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

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Primary bucket allocation code
   4  *
   5  * Copyright 2012 Google, Inc.
   6  *
   7  * Allocation in bcache is done in terms of buckets:
   8  *
   9  * Each bucket has associated an 8 bit gen; this gen corresponds to the gen in
  10  * btree pointers - they must match for the pointer to be considered valid.
  11  *
  12  * Thus (assuming a bucket has no dirty data or metadata in it) we can reuse a
  13  * bucket simply by incrementing its gen.
  14  *
  15  * The gens (along with the priorities; it's really the gens are important but
  16  * the code is named as if it's the priorities) are written in an arbitrary list
  17  * of buckets on disk, with a pointer to them in the journal header.
  18  *
  19  * When we invalidate a bucket, we have to write its new gen to disk and wait
  20  * for that write to complete before we use it - otherwise after a crash we
  21  * could have pointers that appeared to be good but pointed to data that had
  22  * been overwritten.
  23  *
  24  * Since the gens and priorities are all stored contiguously on disk, we can
  25  * batch this up: We fill up the free_inc list with freshly invalidated buckets,
  26  * call prio_write(), and when prio_write() finishes we pull buckets off the
  27  * free_inc list and optionally discard them.
  28  *
  29  * free_inc isn't the only freelist - if it was, we'd often have to sleep while
  30  * priorities and gens were being written before we could allocate. c->free is a
  31  * smaller freelist, and buckets on that list are always ready to be used.
  32  *
  33  * If we've got discards enabled, that happens when a bucket moves from the
  34  * free_inc list to the free list.
  35  *
  36  * It's important to ensure that gens don't wrap around - with respect to
  37  * either the oldest gen in the btree or the gen on disk. This is quite
  38  * difficult to do in practice, but we explicitly guard against it anyways - if
  39  * a bucket is in danger of wrapping around we simply skip invalidating it that
  40  * time around, and we garbage collect or rewrite the priorities sooner than we
  41  * would have otherwise.
  42  *
  43  * bch2_bucket_alloc() allocates a single bucket from a specific device.
  44  *
  45  * bch2_bucket_alloc_set() allocates one or more buckets from different devices
  46  * in a given filesystem.
  47  *
  48  * invalidate_buckets() drives all the processes described above. It's called
  49  * from bch2_bucket_alloc() and a few other places that need to make sure free
  50  * buckets are ready.
  51  *
  52  * invalidate_buckets_(lru|fifo)() find buckets that are available to be
  53  * invalidated, and then invalidate them and stick them on the free_inc list -
  54  * in either lru or fifo order.
  55  */
  56
  57 #include "bcachefs.h"
  58 #include "alloc_background.h"
  59 #include "alloc_foreground.h"
  60 #include "btree_gc.h"
  61 #include "buckets.h"
  62 #include "clock.h"
  63 #include "debug.h"
  64 #include "disk_groups.h"
  65 #include "ec.h"
  66 #include "io.h"
  67
  68 #include <linux/math64.h>
  69 #include <linux/rculist.h>
  70 #include <linux/rcupdate.h>
  71 #include <trace/events/bcachefs.h>
  72
  73 /*
  74  * Open buckets represent a bucket that's currently being allocated from.  They
  75  * serve two purposes:
  76  *
  77  *  - They track buckets that have been partially allocated, allowing for
  78  *    sub-bucket sized allocations - they're used by the sector allocator below
  79  *
  80  *  - They provide a reference to the buckets they own that mark and sweep GC
  81  *    can find, until the new allocation has a pointer to it inserted into the
  82  *    btree
  83  *
  84  * When allocating some space with the sector allocator, the allocation comes
  85  * with a reference to an open bucket - the caller is required to put that
  86  * reference _after_ doing the index update that makes its allocation reachable.
