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
 110         spin_unlock(&ob->lock);
 111         percpu_up_read_preempt_enable(&c->mark_lock);
 112
 113         spin_lock(&c->freelist_lock);
 114         ob->freelist = c->open_buckets_freelist;
 115         c->open_buckets_freelist = ob - c->open_buckets;
 116         c->open_buckets_nr_free++;
 117         spin_unlock(&c->freelist_lock);
 118
 119         closure_wake_up(&c->open_buckets_wait);
 120 }
 121
 122 void bch2_open_bucket_write_error(struct bch_fs *c,
 123                                   struct open_buckets *obs,
 124                                   unsigned dev)
 125 {
 126         struct open_bucket *ob;
 127         unsigned i;
 128
 129         open_bucket_for_each(c, obs, ob, i)
 130                 if (ob->ptr.dev == dev &&
 131                     ob->ec)
 132                         bch2_ec_bucket_cancel(c, ob);
 133 }
 134
 135 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
 136 {
 137         struct open_bucket *ob;
 138
 139         BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
 140
 141         ob = c->open_buckets + c->open_buckets_freelist;
 142         c->open_buckets_freelist = ob->freelist;
 143         atomic_set(&ob->pin, 1);
 144
 145         c->open_buckets_nr_free--;
 146         return ob;
 147 }
 148
 149 static void open_bucket_free_unused(struct bch_fs *c,
 150                                     struct open_bucket *ob,
 151                                     bool may_realloc)
 152 {
 153         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 154
 155         BUG_ON(ca->open_buckets_partial_nr >=
 156                ARRAY_SIZE(ca->open_buckets_partial));
 157
 158         if (ca->open_buckets_partial_nr <
 159             ARRAY_SIZE(ca->open_buckets_partial) &&
 160             may_realloc) {
 161                 spin_lock(&c->freelist_lock);
 162                 ob->on_partial_list = true;
 163                 ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
 164                         ob - c->open_buckets;
 165                 spin_unlock(&c->freelist_lock);
 166
 167                 closure_wake_up(&c->open_buckets_wait);
 168                 closure_wake_up(&c->freelist_wait);
 169         } else {
 170                 bch2_open_bucket_put(c, ob);
 171         }
 172 }
 173
 174 static void verify_not_stale(struct bch_fs *c, const struct open_buckets *obs)
 175 {
 176 #ifdef CONFIG_BCACHEFS_DEBUG
 177         struct open_bucket *ob;
 178         unsigned i;
 179
 180         open_bucket_for_each(c, obs, ob, i) {
 181                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 182
 183                 BUG_ON(ptr_stale(ca, &ob->ptr));
 184         }
 185 #endif
 186 }
 187
 188 /* _only_ for allocating the journal on a new device: */
 189 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
 190 {
 191         struct bucket_array *buckets;
 192         ssize_t b;
 193
 194         rcu_read_lock();
 195         buckets = bucket_array(ca);
 196
 197         for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++)
 198                 if (is_available_bucket(buckets->b[b].mark))
 199                         goto success;
 200         b = -1;
 201 success:
 202         rcu_read_unlock();
 203         return b;
 204 }
 205
 206 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
 207 {
 208         switch (reserve) {
 209         case RESERVE_ALLOC:
 210                 return 0;
 211         case RESERVE_BTREE:
 212                 return BTREE_NODE_RESERVE / 2;
 213         default:
 214                 return BTREE_NODE_RESERVE;
 215         }
 216 }
 217
 218 /**
 219  * bch_bucket_alloc - allocate a single bucket from a specific device
 220  *
 221  * Returns index of bucket on success, 0 on failure
 222  * */
 223 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 224                                       enum alloc_reserve reserve,
 225                                       bool may_alloc_partial,
 226                                       struct closure *cl)
 227 {
 228         struct bucket_array *buckets;
 229         struct open_bucket *ob;
 230         long bucket = 0;
 231
 232         spin_lock(&c->freelist_lock);
 233
 234         if (may_alloc_partial &&
 235             ca->open_buckets_partial_nr) {
 236                 ob = c->open_buckets +
 237                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
 238                 ob->on_partial_list = false;
