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

   1 /*
   2  * Primary bucket allocation code
   3  *
   4  * Copyright 2012 Google, Inc.
   5  *
   6  * Allocation in bcache is done in terms of buckets:
   7  *
   8  * Each bucket has associated an 8 bit gen; this gen corresponds to the gen in
   9  * btree pointers - they must match for the pointer to be considered valid.
  10  *
  11  * Thus (assuming a bucket has no dirty data or metadata in it) we can reuse a
  12  * bucket simply by incrementing its gen.
  13  *
  14  * The gens (along with the priorities; it's really the gens are important but
  15  * the code is named as if it's the priorities) are written in an arbitrary list
  16  * of buckets on disk, with a pointer to them in the journal header.
  17  *
  18  * When we invalidate a bucket, we have to write its new gen to disk and wait
  19  * for that write to complete before we use it - otherwise after a crash we
  20  * could have pointers that appeared to be good but pointed to data that had
  21  * been overwritten.
  22  *
  23  * Since the gens and priorities are all stored contiguously on disk, we can
  24  * batch this up: We fill up the free_inc list with freshly invalidated buckets,
  25  * call prio_write(), and when prio_write() finishes we pull buckets off the
  26  * free_inc list and optionally discard them.
  27  *
  28  * free_inc isn't the only freelist - if it was, we'd often have to sleep while
  29  * priorities and gens were being written before we could allocate. c->free is a
  30  * smaller freelist, and buckets on that list are always ready to be used.
  31  *
  32  * If we've got discards enabled, that happens when a bucket moves from the
  33  * free_inc list to the free list.
  34  *
  35  * It's important to ensure that gens don't wrap around - with respect to
  36  * either the oldest gen in the btree or the gen on disk. This is quite
  37  * difficult to do in practice, but we explicitly guard against it anyways - if
  38  * a bucket is in danger of wrapping around we simply skip invalidating it that
  39  * time around, and we garbage collect or rewrite the priorities sooner than we
  40  * would have otherwise.
  41  *
  42  * bch2_bucket_alloc() allocates a single bucket from a specific device.
  43  *
  44  * bch2_bucket_alloc_set() allocates one or more buckets from different devices
  45  * in a given filesystem.
  46  *
  47  * invalidate_buckets() drives all the processes described above. It's called
  48  * from bch2_bucket_alloc() and a few other places that need to make sure free
  49  * buckets are ready.
  50  *
  51  * invalidate_buckets_(lru|fifo)() find buckets that are available to be
  52  * invalidated, and then invalidate them and stick them on the free_inc list -
  53  * in either lru or fifo order.
  54  */
  55
  56 #include "bcachefs.h"
  57 #include "alloc_background.h"
  58 #include "alloc_foreground.h"
  59 #include "btree_gc.h"
  60 #include "buckets.h"
  61 #include "clock.h"
  62 #include "debug.h"
  63 #include "disk_groups.h"
  64 #include "ec.h"
  65 #include "io.h"
  66
  67 #include <linux/math64.h>
  68 #include <linux/rculist.h>
  69 #include <linux/rcupdate.h>
  70 #include <trace/events/bcachefs.h>
  71
  72 enum bucket_alloc_ret {
  73         ALLOC_SUCCESS,
  74         OPEN_BUCKETS_EMPTY,
  75         FREELIST_EMPTY,         /* Allocator thread not keeping up */
  76 };
  77
  78 /*
  79  * Open buckets represent a bucket that's currently being allocated from.  They
  80  * serve two purposes:
  81  *
  82  *  - They track buckets that have been partially allocated, allowing for
  83  *    sub-bucket sized allocations - they're used by the sector allocator below
  84  *
  85  *  - They provide a reference to the buckets they own that mark and sweep GC
  86  *    can find, until the new allocation has a pointer to it inserted into the
  87  *    btree
  88  *
  89  * When allocating some space with the sector allocator, the allocation comes
  90  * with a reference to an open bucket - the caller is required to put that
  91  * reference _after_ doing the index update that makes its allocation reachable.