  87  */
  88
  89 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
  90 {
  91         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
  92
  93         if (ob->ec) {
  94                 bch2_ec_bucket_written(c, ob);
  95                 return;
  96         }
  97
  98         percpu_down_read(&c->mark_lock);
  99         spin_lock(&ob->lock);
 100
 101         bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr),
 102                                false, gc_pos_alloc(c, ob), 0);
 103         ob->valid = false;
 104         ob->type = 0;
 105
 106         spin_unlock(&ob->lock);
 107         percpu_up_read(&c->mark_lock);
 108
 109         spin_lock(&c->freelist_lock);
 110         ob->freelist = c->open_buckets_freelist;
 111         c->open_buckets_freelist = ob - c->open_buckets;
 112         c->open_buckets_nr_free++;
 113         spin_unlock(&c->freelist_lock);
 114
 115         closure_wake_up(&c->open_buckets_wait);
 116 }
 117
 118 void bch2_open_bucket_write_error(struct bch_fs *c,
 119                                   struct open_buckets *obs,
 120                                   unsigned dev)
 121 {
 122         struct open_bucket *ob;
 123         unsigned i;
 124
 125         open_bucket_for_each(c, obs, ob, i)
 126                 if (ob->ptr.dev == dev &&
 127                     ob->ec)
 128                         bch2_ec_bucket_cancel(c, ob);
 129 }
 130
 131 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
 132 {
 133         struct open_bucket *ob;
 134
 135         BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
 136
 137         ob = c->open_buckets + c->open_buckets_freelist;
 138         c->open_buckets_freelist = ob->freelist;
 139         atomic_set(&ob->pin, 1);
 140         ob->type = 0;
 141
 142         c->open_buckets_nr_free--;
 143         return ob;
 144 }
 145
 146 static void open_bucket_free_unused(struct bch_fs *c,
 147                                     struct write_point *wp,
 148                                     struct open_bucket *ob)
 149 {
 150         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 151         bool may_realloc = wp->type == BCH_DATA_user;
 152
 153         BUG_ON(ca->open_buckets_partial_nr >
 154                ARRAY_SIZE(ca->open_buckets_partial));
 155
 156         if (ca->open_buckets_partial_nr <
 157             ARRAY_SIZE(ca->open_buckets_partial) &&
 158             may_realloc) {
 159                 spin_lock(&c->freelist_lock);
 160                 ob->on_partial_list = true;
 161                 ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
 162                         ob - c->open_buckets;
 163                 spin_unlock(&c->freelist_lock);
 164
 165                 closure_wake_up(&c->open_buckets_wait);
 166                 closure_wake_up(&c->freelist_wait);
 167         } else {
 168                 bch2_open_bucket_put(c, ob);
 169         }
 170 }
 171
 172 static void verify_not_stale(struct bch_fs *c, const struct open_buckets *obs)
 173 {
 174 #ifdef CONFIG_BCACHEFS_DEBUG
 175         struct open_bucket *ob;
 176         unsigned i;
 177
 178         open_bucket_for_each(c, obs, ob, i) {
 179                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 180
 181                 BUG_ON(ptr_stale(ca, &ob->ptr));
 182         }
 183 #endif
 184 }
 185
 186 /* _only_ for allocating the journal on a new device: */
 187 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
 188 {
 189         struct bucket_array *buckets;
 190         ssize_t b;
 191
 192         rcu_read_lock();
 193         buckets = bucket_array(ca);
 194
 195         for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++)
 196                 if (is_available_bucket(buckets->b[b].mark))
 197                         goto success;
 198         b = -1;
 199 success:
 200         rcu_read_unlock();
 201         return b;
 202 }
 203
 204 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
 205 {
 206         switch (reserve) {
 207         case RESERVE_ALLOC:
 208                 return 0;
 209         case RESERVE_BTREE:
 210                 return OPEN_BUCKETS_COUNT / 4;
 211         default:
 212                 return OPEN_BUCKETS_COUNT / 2;
 213         }
 214 }
 215
 216 /**
 217  * bch_bucket_alloc - allocate a single bucket from a specific device
 218  *
 219  * Returns index of bucket on success, 0 on failure
 220  * */
 221 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 222                                       enum alloc_reserve reserve,
 223                                       bool may_alloc_partial,
 224                                       struct closure *cl)
 225 {
 226         struct bucket_array *buckets;
 227         struct open_bucket *ob;
 228         long bucket = 0;
 229
 230         spin_lock(&c->freelist_lock);
 231
 232         if (may_alloc_partial) {
