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 static void verify_not_stale(struct bch_fs *c, const struct open_buckets *obs)
 155 {
 156 #ifdef CONFIG_BCACHEFS_DEBUG
 157         struct open_bucket *ob;
 158         unsigned i;
 159
 160         rcu_read_lock();
 161         open_bucket_for_each(c, obs, ob, i) {
 162                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 163
 164                 BUG_ON(*bucket_gen(ca, ob->bucket) != ob->gen);
 165         }
 166         rcu_read_unlock();
 167 #endif
 168 }
 169
 170 /* _only_ for allocating the journal on a new device: */
 171 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
 172 {
 173         while (ca->new_fs_bucket_idx < ca->mi.nbuckets) {
 174                 u64 b = ca->new_fs_bucket_idx++;
 175
 176                 if (!is_superblock_bucket(ca, b) &&
 177                     (!ca->buckets_nouse || !test_bit(b, ca->buckets_nouse)))
 178                         return b;
 179         }
 180
 181         return -1;
 182 }
 183
 184 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
 185 {
 186         switch (reserve) {
 187         case RESERVE_BTREE:
 188         case RESERVE_BTREE_MOVINGGC:
 189                 return 0;
 190         case RESERVE_MOVINGGC:
 191                 return OPEN_BUCKETS_COUNT / 4;
 192         default:
 193                 return OPEN_BUCKETS_COUNT / 2;
 194         }
 195 }
 196
 197 /**
 198  * bch_bucket_alloc - allocate a single bucket from a specific device
 199  *
 200  * Returns index of bucket on success, 0 on failure
 201  * */
 202 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 203                                       enum alloc_reserve reserve,
 204                                       bool may_alloc_partial,
 205                                       struct closure *cl)
 206 {
 207         struct open_bucket *ob;
 208         long b = 0;
 209
 210         spin_lock(&c->freelist_lock);
 211
 212         if (may_alloc_partial) {
 213                 int i;
 214
 215                 for (i = ca->open_buckets_partial_nr - 1; i >= 0; --i) {
 216                         ob = c->open_buckets + ca->open_buckets_partial[i];
 217
 218                         if (reserve <= ob->alloc_reserve) {
 219                                 array_remove_item(ca->open_buckets_partial,
 220                                                   ca->open_buckets_partial_nr,
 221                                                   i);
 222                                 ob->on_partial_list = false;
 223                                 ob->alloc_reserve = reserve;
 224                                 spin_unlock(&c->freelist_lock);
 225                                 return ob;
 226                         }
 227                 }
 228         }
 229
 230         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
 231                 if (cl)
 232                         closure_wait(&c->open_buckets_wait, cl);
 233
 234                 if (!c->blocked_allocate_open_bucket)
 235                         c->blocked_allocate_open_bucket = local_clock();
 236
 237                 spin_unlock(&c->freelist_lock);
 238                 trace_open_bucket_alloc_fail(ca, reserve);
 239                 return ERR_PTR(-OPEN_BUCKETS_EMPTY);
 240         }
 241
 242         if (likely(fifo_pop(&ca->free[RESERVE_NONE], b)))
 243                 goto out;
 244
 245         switch (reserve) {
 246         case RESERVE_BTREE_MOVINGGC:
 247         case RESERVE_MOVINGGC:
 248                 if (fifo_pop(&ca->free[RESERVE_MOVINGGC], b))
 249                         goto out;
 250                 break;
 251         default:
 252                 break;
 253         }
 254
 255         if (cl)
 256                 closure_wait(&c->freelist_wait, cl);
 257
 258         if (!c->blocked_allocate)
 259                 c->blocked_allocate = local_clock();
 260
 261         spin_unlock(&c->freelist_lock);
 262
 263         trace_bucket_alloc_fail(ca, reserve);
 264         return ERR_PTR(-FREELIST_EMPTY);
 265 out:
 266         verify_not_on_freelist(c, ca, b);
 267
 268         ob = bch2_open_bucket_alloc(c);
 269
 270         spin_lock(&ob->lock);
 271
 272         ob->valid       = true;
 273         ob->sectors_free = ca->mi.bucket_size;
 274         ob->alloc_reserve = reserve;
 275         ob->dev         = ca->dev_idx;
 276         ob->gen         = *bucket_gen(ca, b);
 277         ob->bucket      = b;
 278         spin_unlock(&ob->lock);
 279
 280         ca->nr_open_buckets++;
 281         bch2_open_bucket_hash_add(c, ob);
 282
 283         if (c->blocked_allocate_open_bucket) {