  92  */
  93
  94 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
  95 {
  96         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
  97
  98         if (ob->ec) {
  99                 bch2_ec_bucket_written(c, ob);
 100                 return;
 101         }
 102
 103         percpu_down_read_preempt_disable(&c->mark_lock);
 104         spin_lock(&ob->lock);
 105
 106         bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr),
 107                                false, gc_pos_alloc(c, ob), 0);
 108         ob->valid = false;
 109         ob->type = 0;
 110
 111         spin_unlock(&ob->lock);
 112         percpu_up_read_preempt_enable(&c->mark_lock);
 113
 114         spin_lock(&c->freelist_lock);
 115         ob->freelist = c->open_buckets_freelist;
 116         c->open_buckets_freelist = ob - c->open_buckets;
 117         c->open_buckets_nr_free++;
 118         spin_unlock(&c->freelist_lock);
 119
 120         closure_wake_up(&c->open_buckets_wait);
 121 }
 122
 123 void bch2_open_bucket_write_error(struct bch_fs *c,
 124                                   struct open_buckets *obs,
 125                                   unsigned dev)
 126 {
 127         struct open_bucket *ob;
 128         unsigned i;
 129
 130         open_bucket_for_each(c, obs, ob, i)
 131                 if (ob->ptr.dev == dev &&
 132                     ob->ec)
 133                         bch2_ec_bucket_cancel(c, ob);
 134 }
 135
 136 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
 137 {
 138         struct open_bucket *ob;
 139
 140         BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
 141
 142         ob = c->open_buckets + c->open_buckets_freelist;
 143         c->open_buckets_freelist = ob->freelist;
 144         atomic_set(&ob->pin, 1);
 145         ob->type = 0;
 146
 147         c->open_buckets_nr_free--;
 148         return ob;
 149 }
 150
 151 static void open_bucket_free_unused(struct bch_fs *c,
 152                                     struct open_bucket *ob,
 153                                     bool may_realloc)
 154 {
 155         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 156
 157         BUG_ON(ca->open_buckets_partial_nr >=
 158                ARRAY_SIZE(ca->open_buckets_partial));
 159
 160         if (ca->open_buckets_partial_nr <
 161             ARRAY_SIZE(ca->open_buckets_partial) &&
 162             may_realloc) {
 163                 spin_lock(&c->freelist_lock);
 164                 ob->on_partial_list = true;
 165                 ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
 166                         ob - c->open_buckets;
 167                 spin_unlock(&c->freelist_lock);
 168
 169                 closure_wake_up(&c->open_buckets_wait);
 170                 closure_wake_up(&c->freelist_wait);
 171         } else {
 172                 bch2_open_bucket_put(c, ob);
 173         }
 174 }
 175
 176 static void verify_not_stale(struct bch_fs *c, const struct open_buckets *obs)
 177 {
 178 #ifdef CONFIG_BCACHEFS_DEBUG
 179         struct open_bucket *ob;
 180         unsigned i;
 181
 182         open_bucket_for_each(c, obs, ob, i) {
 183                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 184
 185                 BUG_ON(ptr_stale(ca, &ob->ptr));
 186         }
 187 #endif
 188 }
 189
 190 /* _only_ for allocating the journal on a new device: */
 191 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
 192 {
 193         struct bucket_array *buckets;
 194         ssize_t b;
 195
 196         rcu_read_lock();
 197         buckets = bucket_array(ca);
 198
 199         for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++)
 200                 if (is_available_bucket(buckets->b[b].mark))
 201                         goto success;
 202         b = -1;
 203 success:
 204         rcu_read_unlock();
 205         return b;
 206 }
 207
 208 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
 209 {
 210         switch (reserve) {
 211         case RESERVE_ALLOC:
 212                 return 0;
 213         case RESERVE_BTREE:
 214                 return BTREE_NODE_OPEN_BUCKET_RESERVE;
 215         default:
 216                 return BTREE_NODE_OPEN_BUCKET_RESERVE * 2;
 217         }
 218 }
 219
 220 /**
 221  * bch_bucket_alloc - allocate a single bucket from a specific device
 222  *
 223  * Returns index of bucket on success, 0 on failure
 224  * */
 225 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 226                                       enum alloc_reserve reserve,
 227                                       bool may_alloc_partial,
 228                                       struct closure *cl)
 229 {
 230         struct bucket_array *buckets;
 231         struct open_bucket *ob;
 232         long bucket = 0;
 233
 234         spin_lock(&c->freelist_lock);
 235
 236         if (may_alloc_partial &&
 237             ca->open_buckets_partial_nr) {
 238                 ob = c->open_buckets +