 233                 int i;
 234
 235                 for (i = ca->open_buckets_partial_nr - 1; i >= 0; --i) {
 236                         ob = c->open_buckets + ca->open_buckets_partial[i];
 237
 238                         if (reserve <= ob->alloc_reserve) {
 239                                 array_remove_item(ca->open_buckets_partial,
 240                                                   ca->open_buckets_partial_nr,
 241                                                   i);
 242                                 ob->on_partial_list = false;
 243                                 ob->alloc_reserve = reserve;
 244                                 spin_unlock(&c->freelist_lock);
 245                                 return ob;
 246                         }
 247                 }
 248         }
 249
 250         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
 251                 if (cl)
 252                         closure_wait(&c->open_buckets_wait, cl);
 253
 254                 if (!c->blocked_allocate_open_bucket)
 255                         c->blocked_allocate_open_bucket = local_clock();
 256
 257                 spin_unlock(&c->freelist_lock);
 258                 trace_open_bucket_alloc_fail(ca, reserve);
 259                 return ERR_PTR(-OPEN_BUCKETS_EMPTY);
 260         }
 261
 262         if (likely(fifo_pop(&ca->free[RESERVE_NONE], bucket)))
 263                 goto out;
 264
 265         switch (reserve) {
 266         case RESERVE_ALLOC:
 267                 if (fifo_pop(&ca->free[RESERVE_BTREE], bucket))
 268                         goto out;
 269                 break;
 270         case RESERVE_BTREE:
 271                 if (fifo_used(&ca->free[RESERVE_BTREE]) * 2 >=
 272                     ca->free[RESERVE_BTREE].size &&
 273                     fifo_pop(&ca->free[RESERVE_BTREE], bucket))
 274                         goto out;
 275                 break;
 276         case RESERVE_MOVINGGC:
 277                 if (fifo_pop(&ca->free[RESERVE_MOVINGGC], bucket))
 278                         goto out;
 279                 break;
 280         default:
 281                 break;
 282         }
 283
 284         if (cl)
 285                 closure_wait(&c->freelist_wait, cl);
 286
 287         if (!c->blocked_allocate)
 288                 c->blocked_allocate = local_clock();
 289
 290         spin_unlock(&c->freelist_lock);
 291
 292         trace_bucket_alloc_fail(ca, reserve);
 293         return ERR_PTR(-FREELIST_EMPTY);
 294 out:
 295         verify_not_on_freelist(c, ca, bucket);
 296
 297         ob = bch2_open_bucket_alloc(c);
 298
 299         spin_lock(&ob->lock);
 300         buckets = bucket_array(ca);
 301
 302         ob->valid       = true;
 303         ob->sectors_free = ca->mi.bucket_size;
 304         ob->alloc_reserve = reserve;
 305         ob->ptr         = (struct bch_extent_ptr) {
 306                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
 307                 .gen    = buckets->b[bucket].mark.gen,
 308                 .offset = bucket_to_sector(ca, bucket),
 309                 .dev    = ca->dev_idx,
 310         };
 311
 312         spin_unlock(&ob->lock);
 313
 314         if (c->blocked_allocate_open_bucket) {
 315                 bch2_time_stats_update(
 316                         &c->times[BCH_TIME_blocked_allocate_open_bucket],
 317                         c->blocked_allocate_open_bucket);
 318                 c->blocked_allocate_open_bucket = 0;
 319         }
 320
 321         if (c->blocked_allocate) {
 322                 bch2_time_stats_update(
 323                         &c->times[BCH_TIME_blocked_allocate],
 324                         c->blocked_allocate);
 325                 c->blocked_allocate = 0;
 326         }
 327
 328         spin_unlock(&c->freelist_lock);
 329
 330         bch2_wake_allocator(ca);
 331
 332         trace_bucket_alloc(ca, reserve);
 333         return ob;
 334 }
 335
 336 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 337                             unsigned l, unsigned r)
 338 {
 339         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 340                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 341 }
 342
 343 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 344
 345 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 346                                           struct dev_stripe_state *stripe,
 347                                           struct bch_devs_mask *devs)
 348 {
 349         struct dev_alloc_list ret = { .nr = 0 };
 350         unsigned i;
 351
 352         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 353                 ret.devs[ret.nr++] = i;
 354
 355         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 356         return ret;
 357 }
 358
 359 void bch2_dev_stripe_increment(struct bch_dev *ca,
