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                     !buckets->b[b].mark.owned_by_allocator)
 156                         goto success;
 157         b = -1;
 158 success:
 159         rcu_read_unlock();
 160         return b;
 161 }
 162
 163 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
 164 {
 165         switch (reserve) {
 166         case RESERVE_BTREE:
 167         case RESERVE_BTREE_MOVINGGC:
 168                 return 0;
 169         case RESERVE_MOVINGGC:
 170                 return OPEN_BUCKETS_COUNT / 4;
 171         default:
 172                 return OPEN_BUCKETS_COUNT / 2;
 173         }
 174 }
 175
 176 /**
 177  * bch_bucket_alloc - allocate a single bucket from a specific device
 178  *
 179  * Returns index of bucket on success, 0 on failure
 180  * */
 181 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 182                                       enum alloc_reserve reserve,
 183                                       bool may_alloc_partial,
 184                                       struct closure *cl)
 185 {
 186         struct open_bucket *ob;
 187         long b = 0;
 188
 189         spin_lock(&c->freelist_lock);
 190
 191         if (may_alloc_partial) {
 192                 int i;
 193
 194                 for (i = ca->open_buckets_partial_nr - 1; i >= 0; --i) {
 195                         ob = c->open_buckets + ca->open_buckets_partial[i];
 196
 197                         if (reserve <= ob->alloc_reserve) {
 198                                 array_remove_item(ca->open_buckets_partial,
 199                                                   ca->open_buckets_partial_nr,
 200                                                   i);
 201                                 ob->on_partial_list = false;
 202                                 ob->alloc_reserve = reserve;
 203                                 spin_unlock(&c->freelist_lock);
 204                                 return ob;
 205                         }
 206                 }
 207         }
 208
 209         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
 210                 if (cl)
 211                         closure_wait(&c->open_buckets_wait, cl);
 212
 213                 if (!c->blocked_allocate_open_bucket)
 214                         c->blocked_allocate_open_bucket = local_clock();
 215
 216                 spin_unlock(&c->freelist_lock);
 217                 trace_open_bucket_alloc_fail(ca, reserve);
 218                 return ERR_PTR(-OPEN_BUCKETS_EMPTY);
 219         }
 220
 221         if (likely(fifo_pop(&ca->free[RESERVE_NONE], b)))
 222                 goto out;
 223
 224         switch (reserve) {
 225         case RESERVE_BTREE_MOVINGGC:
 226         case RESERVE_MOVINGGC:
 227                 if (fifo_pop(&ca->free[RESERVE_MOVINGGC], b))
 228                         goto out;
 229                 break;
 230         default:
 231                 break;
 232         }
 233
 234         if (cl)
 235                 closure_wait(&c->freelist_wait, cl);
 236
 237         if (!c->blocked_allocate)
 238                 c->blocked_allocate = local_clock();
 239
 240         spin_unlock(&c->freelist_lock);
 241
 242         trace_bucket_alloc_fail(ca, reserve);
 243         return ERR_PTR(-FREELIST_EMPTY);
 244 out:
 245         verify_not_on_freelist(c, ca, b);
 246
 247         ob = bch2_open_bucket_alloc(c);
 248
 249         spin_lock(&ob->lock);
 250
 251         ob->valid       = true;
 252         ob->sectors_free = ca->mi.bucket_size;
 253         ob->alloc_reserve = reserve;
 254         ob->ptr         = (struct bch_extent_ptr) {
 255                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
 256                 .gen    = bucket(ca, b)->mark.gen,
 257                 .offset = bucket_to_sector(ca, b),
 258                 .dev    = ca->dev_idx,
 259         };
 260
 261         spin_unlock(&ob->lock);
 262
 263         if (c->blocked_allocate_open_bucket) {
 264                 bch2_time_stats_update(
 265                         &c->times[BCH_TIME_blocked_allocate_open_bucket],
 266                         c->blocked_allocate_open_bucket);
 267                 c->blocked_allocate_open_bucket = 0;
 268         }
 269
 270         if (c->blocked_allocate) {
 271                 bch2_time_stats_update(
 272                         &c->times[BCH_TIME_blocked_allocate],
 273                         c->blocked_allocate);
 274                 c->blocked_allocate = 0;
 275         }
 276
 277         ca->nr_open_buckets++;
 278         spin_unlock(&c->freelist_lock);
 279
 280         bch2_wake_allocator(ca);
 281
 282         trace_bucket_alloc(ca, reserve);
 283         return ob;
 284 }
 285
 286 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 287                             unsigned l, unsigned r)
