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         bucket_io_clock_reset(c, ca, bucket, READ);
 313         bucket_io_clock_reset(c, ca, bucket, WRITE);
 314         spin_unlock(&ob->lock);
 315
 316         if (c->blocked_allocate_open_bucket) {
 317                 bch2_time_stats_update(
 318                         &c->times[BCH_TIME_blocked_allocate_open_bucket],
 319                         c->blocked_allocate_open_bucket);
 320                 c->blocked_allocate_open_bucket = 0;
 321         }
 322
 323         if (c->blocked_allocate) {
 324                 bch2_time_stats_update(
 325                         &c->times[BCH_TIME_blocked_allocate],
 326                         c->blocked_allocate);
 327                 c->blocked_allocate = 0;
 328         }
 329
 330         spin_unlock(&c->freelist_lock);
 331
 332         bch2_wake_allocator(ca);
 333
 334         trace_bucket_alloc(ca, reserve);
 335         return ob;
 336 }
 337
 338 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 339                             unsigned l, unsigned r)
 340 {
 341         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 342                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 343 }
 344
 345 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 346
 347 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 348                                           struct dev_stripe_state *stripe,
 349                                           struct bch_devs_mask *devs)
 350 {
 351         struct dev_alloc_list ret = { .nr = 0 };
 352         unsigned i;
 353
 354         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 355                 ret.devs[ret.nr++] = i;
 356
 357         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 358         return ret;
 359 }
 360
 361 void bch2_dev_stripe_increment(struct bch_dev *ca,
 362                                struct dev_stripe_state *stripe)
 363 {
 364         u64 *v = stripe->next_alloc + ca->dev_idx;
 365         u64 free_space = dev_buckets_free(ca);
 366         u64 free_space_inv = free_space
 367                 ? div64_u64(1ULL << 48, free_space)
 368                 : 1ULL << 48;
 369         u64 scale = *v / 4;
 370
 371         if (*v + free_space_inv >= *v)
 372                 *v += free_space_inv;
 373         else
 374                 *v = U64_MAX;
 375
 376         for (v = stripe->next_alloc;
 377              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 378                 *v = *v < scale ? 0 : *v - scale;
 379 }
 380
 381 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 382 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 383
 384 static void add_new_bucket(struct bch_fs *c,
 385                            struct open_buckets *ptrs,
 386                            struct bch_devs_mask *devs_may_alloc,
 387                            unsigned *nr_effective,
 388                            bool *have_cache,
 389                            unsigned flags,
 390                            struct open_bucket *ob)
 391 {
 392         unsigned durability =
 393                 bch_dev_bkey_exists(c, ob->ptr.dev)->mi.durability;
 394
 395         __clear_bit(ob->ptr.dev, devs_may_alloc->d);
 396         *nr_effective   += (flags & BUCKET_ALLOC_USE_DURABILITY)
 397                 ? durability : 1;
 398         *have_cache     |= !durability;
 399
 400         ob_push(c, ptrs, ob);
 401 }
 402
 403 enum bucket_alloc_ret
 404 bch2_bucket_alloc_set(struct bch_fs *c,
 405                       struct open_buckets *ptrs,
 406                       struct dev_stripe_state *stripe,
 407                       struct bch_devs_mask *devs_may_alloc,
 408                       unsigned nr_replicas,
 409                       unsigned *nr_effective,
 410                       bool *have_cache,
 411                       enum alloc_reserve reserve,
 412                       unsigned flags,
 413                       struct closure *cl)
 414 {
 415         struct dev_alloc_list devs_sorted =
 416                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 417         struct bch_dev *ca;
 418         enum bucket_alloc_ret ret = INSUFFICIENT_DEVICES;
 419         unsigned i;
 420
 421         BUG_ON(*nr_effective >= nr_replicas);
 422
 423         for (i = 0; i < devs_sorted.nr; i++) {
 424                 struct open_bucket *ob;
 425
 426                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
 427                 if (!ca)
 428                         continue;
 429
 430                 if (!ca->mi.durability && *have_cache)
 431                         continue;
 432
 433                 ob = bch2_bucket_alloc(c, ca, reserve,
 434                                 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
 435                 if (IS_ERR(ob)) {
 436                         ret = -PTR_ERR(ob);
 437
 438                         if (cl)
 439                                 return ret;
 440                         continue;
 441                 }
 442
 443                 add_new_bucket(c, ptrs, devs_may_alloc,
 444                                nr_effective, have_cache, flags, ob);
 445
 446                 bch2_dev_stripe_increment(ca, stripe);
 447