 360                                struct dev_stripe_state *stripe)
 361 {
 362         u64 *v = stripe->next_alloc + ca->dev_idx;
 363         u64 free_space = dev_buckets_free(ca);
 364         u64 free_space_inv = free_space
 365                 ? div64_u64(1ULL << 48, free_space)
 366                 : 1ULL << 48;
 367         u64 scale = *v / 4;
 368
 369         if (*v + free_space_inv >= *v)
 370                 *v += free_space_inv;
 371         else
 372                 *v = U64_MAX;
 373
 374         for (v = stripe->next_alloc;
 375              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 376                 *v = *v < scale ? 0 : *v - scale;
 377 }
 378
 379 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 380 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 381
 382 static void add_new_bucket(struct bch_fs *c,
 383                            struct open_buckets *ptrs,
 384                            struct bch_devs_mask *devs_may_alloc,
 385                            unsigned *nr_effective,
 386                            bool *have_cache,
 387                            unsigned flags,
 388                            struct open_bucket *ob)
 389 {
 390         unsigned durability =
 391                 bch_dev_bkey_exists(c, ob->ptr.dev)->mi.durability;
 392
 393         __clear_bit(ob->ptr.dev, devs_may_alloc->d);
 394         *nr_effective   += (flags & BUCKET_ALLOC_USE_DURABILITY)
 395                 ? durability : 1;
 396         *have_cache     |= !durability;
 397
 398         ob_push(c, ptrs, ob);
 399 }
 400
 401 enum bucket_alloc_ret
 402 bch2_bucket_alloc_set(struct bch_fs *c,
 403                       struct open_buckets *ptrs,
 404                       struct dev_stripe_state *stripe,
 405                       struct bch_devs_mask *devs_may_alloc,
 406                       unsigned nr_replicas,
 407                       unsigned *nr_effective,
 408                       bool *have_cache,
 409                       enum alloc_reserve reserve,
 410                       unsigned flags,
 411                       struct closure *cl)
 412 {
 413         struct dev_alloc_list devs_sorted =
 414                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 415         struct bch_dev *ca;
 416         enum bucket_alloc_ret ret = INSUFFICIENT_DEVICES;
 417         unsigned i;
 418
 419         BUG_ON(*nr_effective >= nr_replicas);
 420
 421         for (i = 0; i < devs_sorted.nr; i++) {
 422                 struct open_bucket *ob;
 423
 424                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
 425                 if (!ca)
 426                         continue;
 427
 428                 if (!ca->mi.durability && *have_cache)
 429                         continue;
 430
 431                 ob = bch2_bucket_alloc(c, ca, reserve,
 432                                 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
 433                 if (IS_ERR(ob)) {
 434                         ret = -PTR_ERR(ob);
 435
 436                         if (cl)
 437                                 return ret;
 438                         continue;
 439                 }
 440
 441                 add_new_bucket(c, ptrs, devs_may_alloc,
 442                                nr_effective, have_cache, flags, ob);
 443
 444                 bch2_dev_stripe_increment(ca, stripe);
 445
 446                 if (*nr_effective >= nr_replicas)
 447                         return ALLOC_SUCCESS;
 448         }
 449
 450         return ret;
 451 }
 452
 453 /* Allocate from stripes: */
 454
 455 /*
 456  * if we can't allocate a new stripe because there are already too many
 457  * partially filled stripes, force allocating from an existing stripe even when
 458  * it's to a device we don't want:
 459  */
 460
 461 static void bucket_alloc_from_stripe(struct bch_fs *c,
 462                                      struct open_buckets *ptrs,
 463                                      struct write_point *wp,
 464                                      struct bch_devs_mask *devs_may_alloc,
 465                                      u16 target,
 466                                      unsigned erasure_code,
 467                                      unsigned nr_replicas,
 468                                      unsigned *nr_effective,
 469                                      bool *have_cache,
 470                                      unsigned flags)
 471 {
 472         struct dev_alloc_list devs_sorted;
 473         struct ec_stripe_head *h;
 474         struct open_bucket *ob;
 475         struct bch_dev *ca;
 476         unsigned i, ec_idx;
 477
 478         if (!erasure_code)
 479                 return;
 480
 481         if (nr_replicas < 2)
 482                 return;
 483
 484         if (ec_open_bucket(c, ptrs))
 485                 return;
 486
 487         h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1);
