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 = buckets->first_bucket; b < buckets->nbuckets; b++)
 196                 if (is_available_bucket(buckets->b[b].mark) &&
 197                     !buckets->b[b].mark.owned_by_allocator)
 198                         goto success;
 199         b = -1;
 200 success:
 201         rcu_read_unlock();
 202         return b;
 203 }
 204
 205 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
 206 {
 207         switch (reserve) {
 208         case RESERVE_BTREE:
 209         case RESERVE_BTREE_MOVINGGC:
 210                 return 0;
 211         case RESERVE_MOVINGGC:
 212                 return OPEN_BUCKETS_COUNT / 4;
 213         default:
 214                 return OPEN_BUCKETS_COUNT / 2;
 215         }
 216 }
 217
 218 /**
 219  * bch_bucket_alloc - allocate a single bucket from a specific device
 220  *
 221  * Returns index of bucket on success, 0 on failure
 222  * */
 223 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 224                                       enum alloc_reserve reserve,
 225                                       bool may_alloc_partial,
 226                                       struct closure *cl)
 227 {
 228         struct open_bucket *ob;
 229         long b = 0;
 230
 231         spin_lock(&c->freelist_lock);
 232
 233         if (may_alloc_partial) {
 234                 int i;
 235
 236                 for (i = ca->open_buckets_partial_nr - 1; i >= 0; --i) {
 237                         ob = c->open_buckets + ca->open_buckets_partial[i];
 238
 239                         if (reserve <= ob->alloc_reserve) {
 240                                 array_remove_item(ca->open_buckets_partial,
 241                                                   ca->open_buckets_partial_nr,
 242                                                   i);
 243                                 ob->on_partial_list = false;
 244                                 ob->alloc_reserve = reserve;
 245                                 spin_unlock(&c->freelist_lock);
 246                                 return ob;
 247                         }
 248                 }
 249         }
 250
 251         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
 252                 if (cl)
 253                         closure_wait(&c->open_buckets_wait, cl);
 254
 255                 if (!c->blocked_allocate_open_bucket)
 256                         c->blocked_allocate_open_bucket = local_clock();
 257
 258                 spin_unlock(&c->freelist_lock);
 259                 trace_open_bucket_alloc_fail(ca, reserve);
 260                 return ERR_PTR(-OPEN_BUCKETS_EMPTY);
 261         }
 262
 263         if (likely(fifo_pop(&ca->free[RESERVE_NONE], b)))
 264                 goto out;
 265
 266         switch (reserve) {
 267         case RESERVE_BTREE_MOVINGGC:
 268         case RESERVE_MOVINGGC:
 269                 if (fifo_pop(&ca->free[RESERVE_MOVINGGC], b))
 270                         goto out;
 271                 break;
 272         default:
 273                 break;
 274         }
 275
 276         if (cl)
 277                 closure_wait(&c->freelist_wait, cl);
 278
 279         if (!c->blocked_allocate)
 280                 c->blocked_allocate = local_clock();
 281
 282         spin_unlock(&c->freelist_lock);
 283
 284         trace_bucket_alloc_fail(ca, reserve);
 285         return ERR_PTR(-FREELIST_EMPTY);
 286 out:
 287         verify_not_on_freelist(c, ca, b);
 288
 289         ob = bch2_open_bucket_alloc(c);
 290
 291         spin_lock(&ob->lock);
 292
 293         ob->valid       = true;
 294         ob->sectors_free = ca->mi.bucket_size;
 295         ob->alloc_reserve = reserve;
 296         ob->ptr         = (struct bch_extent_ptr) {
 297                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
 298                 .gen    = bucket(ca, b)->mark.gen,
 299                 .offset = bucket_to_sector(ca, b),
 300                 .dev    = ca->dev_idx,
 301         };
 302
 303         spin_unlock(&ob->lock);
 304
 305         if (c->blocked_allocate_open_bucket) {
 306                 bch2_time_stats_update(
 307                         &c->times[BCH_TIME_blocked_allocate_open_bucket],
 308                         c->blocked_allocate_open_bucket);
 309                 c->blocked_allocate_open_bucket = 0;
 310         }
 311
 312         if (c->blocked_allocate) {
 313                 bch2_time_stats_update(
 314                         &c->times[BCH_TIME_blocked_allocate],
