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

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright 2012 Google, Inc.
   4  *
   5  * Foreground allocator code: allocate buckets from freelist, and allocate in
   6  * sector granularity from writepoints.
   7  *
   8  * bch2_bucket_alloc() allocates a single bucket from a specific device.
   9  *
  10  * bch2_bucket_alloc_set() allocates one or more buckets from different devices
  11  * in a given filesystem.
  12  */
  13
  14 #include "bcachefs.h"
  15 #include "alloc_background.h"
  16 #include "alloc_foreground.h"
  17 #include "btree_gc.h"
  18 #include "buckets.h"
  19 #include "clock.h"
  20 #include "debug.h"
  21 #include "disk_groups.h"
  22 #include "ec.h"
  23 #include "io.h"
  24
  25 #include <linux/math64.h>
  26 #include <linux/rculist.h>
  27 #include <linux/rcupdate.h>
  28 #include <trace/events/bcachefs.h>
  29
  30 /*
  31  * Open buckets represent a bucket that's currently being allocated from.  They
  32  * serve two purposes:
  33  *
  34  *  - They track buckets that have been partially allocated, allowing for
  35  *    sub-bucket sized allocations - they're used by the sector allocator below
  36  *
  37  *  - They provide a reference to the buckets they own that mark and sweep GC
  38  *    can find, until the new allocation has a pointer to it inserted into the
  39  *    btree
  40  *
  41  * When allocating some space with the sector allocator, the allocation comes
  42  * with a reference to an open bucket - the caller is required to put that
  43  * reference _after_ doing the index update that makes its allocation reachable.
  44  */
  45
  46 static void bch2_open_bucket_hash_add(struct bch_fs *c, struct open_bucket *ob)
  47 {
  48         open_bucket_idx_t idx = ob - c->open_buckets;
  49         open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
  50
  51         ob->hash = *slot;
  52         *slot = idx;
  53 }
  54
  55 static void bch2_open_bucket_hash_remove(struct bch_fs *c, struct open_bucket *ob)
  56 {
  57         open_bucket_idx_t idx = ob - c->open_buckets;
  58         open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
  59
  60         while (*slot != idx) {
  61                 BUG_ON(!*slot);
  62                 slot = &c->open_buckets[*slot].hash;
  63         }
  64
  65         *slot = ob->hash;
  66         ob->hash = 0;
  67 }
  68
  69 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
  70 {
  71         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
  72
  73         if (ob->ec) {
  74                 bch2_ec_bucket_written(c, ob);
  75                 return;
  76         }
  77
  78         percpu_down_read(&c->mark_lock);
  79         spin_lock(&ob->lock);
  80
  81         bch2_mark_alloc_bucket(c, ca, ob->bucket, false);
  82         ob->valid = false;
  83         ob->data_type = 0;
  84
  85         spin_unlock(&ob->lock);
  86         percpu_up_read(&c->mark_lock);
  87
  88         spin_lock(&c->freelist_lock);
  89         bch2_open_bucket_hash_remove(c, ob);
  90
  91         ob->freelist = c->open_buckets_freelist;
  92         c->open_buckets_freelist = ob - c->open_buckets;
  93
  94         c->open_buckets_nr_free++;
  95         ca->nr_open_buckets--;
  96         spin_unlock(&c->freelist_lock);
  97
  98         closure_wake_up(&c->open_buckets_wait);
  99 }
 100
 101 void bch2_open_bucket_write_error(struct bch_fs *c,
 102                                   struct open_buckets *obs,
 103                                   unsigned dev)
 104 {
 105         struct open_bucket *ob;
 106         unsigned i;
 107
 108         open_bucket_for_each(c, obs, ob, i)
 109                 if (ob->dev == dev && ob->ec)
 110                         bch2_ec_bucket_cancel(c, ob);
 111 }
 112
 113 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
 114 {
 115         struct open_bucket *ob;
 116
 117         BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
 118
 119         ob = c->open_buckets + c->open_buckets_freelist;
 120         c->open_buckets_freelist = ob->freelist;
 121         atomic_set(&ob->pin, 1);
 122         ob->data_type = 0;
 123
 124         c->open_buckets_nr_free--;
 125         return ob;
 126 }
 127
 128 static void open_bucket_free_unused(struct bch_fs *c,
 129                                     struct write_point *wp,
 130                                     struct open_bucket *ob)
 131 {
 132         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 133         bool may_realloc = wp->data_type == BCH_DATA_user;
 134
 135         BUG_ON(ca->open_buckets_partial_nr >