 488         if (!h)
 489                 return;
 490
 491         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 492
 493         for (i = 0; i < devs_sorted.nr; i++)
 494                 open_bucket_for_each(c, &h->s->blocks, ob, ec_idx)
 495                         if (ob->ptr.dev == devs_sorted.devs[i] &&
 496                             !test_and_set_bit(h->s->data_block_idx[ec_idx],
 497                                               h->s->blocks_allocated))
 498                                 goto got_bucket;
 499         goto out_put_head;
 500 got_bucket:
 501         ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 502
 503         ob->ec_idx      = h->s->data_block_idx[ec_idx];
 504         ob->ec          = h->s;
 505
 506         add_new_bucket(c, ptrs, devs_may_alloc,
 507                        nr_effective, have_cache, flags, ob);
 508         atomic_inc(&h->s->pin);
 509 out_put_head:
 510         bch2_ec_stripe_head_put(c, h);
 511 }
 512
 513 /* Sector allocator */
 514
 515 static void get_buckets_from_writepoint(struct bch_fs *c,
 516                                         struct open_buckets *ptrs,
 517                                         struct write_point *wp,
 518                                         struct bch_devs_mask *devs_may_alloc,
 519                                         unsigned nr_replicas,
 520                                         unsigned *nr_effective,
 521                                         bool *have_cache,
 522                                         unsigned flags,
 523                                         bool need_ec)
 524 {
 525         struct open_buckets ptrs_skip = { .nr = 0 };
 526         struct open_bucket *ob;
 527         unsigned i;
 528
 529         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 530                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 531
 532                 if (*nr_effective < nr_replicas &&
 533                     test_bit(ob->ptr.dev, devs_may_alloc->d) &&
 534                     (ca->mi.durability ||
 535                      (wp->type == BCH_DATA_user && !*have_cache)) &&
 536                     (ob->ec || !need_ec)) {
 537                         add_new_bucket(c, ptrs, devs_may_alloc,
 538                                        nr_effective, have_cache,
 539                                        flags, ob);
 540                 } else {
 541                         ob_push(c, &ptrs_skip, ob);
 542                 }
 543         }
 544         wp->ptrs = ptrs_skip;
 545 }
 546
 547 static enum bucket_alloc_ret
 548 open_bucket_add_buckets(struct bch_fs *c,
 549                         struct open_buckets *ptrs,
 550                         struct write_point *wp,
 551                         struct bch_devs_list *devs_have,
 552                         u16 target,
 553                         unsigned erasure_code,
 554                         unsigned nr_replicas,
 555                         unsigned *nr_effective,
 556                         bool *have_cache,
 557                         enum alloc_reserve reserve,
 558                         unsigned flags,
 559                         struct closure *_cl)
 560 {
 561         struct bch_devs_mask devs;
 562         struct open_bucket *ob;
 563         struct closure *cl = NULL;
 564         enum bucket_alloc_ret ret;
 565         unsigned i;
 566
 567         rcu_read_lock();
 568         devs = target_rw_devs(c, wp->type, target);
 569         rcu_read_unlock();
 570
 571         /* Don't allocate from devices we already have pointers to: */
 572         for (i = 0; i < devs_have->nr; i++)
 573                 __clear_bit(devs_have->devs[i], devs.d);
 574
 575         open_bucket_for_each(c, ptrs, ob, i)
 576                 __clear_bit(ob->ptr.dev, devs.d);
 577
 578         if (erasure_code) {
 579                 if (!ec_open_bucket(c, ptrs)) {
 580                         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 581                                                     nr_replicas, nr_effective,
 582                                                     have_cache, flags, true);
 583                         if (*nr_effective >= nr_replicas)
 584                                 return 0;
 585                 }
 586
 587                 if (!ec_open_bucket(c, ptrs)) {
 588                         bucket_alloc_from_stripe(c, ptrs, wp, &devs,
 589                                                  target, erasure_code,
 590                                                  nr_replicas, nr_effective,
 591                                                  have_cache, flags);
 592                         if (*nr_effective >= nr_replicas)
 593                                 return 0;
 594                 }
 595         }
 596