 315                         c->blocked_allocate);
 316                 c->blocked_allocate = 0;
 317         }
 318
 319         spin_unlock(&c->freelist_lock);
 320
 321         bch2_wake_allocator(ca);
 322
 323         trace_bucket_alloc(ca, reserve);
 324         return ob;
 325 }
 326
 327 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 328                             unsigned l, unsigned r)
 329 {
 330         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 331                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 332 }
 333
 334 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 335
 336 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 337                                           struct dev_stripe_state *stripe,
 338                                           struct bch_devs_mask *devs)
 339 {
 340         struct dev_alloc_list ret = { .nr = 0 };
 341         unsigned i;
 342
 343         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 344                 ret.devs[ret.nr++] = i;
 345
 346         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 347         return ret;
 348 }
 349
 350 void bch2_dev_stripe_increment(struct bch_dev *ca,
 351                                struct dev_stripe_state *stripe)
 352 {
 353         u64 *v = stripe->next_alloc + ca->dev_idx;
 354         u64 free_space = dev_buckets_free(ca);
 355         u64 free_space_inv = free_space
 356                 ? div64_u64(1ULL << 48, free_space)
 357                 : 1ULL << 48;
 358         u64 scale = *v / 4;
 359
 360         if (*v + free_space_inv >= *v)
 361                 *v += free_space_inv;
 362         else
 363                 *v = U64_MAX;
 364
 365         for (v = stripe->next_alloc;
 366              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 367                 *v = *v < scale ? 0 : *v - scale;
 368 }
 369
 370 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 371 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 372
 373 static void add_new_bucket(struct bch_fs *c,
 374                            struct open_buckets *ptrs,
 375                            struct bch_devs_mask *devs_may_alloc,
 376                            unsigned *nr_effective,
 377                            bool *have_cache,
 378                            unsigned flags,
 379                            struct open_bucket *ob)
 380 {
 381         unsigned durability =
 382                 bch_dev_bkey_exists(c, ob->ptr.dev)->mi.durability;
 383
 384         __clear_bit(ob->ptr.dev, devs_may_alloc->d);
 385         *nr_effective   += (flags & BUCKET_ALLOC_USE_DURABILITY)
 386                 ? durability : 1;
 387         *have_cache     |= !durability;
 388
 389         ob_push(c, ptrs, ob);
 390 }
 391
 392 enum bucket_alloc_ret
 393 bch2_bucket_alloc_set(struct bch_fs *c,
 394                       struct open_buckets *ptrs,
 395                       struct dev_stripe_state *stripe,
 396                       struct bch_devs_mask *devs_may_alloc,
 397                       unsigned nr_replicas,
 398                       unsigned *nr_effective,
 399                       bool *have_cache,
 400                       enum alloc_reserve reserve,
 401                       unsigned flags,
 402                       struct closure *cl)
 403 {
 404         struct dev_alloc_list devs_sorted =
 405                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 406         struct bch_dev *ca;
 407         enum bucket_alloc_ret ret = INSUFFICIENT_DEVICES;
 408         unsigned i;
 409
 410         BUG_ON(*nr_effective >= nr_replicas);
 411
 412         for (i = 0; i < devs_sorted.nr; i++) {
 413                 struct open_bucket *ob;
 414
 415                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
 416                 if (!ca)
 417                         continue;
 418
 419                 if (!ca->mi.durability && *have_cache)
 420                         continue;
 421
 422                 ob = bch2_bucket_alloc(c, ca, reserve,
 423                                 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
 424                 if (IS_ERR(ob)) {
 425                         ret = -PTR_ERR(ob);
 426
 427                         if (cl)
 428                                 return ret;
 429                         continue;
 430                 }
 431
 432                 add_new_bucket(c, ptrs, devs_may_alloc,
 433                                nr_effective, have_cache, flags, ob);