 136                ARRAY_SIZE(ca->open_buckets_partial));
 137
 138         if (ca->open_buckets_partial_nr <
 139             ARRAY_SIZE(ca->open_buckets_partial) &&
 140             may_realloc) {
 141                 spin_lock(&c->freelist_lock);
 142                 ob->on_partial_list = true;
 143                 ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
 144                         ob - c->open_buckets;
 145                 spin_unlock(&c->freelist_lock);
 146
 147                 closure_wake_up(&c->open_buckets_wait);
 148                 closure_wake_up(&c->freelist_wait);
 149         } else {
 150                 bch2_open_bucket_put(c, ob);
 151         }
 152 }
 153
 154 /* _only_ for allocating the journal on a new device: */
 155 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
 156 {
 157         while (ca->new_fs_bucket_idx < ca->mi.nbuckets) {
 158                 u64 b = ca->new_fs_bucket_idx++;
 159
 160                 if (!is_superblock_bucket(ca, b) &&
 161                     (!ca->buckets_nouse || !test_bit(b, ca->buckets_nouse)))
 162                         return b;
 163         }
 164
 165         return -1;
 166 }
 167
 168 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
 169 {
 170         switch (reserve) {
 171         case RESERVE_BTREE:
 172         case RESERVE_BTREE_MOVINGGC:
 173                 return 0;
 174         case RESERVE_MOVINGGC:
 175                 return OPEN_BUCKETS_COUNT / 4;
 176         default:
 177                 return OPEN_BUCKETS_COUNT / 2;
 178         }
 179 }
 180
 181 /**
 182  * bch_bucket_alloc - allocate a single bucket from a specific device
 183  *
 184  * Returns index of bucket on success, 0 on failure
 185  * */
 186 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 187                                       enum alloc_reserve reserve,
 188                                       bool may_alloc_partial,
 189                                       struct closure *cl)
 190 {
 191         struct open_bucket *ob;
 192         long b = 0;
 193
 194         spin_lock(&c->freelist_lock);
 195
 196         if (may_alloc_partial) {
 197                 int i;
 198
 199                 for (i = ca->open_buckets_partial_nr - 1; i >= 0; --i) {
 200                         ob = c->open_buckets + ca->open_buckets_partial[i];
 201
 202                         if (reserve <= ob->alloc_reserve) {
 203                                 array_remove_item(ca->open_buckets_partial,
 204                                                   ca->open_buckets_partial_nr,
 205                                                   i);
 206                                 ob->on_partial_list = false;
 207                                 ob->alloc_reserve = reserve;
 208                                 spin_unlock(&c->freelist_lock);
 209                                 return ob;
 210                         }
 211                 }
 212         }
 213
 214         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
 215                 if (cl)
 216                         closure_wait(&c->open_buckets_wait, cl);
 217
 218                 if (!c->blocked_allocate_open_bucket)
 219                         c->blocked_allocate_open_bucket = local_clock();
 220
 221                 spin_unlock(&c->freelist_lock);
 222                 trace_open_bucket_alloc_fail(ca, reserve);
 223                 return ERR_PTR(-OPEN_BUCKETS_EMPTY);
 224         }
 225
 226         if (likely(fifo_pop(&ca->free[RESERVE_NONE], b)))
 227                 goto out;
 228
 229         switch (reserve) {
 230         case RESERVE_BTREE_MOVINGGC:
 231         case RESERVE_MOVINGGC:
 232                 if (fifo_pop(&ca->free[RESERVE_MOVINGGC], b))
 233                         goto out;
 234                 break;
 235         default:
 236                 break;
 237         }
 238
 239         if (cl)
 240                 closure_wait(&c->freelist_wait, cl);
 241
 242         if (!c->blocked_allocate)
 243                 c->blocked_allocate = local_clock();
 244
 245         spin_unlock(&c->freelist_lock);
 246
 247         trace_bucket_alloc_fail(ca, reserve);
 248         return ERR_PTR(-FREELIST_EMPTY);
 249 out:
 250         verify_not_on_freelist(c, ca, b);
 251
 252         ob = bch2_open_bucket_alloc(c);
 253
 254         spin_lock(&ob->lock);
 255
 256         ob->valid       = true;
 257         ob->sectors_free = ca->mi.bucket_size;
 258         ob->alloc_reserve = reserve;
 259         ob->dev         = ca->dev_idx;
 260         ob->gen         = *bucket_gen(ca, b);
 261         ob->bucket      = b;
 262         spin_unlock(&ob->lock);
 263