 239                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
 240                 ob->on_partial_list = false;
 241                 spin_unlock(&c->freelist_lock);
 242                 return ob;
 243         }
 244
 245         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
 246                 if (cl)
 247                         closure_wait(&c->open_buckets_wait, cl);
 248                 spin_unlock(&c->freelist_lock);
 249                 trace_open_bucket_alloc_fail(ca, reserve);
 250                 return ERR_PTR(-OPEN_BUCKETS_EMPTY);
 251         }
 252
 253         if (likely(fifo_pop(&ca->free[RESERVE_NONE], bucket)))
 254                 goto out;
 255
 256         switch (reserve) {
 257         case RESERVE_ALLOC:
 258                 if (fifo_pop(&ca->free[RESERVE_BTREE], bucket))
 259                         goto out;
 260                 break;
 261         case RESERVE_BTREE:
 262                 if (fifo_used(&ca->free[RESERVE_BTREE]) * 2 >=
 263                     ca->free[RESERVE_BTREE].size &&
 264                     fifo_pop(&ca->free[RESERVE_BTREE], bucket))
 265                         goto out;
 266                 break;
 267         case RESERVE_MOVINGGC:
 268                 if (fifo_pop(&ca->free[RESERVE_MOVINGGC], bucket))
 269                         goto out;
 270                 break;
 271         default:
 272                 break;
 273         }
 274
 275         if (cl)
 276                 closure_wait(&c->freelist_wait, cl);
 277
 278         spin_unlock(&c->freelist_lock);
 279
 280         trace_bucket_alloc_fail(ca, reserve);
 281         return ERR_PTR(-FREELIST_EMPTY);
 282 out:
 283         verify_not_on_freelist(c, ca, bucket);
 284
 285         ob = bch2_open_bucket_alloc(c);
 286
 287         spin_lock(&ob->lock);
 288         buckets = bucket_array(ca);
 289
 290         ob->valid       = true;
 291         ob->sectors_free = ca->mi.bucket_size;
 292         ob->ptr         = (struct bch_extent_ptr) {
 293                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
 294                 .gen    = buckets->b[bucket].mark.gen,
 295                 .offset = bucket_to_sector(ca, bucket),
 296                 .dev    = ca->dev_idx,
 297         };
 298
 299         bucket_io_clock_reset(c, ca, bucket, READ);
 300         bucket_io_clock_reset(c, ca, bucket, WRITE);
 301         spin_unlock(&ob->lock);
 302
 303         spin_unlock(&c->freelist_lock);
 304
 305         bch2_wake_allocator(ca);
 306
 307         trace_bucket_alloc(ca, reserve);
 308         return ob;
 309 }
 310
 311 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 312                             unsigned l, unsigned r)
 313 {
 314         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 315                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 316 }
 317
 318 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 319
 320 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 321                                           struct dev_stripe_state *stripe,
 322                                           struct bch_devs_mask *devs)
 323 {
 324         struct dev_alloc_list ret = { .nr = 0 };
 325         struct bch_dev *ca;
 326         unsigned i;
 327
 328         for_each_member_device_rcu(ca, c, i, devs)
 329                 ret.devs[ret.nr++] = i;
 330
 331         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 332         return ret;
 333 }
 334
 335 void bch2_dev_stripe_increment(struct bch_fs *c, struct bch_dev *ca,
 336                                struct dev_stripe_state *stripe)
 337 {
 338         u64 *v = stripe->next_alloc + ca->dev_idx;
 339         u64 free_space = dev_buckets_free(c, ca);
 340         u64 free_space_inv = free_space
 341                 ? div64_u64(1ULL << 48, free_space)
 342                 : 1ULL << 48;
 343         u64 scale = *v / 4;
 344
 345         if (*v + free_space_inv >= *v)
 346                 *v += free_space_inv;
 347         else
 348                 *v = U64_MAX;
 349
 350         for (v = stripe->next_alloc;
 351              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 352                 *v = *v < scale ? 0 : *v - scale;
 353 }
 354
 355 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 356 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 357
 358 static int bch2_bucket_alloc_set(struct bch_fs *c,
 359                                  struct open_buckets *ptrs,
 360                                  struct dev_stripe_state *stripe,
 361                                  struct bch_devs_mask *devs_may_alloc,
 362                                  unsigned nr_replicas,
 363                                  unsigned *nr_effective,
 364                                  bool *have_cache,
 365                                  enum alloc_reserve reserve,
 366                                  unsigned flags,