 434
 435                 bch2_dev_stripe_increment(ca, stripe);
 436
 437                 if (*nr_effective >= nr_replicas)
 438                         return ALLOC_SUCCESS;
 439         }
 440
 441         return ret;
 442 }
 443
 444 /* Allocate from stripes: */
 445
 446 /*
 447  * if we can't allocate a new stripe because there are already too many
 448  * partially filled stripes, force allocating from an existing stripe even when
 449  * it's to a device we don't want:
 450  */
 451
 452 static enum bucket_alloc_ret
 453 bucket_alloc_from_stripe(struct bch_fs *c,
 454                          struct open_buckets *ptrs,
 455                          struct write_point *wp,
 456                          struct bch_devs_mask *devs_may_alloc,
 457                          u16 target,
 458                          unsigned erasure_code,
 459                          unsigned nr_replicas,
 460                          unsigned *nr_effective,
 461                          bool *have_cache,
 462                          unsigned flags,
 463                          struct closure *cl)
 464 {
 465         struct dev_alloc_list devs_sorted;
 466         struct ec_stripe_head *h;
 467         struct open_bucket *ob;
 468         struct bch_dev *ca;
 469         unsigned i, ec_idx;
 470
 471         if (!erasure_code)
 472                 return 0;
 473
 474         if (nr_replicas < 2)
 475                 return 0;
 476
 477         if (ec_open_bucket(c, ptrs))
 478                 return 0;
 479
 480         h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1,
 481                                     wp == &c->copygc_write_point,
 482                                     cl);
 483         if (IS_ERR(h))
 484                 return -PTR_ERR(h);
 485         if (!h)
 486                 return 0;
 487
 488         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 489
 490         for (i = 0; i < devs_sorted.nr; i++)
 491                 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 492                         if (!h->s->blocks[ec_idx])
 493                                 continue;
 494
 495                         ob = c->open_buckets + h->s->blocks[ec_idx];
 496                         if (ob->ptr.dev == devs_sorted.devs[i] &&
 497                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 498                                 goto got_bucket;
 499                 }
 500         goto out_put_head;
 501 got_bucket:
 502         ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 503
 504         ob->ec_idx      = ec_idx;
 505         ob->ec          = h->s;
 506
 507         add_new_bucket(c, ptrs, devs_may_alloc,
 508                        nr_effective, have_cache, flags, ob);
 509         atomic_inc(&h->s->pin);
 510 out_put_head:
 511         bch2_ec_stripe_head_put(c, h);
 512         return 0;
 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                         ret = bucket_alloc_from_stripe(c, ptrs, wp, &devs,
 591                                                  target, erasure_code,
 592                                                  nr_replicas, nr_effective,
 593                                                  have_cache, flags, _cl);
 594                         if (ret == FREELIST_EMPTY ||
 595                             ret == OPEN_BUCKETS_EMPTY)
 596                                 return ret;
 597                         if (*nr_effective >= nr_replicas)
 598                                 return 0;
 599                 }
 600         }
 601
 602         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 603                                     nr_replicas, nr_effective,
 604                                     have_cache, flags, false);
 605         if (*nr_effective >= nr_replicas)
 606                 return 0;
 607
 608         percpu_down_read(&c->mark_lock);
 609         rcu_read_lock();
 610
 611 retry_blocking:
 612         /*
 613          * Try nonblocking first, so that if one device is full we'll try from
 614          * other devices:
 615          */
 616         ret = bch2_bucket_alloc_set(c, ptrs, &wp->stripe, &devs,
 617                                 nr_replicas, nr_effective, have_cache,
 618                                 reserve, flags, cl);
 619         if (ret && ret != INSUFFICIENT_DEVICES && !cl && _cl) {
 620                 cl = _cl;
 621                 goto retry_blocking;
 622         }
 623
 624         rcu_read_unlock();
 625         percpu_up_read(&c->mark_lock);