 284                 bch2_time_stats_update(
 285                         &c->times[BCH_TIME_blocked_allocate_open_bucket],
 286                         c->blocked_allocate_open_bucket);
 287                 c->blocked_allocate_open_bucket = 0;
 288         }
 289
 290         if (c->blocked_allocate) {
 291                 bch2_time_stats_update(
 292                         &c->times[BCH_TIME_blocked_allocate],
 293                         c->blocked_allocate);
 294                 c->blocked_allocate = 0;
 295         }
 296
 297         spin_unlock(&c->freelist_lock);
 298
 299         bch2_wake_allocator(ca);
 300
 301         trace_bucket_alloc(ca, reserve);
 302         return ob;
 303 }
 304
 305 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 306                             unsigned l, unsigned r)
 307 {
 308         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 309                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 310 }
 311
 312 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 313
 314 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 315                                           struct dev_stripe_state *stripe,
 316                                           struct bch_devs_mask *devs)
 317 {
 318         struct dev_alloc_list ret = { .nr = 0 };
 319         unsigned i;
 320
 321         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 322                 ret.devs[ret.nr++] = i;
 323
 324         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 325         return ret;
 326 }
 327
 328 void bch2_dev_stripe_increment(struct bch_dev *ca,
 329                                struct dev_stripe_state *stripe)
 330 {
 331         u64 *v = stripe->next_alloc + ca->dev_idx;
 332         u64 free_space = dev_buckets_available(ca);
 333         u64 free_space_inv = free_space
 334                 ? div64_u64(1ULL << 48, free_space)
 335                 : 1ULL << 48;
 336         u64 scale = *v / 4;
 337
 338         if (*v + free_space_inv >= *v)
 339                 *v += free_space_inv;
 340         else
 341                 *v = U64_MAX;
 342
 343         for (v = stripe->next_alloc;
 344              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 345                 *v = *v < scale ? 0 : *v - scale;
 346 }
 347
 348 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 349 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 350
 351 static void add_new_bucket(struct bch_fs *c,
 352                            struct open_buckets *ptrs,
 353                            struct bch_devs_mask *devs_may_alloc,
 354                            unsigned *nr_effective,
 355                            bool *have_cache,
 356                            unsigned flags,
 357                            struct open_bucket *ob)
 358 {
 359         unsigned durability =
 360                 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
 361
 362         __clear_bit(ob->dev, devs_may_alloc->d);
 363         *nr_effective   += (flags & BUCKET_ALLOC_USE_DURABILITY)
 364                 ? durability : 1;
 365         *have_cache     |= !durability;
 366
 367         ob_push(c, ptrs, ob);
 368 }
 369
 370 int bch2_bucket_alloc_set(struct bch_fs *c,
 371                       struct open_buckets *ptrs,
 372                       struct dev_stripe_state *stripe,
 373                       struct bch_devs_mask *devs_may_alloc,
 374                       unsigned nr_replicas,
 375                       unsigned *nr_effective,
 376                       bool *have_cache,
 377                       enum alloc_reserve reserve,
 378                       unsigned flags,
 379                       struct closure *cl)
 380 {
 381         struct dev_alloc_list devs_sorted =
 382                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 383         struct bch_dev *ca;
 384         int ret = -INSUFFICIENT_DEVICES;
 385         unsigned i;
 386
 387         BUG_ON(*nr_effective >= nr_replicas);
 388
 389         for (i = 0; i < devs_sorted.nr; i++) {
 390                 struct open_bucket *ob;
 391
 392                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
 393                 if (!ca)
 394                         continue;
 395
 396                 if (!ca->mi.durability && *have_cache)
 397                         continue;
 398
 399                 ob = bch2_bucket_alloc(c, ca, reserve,
 400                                 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
 401                 if (IS_ERR(ob)) {
 402                         ret = PTR_ERR(ob);
 403
 404                         if (cl)