 239                 spin_unlock(&c->freelist_lock);
 240                 return ob;
 241         }
 242
 243         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
 244                 if (cl)
 245                         closure_wait(&c->open_buckets_wait, cl);
 246                 spin_unlock(&c->freelist_lock);
 247                 trace_open_bucket_alloc_fail(ca, reserve);
 248                 return ERR_PTR(-OPEN_BUCKETS_EMPTY);
 249         }
 250
 251         if (likely(fifo_pop(&ca->free[RESERVE_NONE], bucket)))
 252                 goto out;
 253
 254         switch (reserve) {
 255         case RESERVE_ALLOC:
 256                 if (fifo_pop(&ca->free[RESERVE_BTREE], bucket))
 257                         goto out;
 258                 break;
 259         case RESERVE_BTREE:
 260                 if (fifo_used(&ca->free[RESERVE_BTREE]) * 2 >=
 261                     ca->free[RESERVE_BTREE].size &&
 262                     fifo_pop(&ca->free[RESERVE_BTREE], bucket))
 263                         goto out;
 264                 break;
 265         case RESERVE_MOVINGGC:
 266                 if (fifo_pop(&ca->free[RESERVE_MOVINGGC], bucket))
 267                         goto out;
 268                 break;
 269         default:
 270                 break;
 271         }
 272
 273         if (cl)
 274                 closure_wait(&c->freelist_wait, cl);
 275
 276         spin_unlock(&c->freelist_lock);
 277
 278         trace_bucket_alloc_fail(ca, reserve);
 279         return ERR_PTR(-FREELIST_EMPTY);
 280 out:
 281         verify_not_on_freelist(c, ca, bucket);
 282
 283         ob = bch2_open_bucket_alloc(c);
 284
 285         spin_lock(&ob->lock);
 286         buckets = bucket_array(ca);
 287
 288         ob->valid       = true;
 289         ob->sectors_free = ca->mi.bucket_size;
 290         ob->ptr         = (struct bch_extent_ptr) {
 291                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
 292                 .gen    = buckets->b[bucket].mark.gen,
 293                 .offset = bucket_to_sector(ca, bucket),
 294                 .dev    = ca->dev_idx,
 295         };
 296
 297         bucket_io_clock_reset(c, ca, bucket, READ);
 298         bucket_io_clock_reset(c, ca, bucket, WRITE);
 299         spin_unlock(&ob->lock);
 300
 301         spin_unlock(&c->freelist_lock);
 302
 303         bch2_wake_allocator(ca);
 304
 305         trace_bucket_alloc(ca, reserve);
 306         return ob;
 307 }
 308
 309 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 310                             unsigned l, unsigned r)
 311 {
 312         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 313                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 314 }
 315
 316 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 317
 318 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 319                                           struct dev_stripe_state *stripe,
 320                                           struct bch_devs_mask *devs)
 321 {
 322         struct dev_alloc_list ret = { .nr = 0 };
 323         struct bch_dev *ca;
 324         unsigned i;
 325
 326         for_each_member_device_rcu(ca, c, i, devs)
 327                 ret.devs[ret.nr++] = i;
 328
 329         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 330         return ret;
 331 }
 332
 333 void bch2_dev_stripe_increment(struct bch_fs *c, struct bch_dev *ca,
 334                                struct dev_stripe_state *stripe)
 335 {
 336         u64 *v = stripe->next_alloc + ca->dev_idx;
 337         u64 free_space = dev_buckets_free(c, ca);
 338         u64 free_space_inv = free_space
 339                 ? div64_u64(1ULL << 48, free_space)
 340                 : 1ULL << 48;
 341         u64 scale = *v / 4;
 342
 343         if (*v + free_space_inv >= *v)
 344                 *v += free_space_inv;
 345         else
 346                 *v = U64_MAX;
 347
 348         for (v = stripe->next_alloc;
 349              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 350                 *v = *v < scale ? 0 : *v - scale;
 351 }
 352
 353 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 354 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 355
 356 static int bch2_bucket_alloc_set(struct bch_fs *c,
 357                                  struct open_buckets *ptrs,
 358                                  struct dev_stripe_state *stripe,
 359                                  struct bch_devs_mask *devs_may_alloc,
 360                                  unsigned nr_replicas,