 626
 627         return ret;
 628 }
 629
 630 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 631                                 struct open_buckets *obs)
 632 {
 633         struct open_buckets ptrs = { .nr = 0 };
 634         struct open_bucket *ob, *ob2;
 635         unsigned i, j;
 636
 637         open_bucket_for_each(c, obs, ob, i) {
 638                 bool drop = !ca || ob->ptr.dev == ca->dev_idx;
 639
 640                 if (!drop && ob->ec) {
 641                         mutex_lock(&ob->ec->lock);
 642                         for (j = 0; j < ob->ec->new_stripe.key.v.nr_blocks; j++) {
 643                                 if (!ob->ec->blocks[j])
 644                                         continue;
 645
 646                                 ob2 = c->open_buckets + ob->ec->blocks[j];
 647                                 drop |= ob2->ptr.dev == ca->dev_idx;
 648                         }
 649                         mutex_unlock(&ob->ec->lock);
 650                 }
 651
 652                 if (drop)
 653                         bch2_open_bucket_put(c, ob);
 654                 else
 655                         ob_push(c, &ptrs, ob);
 656         }
 657
 658         *obs = ptrs;
 659 }
 660
 661 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 662                           struct write_point *wp)
 663 {
 664         mutex_lock(&wp->lock);
 665         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
 666         mutex_unlock(&wp->lock);
 667 }
 668
 669 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 670                                                  unsigned long write_point)
 671 {
 672         unsigned hash =
 673                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
 674
 675         return &c->write_points_hash[hash];
 676 }
 677
 678 static struct write_point *__writepoint_find(struct hlist_head *head,
 679                                              unsigned long write_point)
 680 {
 681         struct write_point *wp;
 682
 683         hlist_for_each_entry_rcu(wp, head, node)
 684                 if (wp->write_point == write_point)
 685                         return wp;
 686
 687         return NULL;
 688 }
 689
 690 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
 691 {
 692         u64 stranded    = c->write_points_nr * c->bucket_size_max;
 693         u64 free        = bch2_fs_usage_read_short(c).free;
 694
 695         return stranded * factor > free;
 696 }
 697
 698 static bool try_increase_writepoints(struct bch_fs *c)
 699 {
 700         struct write_point *wp;
 701
 702         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
 703             too_many_writepoints(c, 32))
 704                 return false;
 705
 706         wp = c->write_points + c->write_points_nr++;
 707         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
 708         return true;
 709 }
 710
 711 static bool try_decrease_writepoints(struct bch_fs *c,
 712                                      unsigned old_nr)
 713 {
 714         struct write_point *wp;
 715
 716         mutex_lock(&c->write_points_hash_lock);
 717         if (c->write_points_nr < old_nr) {
 718                 mutex_unlock(&c->write_points_hash_lock);
 719                 return true;
 720         }
 721
 722         if (c->write_points_nr == 1 ||
 723             !too_many_writepoints(c, 8)) {
 724                 mutex_unlock(&c->write_points_hash_lock);
 725                 return false;
 726         }
 727
 728         wp = c->write_points + --c->write_points_nr;
 729
 730         hlist_del_rcu(&wp->node);
 731         mutex_unlock(&c->write_points_hash_lock);
 732
 733         bch2_writepoint_stop(c, NULL, wp);
 734         return true;
 735 }
 736
 737 static struct write_point *writepoint_find(struct bch_fs *c,
 738                                            unsigned long write_point)
 739 {
 740         struct write_point *wp, *oldest;
 741         struct hlist_head *head;
 742
 743         if (!(write_point & 1UL)) {
 744                 wp = (struct write_point *) write_point;
 745                 mutex_lock(&wp->lock);
 746                 return wp;
 747         }
 748
 749         head = writepoint_hash(c, write_point);
 750 restart_find:
 751         wp = __writepoint_find(head, write_point);
 752         if (wp) {
 753 lock_wp:
 754                 mutex_lock(&wp->lock);
 755                 if (wp->write_point == write_point)
 756                         goto out;