 448                 if (*nr_effective >= nr_replicas)
 449                         return ALLOC_SUCCESS;
 450         }
 451
 452         return ret;
 453 }
 454
 455 /* Allocate from stripes: */
 456
 457 /*
 458  * if we can't allocate a new stripe because there are already too many
 459  * partially filled stripes, force allocating from an existing stripe even when
 460  * it's to a device we don't want:
 461  */
 462
 463 static void bucket_alloc_from_stripe(struct bch_fs *c,
 464                                      struct open_buckets *ptrs,
 465                                      struct write_point *wp,
 466                                      struct bch_devs_mask *devs_may_alloc,
 467                                      u16 target,
 468                                      unsigned erasure_code,
 469                                      unsigned nr_replicas,
 470                                      unsigned *nr_effective,
 471                                      bool *have_cache,
 472                                      unsigned flags)
 473 {
 474         struct dev_alloc_list devs_sorted;
 475         struct ec_stripe_head *h;
 476         struct open_bucket *ob;
 477         struct bch_dev *ca;
 478         unsigned i, ec_idx;
 479
 480         if (!erasure_code)
 481                 return;
 482
 483         if (nr_replicas < 2)
 484                 return;
 485
 486         if (ec_open_bucket(c, ptrs))
 487                 return;
 488
 489         h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1);
 490         if (!h)
 491                 return;
 492
 493         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 494
 495         for (i = 0; i < devs_sorted.nr; i++)
 496                 open_bucket_for_each(c, &h->s->blocks, ob, ec_idx)
 497                         if (ob->ptr.dev == devs_sorted.devs[i] &&
 498                             !test_and_set_bit(h->s->data_block_idx[ec_idx],
 499                                               h->s->blocks_allocated))
 500                                 goto got_bucket;
 501         goto out_put_head;
 502 got_bucket:
 503         ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 504
 505         ob->ec_idx      = h->s->data_block_idx[ec_idx];
 506         ob->ec          = h->s;
 507
 508         add_new_bucket(c, ptrs, devs_may_alloc,
 509                        nr_effective, have_cache, flags, ob);
 510         atomic_inc(&h->s->pin);
 511 out_put_head:
 512         bch2_ec_stripe_head_put(c, h);
 513 }
 514
 515 /* Sector allocator */
 516
 517 static void get_buckets_from_writepoint(struct bch_fs *c,
 518                                         struct open_buckets *ptrs,
 519                                         struct write_point *wp,
 520                                         struct bch_devs_mask *devs_may_alloc,
 521                                         unsigned nr_replicas,
 522                                         unsigned *nr_effective,
 523                                         bool *have_cache,
 524                                         unsigned flags,
 525                                         bool need_ec)
 526 {
 527         struct open_buckets ptrs_skip = { .nr = 0 };
 528         struct open_bucket *ob;
 529         unsigned i;
 530
 531         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 532                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 533
 534                 if (*nr_effective < nr_replicas &&
 535                     test_bit(ob->ptr.dev, devs_may_alloc->d) &&
 536                     (ca->mi.durability ||
 537                      (wp->type == BCH_DATA_user && !*have_cache)) &&
 538                     (ob->ec || !need_ec)) {
 539                         add_new_bucket(c, ptrs, devs_may_alloc,
 540                                        nr_effective, have_cache,
 541                                        flags, ob);
 542                 } else {
 543                         ob_push(c, &ptrs_skip, ob);
 544                 }
 545         }
 546         wp->ptrs = ptrs_skip;
 547 }
 548
 549 static enum bucket_alloc_ret
 550 open_bucket_add_buckets(struct bch_fs *c,
 551                         struct open_buckets *ptrs,
 552                         struct write_point *wp,
 553                         struct bch_devs_list *devs_have,
 554                         u16 target,
 555                         unsigned erasure_code,
 556                         unsigned nr_replicas,
 557                         unsigned *nr_effective,
 558                         bool *have_cache,
 559                         enum alloc_reserve reserve,
 560                         unsigned flags,
 561                         struct closure *_cl)
 562 {
 563         struct bch_devs_mask devs;
 564         struct open_bucket *ob;
 565         struct closure *cl = NULL;
 566         enum bucket_alloc_ret ret;
 567         unsigned i;
 568
 569         rcu_read_lock();
 570         devs = target_rw_devs(c, wp->type, target);
 571         rcu_read_unlock();
 572
 573         /* Don't allocate from devices we already have pointers to: */
 574         for (i = 0; i < devs_have->nr; i++)
 575                 __clear_bit(devs_have->devs[i], devs.d);
 576