 367                                  struct closure *cl)
 368 {
 369         struct dev_alloc_list devs_sorted =
 370                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 371         struct bch_dev *ca;
 372         bool alloc_failure = false;
 373         unsigned i, durability;
 374
 375         BUG_ON(*nr_effective >= nr_replicas);
 376
 377         for (i = 0; i < devs_sorted.nr; i++) {
 378                 struct open_bucket *ob;
 379
 380                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
 381                 if (!ca)
 382                         continue;
 383
 384                 if (!ca->mi.durability && *have_cache)
 385                         continue;
 386
 387                 ob = bch2_bucket_alloc(c, ca, reserve,
 388                                 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
 389                 if (IS_ERR(ob)) {
 390                         enum bucket_alloc_ret ret = -PTR_ERR(ob);
 391
 392                         WARN_ON(reserve == RESERVE_MOVINGGC &&
 393                                 ret != OPEN_BUCKETS_EMPTY);
 394
 395                         if (cl)
 396                                 return -EAGAIN;
 397                         if (ret == OPEN_BUCKETS_EMPTY)
 398                                 return -ENOSPC;
 399                         alloc_failure = true;
 400                         continue;
 401                 }
 402
 403                 durability = (flags & BUCKET_ALLOC_USE_DURABILITY)
 404                         ? ca->mi.durability : 1;
 405
 406                 __clear_bit(ca->dev_idx, devs_may_alloc->d);
 407                 *nr_effective   += durability;
 408                 *have_cache     |= !durability;
 409
 410                 ob_push(c, ptrs, ob);
 411
 412                 bch2_dev_stripe_increment(c, ca, stripe);
 413
 414                 if (*nr_effective >= nr_replicas)
 415                         return 0;
 416         }
 417
 418         return alloc_failure ? -ENOSPC : -EROFS;
 419 }
 420
 421 /* Allocate from stripes: */
 422
 423 /*
 424  * XXX: use a higher watermark for allocating open buckets here:
 425  */
 426 static int ec_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
 427 {
 428         struct bch_devs_mask devs;
 429         struct open_bucket *ob;
 430         unsigned i, nr_have = 0, nr_data =
 431                 min_t(unsigned, h->nr_active_devs,
 432                       EC_STRIPE_MAX) - h->redundancy;
 433         bool have_cache = true;
 434         int ret = 0;
 435
 436         BUG_ON(h->blocks.nr > nr_data);
 437         BUG_ON(h->parity.nr > h->redundancy);
 438
 439         devs = h->devs;
 440
 441         open_bucket_for_each(c, &h->parity, ob, i)
 442                 __clear_bit(ob->ptr.dev, devs.d);
 443         open_bucket_for_each(c, &h->blocks, ob, i)
 444                 __clear_bit(ob->ptr.dev, devs.d);
 445
 446         percpu_down_read_preempt_disable(&c->mark_lock);
 447         rcu_read_lock();
 448
 449         if (h->parity.nr < h->redundancy) {
 450                 nr_have = h->parity.nr;
 451
 452                 ret = bch2_bucket_alloc_set(c, &h->parity,
 453                                             &h->parity_stripe,
 454                                             &devs,
 455                                             h->redundancy,
 456                                             &nr_have,
 457                                             &have_cache,
 458                                             RESERVE_NONE,
 459                                             0,
 460                                             NULL);
 461                 if (ret)
 462                         goto err;
 463         }
 464
 465         if (h->blocks.nr < nr_data) {
 466                 nr_have = h->blocks.nr;
 467
 468                 ret = bch2_bucket_alloc_set(c, &h->blocks,
 469                                             &h->block_stripe,
 470                                             &devs,
 471                                             nr_data,
 472                                             &nr_have,
 473                                             &have_cache,
 474                                             RESERVE_NONE,
 475                                             0,
 476                                             NULL);
 477                 if (ret)
 478                         goto err;
 479         }
 480
 481         rcu_read_unlock();
 482         percpu_up_read_preempt_enable(&c->mark_lock);
 483
 484         return bch2_ec_stripe_new_alloc(c, h);
 485 err:
 486         rcu_read_unlock();
 487         percpu_up_read_preempt_enable(&c->mark_lock);
 488         return -1;
 489 }
 490
 491 /*
 492  * if we can't allocate a new stripe because there are already too many
 493  * partially filled stripes, force allocating from an existing stripe even when