 757                 mutex_unlock(&wp->lock);
 758                 goto restart_find;
 759         }
 760 restart_find_oldest:
 761         oldest = NULL;
 762         for (wp = c->write_points;
 763              wp < c->write_points + c->write_points_nr; wp++)
 764                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
 765                         oldest = wp;
 766
 767         mutex_lock(&oldest->lock);
 768         mutex_lock(&c->write_points_hash_lock);
 769         if (oldest >= c->write_points + c->write_points_nr ||
 770             try_increase_writepoints(c)) {
 771                 mutex_unlock(&c->write_points_hash_lock);
 772                 mutex_unlock(&oldest->lock);
 773                 goto restart_find_oldest;
 774         }
 775
 776         wp = __writepoint_find(head, write_point);
 777         if (wp && wp != oldest) {
 778                 mutex_unlock(&c->write_points_hash_lock);
 779                 mutex_unlock(&oldest->lock);
 780                 goto lock_wp;
 781         }
 782
 783         wp = oldest;
 784         hlist_del_rcu(&wp->node);
 785         wp->write_point = write_point;
 786         hlist_add_head_rcu(&wp->node, head);
 787         mutex_unlock(&c->write_points_hash_lock);
 788 out:
 789         wp->last_used = sched_clock();
 790         return wp;
 791 }
 792
 793 /*
 794  * Get us an open_bucket we can allocate from, return with it locked:
 795  */
 796 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
 797                                 unsigned target,
 798                                 unsigned erasure_code,
 799                                 struct write_point_specifier write_point,
 800                                 struct bch_devs_list *devs_have,
 801                                 unsigned nr_replicas,
 802                                 unsigned nr_replicas_required,
 803                                 enum alloc_reserve reserve,
 804                                 unsigned flags,
 805                                 struct closure *cl)
 806 {
 807         struct write_point *wp;
 808         struct open_bucket *ob;
 809         struct open_buckets ptrs;
 810         unsigned nr_effective, write_points_nr;
 811         unsigned ob_flags = 0;
 812         bool have_cache;
 813         enum bucket_alloc_ret ret;
 814         int i;
 815
 816         if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
 817                 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
 818
 819         BUG_ON(!nr_replicas || !nr_replicas_required);
 820 retry:
 821         ptrs.nr         = 0;
 822         nr_effective    = 0;
 823         write_points_nr = c->write_points_nr;
 824         have_cache      = false;
 825
 826         wp = writepoint_find(c, write_point.v);
 827
 828         if (wp->type == BCH_DATA_user)
 829                 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
 830
 831         /* metadata may not allocate on cache devices: */
 832         if (wp->type != BCH_DATA_user)
 833                 have_cache = true;
 834
 835         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
 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, cl);
 841         } else {
 842                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 843                                               target, erasure_code,
 844                                               nr_replicas, &nr_effective,
 845                                               &have_cache, reserve,
 846                                               ob_flags, NULL);
 847                 if (!ret)
 848                         goto alloc_done;
 849
 850                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 851                                               0, erasure_code,
 852                                               nr_replicas, &nr_effective,
 853                                               &have_cache, reserve,
 854                                               ob_flags, cl);
 855         }
 856 alloc_done:
 857         BUG_ON(!ret && nr_effective < nr_replicas);
 858
 859         if (erasure_code && !ec_open_bucket(c, &ptrs))
 860                 pr_debug("failed to get ec bucket: ret %u", ret);
 861
 862         if (ret == INSUFFICIENT_DEVICES &&
 863             nr_effective >= nr_replicas_required)
 864                 ret = 0;
 865
 866         if (ret)
 867                 goto err;
 868
 869         /* Free buckets we didn't use: */
 870         open_bucket_for_each(c, &wp->ptrs, ob, i)