 577         open_bucket_for_each(c, ptrs, ob, i)
 578                 __clear_bit(ob->ptr.dev, devs.d);
 579
 580         if (erasure_code) {
 581                 if (!ec_open_bucket(c, ptrs)) {
 582                         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 583                                                     nr_replicas, nr_effective,
 584                                                     have_cache, flags, true);
 585                         if (*nr_effective >= nr_replicas)
 586                                 return 0;
 587                 }
 588
 589                 if (!ec_open_bucket(c, ptrs)) {
 590                         bucket_alloc_from_stripe(c, ptrs, wp, &devs,
 591                                                  target, erasure_code,
 592                                                  nr_replicas, nr_effective,
 593                                                  have_cache, flags);
 594                         if (*nr_effective >= nr_replicas)
 595                                 return 0;
 596                 }
 597         }
 598
 599         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 600                                     nr_replicas, nr_effective,
 601                                     have_cache, flags, false);
 602         if (*nr_effective >= nr_replicas)
 603                 return 0;
 604
 605         percpu_down_read(&c->mark_lock);
 606         rcu_read_lock();
 607
 608 retry_blocking:
 609         /*
 610          * Try nonblocking first, so that if one device is full we'll try from
 611          * other devices:
 612          */
 613         ret = bch2_bucket_alloc_set(c, ptrs, &wp->stripe, &devs,
 614                                 nr_replicas, nr_effective, have_cache,
 615                                 reserve, flags, cl);
 616         if (ret && ret != INSUFFICIENT_DEVICES && !cl && _cl) {
 617                 cl = _cl;
 618                 goto retry_blocking;
 619         }
 620
 621         rcu_read_unlock();
 622         percpu_up_read(&c->mark_lock);
 623
 624         return ret;
 625 }
 626
 627 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 628                                 struct open_buckets *obs)
 629 {
 630         struct open_buckets ptrs = { .nr = 0 };
 631         struct open_bucket *ob, *ob2;
 632         unsigned i, j;
 633
 634         open_bucket_for_each(c, obs, ob, i) {
 635                 bool drop = !ca || ob->ptr.dev == ca->dev_idx;
 636
 637                 if (!drop && ob->ec) {
 638                         mutex_lock(&ob->ec->lock);
 639                         open_bucket_for_each(c, &ob->ec->blocks, ob2, j)
 640                                 drop |= ob2->ptr.dev == ca->dev_idx;
 641                         open_bucket_for_each(c, &ob->ec->parity, ob2, j)
 642                                 drop |= ob2->ptr.dev == ca->dev_idx;
 643                         mutex_unlock(&ob->ec->lock);
 644                 }
 645
 646                 if (drop)
 647                         bch2_open_bucket_put(c, ob);
 648                 else
 649                         ob_push(c, &ptrs, ob);
 650         }
 651
 652         *obs = ptrs;
 653 }
 654
 655 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 656                           struct write_point *wp)
 657 {
 658         mutex_lock(&wp->lock);
 659         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
 660         mutex_unlock(&wp->lock);
 661 }
 662
 663 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 664                                                  unsigned long write_point)
 665 {
 666         unsigned hash =
 667                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
 668
 669         return &c->write_points_hash[hash];
 670 }
 671
 672 static struct write_point *__writepoint_find(struct hlist_head *head,
 673                                              unsigned long write_point)
 674 {
 675         struct write_point *wp;
 676
 677         hlist_for_each_entry_rcu(wp, head, node)
 678                 if (wp->write_point == write_point)
 679                         return wp;
 680
 681         return NULL;
 682 }
 683
 684 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
 685 {
 686         u64 stranded    = c->write_points_nr * c->bucket_size_max;
 687         u64 free        = bch2_fs_usage_read_short(c).free;
 688
 689         return stranded * factor > free;
 690 }
 691
 692 static bool try_increase_writepoints(struct bch_fs *c)
 693 {
 694         struct write_point *wp;
 695
 696         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
 697             too_many_writepoints(c, 32))
 698                 return false;
 699
 700         wp = c->write_points + c->write_points_nr++;
 701         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
 702         return true;
 703 }
 704
 705 static bool try_decrease_writepoints(struct bch_fs *c,
 706                                      unsigned old_nr)
 707 {
 708         struct write_point *wp;
 709
 710         mutex_lock(&c->write_points_hash_lock);
 711         if (c->write_points_nr < old_nr) {
 712                 mutex_unlock(&c->write_points_hash_lock);
 713                 return true;
 714         }
 715