 494  * it's to a device we don't want:
 495  */
 496
 497 static void bucket_alloc_from_stripe(struct bch_fs *c,
 498                                      struct open_buckets *ptrs,
 499                                      struct write_point *wp,
 500                                      struct bch_devs_mask *devs_may_alloc,
 501                                      u16 target,
 502                                      unsigned erasure_code,
 503                                      unsigned nr_replicas,
 504                                      unsigned *nr_effective,
 505                                      bool *have_cache)
 506 {
 507         struct dev_alloc_list devs_sorted;
 508         struct ec_stripe_head *h;
 509         struct open_bucket *ob;
 510         struct bch_dev *ca;
 511         unsigned i, ec_idx;
 512
 513         if (!erasure_code)
 514                 return;
 515
 516         if (nr_replicas < 2)
 517                 return;
 518
 519         if (ec_open_bucket(c, ptrs))
 520                 return;
 521
 522         h = bch2_ec_stripe_head_get(c, target, erasure_code, nr_replicas - 1);
 523         if (!h)
 524                 return;
 525
 526         if (!h->s && ec_stripe_alloc(c, h))
 527                 goto out_put_head;
 528
 529         rcu_read_lock();
 530         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 531         rcu_read_unlock();
 532
 533         for (i = 0; i < devs_sorted.nr; i++)
 534                 open_bucket_for_each(c, &h->s->blocks, ob, ec_idx)
 535                         if (ob->ptr.dev == devs_sorted.devs[i] &&
 536                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 537                                 goto got_bucket;
 538         goto out_put_head;
 539 got_bucket:
 540         ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 541
 542         ob->ec_idx      = ec_idx;
 543         ob->ec          = h->s;
 544
 545         __clear_bit(ob->ptr.dev, devs_may_alloc->d);
 546         *nr_effective   += ca->mi.durability;
 547         *have_cache     |= !ca->mi.durability;
 548
 549         ob_push(c, ptrs, ob);
 550         atomic_inc(&h->s->pin);
 551 out_put_head:
 552         bch2_ec_stripe_head_put(h);
 553 }
 554
 555 /* Sector allocator */
 556
 557 static void get_buckets_from_writepoint(struct bch_fs *c,
 558                                         struct open_buckets *ptrs,
 559                                         struct write_point *wp,
 560                                         struct bch_devs_mask *devs_may_alloc,
 561                                         unsigned nr_replicas,
 562                                         unsigned *nr_effective,
 563                                         bool *have_cache,
 564                                         bool need_ec)
 565 {
 566         struct open_buckets ptrs_skip = { .nr = 0 };
 567         struct open_bucket *ob;
 568         unsigned i;
 569
 570         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 571                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 572
 573                 if (*nr_effective < nr_replicas &&
 574                     test_bit(ob->ptr.dev, devs_may_alloc->d) &&
 575                     (ca->mi.durability ||
 576                      (wp->type == BCH_DATA_USER && !*have_cache)) &&
 577                     (ob->ec || !need_ec)) {
 578                         __clear_bit(ob->ptr.dev, devs_may_alloc->d);
 579                         *nr_effective   += ca->mi.durability;
 580                         *have_cache     |= !ca->mi.durability;
 581
 582                         ob_push(c, ptrs, ob);
 583                 } else {
 584                         ob_push(c, &ptrs_skip, ob);
 585                 }
 586         }
 587         wp->ptrs = ptrs_skip;
 588 }
 589
 590 static int open_bucket_add_buckets(struct bch_fs *c,
 591                                    struct open_buckets *ptrs,
 592                                    struct write_point *wp,
 593                                    struct bch_devs_list *devs_have,
 594                                    u16 target,
 595                                    unsigned erasure_code,
 596                                    unsigned nr_replicas,
 597                                    unsigned *nr_effective,
 598                                    bool *have_cache,
 599                                    enum alloc_reserve reserve,
 600                                    struct closure *_cl)
 601 {
 602         struct bch_devs_mask devs;
 603         struct open_bucket *ob;
 604         struct closure *cl = NULL;
 605         unsigned i, flags = BUCKET_ALLOC_USE_DURABILITY;
 606         int ret;
 607
 608         if (wp->type == BCH_DATA_USER)
 609                 flags |= BUCKET_MAY_ALLOC_PARTIAL;
 610
 611         rcu_read_lock();
 612         devs = target_rw_devs(c, wp->type, target);
 613         rcu_read_unlock();