 405                                 return ret;
 406                         continue;
 407                 }
 408
 409                 add_new_bucket(c, ptrs, devs_may_alloc,
 410                                nr_effective, have_cache, flags, ob);
 411
 412                 bch2_dev_stripe_increment(ca, stripe);
 413
 414                 if (*nr_effective >= nr_replicas)
 415                         return 0;
 416         }
 417
 418         return ret;
 419 }
 420
 421 /* Allocate from stripes: */
 422
 423 /*
 424  * if we can't allocate a new stripe because there are already too many
 425  * partially filled stripes, force allocating from an existing stripe even when
 426  * it's to a device we don't want:
 427  */
 428
 429 static int bucket_alloc_from_stripe(struct bch_fs *c,
 430                          struct open_buckets *ptrs,
 431                          struct write_point *wp,
 432                          struct bch_devs_mask *devs_may_alloc,
 433                          u16 target,
 434                          unsigned erasure_code,
 435                          unsigned nr_replicas,
 436                          unsigned *nr_effective,
 437                          bool *have_cache,
 438                          unsigned flags,
 439                          struct closure *cl)
 440 {
 441         struct dev_alloc_list devs_sorted;
 442         struct ec_stripe_head *h;
 443         struct open_bucket *ob;
 444         struct bch_dev *ca;
 445         unsigned i, ec_idx;
 446
 447         if (!erasure_code)
 448                 return 0;
 449
 450         if (nr_replicas < 2)
 451                 return 0;
 452
 453         if (ec_open_bucket(c, ptrs))
 454                 return 0;
 455
 456         h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1,
 457                                     wp == &c->copygc_write_point,
 458                                     cl);
 459         if (IS_ERR(h))
 460                 return -PTR_ERR(h);
 461         if (!h)
 462                 return 0;
 463
 464         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 465
 466         for (i = 0; i < devs_sorted.nr; i++)
 467                 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 468                         if (!h->s->blocks[ec_idx])
 469                                 continue;
 470
 471                         ob = c->open_buckets + h->s->blocks[ec_idx];
 472                         if (ob->dev == devs_sorted.devs[i] &&
 473                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 474                                 goto got_bucket;
 475                 }
 476         goto out_put_head;
 477 got_bucket:
 478         ca = bch_dev_bkey_exists(c, ob->dev);
 479
 480         ob->ec_idx      = ec_idx;
 481         ob->ec          = h->s;
 482
 483         add_new_bucket(c, ptrs, devs_may_alloc,
 484                        nr_effective, have_cache, flags, ob);
 485         atomic_inc(&h->s->pin);
 486 out_put_head:
 487         bch2_ec_stripe_head_put(c, h);
 488         return 0;
 489 }
 490
 491 /* Sector allocator */
 492
 493 static void get_buckets_from_writepoint(struct bch_fs *c,
 494                                         struct open_buckets *ptrs,
 495                                         struct write_point *wp,
 496                                         struct bch_devs_mask *devs_may_alloc,
 497                                         unsigned nr_replicas,
 498                                         unsigned *nr_effective,
 499                                         bool *have_cache,
 500                                         unsigned flags,
 501                                         bool need_ec)
 502 {
 503         struct open_buckets ptrs_skip = { .nr = 0 };
 504         struct open_bucket *ob;
 505         unsigned i;
 506
 507         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 508                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 509
 510                 if (*nr_effective < nr_replicas &&
 511                     test_bit(ob->dev, devs_may_alloc->d) &&
 512                     (ca->mi.durability ||
 513                      (wp->data_type == BCH_DATA_user && !*have_cache)) &&
 514                     (ob->ec || !need_ec)) {
 515                         add_new_bucket(c, ptrs, devs_may_alloc,
 516                                        nr_effective, have_cache,
 517                                        flags, ob);
 518                 } else {
 519                         ob_push(c, &ptrs_skip, ob);
 520                 }
 521         }
 522         wp->ptrs = ptrs_skip;
