_buckets;
 943         }
 944
 945         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
 946         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
 947         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
 948
 949         for (wp = c->write_points;
 950              wp < c->write_points + c->write_points_nr; wp++) {
 951                 writepoint_init(wp, BCH_DATA_user);
 952
 953                 wp->last_used   = sched_clock();
 954                 wp->write_point = (unsigned long) wp;
 955                 hlist_add_head_rcu(&wp->node,
 956                                    writepoint_hash(c, wp->write_point));
 957         }
 958 }