 264         ca->nr_open_buckets++;
 265         bch2_open_bucket_hash_add(c, ob);
 266
 267         if (c->blocked_allocate_open_bucket) {
 268                 bch2_time_stats_update(
 269                         &c->times[BCH_TIME_blocked_allocate_open_bucket],
 270                         c->blocked_allocate_open_bucket);
 271                 c->blocked_allocate_open_bucket = 0;
 272         }
 273
 274         if (c->blocked_allocate) {
 275                 bch2_time_stats_update(
 276                         &c->times[BCH_TIME_blocked_allocate],
 277                         c->blocked_allocate);
 278                 c->blocked_allocate = 0;
 279         }
 280
 281         spin_unlock(&c->freelist_lock);
 282
 283         bch2_wake_allocator(ca);
 284
 285         trace_bucket_alloc(ca, reserve);
 286         return ob;
 287 }
 288
 289 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 290                             unsigned l, unsigned r)
 291 {
 292         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 293                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 294 }
 295
 296 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 297
 298 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 299                                           struct dev_stripe_state *stripe,
 300                                           struct bch_devs_mask *devs)
 301 {
 302         struct dev_alloc_list ret = { .nr = 0 };
 303         unsigned i;
 304
 305         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 306                 ret.devs[ret.nr++] = i;
 307
 308         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 309         return ret;
 310 }
 311
 312 void bch2_dev_stripe_increment(struct bch_dev *ca,
 313                                struct dev_stripe_state *stripe)
 314 {
 315         u64 *v = stripe->next_alloc + ca->dev_idx;
 316         u64 free_space = dev_buckets_available(ca);
 317         u64 free_space_inv = free_space
 318                 ? div64_u64(1ULL << 48, free_space)
 319                 : 1ULL << 48;
 320         u64 scale = *v / 4;
 321
 322         if (*v + free_space_inv >= *v)
 323                 *v += free_space_inv;
 324         else
 325                 *v = U64_MAX;
 326
 327         for (v = stripe->next_alloc;
 328              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 329                 *v = *v < scale ? 0 : *v - scale;
 330 }
 331
 332 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 333 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 334
 335 static void add_new_bucket(struct bch_fs *c,
 336                            struct open_buckets *ptrs,
 337                            struct bch_devs_mask *devs_may_alloc,
 338                            unsigned *nr_effective,
 339                            bool *have_cache,
 340                            unsigned flags,
 341                            struct open_bucket *ob)
 342 {
 343         unsigned durability =
 344                 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
 345
 346         __clear_bit(ob->dev, devs_may_alloc->d);
 347         *nr_effective   += (flags & BUCKET_ALLOC_USE_DURABILITY)
 348                 ? durability : 1;
 349         *have_cache     |= !durability;
 350
 351         ob_push(c, ptrs, ob);
 352 }
 353
 354 int bch2_bucket_alloc_set(struct bch_fs *c,
 355                       struct open_buckets *ptrs,
 356                       struct dev_stripe_state *stripe,
 357                       struct bch_devs_mask *devs_may_alloc,
 358                       unsigned nr_replicas,
 359                       unsigned *nr_effective,
 360                       bool *have_cache,
 361                       enum alloc_reserve reserve,
 362                       unsigned flags,
 363                       struct closure *cl)
 364 {
 365         struct dev_alloc_list devs_sorted =
 366                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 367         struct bch_dev *ca;
 368         int ret = -INSUFFICIENT_DEVICES;
 369         unsigned i;
 370
 371         BUG_ON(*nr_effective >= nr_replicas);
 372
 373         for (i = 0; i < devs_sorted.nr; i++) {
 374                 struct open_bucket *ob;
 375
 376                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
 377                 if (!ca)
 378                         continue;
 379
 380                 if (!ca->mi.durability && *have_cache)
 381                         continue;
 382
 383                 ob = bch2_bucket_alloc(c, ca, reserve,
 384                                 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