 716         if (c->write_points_nr == 1 ||
 717             !too_many_writepoints(c, 8)) {
 718                 mutex_unlock(&c->write_points_hash_lock);
 719                 return false;
 720         }
 721
 722         wp = c->write_points + --c->write_points_nr;
 723
 724         hlist_del_rcu(&wp->node);
 725         mutex_unlock(&c->write_points_hash_lock);
 726
 727         bch2_writepoint_stop(c, NULL, wp);
 728         return true;
 729 }
 730
 731 static struct write_point *writepoint_find(struct bch_fs *c,
 732                                            unsigned long write_point)
 733 {
 734         struct write_point *wp, *oldest;
 735         struct hlist_head *head;
 736
 737         if (!(write_point & 1UL)) {
 738                 wp = (struct write_point *) write_point;
 739                 mutex_lock(&wp->lock);
 740                 return wp;
 741         }
 742
 743         head = writepoint_hash(c, write_point);
 744 restart_find:
 745         wp = __writepoint_find(head, write_point);
 746         if (wp) {
 747 lock_wp:
 748                 mutex_lock(&wp->lock);
 749                 if (wp->write_point == write_point)
 750                         goto out;
 751                 mutex_unlock(&wp->lock);
 752                 goto restart_find;
 753         }
 754 restart_find_oldest:
 755         oldest = NULL;
 756         for (wp = c->write_points;
 757              wp < c->write_points + c->write_points_nr; wp++)
 758                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
 759                         oldest = wp;
 760
 761         mutex_lock(&oldest->lock);
 762         mutex_lock(&c->write_points_hash_lock);
 763         if (oldest >= c->write_points + c->write_points_nr ||
 764             try_increase_writepoints(c)) {
 765                 mutex_unlock(&c->write_points_hash_lock);
 766                 mutex_unlock(&oldest->lock);
 767                 goto restart_find_oldest;
 768         }
 769
 770         wp = __writepoint_find(head, write_point);
 771         if (wp && wp != oldest) {
 772                 mutex_unlock(&c->write_points_hash_lock);
 773                 mutex_unlock(&oldest->lock);
 774                 goto lock_wp;
 775         }
 776
 777         wp = oldest;
 778         hlist_del_rcu(&wp->node);
 779         wp->write_point = write_point;
 780         hlist_add_head_rcu(&wp->node, head);
 781         mutex_unlock(&c->write_points_hash_lock);
 782 out:
 783         wp->last_used = sched_clock();
 784         return wp;
 785 }
 786
 787 /*
 788  * Get us an open_bucket we can allocate from, return with it locked:
 789  */
 790 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
 791                                 unsigned target,
 792                                 unsigned erasure_code,
 793                                 struct write_point_specifier write_point,
 794                                 struct bch_devs_list *devs_have,
 795                                 unsigned nr_replicas,
 796                                 unsigned nr_replicas_required,
 797                                 enum alloc_reserve reserve,
 798                                 unsigned flags,
 799                                 struct closure *cl)
 800 {
 801         struct write_point *wp;
 802         struct open_bucket *ob;
 803         struct open_buckets ptrs;
 804         unsigned nr_effective, write_points_nr;
 805         unsigned ob_flags = 0;
 806         bool have_cache;
 807         enum bucket_alloc_ret ret;
 808         int i;
 809
 810         if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
 811                 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
 812
 813         BUG_ON(!nr_replicas || !nr_replicas_required);
 814 retry:
 815         ptrs.nr         = 0;
 816         nr_effective    = 0;
 817         write_points_nr = c->write_points_nr;
 818         have_cache      = false;
 819
 820         wp = writepoint_find(c, write_point.v);
 821
 822         if (wp->type == BCH_DATA_user)
 823                 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
 824
 825         /* metadata may not allocate on cache devices: */
 826         if (wp->type != BCH_DATA_user)
 827                 have_cache = true;
 828
 829         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
 830                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 831                                               target, erasure_code,
 832                                               nr_replicas, &nr_effective,
 833                                               &have_cache, reserve,
 834                                               ob_flags, cl);
 835         } else {
 836                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 837                                               target, erasure_code,
 838                                               nr_replicas, &nr_effective,
 839                                               &have_cache, reserve,
 840                                               ob_flags, NULL);
 841                 if (!ret)
 842                         goto alloc_done;
 843
 844                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 845                                               0, erasure_code,