 361                                  unsigned *nr_effective,
 362                                  bool *have_cache,
 363                                  enum alloc_reserve reserve,
 364                                  unsigned flags,
 365                                  struct closure *cl)
 366 {
 367         struct dev_alloc_list devs_sorted =
 368                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 369         struct bch_dev *ca;
 370         bool alloc_failure = false;
 371         unsigned i, durability;
 372
 373         BUG_ON(*nr_effective >= nr_replicas);
 374
 375         for (i = 0; i < devs_sorted.nr; i++) {
 376                 struct open_bucket *ob;
 377
 378                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
 379                 if (!ca)
 380                         continue;
 381
 382                 if (!ca->mi.durability && *have_cache)
 383                         continue;
 384
 385                 ob = bch2_bucket_alloc(c, ca, reserve,
 386                                 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
 387                 if (IS_ERR(ob)) {
 388                         enum bucket_alloc_ret ret = -PTR_ERR(ob);
 389
 390                         WARN_ON(reserve == RESERVE_MOVINGGC &&
 391                                 ret != OPEN_BUCKETS_EMPTY);
 392
 393                         if (cl)
 394                                 return -EAGAIN;
 395                         if (ret == OPEN_BUCKETS_EMPTY)
 396                                 return -ENOSPC;
 397                         alloc_failure = true;
 398                         continue;
 399                 }
 400
 401                 durability = (flags & BUCKET_ALLOC_USE_DURABILITY)
 402                         ? ca->mi.durability : 1;
 403
 404                 __clear_bit(ca->dev_idx, devs_may_alloc->d);
 405                 *nr_effective   += durability;
 406                 *have_cache     |= !durability;
 407
 408                 ob_push(c, ptrs, ob);
 409
 410                 bch2_dev_stripe_increment(c, ca, stripe);
 411
 412                 if (*nr_effective >= nr_replicas)
 413                         return 0;
 414         }
 415
 416         return alloc_failure ? -ENOSPC : -EROFS;
 417 }
 418
 419 /* Allocate from stripes: */
 420
 421 /*
 422  * XXX: use a higher watermark for allocating open buckets here:
 423  */
 424 static int ec_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
 425 {
 426         struct bch_devs_mask devs;
 427         struct open_bucket *ob;
 428         unsigned i, nr_have = 0, nr_data =
 429                 min_t(unsigned, h->nr_active_devs,
 430                       EC_STRIPE_MAX) - h->redundancy;
 431         bool have_cache = true;
 432         int ret = 0;
 433
 434         BUG_ON(h->blocks.nr > nr_data);
 435         BUG_ON(h->parity.nr > h->redundancy);
 436
 437         devs = h->devs;
 438
 439         open_bucket_for_each(c, &h->parity, ob, i)
 440                 __clear_bit(ob->ptr.dev, devs.d);
 441         open_bucket_for_each(c, &h->blocks, ob, i)
 442                 __clear_bit(ob->ptr.dev, devs.d);
 443
 444         percpu_down_read_preempt_disable(&c->mark_lock);
 445         rcu_read_lock();
 446
 447         if (h->parity.nr < h->redundancy) {
 448                 nr_have = h->parity.nr;
 449
 450                 ret = bch2_bucket_alloc_set(c, &h->parity,
 451                                             &h->parity_stripe,
 452                                             &devs,
 453                                             h->redundancy,
 454                                             &nr_have,
 455                                             &have_cache,
 456                                             RESERVE_NONE,
 457                                             0,
 458                                             NULL);
 459                 if (ret)
 460                         goto err;
 461         }
 462
 463         if (h->blocks.nr < nr_data) {
 464                 nr_have = h->blocks.nr;
 465
 466                 ret = bch2_bucket_alloc_set(c, &h->blocks,
 467                                             &h->block_stripe,
 468                                             &devs,
 469                                             nr_data,
 470                                             &nr_have,
 471                                             &have_cache,
 472                                             RESERVE_NONE,