 523 }
 524
 525 static int open_bucket_add_buckets(struct bch_fs *c,
 526                         struct open_buckets *ptrs,
 527                         struct write_point *wp,
 528                         struct bch_devs_list *devs_have,
 529                         u16 target,
 530                         unsigned erasure_code,
 531                         unsigned nr_replicas,
 532                         unsigned *nr_effective,
 533                         bool *have_cache,
 534                         enum alloc_reserve reserve,
 535                         unsigned flags,
 536                         struct closure *_cl)
 537 {
 538         struct bch_devs_mask devs;
 539         struct open_bucket *ob;
 540         struct closure *cl = NULL;
 541         int ret;
 542         unsigned i;
 543
 544         rcu_read_lock();
 545         devs = target_rw_devs(c, wp->data_type, target);
 546         rcu_read_unlock();
 547
 548         /* Don't allocate from devices we already have pointers to: */
 549         for (i = 0; i < devs_have->nr; i++)
 550                 __clear_bit(devs_have->devs[i], devs.d);
 551
 552         open_bucket_for_each(c, ptrs, ob, i)
 553                 __clear_bit(ob->dev, devs.d);
 554
 555         if (erasure_code) {
 556                 if (!ec_open_bucket(c, ptrs)) {
 557                         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 558                                                     nr_replicas, nr_effective,
 559                                                     have_cache, flags, true);
 560                         if (*nr_effective >= nr_replicas)
 561                                 return 0;
 562                 }
 563
 564                 if (!ec_open_bucket(c, ptrs)) {
 565                         ret = bucket_alloc_from_stripe(c, ptrs, wp, &devs,
 566                                                  target, erasure_code,
 567                                                  nr_replicas, nr_effective,
 568                                                  have_cache, flags, _cl);
 569                         if (ret == -FREELIST_EMPTY ||
 570                             ret == -OPEN_BUCKETS_EMPTY)
 571                                 return ret;
 572                         if (*nr_effective >= nr_replicas)
 573                                 return 0;
 574                 }
 575         }
 576
 577         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 578                                     nr_replicas, nr_effective,
 579                                     have_cache, flags, false);
 580         if (*nr_effective >= nr_replicas)
 581                 return 0;
 582
 583         percpu_down_read(&c->mark_lock);
 584         rcu_read_lock();
 585
 586 retry_blocking:
 587         /*
 588          * Try nonblocking first, so that if one device is full we'll try from
 589          * other devices:
 590          */
 591         ret = bch2_bucket_alloc_set(c, ptrs, &wp->stripe, &devs,
 592                                 nr_replicas, nr_effective, have_cache,
 593                                 reserve, flags, cl);
 594         if (ret && ret != -INSUFFICIENT_DEVICES && !cl && _cl) {
 595                 cl = _cl;
 596                 goto retry_blocking;
 597         }
 598
 599         rcu_read_unlock();
 600         percpu_up_read(&c->mark_lock);
 601
 602         return ret;
 603 }
 604
 605 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 606                                 struct open_buckets *obs)
 607 {
 608         struct open_buckets ptrs = { .nr = 0 };
 609         struct open_bucket *ob, *ob2;
 610         unsigned i, j;
 611
 612         open_bucket_for_each(c, obs, ob, i) {
 613                 bool drop = !ca || ob->dev == ca->dev_idx;
 614
 615                 if (!drop && ob->ec) {
 616                         mutex_lock(&ob->ec->lock);
 617                         for (j = 0; j < ob->ec->new_stripe.key.v.nr_blocks; j++) {
 618                                 if (!ob->ec->blocks[j])
 619                                         continue;
 620
 621                                 ob2 = c->open_buckets + ob->ec->blocks[j];
 622                                 drop |= ob2->dev == ca->dev_idx;
 623                         }
 624                         mutex_unlock(&ob->ec->lock);
 625                 }
 626
 627                 if (drop)
 628                         bch2_open_bucket_put(c, ob);
 629                 else
 630                         ob_push(c, &ptrs, ob);
 631         }
 632
 633         *obs = ptrs;
 634 }
 635