 846                                               nr_replicas, &nr_effective,
 847                                               &have_cache, reserve,
 848                                               ob_flags, cl);
 849         }
 850 alloc_done:
 851         BUG_ON(!ret && nr_effective < nr_replicas);
 852
 853         if (erasure_code && !ec_open_bucket(c, &ptrs))
 854                 pr_debug("failed to get ec bucket: ret %u", ret);
 855
 856         if (ret == INSUFFICIENT_DEVICES &&
 857             nr_effective >= nr_replicas_required)
 858                 ret = 0;
 859
 860         if (ret)
 861                 goto err;
 862
 863         /* Free buckets we didn't use: */
 864         open_bucket_for_each(c, &wp->ptrs, ob, i)
 865                 open_bucket_free_unused(c, wp, ob);
 866
 867         wp->ptrs = ptrs;
 868
 869         wp->sectors_free = UINT_MAX;
 870
 871         open_bucket_for_each(c, &wp->ptrs, ob, i)
 872                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
 873
 874         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
 875
 876         verify_not_stale(c, &wp->ptrs);
 877
 878         return wp;
 879 err:
 880         open_bucket_for_each(c, &wp->ptrs, ob, i)
 881                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
 882                         ob_push(c, &ptrs, ob);
 883                 else
 884                         open_bucket_free_unused(c, wp, ob);
 885         wp->ptrs = ptrs;
 886
 887         mutex_unlock(&wp->lock);
 888
 889         if (ret == FREELIST_EMPTY &&
 890             try_decrease_writepoints(c, write_points_nr))
 891                 goto retry;
 892
 893         switch (ret) {
 894         case OPEN_BUCKETS_EMPTY:
 895         case FREELIST_EMPTY:
 896                 return cl ? ERR_PTR(-EAGAIN) : ERR_PTR(-ENOSPC);
 897         case INSUFFICIENT_DEVICES:
 898                 return ERR_PTR(-EROFS);
 899         default:
 900                 BUG();
 901         }
 902 }
 903
 904 /*
 905  * Append pointers to the space we just allocated to @k, and mark @sectors space
 906  * as allocated out of @ob
 907  */
 908 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
 909                                     struct bkey_i *k, unsigned sectors)
 910
 911 {
 912         struct open_bucket *ob;
 913         unsigned i;
 914
 915         BUG_ON(sectors > wp->sectors_free);
 916         wp->sectors_free -= sectors;
 917
 918         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 919                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 920                 struct bch_extent_ptr tmp = ob->ptr;
 921
 922                 tmp.cached = !ca->mi.durability &&
 923                         wp->type == BCH_DATA_user;
 924
 925                 tmp.offset += ca->mi.bucket_size - ob->sectors_free;
 926                 bch2_bkey_append_ptr(k, tmp);
 927
 928                 BUG_ON(sectors > ob->sectors_free);
 929                 ob->sectors_free -= sectors;
 930         }
 931 }
 932
 933 /*
 934  * Append pointers to the space we just allocated to @k, and mark @sectors space
 935  * as allocated out of @ob
 936  */
 937 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
 938 {
 939         struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
 940         struct open_bucket *ob;
 941         unsigned i;
 942
 943         open_bucket_for_each(c, &wp->ptrs, ob, i)
 944                 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
 945         wp->ptrs = keep;
 946
 947         mutex_unlock(&wp->lock);
 948
 949         bch2_open_buckets_put(c, &ptrs);
 950 }
 951
 952 static inline void writepoint_init(struct write_point *wp,
 953                                    enum bch_data_type type)
 954 {
 955         mutex_init(&wp->lock);
 956         wp->type = type;
 957 }
 958
 959 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
 960 {
 961         struct open_bucket *ob;
 962         struct write_point *wp;
 963
 964         mutex_init(&c->write_points_hash_lock);
 965         c->write_points_nr = ARRAY_SIZE(c->write_points);
 966
 967         /* open bucket 0 is a sentinal NULL: */
 968         spin_lock_init(&c->open_buckets[0].lock);
 969
 970         for (ob = c->open_buckets + 1;
 971              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
 972                 spin_lock_init(&ob->lock);
 973                 c->open_buckets_nr_free++;
 974
 975                 ob->freelist = c->open_buckets_freelist;
 976                 c->open_buckets_freelist = ob - c->open_buckets;
 977         }
 978
 979         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
 980         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
 981         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
 982
 983         for (wp = c->write_points;
 984              wp < c->write_points + c->write_points_nr; wp++) {
 985                 writepoint_init(wp, BCH_DATA_user);
 986
 987                 wp->last_used   = sched_clock();
 988                 wp->write_point = (unsigned long) wp;
 989                 hlist_add_head_rcu(&wp->node,
 990                                    writepoint_hash(c, wp->write_point));
 991         }
 992 }