 473                                             0,
 474                                             NULL);
 475                 if (ret)
 476                         goto err;
 477         }
 478
 479         rcu_read_unlock();
 480         percpu_up_read_preempt_enable(&c->mark_lock);
 481
 482         return bch2_ec_stripe_new_alloc(c, h);
 483 err:
 484         rcu_read_unlock();
 485         percpu_up_read_preempt_enable(&c->mark_lock);
 486         return -1;
 487 }
 488
 489 /*
 490  * if we can't allocate a new stripe because there are already too many
 491  * partially filled stripes, force allocating from an existing stripe even when
 492  * it's to a device we don't want:
 493  */
 494
 495 static void bucket_alloc_from_stripe(struct bch_fs *c,
 496                                      struct open_buckets *ptrs,
 497                                      struct write_point *wp,
 498                                      struct bch_devs_mask *devs_may_alloc,
 499                                      u16 target,
 500                                      unsigned erasure_code,
 501                                      unsigned nr_replicas,
 502                                      unsigned *nr_effective,
 503                                      bool *have_cache)
 504 {
 505         struct dev_alloc_list devs_sorted;
 506         struct ec_stripe_head *h;
 507         struct open_bucket *ob;
 508         struct bch_dev *ca;
 509         unsigned i, ec_idx;
 510
 511         if (!erasure_code)
 512                 return;
 513
 514         if (nr_replicas < 2)
 515                 return;
 516
 517         if (ec_open_bucket(c, ptrs))
 518                 return;
 519
 520         h = bch2_ec_stripe_head_get(c, target, erasure_code, nr_replicas - 1);
 521         if (!h)
 522                 return;
 523
 524         if (!h->s && ec_stripe_alloc(c, h))
 525                 goto out_put_head;
 526
 527         rcu_read_lock();
 528         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 529         rcu_read_unlock();
 530
 531         for (i = 0; i < devs_sorted.nr; i++)
 532                 open_bucket_for_each(c, &h->s->blocks, ob, ec_idx)
 533                         if (ob->ptr.dev == devs_sorted.devs[i] &&
 534                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 535                                 goto got_bucket;
 536         goto out_put_head;
 537 got_bucket:
 538         ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 539
 540         ob->ec_idx      = ec_idx;
 541         ob->ec          = h->s;
 542
 543         __clear_bit(ob->ptr.dev, devs_may_alloc->d);
 544         *nr_effective   += ca->mi.durability;
 545         *have_cache     |= !ca->mi.durability;
 546
 547         ob_push(c, ptrs, ob);
 548         atomic_inc(&h->s->pin);
 549 out_put_head:
 550         bch2_ec_stripe_head_put(h);
 551 }
 552
 553 /* Sector allocator */
 554
 555 static void get_buckets_from_writepoint(struct bch_fs *c,
 556                                         struct open_buckets *ptrs,
 557                                         struct write_point *wp,
 558                                         struct bch_devs_mask *devs_may_alloc,
 559                                         unsigned nr_replicas,
 560                                         unsigned *nr_effective,
 561                                         bool *have_cache,
 562                                         bool need_ec)
 563 {
 564         struct open_buckets ptrs_skip = { .nr = 0 };
 565         struct open_bucket *ob;
 566         unsigned i;
 567
 568         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 569                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 570
 571                 if (*nr_effective < nr_replicas &&
 572                     test_bit(ob->ptr.dev, devs_may_alloc->d) &&
 573                     (ca->mi.durability ||
 574                      (wp->type == BCH_DATA_USER && !*have_cache)) &&
 575                     (ob->ec || !need_ec)) {
 576                         __clear_bit(ob->ptr.dev, devs_may_alloc->d);
 577                         *nr_effective   += ca->mi.durability;
 578                         *have_cache     |= !ca->mi.durability;
 579
 580                         ob_push(c, ptrs, ob);
 581                 } else {
 582                         ob_push(c, &ptrs_skip, ob);
 583                 }
 584         }
 585         wp->ptrs = ptrs_skip;
 586 }
 587
 588 static int open_bucket_add_buckets(struct bch_fs *c,