 288 {
 289         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 290                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 291 }
 292
 293 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 294
 295 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 296                                           struct dev_stripe_state *stripe,
 297                                           struct bch_devs_mask *devs)
 298 {
 299         struct dev_alloc_list ret = { .nr = 0 };
 300         unsigned i;
 301
 302         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 303                 ret.devs[ret.nr++] = i;
 304
 305         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 306         return ret;
 307 }
 308
 309 void bch2_dev_stripe_increment(struct bch_dev *ca,
 310                                struct dev_stripe_state *stripe)
 311 {
 312         u64 *v = stripe->next_alloc + ca->dev_idx;
 313         u64 free_space = dev_buckets_available(ca);
 314         u64 free_space_inv = free_space
 315                 ? div64_u64(1ULL << 48, free_space)
 316                 : 1ULL << 48;
 317         u64 scale = *v / 4;
 318
 319         if (*v + free_space_inv >= *v)
 320                 *v += free_space_inv;
 321         else
 322                 *v = U64_MAX;
 323
 324         for (v = stripe->next_alloc;
 325              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 326                 *v = *v < scale ? 0 : *v - scale;
 327 }
 328
 329 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 330 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 331
 332 static void add_new_bucket(struct bch_fs *c,
 333                            struct open_buckets *ptrs,
 334                            struct bch_devs_mask *devs_may_alloc,
 335                            unsigned *nr_effective,
 336                            bool *have_cache,
 337                            unsigned flags,
 338                            struct open_bucket *ob)
 339 {
 340         unsigned durability =
 341                 bch_dev_bkey_exists(c, ob->ptr.dev)->mi.durability;
 342
 343         __clear_bit(ob->ptr.dev, devs_may_alloc->d);
 344         *nr_effective   += (flags & BUCKET_ALLOC_USE_DURABILITY)
 345                 ? durability : 1;
 346         *have_cache     |= !durability;
 347
 348         ob_push(c, ptrs, ob);
 349 }
 350
 351 enum bucket_alloc_ret
 352 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         enum bucket_alloc_ret 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 ALLOC_SUCCESS;
 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 enum bucket_alloc_ret
 412 bucket_alloc_from_stripe(struct bch_fs *c,
 413                          struct open_buckets *ptrs,
 414                          struct write_point *wp,
 415                          struct bch_devs_mask *devs_may_alloc,
 416                          u16 target,
 417                          unsigned erasure_code,
 418                          unsigned nr_replicas,
 419                          unsigned *nr_effective,
 420                          bool *have_cache,
 421                          unsigned flags,
 422                          struct closure *cl)
 423 {
 424         struct dev_alloc_list devs_sorted;
 425         struct ec_stripe_head *h;
 426         struct open_bucket *ob;
 427         struct bch_dev *ca;
 428         unsigned i, ec_idx;
 429
 430         if (!erasure_code)
 431                 return 0;
 432
 433         if (nr_replicas < 2)
 434                 return 0;
 435
 436         if (ec_open_bucket(c, ptrs))
 437                 return 0;
 438
 439         h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1,
 440                                     wp == &c->copygc_write_point,
 441                                     cl);
 442         if (IS_ERR(h))
 443                 return -PTR_ERR(h);
 444         if (!h)
 445                 return 0;
 446
 447         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 448
 449         for (i = 0; i < devs_sorted.nr; i++)
 450                 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 451                         if (!h->s->blocks[ec_idx])
 452                                 continue;
 453
 454                         ob = c->open_buckets + h->s->blocks[ec_idx];
 455                         if (ob->ptr.dev == devs_sorted.devs[i] &&
 456                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 457                                 goto got_bucket;
 458                 }
 459         goto out_put_head;
 460 got_bucket:
 461         ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 462
 463         ob->ec_idx      = ec_idx;
 464         ob->ec          = h->s;
 465