 385                 if (IS_ERR(ob)) {
 386                         ret = PTR_ERR(ob);
 387
 388                         if (cl)
 389                                 return ret;
 390                         continue;
 391                 }
 392
 393                 add_new_bucket(c, ptrs, devs_may_alloc,
 394                                nr_effective, have_cache, flags, ob);
 395
 396                 bch2_dev_stripe_increment(ca, stripe);
 397
 398                 if (*nr_effective >= nr_replicas)
 399                         return 0;
 400         }
 401
 402         return ret;
 403 }
 404
 405 /* Allocate from stripes: */
 406
 407 /*
 408  * if we can't allocate a new stripe because there are already too many
 409  * partially filled stripes, force allocating from an existing stripe even when
 410  * it's to a device we don't want:
 411  */
 412
 413 static int bucket_alloc_from_stripe(struct bch_fs *c,
 414                          struct open_buckets *ptrs,
 415                          struct write_point *wp,
 416                          struct bch_devs_mask *devs_may_alloc,
 417                          u16 target,
 418                          unsigned erasure_code,
 419                          unsigned nr_replicas,
 420                          unsigned *nr_effective,
 421                          bool *have_cache,
 422                          unsigned flags,
 423                          struct closure *cl)
 424 {
 425         struct dev_alloc_list devs_sorted;
 426         struct ec_stripe_head *h;
 427         struct open_bucket *ob;
 428         struct bch_dev *ca;
 429         unsigned i, ec_idx;
 430
 431         if (!erasure_code)
 432                 return 0;
 433
 434         if (nr_replicas < 2)
 435                 return 0;
 436
 437         if (ec_open_bucket(c, ptrs))
 438                 return 0;
 439
 440         h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1,
 441                                     wp == &c->copygc_write_point,
 442                                     cl);
 443         if (IS_ERR(h))
 444                 return -PTR_ERR(h);
 445         if (!h)
 446                 return 0;
 447
 448         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 449
 450         for (i = 0; i < devs_sorted.nr; i++)
 451                 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 452                         if (!h->s->blocks[ec_idx])
 453                                 continue;
 454
 455                         ob = c->open_buckets + h->s->blocks[ec_idx];
 456                         if (ob->dev == devs_sorted.devs[i] &&
 457                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 458                                 goto got_bucket;
 459                 }
 460         goto out_put_head;
 461 got_bucket:
 462         ca = bch_dev_bkey_exists(c, ob->dev);
 463
 464         ob->ec_idx      = ec_idx;
 465         ob->ec          = h->s;
 466
 467         add_new_bucket(c, ptrs, devs_may_alloc,
 468                        nr_effective, have_cache, flags, ob);
 469         atomic_inc(&h->s->pin);
 470 out_put_head:
 471         bch2_ec_stripe_head_put(c, h);
 472         return 0;
 473 }
 474
 475 /* Sector allocator */
 476
 477 static void get_buckets_from_writepoint(struct bch_fs *c,
 478                                         struct open_buckets *ptrs,
 479                                         struct write_point *wp,
 480                                         struct bch_devs_mask *devs_may_alloc,
 481                                         unsigned nr_replicas,
 482                                         unsigned *nr_effective,
 483                                         bool *have_cache,
 484                                         unsigned flags,
 485                                         bool need_ec)
 486 {
 487         struct open_buckets ptrs_skip = { .nr = 0 };
 488         struct open_bucket *ob;
 489         unsigned i;
 490
 491         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 492                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 493
 494                 if (*nr_effective < nr_replicas &&
 495                     test_bit(ob->dev, devs_may_alloc->d) &&
 496                     (ca->mi.durability ||
 497                      (wp->data_type == BCH_DATA_user && !*have_cache)) &&
 498                     (ob->ec || !need_ec)) {
 499                         add_new_bucket(c, ptrs, devs_may_alloc,
 500                                        nr_effective, have_cache,
 501                                        flags, ob);
 502                 } else {