 636 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 637                           struct write_point *wp)
 638 {
 639         mutex_lock(&wp->lock);
 640         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
 641         mutex_unlock(&wp->lock);
 642 }
 643
 644 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 645                                                  unsigned long write_point)
 646 {
 647         unsigned hash =
 648                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
 649
 650         return &c->write_points_hash[hash];
 651 }
 652
 653 static struct write_point *__writepoint_find(struct hlist_head *head,
 654                                              unsigned long write_point)
 655 {
 656         struct write_point *wp;
 657
 658         rcu_read_lock();
 659         hlist_for_each_entry_rcu(wp, head, node)
 660                 if (wp->write_point == write_point)
 661                         goto out;
 662         wp = NULL;
 663 out:
 664         rcu_read_unlock();
 665         return wp;
 666 }
 667
 668 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
 669 {
 670         u64 stranded    = c->write_points_nr * c->bucket_size_max;
 671         u64 free        = bch2_fs_usage_read_short(c).free;
 672
 673         return stranded * factor > free;
 674 }
 675
 676 static bool try_increase_writepoints(struct bch_fs *c)
 677 {
 678         struct write_point *wp;
 679
 680         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
 681             too_many_writepoints(c, 32))
 682                 return false;
 683
 684         wp = c->write_points + c->write_points_nr++;
 685         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
 686         return true;
 687 }
 688
 689 static bool try_decrease_writepoints(struct bch_fs *c,
 690                                      unsigned old_nr)
 691 {
 692         struct write_point *wp;
 693
 694         mutex_lock(&c->write_points_hash_lock);
 695         if (c->write_points_nr < old_nr) {
 696                 mutex_unlock(&c->write_points_hash_lock);
 697                 return true;
 698         }
 699
 700         if (c->write_points_nr == 1 ||
 701             !too_many_writepoints(c, 8)) {
 702                 mutex_unlock(&c->write_points_hash_lock);
 703                 return false;
 704         }
 705
 706         wp = c->write_points + --c->write_points_nr;
 707
 708         hlist_del_rcu(&wp->node);
 709         mutex_unlock(&c->write_points_hash_lock);
 710
 711         bch2_writepoint_stop(c, NULL, wp);
 712         return true;
 713 }
 714
 715 static struct write_point *writepoint_find(struct bch_fs *c,
 716                                            unsigned long write_point)
 717 {
 718         struct write_point *wp, *oldest;
 719         struct hlist_head *head;
 720
 721         if (!(write_point & 1UL)) {
 722                 wp = (struct write_point *) write_point;
 723                 mutex_lock(&wp->lock);
 724                 return wp;
 725         }
 726
 727         head = writepoint_hash(c, write_point);
 728 restart_find:
 729         wp = __writepoint_find(head, write_point);
 730         if (wp) {
 731 lock_wp:
 732                 mutex_lock(&wp->lock);
 733                 if (wp->write_point == write_point)
 734                         goto out;
 735                 mutex_unlock(&wp->lock);
 736                 goto restart_find;
 737         }
 738 restart_find_oldest:
 739         oldest = NULL;
 740         for (wp = c->write_points;
 741              wp < c->write_points + c->write_points_nr; wp++)
 742                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
 743                         oldest = wp;
 744
 745         mutex_lock(&oldest->lock);
 746         mutex_lock(&c->write_points_hash_lock);
 747         if (oldest >= c->write_points + c->write_points_nr ||
 748             try_increase_writepoints(c)) {
 749                 mutex_unlock(&c->write_points_hash_lock);
 750                 mutex_unlock(&oldest->lock);
 751                 goto restart_find_oldest;
 752         }
 753
 754         wp = __writepoint_find(head, write_point);
 755         if (wp && wp != oldest) {
 756                 mutex_unlock(&c->write_points_hash_lock);
 757                 mutex_unlock(&oldest->lock);
 758                 goto lock_wp;
 759         }
 760
 761         wp = oldest;
 762         hlist_del_rcu(&wp->node);
 763         wp->write_point = write_point;
 764         hlist_add_head_rcu(&wp->node, head);
 765         mutex_unlock(&c->write_points_hash_lock);
 766 out:
 767         wp->last_used = sched_clock();
 768         return wp;
 769 }