 466         add_new_bucket(c, ptrs, devs_may_alloc,
 467                        nr_effective, have_cache, flags, ob);
 468         atomic_inc(&h->s->pin);
 469 out_put_head:
 470         bch2_ec_stripe_head_put(c, h);
 471         return 0;
 472 }
 473
 474 /* Sector allocator */
 475
 476 static void get_buckets_from_writepoint(struct bch_fs *c,
 477                                         struct open_buckets *ptrs,
 478                                         struct write_point *wp,
 479                                         struct bch_devs_mask *devs_may_alloc,
 480                                         unsigned nr_replicas,
 481                                         unsigned *nr_effective,
 482                                         bool *have_cache,
 483                                         unsigned flags,
 484                                         bool need_ec)
 485 {
 486         struct open_buckets ptrs_skip = { .nr = 0 };
 487         struct open_bucket *ob;
 488         unsigned i;
 489
 490         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 491                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 492
 493                 if (*nr_effective < nr_replicas &&
 494                     test_bit(ob->ptr.dev, devs_may_alloc->d) &&
 495                     (ca->mi.durability ||
 496                      (wp->type == BCH_DATA_user && !*have_cache)) &&
 497                     (ob->ec || !need_ec)) {
 498                         add_new_bucket(c, ptrs, devs_may_alloc,
 499                                        nr_effective, have_cache,
 500                                        flags, ob);
 501                 } else {
 502                         ob_push(c, &ptrs_skip, ob);
 503                 }
 504         }
 505         wp->ptrs = ptrs_skip;
 506 }
 507
 508 static enum bucket_alloc_ret
 509 open_bucket_add_buckets(struct bch_fs *c,
 510                         struct open_buckets *ptrs,
 511                         struct write_point *wp,
 512                         struct bch_devs_list *devs_have,
 513                         u16 target,
 514                         unsigned erasure_code,
 515                         unsigned nr_replicas,
 516                         unsigned *nr_effective,
 517                         bool *have_cache,
 518                         enum alloc_reserve reserve,
 519                         unsigned flags,
 520                         struct closure *_cl)
 521 {
 522         struct bch_devs_mask devs;
 523         struct open_bucket *ob;
 524         struct closure *cl = NULL;
 525         enum bucket_alloc_ret ret;
 526         unsigned i;
 527
 528         rcu_read_lock();
 529         devs = target_rw_devs(c, wp->type, target);
 530         rcu_read_unlock();
 531
 532         /* Don't allocate from devices we already have pointers to: */
 533         for (i = 0; i < devs_have->nr; i++)
 534                 __clear_bit(devs_have->devs[i], devs.d);
 535
 536         open_bucket_for_each(c, ptrs, ob, i)
 537                 __clear_bit(ob->ptr.dev, devs.d);
 538
 539         if (erasure_code) {
 540                 if (!ec_open_bucket(c, ptrs)) {
 541                         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 542                                                     nr_replicas, nr_effective,
 543                                                     have_cache, flags, true);
 544                         if (*nr_effective >= nr_replicas)
 545                                 return 0;
 546                 }
 547
 548                 if (!ec_open_bucket(c, ptrs)) {
 549                         ret = bucket_alloc_from_stripe(c, ptrs, wp, &devs,
 550                                                  target, erasure_code,
 551                                                  nr_replicas, nr_effective,
 552                                                  have_cache, flags, _cl);
 553                         if (ret == FREELIST_EMPTY ||
 554                             ret == OPEN_BUCKETS_EMPTY)
 555                                 return ret;
 556                         if (*nr_effective >= nr_replicas)
 557                                 return 0;
 558                 }
 559         }
 560
 561         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 562                                     nr_replicas, nr_effective,
 563                                     have_cache, flags, false);
 564         if (*nr_effective >= nr_replicas)
 565                 return 0;
 566
 567         percpu_down_read(&c->mark_lock);
 568         rcu_read_lock();
 569
 570 retry_blocking:
 571         /*
 572          * Try nonblocking first, so that if one device is full we'll try from
 573          * other devices:
 574          */