 503                         ob_push(c, &ptrs_skip, ob);
 504                 }
 505         }
 506         wp->ptrs = ptrs_skip;
 507 }
 508
 509 static int open_bucket_add_buckets(struct bch_fs *c,
 510                         struct open_buckets *ptrs,
 511                         struct write_point *wp,
 512                         struct bch_devs_list *devs_have,
 513                         u16 target,
 514                         unsigned erasure_code,
 515                         unsigned nr_replicas,
 516                         unsigned *nr_effective,
 517                         bool *have_cache,
 518                         enum alloc_reserve reserve,
 519                         unsigned flags,
 520                         struct closure *_cl)
 521 {
 522         struct bch_devs_mask devs;
 523         struct open_bucket *ob;
 524         struct closure *cl = NULL;
 525         int ret;
 526         unsigned i;
 527
 528         rcu_read_lock();
 529         devs = target_rw_devs(c, wp->data_type, target);
 530         rcu_read_unlock();
 531
 532         /* Don't allocate from devices we already have pointers to: */
 533         for (i = 0; i < devs_have->nr; i++)
 534                 __clear_bit(devs_have->devs[i], devs.d);
 535
 536         open_bucket_for_each(c, ptrs, ob, i)
 537                 __clear_bit(ob->dev, devs.d);
 538
 539         if (erasure_code) {
 540                 if (!ec_open_bucket(c, ptrs)) {
 541                         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 542                                                     nr_replicas, nr_effective,
 543                                                     have_cache, flags, true);
 544                         if (*nr_effective >= nr_replicas)
 545                                 return 0;
 546                 }
 547
 548                 if (!ec_open_bucket(c, ptrs)) {
 549                         ret = bucket_alloc_from_stripe(c, ptrs, wp, &devs,
 550                                                  target, erasure_code,
 551                                                  nr_replicas, nr_effective,
 552                                                  have_cache, flags, _cl);
 553                         if (ret == -FREELIST_EMPTY ||
 554                             ret == -OPEN_BUCKETS_EMPTY)
 555                                 return ret;
 556                         if (*nr_effective >= nr_replicas)
 557                                 return 0;
 558                 }
 559         }
 560
 561         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 562                                     nr_replicas, nr_effective,
 563                                     have_cache, flags, false);
 564         if (*nr_effective >= nr_replicas)
 565                 return 0;
 566
 567         percpu_down_read(&c->mark_lock);
 568         rcu_read_lock();
 569
 570 retry_blocking:
 571         /*
 572          * Try nonblocking first, so that if one device is full we'll try from
 573          * other devices:
 574          */
 575         ret = bch2_bucket_alloc_set(c, ptrs, &wp->stripe, &devs,
 576                                 nr_replicas, nr_effective, have_cache,
 577                                 reserve, flags, cl);
 578         if (ret && ret != -INSUFFICIENT_DEVICES && !cl && _cl) {
 579                 cl = _cl;
 580                 goto retry_blocking;
 581         }
 582
 583         rcu_read_unlock();
 584         percpu_up_read(&c->mark_lock);
 585
 586         return ret;
 587 }
 588
 589 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 590                                 struct open_buckets *obs)
 591 {
 592         struct open_buckets ptrs = { .nr = 0 };
 593         struct open_bucket *ob, *ob2;
 594         unsigned i, j;
 595
 596         open_bucket_for_each(c, obs, ob, i) {
 597                 bool drop = !ca || ob->dev == ca->dev_idx;
 598
 599                 if (!drop && ob->ec) {
 600                         mutex_lock(&ob->ec->lock);
 601                         for (j = 0; j < ob->ec->new_stripe.key.v.nr_blocks; j++) {
 602                                 if (!ob->ec->blocks[j])
 603                                         continue;
 604
 605                                 ob2 = c->open_buckets + ob->ec->blocks[j];
 606                                 drop |= ob2->dev == ca->dev_idx;
 607                         }
 608                         mutex_unlock(&ob->ec->lock);
 609                 }
 610
 611                 if (drop)
 612                         bch2_open_bucket_put(c, ob);
 613                 else
 614                         ob_push(c, &ptrs, ob);