 770
 771 /*
 772  * Get us an open_bucket we can allocate from, return with it locked:
 773  */
 774 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
 775                                 unsigned target,
 776                                 unsigned erasure_code,
 777                                 struct write_point_specifier write_point,
 778                                 struct bch_devs_list *devs_have,
 779                                 unsigned nr_replicas,
 780                                 unsigned nr_replicas_required,
 781                                 enum alloc_reserve reserve,
 782                                 unsigned flags,
 783                                 struct closure *cl)
 784 {
 785         struct write_point *wp;
 786         struct open_bucket *ob;
 787         struct open_buckets ptrs;
 788         unsigned nr_effective, write_points_nr;
 789         unsigned ob_flags = 0;
 790         bool have_cache;
 791         int ret;
 792         int i;
 793
 794         if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
 795                 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
 796
 797         BUG_ON(!nr_replicas || !nr_replicas_required);
 798 retry:
 799         ptrs.nr         = 0;
 800         nr_effective    = 0;
 801         write_points_nr = c->write_points_nr;
 802         have_cache      = false;
 803
 804         wp = writepoint_find(c, write_point.v);
 805
 806         if (wp->data_type == BCH_DATA_user)
 807                 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
 808
 809         /* metadata may not allocate on cache devices: */
 810         if (wp->data_type != BCH_DATA_user)
 811                 have_cache = true;
 812
 813         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
 814                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 815                                               target, erasure_code,
 816                                               nr_replicas, &nr_effective,
 817                                               &have_cache, reserve,
 818                                               ob_flags, cl);
 819         } else {
 820                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 821                                               target, erasure_code,
 822                                               nr_replicas, &nr_effective,
 823                                               &have_cache, reserve,
 824                                               ob_flags, NULL);
 825                 if (!ret)
 826                         goto alloc_done;
 827
 828                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 829                                               0, erasure_code,
 830                                               nr_replicas, &nr_effective,
 831                                               &have_cache, reserve,
 832                                               ob_flags, cl);
 833         }
 834 alloc_done:
 835         BUG_ON(!ret && nr_effective < nr_replicas);
 836
 837         if (erasure_code && !ec_open_bucket(c, &ptrs))
 838                 pr_debug("failed to get ec bucket: ret %u", ret);
 839
 840         if (ret == -INSUFFICIENT_DEVICES &&
 841             nr_effective >= nr_replicas_required)
 842                 ret = 0;
 843
 844         if (ret)
 845                 goto err;
 846
 847         /* Free buckets we didn't use: */
 848         open_bucket_for_each(c, &wp->ptrs, ob, i)
 849                 open_bucket_free_unused(c, wp, ob);
 850
 851         wp->ptrs = ptrs;
 852
 853         wp->sectors_free = UINT_MAX;
 854
 855         open_bucket_for_each(c, &wp->ptrs, ob, i)
 856                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
 857
 858         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
 859
 860         verify_not_stale(c, &wp->ptrs);
 861
 862         return wp;
 863 err:
 864         open_bucket_for_each(c, &wp->ptrs, ob, i)
 865                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
 866                         ob_push(c, &ptrs, ob);
 867                 else
 868                         open_bucket_free_unused(c, wp, ob);
 869         wp->ptrs = ptrs;
 870
 871         mutex_unlock(&wp->lock);
 872
 873         if (ret == -FREELIST_EMPTY &&
 874             try_decrease_writepoints(c, write_points_nr))
 875                 goto retry;
 876
 877         switch (ret) {
 878         case -OPEN_BUCKETS_EMPTY:
 879         case -FREELIST_EMPTY:
 880                 return cl ? ERR_PTR(-EAGAIN) : ERR_PTR(-ENOSPC);
 881         case -INSUFFICIENT_DEVICES:
 882                 return ERR_PTR(-EROFS);
 883         default:
 884                 BUG();
 885         }
 886 }
 887
 888 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
 889 {
 890         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 891
 892         return (struct bch_extent_ptr) {
 893                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
 894                 .gen    = ob->gen,
 895                 .dev    = ob->dev,
 896                 .offset = bucket_to_sector(ca, ob->bucket) +
 897                         ca->mi.bucket_size -
 898                         ob->sectors_free,
 899         };
 900 }
 901
 902 /*
 903  * Append pointers to the space we just allocated to @k, and mark @sectors space
 904  * as allocated out of @ob
 905  */
 906 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
 907                                     struct bkey_i *k, unsigned sectors,
 908                                     bool cached)
 909
 910 {
 911         struct open_bucket *ob;
 912         unsigned i;
 913
 914         BUG_ON(sectors > wp->sectors_free);
 915         wp->sectors_free -= sectors;
 916
 917         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 918                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 919                 struct bch_extent_ptr ptr = bch2_ob_ptr(c, ob);
 920
 921                 ptr.cached = cached ||
 922                         (!ca->mi.durability &&
 923                          wp->data_type == BCH_DATA_user);
 924
 925                 bch2_bkey_append_ptr(k, ptr);
 926
 927                 BUG_ON(sectors > ob->sectors_free);
 928                 ob->sectors_free -= sectors;
 929         }
 930 }
 931
 932 /*
 933  * Append pointers to the space we just allocated to @k, and mark @sectors space
 934  * as allocated out of @ob
 935  */
 936 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
 937 {
 938         struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
 939         struct open_bucket *ob;
 940         unsigned i;
 941
 942         open_bucket_for_each(c, &wp->ptrs, ob, i)
 943                 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
 944         wp->ptrs = keep;
 945
 946         mutex_unlock(&wp->lock);
 947
 948         bch2_open_buckets_put(c, &ptrs);
 949 }
 950
 951 static inline void writepoint_init(struct write_point *wp,
 952                                    enum bch_data_type type)
 953 {
 954         mutex_init(&wp->lock);
 955         wp->data_type = type;
 956 }
 957
 958 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
 959 {
 960         struct open_bucket *ob;
 961         struct write_point *wp;
 962
 963         mutex_init(&c->write_points_hash_lock);
 964         c->write_points_nr = ARRAY_SIZE(c->write_points);
 965
 966         /* open bucket 0 is a sentinal NULL: */
 967         spin_lock_init(&c->open_buckets[0].lock);
 968
 969         for (ob = c->open_buckets + 1;
 970              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
 971                 spin_lock_init(&ob->lock);
 972                 c->open_buckets_nr_free++;
 973
 974                 ob->freelist = c->open_buckets_freelist;
 975                 c->open_buckets_freelist = ob - c->open_buckets;
 976         }
 977
 978         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
 979         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
 980         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
 981
 982         for (wp = c->write_points;
 983              wp < c->write_points + c->write_points_nr; wp++) {
 984                 writepoint_init(wp, BCH_DATA_user);
 985
 986                 wp->last_used   = sched_clock();
 987                 wp->write_point = (unsigned long) wp;
 988                 hlist_add_head_rcu(&wp->node,
 989                                    writepoint_hash(c, wp->write_point));
 990         }
 991 }
 992
 993 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
 994 {
 995         struct open_bucket *ob;
 996
 997         for (ob = c->open_buckets;
 998              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
 999              ob++) {
1000                 spin_lock(&ob->lock);
1001                 if (ob->valid && !ob->on_partial_list) {
1002                         pr_buf(out, "%zu ref %u type %s\n",
1003                                ob - c->open_buckets,
1004                                atomic_read(&ob->pin),
1005                                bch2_data_types[ob->data_type]);
1006                 }
1007                 spin_unlock(&ob->lock);
1008         }
1009
1010 }