 597         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 598                                     nr_replicas, nr_effective,
 599                                     have_cache, flags, false);
 600         if (*nr_effective >= nr_replicas)
 601                 return 0;
 602
 603         percpu_down_read(&c->mark_lock);
 604         rcu_read_lock();
 605
 606 retry_blocking:
 607         /*
 608          * Try nonblocking first, so that if one device is full we'll try from
 609          * other devices:
 610          */
 611         ret = bch2_bucket_alloc_set(c, ptrs, &wp->stripe, &devs,
 612                                 nr_replicas, nr_effective, have_cache,
 613                                 reserve, flags, cl);
 614         if (ret && ret != INSUFFICIENT_DEVICES && !cl && _cl) {
 615                 cl = _cl;
 616                 goto retry_blocking;
 617         }
 618
 619         rcu_read_unlock();
 620         percpu_up_read(&c->mark_lock);
 621
 622         return ret;
 623 }
 624
 625 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 626                                 struct open_buckets *obs)
 627 {
 628         struct open_buckets ptrs = { .nr = 0 };
 629         struct open_bucket *ob, *ob2;
 630         unsigned i, j;
 631
 632         open_bucket_for_each(c, obs, ob, i) {
 633                 bool drop = !ca || ob->ptr.dev == ca->dev_idx;
 634
 635                 if (!drop && ob->ec) {
 636                         mutex_lock(&ob->ec->lock);
 637                         open_bucket_for_each(c, &ob->ec->blocks, ob2, j)
 638                                 drop |= ob2->ptr.dev == ca->dev_idx;
 639                         open_bucket_for_each(c, &ob->ec->parity, ob2, j)
 640                                 drop |= ob2->ptr.dev == ca->dev_idx;
 641                         mutex_unlock(&ob->ec->lock);
 642                 }
 643
 644                 if (drop)
 645                         bch2_open_bucket_put(c, ob);
 646                 else
 647                         ob_push(c, &ptrs, ob);
 648         }
 649
 650         *obs = ptrs;
 651 }
 652
 653 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 654                           struct write_point *wp)
 655 {
 656         mutex_lock(&wp->lock);
 657         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
 658         mutex_unlock(&wp->lock);
 659 }
 660
 661 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 662                                                  unsigned long write_point)
 663 {
 664         unsigned hash =
 665                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
 666
 667         return &c->write_points_hash[hash];
 668 }
 669
 670 static struct write_point *__writepoint_find(struct hlist_head *head,
 671                                              unsigned long write_point)
 672 {
 673         struct write_point *wp;
 674
 675         hlist_for_each_entry_rcu(wp, head, node)
 676                 if (wp->write_point == write_point)
 677                         return wp;
 678
 679         return NULL;
 680 }
 681
 682 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
 683 {
 684         u64 stranded    = c->write_points_nr * c->bucket_size_max;
 685         u64 free        = bch2_fs_usage_read_short(c).free;
 686
 687         return stranded * factor > free;
 688 }
 689
 690 static bool try_increase_writepoints(struct bch_fs *c)
 691 {
 692         struct write_point *wp;
 693
 694         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
 695             too_many_writepoints(c, 32))
 696                 return false;
 697
 698         wp = c->write_points + c->write_points_nr++;
 699         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
 700         return true;
 701 }
 702
 703 static bool try_decrease_writepoints(struct bch_fs *c,
 704                                      unsigned old_nr)
 705 {
 706         struct write_point *wp;
 707
 708         mutex_lock(&c->write_points_hash_lock);
 709         if (c->write_points_nr < old_nr) {
 710                 mutex_unlock(&c->write_points_hash_lock);
 711                 return true;
 712         }
 713
 714         if (c->write_points_nr == 1 ||
 715             !too_many_writepoints(c, 8)) {
 716                 mutex_unlock(&c->write_points_hash_lock);
 717                 return false;
 718         }
 719
 720         wp = c->write_points + --c->write_points_nr;
 721
 722         hlist_del_rcu(&wp->node);
 723         mutex_unlock(&c->write_points_hash_lock);
 724
 725         bch2_writepoint_stop(c, NULL, wp);
 726         return true;
 727 }
 728
 729 static struct write_point *writepoint_find(struct bch_fs *c,
 730                                            unsigned long write_point)
 731 {
 732         struct write_point *wp, *oldest;
 733         struct hlist_head *head;
 734
 735         if (!(write_point & 1UL)) {
 736                 wp = (struct write_point *) write_point;