 615         }
 616
 617         *obs = ptrs;
 618 }
 619
 620 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 621                           struct write_point *wp)
 622 {
 623         mutex_lock(&wp->lock);
 624         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
 625         mutex_unlock(&wp->lock);
 626 }
 627
 628 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 629                                                  unsigned long write_point)
 630 {
 631         unsigned hash =
 632                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
 633
 634         return &c->write_points_hash[hash];
 635 }
 636
 637 static struct write_point *__writepoint_find(struct hlist_head *head,
 638                                              unsigned long write_point)
 639 {
 640         struct write_point *wp;
 641
 642         rcu_read_lock();
 643         hlist_for_each_entry_rcu(wp, head, node)
 644                 if (wp->write_point == write_point)
 645                         goto out;
 646         wp = NULL;
 647 out:
 648         rcu_read_unlock();
 649         return wp;
 650 }
 651
 652 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
 653 {
 654         u64 stranded    = c->write_points_nr * c->bucket_size_max;
 655         u64 free        = bch2_fs_usage_read_short(c).free;
 656
 657         return stranded * factor > free;
 658 }
 659
 660 static bool try_increase_writepoints(struct bch_fs *c)
 661 {
 662         struct write_point *wp;
 663
 664         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
 665             too_many_writepoints(c, 32))
 666                 return false;
 667
 668         wp = c->write_points + c->write_points_nr++;
 669         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
 670         return true;
 671 }
 672
 673 static bool try_decrease_writepoints(struct bch_fs *c,
 674                                      unsigned old_nr)
 675 {
 676         struct write_point *wp;
 677
 678         mutex_lock(&c->write_points_hash_lock);
 679         if (c->write_points_nr < old_nr) {
 680                 mutex_unlock(&c->write_points_hash_lock);
 681                 return true;
 682         }
 683
 684         if (c->write_points_nr == 1 ||
 685             !too_many_writepoints(c, 8)) {
 686                 mutex_unlock(&c->write_points_hash_lock);
 687                 return false;
 688         }
 689
 690         wp = c->write_points + --c->write_points_nr;
 691
 692         hlist_del_rcu(&wp->node);
 693         mutex_unlock(&c->write_points_hash_lock);
 694
 695         bch2_writepoint_stop(c, NULL, wp);
 696         return true;
 697 }
 698
 699 static struct write_point *writepoint_find(struct bch_fs *c,
 700                                            unsigned long write_point)
 701 {
 702         struct write_point *wp, *oldest;
 703         struct hlist_head *head;
 704
 705         if (!(write_point & 1UL)) {
 706                 wp = (struct write_point *) write_point;
 707                 mutex_lock(&wp->lock);
 708                 return wp;
 709         }
 710
 711         head = writepoint_hash(c, write_point);
 712 restart_find:
 713         wp = __writepoint_find(head, write_point);
 714         if (wp) {
 715 lock_wp:
 716                 mutex_lock(&wp->lock);
 717                 if (wp->write_point == write_point)
 718                         goto out;
 719                 mutex_unlock(&wp->lock);
 720                 goto restart_find;
 721         }
 722 restart_find_oldest:
 723         oldest = NULL;
 724         for (wp = c->write_points;
 725              wp < c->write_points + c->write_points_nr; wp++)
 726                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
 727                         oldest = wp;
 728
 729         mutex_lock(&oldest->lock);
 730         mutex_lock(&c->write_points_hash_lock);
 731         if (oldest >= c->write_points + c->write_points_nr ||
 732             try_increase_writepoints(c)) {
 733                 mutex_unlock(&c->write_points_hash_lock);
 734                 mutex_unlock(&oldest->lock);
 735                 goto restart_find_oldest;
 736         }
 737
 738         wp = __writepoint_find(head, write_point);
 739         if (wp && wp != oldest) {
 740                 mutex_unlock(&c->write_points_hash_lock);
 741                 mutex_unlock(&oldest->lock);
 742                 goto lock_wp;
 743         }
 744
 745         wp = oldest;
 746         hlist_del_rcu(&wp->node);
 747         wp->write_point = write_point;
 748         hlist_add_head_rcu(&wp->node, head);
 749         mutex_unlock(&c->write_points_hash_lock);
 750 out:
 751         wp->last_used = sched_clock();
 752         return wp;
 753 }
 754
 755 /*
 756  * Get us an open_bucket we can allocate from, return with it locked:
 757  */
 758 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
 759                                 unsigned target,
 760                                 unsigned erasure_code,
 761                                 struct write_point_specifier write_point,
 762                                 struct bch_devs_list *devs_have,
 763                                 unsigned nr_replicas,
 764                                 unsigned nr_replicas_required,
 765                                 enum alloc_reserve reserve,
 766                                 unsigned flags,
 767                                 struct closure *cl)
 768 {
 769         struct write_point *wp;
 770         struct open_bucket *ob;
 771         struct open_buckets ptrs;
 772         unsigned nr_effective, write_points_nr;
 773         unsigned ob_flags = 0;
 774         bool have_cache;
 775         int ret;
 776         int i;
 777
 778         if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
 779                 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
 780
 781         BUG_ON(!nr_replicas || !nr_replicas_required);
 782 retry:
 783         ptrs.nr         = 0;
 784         nr_effective    = 0;
 785         write_points_nr = c->write_points_nr;
 786         have_cache      = false;
 787
 788         wp = writepoint_find(c, write_point.v);
 789
 790         if (wp->data_type == BCH_DATA_user)
 791                 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
 792
 793         /* metadata may not allocate on cache devices: */
 794         if (wp->data_type != BCH_DATA_user)
 795                 have_cache = true;
 796
 797         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
 798                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 799                                               target, erasure_code,
 800                                               nr_replicas, &nr_effective,
 801                                               &have_cache, reserve,
 802                                               ob_flags, cl);
 803         } else {
 804                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 805                                               target, erasure_code,
 806                                               nr_replicas, &nr_effective,
 807                                               &have_cache, reserve,
 808                                               ob_flags, NULL);
 809                 if (!ret)
 810                         goto alloc_done;
 811
 812                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 813                                               0, erasure_code,
 814                                               nr_replicas, &nr_effective,
 815                                               &have_cache, reserve,
 816                                               ob_flags, cl);
 817         }
 818 alloc_done:
 819         BUG_ON(!ret && nr_effective < nr_replicas);
 820
 821         if (erasure_code && !ec_open_bucket(c, &ptrs))
 822                 pr_debug("failed to get ec bucket: ret %u", ret);
 823
 824         if (ret == -INSUFFICIENT_DEVICES &&
 825             nr_effective >= nr_replicas_required)
 826                 ret = 0;
 827
 828         if (ret)
 829                 goto err;
 830
 831         /* Free buckets we didn't use: */
 832         open_bucket_for_each(c, &wp->ptrs, ob, i)
 833                 open_bucket_free_unused(c, wp, ob);
 834
 835         wp->ptrs = ptrs;
 836
 837         wp->sectors_free = UINT_MAX;
 838
 839         open_bucket_for_each(c, &wp->ptrs, ob, i)
 840                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
 841
 842         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
 843
 844         return wp;
 845 err:
 846         open_bucket_for_each(c, &wp->ptrs, ob, i)
 847                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
 848                         ob_push(c, &ptrs, ob);
 849                 else
 850                         open_bucket_free_unused(c, wp, ob);
 851         wp->ptrs = ptrs;
 852
 853         mutex_unlock(&wp->lock);
 854
 855         if (ret == -FREELIST_EMPTY &&
 856             try_decrease_writepoints(c, write_points_nr))
 857                 goto retry;
 858
 859         switch (ret) {
 860         case -OPEN_BUCKETS_EMPTY:
 861         case -FREELIST_EMPTY:
 862                 return cl ? ERR_PTR(-EAGAIN) : ERR_PTR(-ENOSPC);