 614
 615         /* Don't allocate from devices we already have pointers to: */
 616         for (i = 0; i < devs_have->nr; i++)
 617                 __clear_bit(devs_have->devs[i], devs.d);
 618
 619         open_bucket_for_each(c, ptrs, ob, i)
 620                 __clear_bit(ob->ptr.dev, devs.d);
 621
 622         if (erasure_code) {
 623                 get_buckets_from_writepoint(c, ptrs, wp, &devs,
 624                                             nr_replicas, nr_effective,
 625                                             have_cache, true);
 626                 if (*nr_effective >= nr_replicas)
 627                         return 0;
 628
 629                 bucket_alloc_from_stripe(c, ptrs, wp, &devs,
 630                                          target, erasure_code,
 631                                          nr_replicas, nr_effective,
 632                                          have_cache);
 633                 if (*nr_effective >= nr_replicas)
 634                         return 0;
 635         }
 636
 637         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 638                                     nr_replicas, nr_effective,
 639                                     have_cache, false);
 640         if (*nr_effective >= nr_replicas)
 641                 return 0;
 642
 643         percpu_down_read_preempt_disable(&c->mark_lock);
 644         rcu_read_lock();
 645
 646 retry_blocking:
 647         /*
 648          * Try nonblocking first, so that if one device is full we'll try from
 649          * other devices:
 650          */
 651         ret = bch2_bucket_alloc_set(c, ptrs, &wp->stripe, &devs,
 652                                 nr_replicas, nr_effective, have_cache,
 653                                 reserve, flags, cl);
 654         if (ret && ret != -EROFS && !cl && _cl) {
 655                 cl = _cl;
 656                 goto retry_blocking;
 657         }
 658
 659         rcu_read_unlock();
 660         percpu_up_read_preempt_enable(&c->mark_lock);
 661
 662         return ret;
 663 }
 664
 665 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 666                                 struct open_buckets *obs,
 667                                 enum bch_data_type data_type)
 668 {
 669         struct open_buckets ptrs = { .nr = 0 };
 670         struct open_bucket *ob, *ob2;
 671         unsigned i, j;
 672
 673         open_bucket_for_each(c, obs, ob, i) {
 674                 bool drop = !ca || ob->ptr.dev == ca->dev_idx;
 675
 676                 if (!drop && ob->ec) {
 677                         mutex_lock(&ob->ec->lock);
 678                         open_bucket_for_each(c, &ob->ec->blocks, ob2, j)
 679                                 drop |= ob2->ptr.dev == ca->dev_idx;
 680                         open_bucket_for_each(c, &ob->ec->parity, ob2, j)
 681                                 drop |= ob2->ptr.dev == ca->dev_idx;
 682                         mutex_unlock(&ob->ec->lock);
 683                 }
 684
 685                 if (drop)
 686                         bch2_open_bucket_put(c, ob);
 687                 else
 688                         ob_push(c, &ptrs, ob);
 689         }
 690
 691         *obs = ptrs;
 692 }
 693
 694 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 695                           struct write_point *wp)
 696 {
 697         mutex_lock(&wp->lock);
 698         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs, wp->type);
 699         mutex_unlock(&wp->lock);
 700 }
 701
 702 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 703                                                  unsigned long write_point)
 704 {
 705         unsigned hash =
 706                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
 707
 708         return &c->write_points_hash[hash];
 709 }
 710
 711 static struct write_point *__writepoint_find(struct hlist_head *head,
 712                                              unsigned long write_point)
 713 {
 714         struct write_point *wp;
 715
 716         hlist_for_each_entry_rcu(wp, head, node)
 717                 if (wp->write_point == write_point)
 718                         return wp;
 719
 720         return NULL;
 721 }
 722
 723 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
 724 {
 725         u64 stranded    = c->write_points_nr * c->bucket_size_max;
 726         u64 free        = bch2_fs_usage_read_short(c).free;
 727
 728         return stranded * factor > free;
 729 }
 730
 731 static bool try_increase_writepoints(struct bch_fs *c)
 732 {
 733         struct write_point *wp;
 734
 735         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
 736             too_many_writepoints(c, 32))
 737                 return false;
 738
 739         wp = c->write_points + c->write_points_nr++;
 740         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
 741         return true;
 742 }
 743
 744 static bool try_decrease_writepoints(struct bch_fs *c,