 575         ret = bch2_bucket_alloc_set(c, ptrs, &wp->stripe, &devs,
 576                                 nr_replicas, nr_effective, have_cache,
 577                                 reserve, flags, cl);
 578         if (ret && ret != INSUFFICIENT_DEVICES && !cl && _cl) {
 579                 cl = _cl;
 580                 goto retry_blocking;
 581         }
 582
 583         rcu_read_unlock();
 584         percpu_up_read(&c->mark_lock);
 585
 586         return ret;
 587 }
 588
 589 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 590                                 struct open_buckets *obs)
 591 {
 592         struct open_buckets ptrs = { .nr = 0 };
 593         struct open_bucket *ob, *ob2;
 594         unsigned i, j;
 595
 596         open_bucket_for_each(c, obs, ob, i) {
 597                 bool drop = !ca || ob->ptr.dev == ca->dev_idx;
 598
 599                 if (!drop && ob->ec) {
 600                         mutex_lock(&ob->ec->lock);
 601                         for (j = 0; j < ob->ec->new_stripe.key.v.nr_blocks; j++) {
 602                                 if (!ob->ec->blocks[j])
 603                                         continue;
 604
 605                                 ob2 = c->open_buckets + ob->ec->blocks[j];
 606                                 drop |= ob2->ptr.dev == ca->dev_idx;
 607                         }
 608                         mutex_unlock(&ob->ec->lock);
 609                 }
 610
 611                 if (drop)
 612                         bch2_open_bucket_put(c, ob);
 613                 else
 614                         ob_push(c, &ptrs, ob);
 615         }
 616
 617         *obs = ptrs;
 618 }
 619
 620 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 621                           struct write_point *wp)
 622 {
 623         mutex_lock(&wp->lock);
 624         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
 625         mutex_unlock(&wp->lock);
 626 }
 627
 628 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 629                                                  unsigned long write_point)
 630 {
 631         unsigned hash =
 632                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
 633
 634         return &c->write_points_hash[hash];
 635 }
 636
 637 static struct write_point *__writepoint_find(struct hlist_head *head,
 638                                              unsigned long write_point)
 639 {
 640         struct write_point *wp;
 641
 642         rcu_read_lock();
 643         hlist_for_each_entry_rcu(wp, head, node)
 644                 if (wp->write_point == write_point)
 645                         goto out;
 646         wp = NULL;
 647 out:
 648         rcu_read_unlock();
 649         return wp;
 650 }
 651
 652 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
 653 {
 654         u64 stranded    = c->write_points_nr * c->bucket_size_max;
 655         u64 free        = bch2_fs_usage_read_short(c).free;
 656
 657         return stranded * factor > free;
 658 }
 659
 660 static bool try_increase_writepoints(struct bch_fs *c)
 661 {
 662         struct write_point *wp;
 663
 664         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
 665             too_many_writepoints(c, 32))
 666                 return false;
 667
 668         wp = c->write_points + c->write_points_nr++;
 669         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
 670         return true;
 671 }
 672
 673 static bool try_decrease_writepoints(struct bch_fs *c,
 674                                      unsigned old_nr)
 675 {
 676         struct write_point *wp;
 677
 678         mutex_lock(&c->write_points_hash_lock);
 679         if (c->write_points_nr < old_nr) {
 680                 mutex_unlock(&c->write_points_hash_lock);
 681                 return true;
 682         }
 683
 684         if (c->write_points_nr == 1 ||
 685             !too_many_writepoints(c, 8)) {
 686                 mutex_unlock(&c->write_points_hash_lock);
 687                 return false;
 688         }
 689
 690         wp = c->write_points + --c->write_points_nr;
 691
 692         hlist_del_rcu(&wp->node);
 693         mutex_unlock(&c->write_points_hash_lock);
 694
 695         bch2_writepoint_stop(c, NULL, wp);
 696         return true;
 697 }
 698
 699 static struct write_point *writepoint_find(struct bch_fs *c,
 700                                            unsigned long write_point)
 701 {
 702         struct write_point *wp, *oldest;
 703         struct hlist_head *head;
 704
 705         if (!(write_point & 1UL)) {
 706                 wp = (struct write_point *) write_point;
 707                 mutex_lock(&wp->lock);
 708                 return wp;
 709         }
 710
 711         head = writepoint_hash(c, write_point);
 712 restart_find:
 713         wp = __writepoint_find(head, write_point);
 714         if (wp) {
 715 lock_wp:
 716                 mutex_lock(&wp->lock);
 717                 if (wp->write_point == write_point)
 718                         goto out;
 719                 mutex_unlock(&wp->lock);
 720                 goto restart_find;
 721         }
 722 restart_find_oldest:
 723         oldest = NULL;
 724         for (wp = c->write_points;
 725              wp < c->write_points + c->write_points_nr; wp++)
 726                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
 727                         oldest = wp;
 728
 729         mutex_lock(&oldest->lock);
 730         mutex_lock(&c->write_points_hash_lock);
 731         if (oldest >= c->write_points + c->write_points_nr ||
 732             try_increase_writepoints(c)) {
 733                 mutex_unlock(&c->write_points_hash_lock);
 734                 mutex_unlock(&oldest->lock);
 735                 goto restart_find_oldest;
 736         }
 737
 738         wp = __writepoint_find(head, write_point);
 739         if (wp && wp != oldest) {
 740                 mutex_unlock(&c->write_points_hash_lock);
 741                 mutex_unlock(&oldest->lock);
 742                 goto lock_wp;
 743         }
 744
 745         wp = oldest;
 746         hlist_del_rcu(&wp->node);
 747         wp->write_point = write_point;
 748         hlist_add_head_rcu(&wp->node, head);
 749         mutex_unlock(&c->write_points_hash_lock);
 750 out:
 751         wp->last_used = sched_clock();
 752         return wp;
 753 }
 754
 755 /*
 756  * Get us an open_bucket we can allocate from, return with it locked:
 757  */
 758 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
 759                                 unsigned target,
 760                                 unsigned erasure_code,
 761                                 struct write_point_specifier write_point,
 762                                 struct bch_devs_list *devs_have,
 763                                 unsigned nr_replicas,
 764                                 unsigned nr_replicas_required,
 765                                 enum alloc_reserve reserve,
 766                                 unsigned flags,
 767                                 struct closure *cl)
 768 {
 769         struct write_point *wp;
 770         struct open_bucket *ob;
 771         struct open_buckets ptrs;
 772         unsigned nr_effective, write_points_nr;
 773         unsigned ob_flags = 0;
 774         bool have_cache;
 775         enum bucket_alloc_ret ret;
 776         int i;
 777
 778         if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
 779                 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
 780
 781         BUG_ON(!nr_replicas || !nr_replicas_required);
 782 retry:
 783         ptrs.nr         = 0;
 784         nr_effective    = 0;
 785         write_points_nr = c->write_points_nr;
 786         have_cache      = false;
 787
 788         wp = writepoint_find(c, write_point.v);
 789
 790         if (wp->type == BCH_DATA_user)
 791                 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
 792
 793         /* metadata may not allocate on cache devices: */
 794         if (wp->type != BCH_DATA_user)
 795                 have_cache = true;
 796
 797         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
 798                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 799                                               target, erasure_code,
 800                                               nr_replicas, &nr_effective,
 801                                               &have_cache, reserve,
 802                                               ob_flags, cl);
 803         } else {
 804                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 805                                               target, erasure_code,
 806                                               nr_replicas, &nr_effective,
 807                                               &have_cache, reserve,
 808                                               ob_flags, NULL);
 809                 if (!ret)
 810                         goto alloc_done;
 811
 812                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 813                                               0, erasure_code,
 814                                               nr_replicas, &nr_effective,
 815                                               &have_cache, reserve,
 816                                               ob_flags, cl);
 817         }
 818 alloc_done:
 819         BUG_ON(!ret && nr_effective < nr_replicas);
 820
 821         if (erasure_code && !ec_open_bucket(c, &ptrs))
 822                 pr_debug("failed to get ec bucket: ret %u", ret);
 823