 737                 mutex_lock(&wp->lock);
 738                 return wp;
 739         }
 740
 741         head = writepoint_hash(c, write_point);
 742 restart_find:
 743         wp = __writepoint_find(head, write_point);
 744         if (wp) {
 745 lock_wp:
 746                 mutex_lock(&wp->lock);
 747                 if (wp->write_point == write_point)
 748                         goto out;
 749                 mutex_unlock(&wp->lock);
 750                 goto restart_find;
 751         }
 752 restart_find_oldest:
 753         oldest = NULL;
 754         for (wp = c->write_points;
 755              wp < c->write_points + c->write_points_nr; wp++)
 756                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
 757                         oldest = wp;
 758
 759         mutex_lock(&oldest->lock);
 760         mutex_lock(&c->write_points_hash_lock);
 761         if (oldest >= c->write_points + c->write_points_nr ||
 762             try_increase_writepoints(c)) {
 763                 mutex_unlock(&c->write_points_hash_lock);
 764                 mutex_unlock(&oldest->lock);
 765                 goto restart_find_oldest;
 766         }
 767
 768         wp = __writepoint_find(head, write_point);
 769         if (wp && wp != oldest) {
 770                 mutex_unlock(&c->write_points_hash_lock);
 771                 mutex_unlock(&oldest->lock);
 772                 goto lock_wp;
 773         }
 774
 775         wp = oldest;
 776         hlist_del_rcu(&wp->node);
 777         wp->write_point = write_point;
 778         hlist_add_head_rcu(&wp->node, head);
 779         mutex_unlock(&c->write_points_hash_lock);
 780 out:
 781         wp->last_used = sched_clock();
 782         return wp;
 783 }
 784
 785 /*
 786  * Get us an open_bucket we can allocate from, return with it locked:
 787  */
 788 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
 789                                 unsigned target,
 790                                 unsigned erasure_code,
 791                                 struct write_point_specifier write_point,
 792                                 struct bch_devs_list *devs_have,
 793                                 unsigned nr_replicas,
 794                                 unsigned nr_replicas_required,
 795                                 enum alloc_reserve reserve,
 796                                 unsigned flags,
 797                                 struct closure *cl)
 798 {
 799         struct write_point *wp;
 800         struct open_bucket *ob;
 801         struct open_buckets ptrs;
 802         unsigned nr_effective, write_points_nr;
 803         unsigned ob_flags = 0;
 804         bool have_cache;
 805         enum bucket_alloc_ret ret;
 806         int i;
 807
 808         if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
 809                 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
 810
 811         BUG_ON(!nr_replicas || !nr_replicas_required);
 812 retry:
 813         ptrs.nr         = 0;
 814         nr_effective    = 0;
 815         write_points_nr = c->write_points_nr;
 816         have_cache      = false;
 817
 818         wp = writepoint_find(c, write_point.v);
 819
 820         if (wp->type == BCH_DATA_user)
 821                 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
 822
 823         /* metadata may not allocate on cache devices: */
 824         if (wp->type != BCH_DATA_user)
 825                 have_cache = true;
 826
 827         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
 828                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 829                                               target, erasure_code,
 830                                               nr_replicas, &nr_effective,
 831                                               &have_cache, reserve,
 832                                               ob_flags, cl);
 833         } else {
 834                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 835                                               target, erasure_code,
 836                                               nr_replicas, &nr_effective,
 837                                               &have_cache, reserve,
 838                                               ob_flags, NULL);
 839                 if (!ret)
 840                         goto alloc_done;
 841
 842                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 843                                               0, erasure_code,
 844                                               nr_replicas, &nr_effective,
 845                                               &have_cache, reserve,
 846                                               ob_flags, cl);
 847         }
 848 alloc_done:
 849         BUG_ON(!ret && nr_effective < nr_replicas);
 850
 851         if (erasure_code && !ec_open_bucket(c, &ptrs))