 863         case -INSUFFICIENT_DEVICES:
 864                 return ERR_PTR(-EROFS);
 865         default:
 866                 BUG();
 867         }
 868 }
 869
 870 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
 871 {
 872         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 873
 874         return (struct bch_extent_ptr) {
 875                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
 876                 .gen    = ob->gen,
 877                 .dev    = ob->dev,
 878                 .offset = bucket_to_sector(ca, ob->bucket) +
 879                         ca->mi.bucket_size -
 880                         ob->sectors_free,
 881         };
 882 }
 883
 884 /*
 885  * Append pointers to the space we just allocated to @k, and mark @sectors space
 886  * as allocated out of @ob
 887  */
 888 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
 889                                     struct bkey_i *k, unsigned sectors,
 890                                     bool cached)
 891
 892 {
 893         struct open_bucket *ob;
 894         unsigned i;
 895
 896         BUG_ON(sectors > wp->sectors_free);
 897         wp->sectors_free -= sectors;
 898
 899         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 900                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 901                 struct bch_extent_ptr ptr = bch2_ob_ptr(c, ob);
 902
 903                 ptr.cached = cached ||
 904                         (!ca->mi.durability &&
 905                          wp->data_type == BCH_DATA_user);
 906
 907                 bch2_bkey_append_ptr(k, ptr);
 908
 909                 BUG_ON(sectors > ob->sectors_free);
 910                 ob->sectors_free -= sectors;
 911         }
 912 }
 913
 914 /*
 915  * Append pointers to the space we just allocated to @k, and mark @sectors space
 916  * as allocated out of @ob
 917  */
 918 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
 919 {
 920         struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
 921         struct open_bucket *ob;
 922         unsigned i;
 923
 924         open_bucket_for_each(c, &wp->ptrs, ob, i)
 925                 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
 926         wp->ptrs = keep;
 927
 928         mutex_unlock(&wp->lock);
 929
 930         bch2_open_buckets_put(c, &ptrs);
 931 }
 932
 933 static inline void writepoint_init(struct write_point *wp,
 934                                    enum bch_data_type type)
 935 {
 936         mutex_init(&wp->lock);
 937         wp->data_type = type;
 938 }
 939
 940 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
 941 {
 942         struct open_bucket *ob;
 943         struct write_point *wp;
 944
 945         mutex_init(&c->write_points_hash_lock);
 946         c->write_points_nr = ARRAY_SIZE(c->write_points);
 947
 948         /* open bucket 0 is a sentinal NULL: */
 949         spin_lock_init(&c->open_buckets[0].lock);
 950
 951         for (ob = c->open_buckets + 1;
 952              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
 953                 spin_lock_init(&ob->lock);
 954                 c->open_buckets_nr_free++;
 955
 956                 ob->freelist = c->open_buckets_freelist;
 957                 c->open_buckets_freelist = ob - c->open_buckets;
 958         }
 959
 960         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
 961         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
 962         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
 963
 964         for (wp = c->write_points;
 965              wp < c->write_points + c->write_points_nr; wp++) {
 966                 writepoint_init(wp, BCH_DATA_user);
 967
 968                 wp->last_used   = sched_clock();
 969                 wp->write_point = (unsigned long) wp;
 970                 hlist_add_head_rcu(&wp->node,
 971                                    writepoint_hash(c, wp->write_point));
 972         }
 973 }
 974
 975 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
 976 {
 977         struct open_bucket *ob;
 978
 979         for (ob = c->open_buckets;
 980              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
 981              ob++) {
 982                 spin_lock(&ob->lock);
 983                 if (ob->valid && !ob->on_partial_list) {
 984                         pr_buf(out, "%zu ref %u type %s\n",
 985                                ob - c->open_buckets,
 986                                atomic_read(&ob->pin),
 987                                bch2_data_types[ob->data_type]);
 988                 }
 989                 spin_unlock(&ob->lock);
 990         }
 991
 992 }