 745                                      unsigned old_nr)
 746 {
 747         struct write_point *wp;
 748
 749         mutex_lock(&c->write_points_hash_lock);
 750         if (c->write_points_nr < old_nr) {
 751                 mutex_unlock(&c->write_points_hash_lock);
 752                 return true;
 753         }
 754
 755         if (c->write_points_nr == 1 ||
 756             !too_many_writepoints(c, 8)) {
 757                 mutex_unlock(&c->write_points_hash_lock);
 758                 return false;
 759         }
 760
 761         wp = c->write_points + --c->write_points_nr;
 762
 763         hlist_del_rcu(&wp->node);
 764         mutex_unlock(&c->write_points_hash_lock);
 765
 766         bch2_writepoint_stop(c, NULL, wp);
 767         return true;
 768 }
 769
 770 static struct write_point *writepoint_find(struct bch_fs *c,
 771                                            unsigned long write_point)
 772 {
 773         struct write_point *wp, *oldest;
 774         struct hlist_head *head;
 775
 776         if (!(write_point & 1UL)) {
 777                 wp = (struct write_point *) write_point;
 778                 mutex_lock(&wp->lock);
 779                 return wp;
 780         }
 781
 782         head = writepoint_hash(c, write_point);
 783 restart_find:
 784         wp = __writepoint_find(head, write_point);
 785         if (wp) {
 786 lock_wp:
 787                 mutex_lock(&wp->lock);
 788                 if (wp->write_point == write_point)
 789                         goto out;
 790                 mutex_unlock(&wp->lock);
 791                 goto restart_find;
 792         }
 793 restart_find_oldest:
 794         oldest = NULL;
 795         for (wp = c->write_points;
 796              wp < c->write_points + c->write_points_nr; wp++)
 797                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
 798                         oldest = wp;
 799
 800         mutex_lock(&oldest->lock);
 801         mutex_lock(&c->write_points_hash_lock);
 802         if (oldest >= c->write_points + c->write_points_nr ||
 803             try_increase_writepoints(c)) {
 804                 mutex_unlock(&c->write_points_hash_lock);
 805                 mutex_unlock(&oldest->lock);
 806                 goto restart_find_oldest;
 807         }
 808
 809         wp = __writepoint_find(head, write_point);
 810         if (wp && wp != oldest) {
 811                 mutex_unlock(&c->write_points_hash_lock);
 812                 mutex_unlock(&oldest->lock);
 813                 goto lock_wp;
 814         }
 815
 816         wp = oldest;
 817         hlist_del_rcu(&wp->node);
 818         wp->write_point = write_point;
 819         hlist_add_head_rcu(&wp->node, head);
 820         mutex_unlock(&c->write_points_hash_lock);
 821 out:
 822         wp->last_used = sched_clock();
 823         return wp;
 824 }
 825
 826 /*
 827  * Get us an open_bucket we can allocate from, return with it locked:
 828  */
 829 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
 830                                 unsigned target,
 831                                 unsigned erasure_code,
 832                                 struct write_point_specifier write_point,
 833                                 struct bch_devs_list *devs_have,
 834                                 unsigned nr_replicas,
 835                                 unsigned nr_replicas_required,
 836                                 enum alloc_reserve reserve,
 837                                 unsigned flags,
 838                                 struct closure *cl)
 839 {
 840         struct write_point *wp;
 841         struct open_bucket *ob;
 842         struct open_buckets ptrs;
 843         unsigned nr_effective, write_points_nr;
 844         bool have_cache;
 845         int ret, i;
 846
 847         BUG_ON(!nr_replicas || !nr_replicas_required);
 848 retry:
 849         ptrs.nr         = 0;
 850         nr_effective    = 0;
 851         write_points_nr = c->write_points_nr;
 852         have_cache      = false;
 853
 854         wp = writepoint_find(c, write_point.v);
 855
 856         /* metadata may not allocate on cache devices: */
 857         if (wp->type != BCH_DATA_USER)
 858                 have_cache = true;
 859
 860         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
 861                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 862                                               target, erasure_code,
 863                                               nr_replicas, &nr_effective,
 864                                               &have_cache, reserve, cl);
 865         } else {
 866                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 867                                               target, erasure_code,
 868                                               nr_replicas, &nr_effective,
 869                                               &have_cache, reserve, NULL);