 824         if (ret == INSUFFICIENT_DEVICES &&
 825             nr_effective >= nr_replicas_required)
 826                 ret = 0;
 827
 828         if (ret)
 829                 goto err;
 830
 831         /* Free buckets we didn't use: */
 832         open_bucket_for_each(c, &wp->ptrs, ob, i)
 833                 open_bucket_free_unused(c, wp, ob);
 834
 835         wp->ptrs = ptrs;
 836
 837         wp->sectors_free = UINT_MAX;
 838
 839         open_bucket_for_each(c, &wp->ptrs, ob, i)
 840                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
 841
 842         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
 843
 844         verify_not_stale(c, &wp->ptrs);
 845
 846         return wp;
 847 err:
 848         open_bucket_for_each(c, &wp->ptrs, ob, i)
 849                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
 850                         ob_push(c, &ptrs, ob);
 851                 else
 852                         open_bucket_free_unused(c, wp, ob);
 853         wp->ptrs = ptrs;
 854
 855         mutex_unlock(&wp->lock);
 856
 857         if (ret == FREELIST_EMPTY &&
 858             try_decrease_writepoints(c, write_points_nr))
 859                 goto retry;
 860
 861         switch (ret) {
 862         case OPEN_BUCKETS_EMPTY:
 863         case FREELIST_EMPTY:
 864                 return cl ? ERR_PTR(-EAGAIN) : ERR_PTR(-ENOSPC);
 865         case INSUFFICIENT_DEVICES:
 866                 return ERR_PTR(-EROFS);
 867         default:
 868                 BUG();
 869         }
 870 }
 871
 872 /*
 873  * Append pointers to the space we just allocated to @k, and mark @sectors space
 874  * as allocated out of @ob
 875  */
 876 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
 877                                     struct bkey_i *k, unsigned sectors)
 878
 879 {
 880         struct open_bucket *ob;
 881         unsigned i;
 882
 883         BUG_ON(sectors > wp->sectors_free);
 884         wp->sectors_free -= sectors;
 885
 886         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 887                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 888                 struct bch_extent_ptr tmp = ob->ptr;
 889
 890                 tmp.cached = !ca->mi.durability &&
 891                         wp->type == BCH_DATA_user;
 892
 893                 tmp.offset += ca->mi.bucket_size - ob->sectors_free;
 894                 bch2_bkey_append_ptr(k, tmp);
 895
 896                 BUG_ON(sectors > ob->sectors_free);
 897                 ob->sectors_free -= sectors;
 898         }
 899 }
 900
 901 /*
 902  * Append pointers to the space we just allocated to @k, and mark @sectors space
 903  * as allocated out of @ob
 904  */
 905 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
 906 {
 907         struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
 908         struct open_bucket *ob;
 909         unsigned i;
 910
 911         open_bucket_for_each(c, &wp->ptrs, ob, i)
 912                 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
 913         wp->ptrs = keep;
 914
 915         mutex_unlock(&wp->lock);
 916
 917         bch2_open_buckets_put(c, &ptrs);
 918 }
 919
 920 static inline void writepoint_init(struct write_point *wp,
 921                                    enum bch_data_type type)
 922 {
 923         mutex_init(&wp->lock);
 924         wp->type = type;
 925 }
 926
 927 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
 928 {
 929         struct open_bucket *ob;
 930         struct write_point *wp;
 931
 932         mutex_init(&c->write_points_hash_lock);
 933         c->write_points_nr = ARRAY_SIZE(c->write_points);
 934
 935         /* open bucket 0 is a sentinal NULL: */
 936         spin_lock_init(&c->open_buckets[0].lock);
 937
 938         for (ob = c->open_buckets + 1;
 939              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
 940                 spin_lock_init(&ob->lock);
 941                 c->open_buckets_nr_free++;
 942
 943                 ob->freelist = c->open_buckets_freelist;
 944                 c->open_buckets_freelist = ob - c->open_buckets;
 945         }
 946
 947         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
 948         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
 949         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
 950
 951         for (wp = c->write_points;
 952              wp < c->write_points + c->write_points_nr; wp++) {
 953                 writepoint_init(wp, BCH_DATA_user);
 954
 955                 wp->last_used   = sched_clock();
 956                 wp->write_point = (unsigned long) wp;
 957                 hlist_add_head_rcu(&wp->node,
 958                                    writepoint_hash(c, wp->write_point));
 959         }
 960 }