 852                 pr_debug("failed to get ec bucket: ret %u", ret);
 853
 854         if (ret == INSUFFICIENT_DEVICES &&
 855             nr_effective >= nr_replicas_required)
 856                 ret = 0;
 857
 858         if (ret)
 859                 goto err;
 860
 861         /* Free buckets we didn't use: */
 862         open_bucket_for_each(c, &wp->ptrs, ob, i)
 863                 open_bucket_free_unused(c, wp, ob);
 864
 865         wp->ptrs = ptrs;
 866
 867         wp->sectors_free = UINT_MAX;
 868
 869         open_bucket_for_each(c, &wp->ptrs, ob, i)
 870                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
 871
 872         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
 873
 874         verify_not_stale(c, &wp->ptrs);
 875
 876         return wp;
 877 err:
 878         open_bucket_for_each(c, &wp->ptrs, ob, i)
 879                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
 880                         ob_push(c, &ptrs, ob);
 881                 else
 882                         open_bucket_free_unused(c, wp, ob);
 883         wp->ptrs = ptrs;
 884
 885         mutex_unlock(&wp->lock);
 886
 887         if (ret == FREELIST_EMPTY &&
 888             try_decrease_writepoints(c, write_points_nr))
 889                 goto retry;
 890
 891         switch (ret) {
 892         case OPEN_BUCKETS_EMPTY:
 893         case FREELIST_EMPTY:
 894                 return cl ? ERR_PTR(-EAGAIN) : ERR_PTR(-ENOSPC);
 895         case INSUFFICIENT_DEVICES:
 896                 return ERR_PTR(-EROFS);
 897         default:
 898                 BUG();
 899         }
 900 }
 901
 902 /*
 903  * Append pointers to the space we just allocated to @k, and mark @sectors space
 904  * as allocated out of @ob
 905  */
 906 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
 907                                     struct bkey_i *k, unsigned sectors)
 908
 909 {
 910         struct open_bucket *ob;
 911         unsigned i;
 912
 913         BUG_ON(sectors > wp->sectors_free);
 914         wp->sectors_free -= sectors;
 915
 916         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 917                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 918                 struct bch_extent_ptr tmp = ob->ptr;
 919
 920                 tmp.cached = !ca->mi.durability &&
 921                         wp->type == BCH_DATA_user;
 922
 923                 tmp.offset += ca->mi.bucket_size - ob->sectors_free;
 924                 bch2_bkey_append_ptr(k, tmp);
 925
 926                 BUG_ON(sectors > ob->sectors_free);
 927                 ob->sectors_free -= sectors;
 928         }
 929 }
 930
 931 /*
 932  * Append pointers to the space we just allocated to @k, and mark @sectors space
 933  * as allocated out of @ob
 934  */
 935 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
 936 {
 937         struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
 938         struct open_bucket *ob;
 939         unsigned i;
 940
 941         open_bucket_for_each(c, &wp->ptrs, ob, i)
 942                 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
 943         wp->ptrs = keep;
 944
 945         mutex_unlock(&wp->lock);
 946
 947         bch2_open_buckets_put(c, &ptrs);
 948 }
 949
 950 static inline void writepoint_init(struct write_point *wp,
 951                                    enum bch_data_type type)
 952 {
 953         mutex_init(&wp->lock);
 954         wp->type = type;
 955 }
 956
 957 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
 958 {
 959         struct open_bucket *ob;
 960         struct write_point *wp;
 961
 962         mutex_init(&c->write_points_hash_lock);
 963         c->write_points_nr = ARRAY_SIZE(c->write_points);
 964
 965         /* open bucket 0 is a sentinal NULL: */
 966         spin_lock_init(&c->open_buckets[0].lock);
 967
 968         for (ob = c->open_buckets + 1;
 969              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
 970                 spin_lock_init(&ob->lock);
 971                 c->open_buckets_nr_free++;
 972
 973                 ob->freelist = c->open_buckets_freelist;
 974                 c->open_buckets_freelist = ob - c->open_buckets;
 975         }
 976
 977         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
 978         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
 979         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
 980
 981         for (wp = c->write_points;
 982              wp < c->write_points + c->write_points_nr; wp++) {
 983                 writepoint_init(wp, BCH_DATA_user);
 984
 985                 wp->last_used   = sched_clock();
 986                 wp->write_point = (unsigned long) wp;
 987                 hlist_add_head_rcu(&wp->node,
 988                                    writepoint_hash(c, wp->write_point));
 989         }
 990 }