 870                 if (!ret)
 871                         goto alloc_done;
 872
 873                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 874                                               0, erasure_code,
 875                                               nr_replicas, &nr_effective,
 876                                               &have_cache, reserve, cl);
 877         }
 878 alloc_done:
 879         BUG_ON(!ret && nr_effective < nr_replicas);
 880
 881         if (erasure_code && !ec_open_bucket(c, &ptrs))
 882                 pr_debug("failed to get ec bucket: ret %u", ret);
 883
 884         if (ret == -EROFS &&
 885             nr_effective >= nr_replicas_required)
 886                 ret = 0;
 887
 888         if (ret)
 889                 goto err;
 890
 891         /* Free buckets we didn't use: */
 892         open_bucket_for_each(c, &wp->ptrs, ob, i)
 893                 open_bucket_free_unused(c, ob, wp->type == BCH_DATA_USER);
 894
 895         wp->ptrs = ptrs;
 896
 897         wp->sectors_free = UINT_MAX;
 898
 899         open_bucket_for_each(c, &wp->ptrs, ob, i)
 900                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
 901
 902         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
 903
 904         verify_not_stale(c, &wp->ptrs);
 905
 906         return wp;
 907 err:
 908         open_bucket_for_each(c, &wp->ptrs, ob, i)
 909                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
 910                         ob_push(c, &ptrs, ob);
 911                 else
 912                         open_bucket_free_unused(c, ob,
 913                                         wp->type == BCH_DATA_USER);
 914         wp->ptrs = ptrs;
 915
 916         mutex_unlock(&wp->lock);
 917
 918         if (ret == -ENOSPC &&
 919             try_decrease_writepoints(c, write_points_nr))
 920                 goto retry;
 921
 922         return ERR_PTR(ret);
 923 }
 924
 925 /*
 926  * Append pointers to the space we just allocated to @k, and mark @sectors space
 927  * as allocated out of @ob
 928  */
 929 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
 930                                     struct bkey_i *k, unsigned sectors)
 931
 932 {
 933         struct open_bucket *ob;
 934         unsigned i;
 935
 936         BUG_ON(sectors > wp->sectors_free);
 937         wp->sectors_free -= sectors;
 938
 939         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 940                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 941                 struct bch_extent_ptr tmp = ob->ptr;
 942
 943                 tmp.cached = !ca->mi.durability &&
 944                         wp->type == BCH_DATA_USER;
 945
 946                 tmp.offset += ca->mi.bucket_size - ob->sectors_free;
 947                 bch2_bkey_append_ptr(k, tmp);
 948
 949                 BUG_ON(sectors > ob->sectors_free);
 950                 ob->sectors_free -= sectors;
 951         }
 952 }
 953
 954 /*
 955  * Append pointers to the space we just allocated to @k, and mark @sectors space
 956  * as allocated out of @ob
 957  */
 958 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
 959 {
 960         struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
 961         struct open_bucket *ob;
 962         unsigned i;
 963
 964         open_bucket_for_each(c, &wp->ptrs, ob, i)
 965                 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
 966         wp->ptrs = keep;
 967
 968         mutex_unlock(&wp->lock);
 969
 970         bch2_open_buckets_put(c, &ptrs);
 971 }
 972
 973 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
 974 {
 975         struct open_bucket *ob;
 976         struct write_point *wp;
 977
 978         mutex_init(&c->write_points_hash_lock);
 979         c->write_points_nr = ARRAY_SIZE(c->write_points);
 980
 981         /* open bucket 0 is a sentinal NULL: */
 982         spin_lock_init(&c->open_buckets[0].lock);
 983
 984         for (ob = c->open_buckets + 1;
 985              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
 986                 spin_lock_init(&ob->lock);
 987                 c->open_buckets_nr_free++;
 988
 989                 ob->freelist = c->open_buckets_freelist;
 990                 c->open_buckets_freelist = ob - c->open_buckets;
 991         }
 992
 993         writepoint_init(&c->btree_write_point, BCH_DATA_BTREE);
 994         writepoint_init(&c->rebalance_write_point, BCH_DATA_USER);
 995
 996         for (wp = c->write_points;
 997              wp < c->write_points + c->write_points_nr; wp++) {
 998                 writepoint_init(wp, BCH_DATA_USER);
 999
1000                 wp->last_used   = sched_clock();
1001                 wp->write_point = (unsigned long) wp;
1002                 hlist_add_head_rcu(&wp->node,
1003                                    writepoint_hash(c, wp->write_point));
1004         }
1005 }