 871                 open_bucket_free_unused(c, wp, ob);
 872
 873         wp->ptrs = ptrs;
 874
 875         wp->sectors_free = UINT_MAX;
 876
 877         open_bucket_for_each(c, &wp->ptrs, ob, i)
 878                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
 879
 880         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
 881
 882         verify_not_stale(c, &wp->ptrs);
 883
 884         return wp;
 885 err:
 886         open_bucket_for_each(c, &wp->ptrs, ob, i)
 887                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
 888                         ob_push(c, &ptrs, ob);
 889                 else
 890                         open_bucket_free_unused(c, wp, ob);
 891         wp->ptrs = ptrs;
 892
 893         mutex_unlock(&wp->lock);
 894
 895         if (ret == FREELIST_EMPTY &&
 896             try_decrease_writepoints(c, write_points_nr))
 897                 goto retry;
 898
 899         switch (ret) {
 900         case OPEN_BUCKETS_EMPTY:
 901         case FREELIST_EMPTY:
 902                 return cl ? ERR_PTR(-EAGAIN) : ERR_PTR(-ENOSPC);
 903         case INSUFFICIENT_DEVICES:
 904                 return ERR_PTR(-EROFS);
 905         default:
 906                 BUG();
 907         }
 908 }
 909
 910 /*
 911  * Append pointers to the space we just allocated to @k, and mark @sectors space
 912  * as allocated out of @ob
 913  */
 914 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
 915                                     struct bkey_i *k, unsigned sectors)
 916
 917 {
 918         struct open_bucket *ob;
 919         unsigned i;
 920
 921         BUG_ON(sectors > wp->sectors_free);
 922         wp->sectors_free -= sectors;
 923
 924         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 925                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 926                 struct bch_extent_ptr tmp = ob->ptr;
 927
 928                 tmp.cached = !ca->mi.durability &&
 929                         wp->type == BCH_DATA_user;
 930
 931                 tmp.offset += ca->mi.bucket_size - ob->sectors_free;
 932                 bch2_bkey_append_ptr(k, tmp);
 933
 934                 BUG_ON(sectors > ob->sectors_free);
 935                 ob->sectors_free -= sectors;
 936         }
 937 }
 938
 939 /*
 940  * Append pointers to the space we just allocated to @k, and mark @sectors space
 941  * as allocated out of @ob
 942  */
 943 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
 944 {
 945         struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
 946         struct open_bucket *ob;
 947         unsigned i;
 948
 949         open_bucket_for_each(c, &wp->ptrs, ob, i)
 950                 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
 951         wp->ptrs = keep;
 952
 953         mutex_unlock(&wp->lock);
 954
 955         bch2_open_buckets_put(c, &ptrs);
 956 }
 957
 958 static inline void writepoint_init(struct write_point *wp,
 959                                    enum bch_data_type type)
 960 {
 961         mutex_init(&wp->lock);
 962         wp->type = type;
 963 }
 964
 965 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
 966 {
 967         struct open_bucket *ob;
 968         struct write_point *wp;
 969
 970         mutex_init(&c->write_points_hash_lock);
 971         c->write_points_nr = ARRAY_SIZE(c->write_points);
 972
 973         /* open bucket 0 is a sentinal NULL: */
 974         spin_lock_init(&c->open_buckets[0].lock);
 975
 976         for (ob = c->open_buckets + 1;
 977              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
 978                 spin_lock_init(&ob->lock);
 979                 c->open_buckets_nr_free++;
 980
 981                 ob->freelist = c->open_buckets_freelist;
 982                 c->open_buckets_freelist = ob - c->open_buckets;
 983         }
 984
 985         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
 986         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
 987         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
 988
 989         for (wp = c->write_points;
 990              wp < c->write_points + c->write_points_nr; wp++) {
 991                 writepoint_init(wp, BCH_DATA_user);
 992
 993                 wp->last_used   = sched_clock();
 994                 wp->write_point = (unsigned long) wp;
 995                 hlist_add_head_rcu(&wp->node,
 996                                    writepoint_hash(c, wp->write_point));
 997         }
 998 }