 589                                    struct open_buckets *ptrs,
 590                                    struct write_point *wp,
 591                                    struct bch_devs_list *devs_have,
 592                                    u16 target,
 593                                    unsigned erasure_code,
 594                                    unsigned nr_replicas,
 595                                    unsigned *nr_effective,
 596                                    bool *have_cache,
 597                                    enum alloc_reserve reserve,
 598                                    struct closure *_cl)
 599 {
 600         struct bch_devs_mask devs;
 601         struct open_bucket *ob;
 602         struct closure *cl = NULL;
 603         unsigned i, flags = BUCKET_ALLOC_USE_DURABILITY;
 604         int ret;
 605
 606         if (wp->type == BCH_DATA_USER)
 607                 flags |= BUCKET_MAY_ALLOC_PARTIAL;
 608
 609         rcu_read_lock();
 610         devs = target_rw_devs(c, wp->type, target);
 611         rcu_read_unlock();
 612
 613         /* Don't allocate from devices we already have pointers to: */
 614         for (i = 0; i < devs_have->nr; i++)
 615                 __clear_bit(devs_have->devs[i], devs.d);
 616
 617         open_bucket_for_each(c, ptrs, ob, i)
 618                 __clear_bit(ob->ptr.dev, devs.d);
 619
 620         if (erasure_code) {
 621                 get_buckets_from_writepoint(c, ptrs, wp, &devs,
 622                                             nr_replicas, nr_effective,
 623                                             have_cache, true);
 624                 if (*nr_effective >= nr_replicas)
 625                         return 0;
 626
 627                 bucket_alloc_from_stripe(c, ptrs, wp, &devs,
 628                                          target, erasure_code,
 629                                          nr_replicas, nr_effective,
 630                                          have_cache);
 631                 if (*nr_effective >= nr_replicas)
 632                         return 0;
 633         }
 634
 635         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 636                                     nr_replicas, nr_effective,
 637                                     have_cache, false);
 638         if (*nr_effective >= nr_replicas)
 639                 return 0;
 640
 641         percpu_down_read_preempt_disable(&c->mark_lock);
 642         rcu_read_lock();
 643
 644 retry_blocking:
 645         /*
 646          * Try nonblocking first, so that if one device is full we'll try from
 647          * other devices:
 648          */
 649         ret = bch2_bucket_alloc_set(c, ptrs, &wp->stripe, &devs,
 650                                 nr_replicas, nr_effective, have_cache,
 651                                 reserve, flags, cl);
 652         if (ret && ret != -EROFS && !cl && _cl) {
 653                 cl = _cl;
 654                 goto retry_blocking;
 655         }
 656
 657         rcu_read_unlock();
 658         percpu_up_read_preempt_enable(&c->mark_lock);
 659
 660         return ret;
 661 }
 662
 663 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 664                                 struct open_buckets *obs,
 665                                 enum bch_data_type data_type)
 666 {
 667         struct open_buckets ptrs = { .nr = 0 };
 668         struct open_bucket *ob, *ob2;
 669         unsigned i, j;
 670
 671         open_bucket_for_each(c, obs, ob, i) {
 672                 bool drop = !ca || ob->ptr.dev == ca->dev_idx;
 673
 674                 if (!drop && ob->ec) {
 675                         mutex_lock(&ob->ec->lock);
 676                         open_bucket_for_each(c, &ob->ec->blocks, ob2, j)
 677                                 drop |= ob2->ptr.dev == ca->dev_idx;
 678                         open_bucket_for_each(c, &ob->ec->parity, ob2, j)
 679                                 drop |= ob2->ptr.dev == ca->dev_idx;
 680                         mutex_unlock(&ob->ec->lock);
 681                 }
 682
 683                 if (drop)
 684                         bch2_open_bucket_put(c, ob);
 685                 else
 686                         ob_push(c, &ptrs, ob);
 687         }
 688
 689         *obs = ptrs;
 690 }
 691
 692 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 693                           struct write_point *wp)
 694 {
 695         mutex_lock(&wp->lock);
 696         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs, wp->type);
 697         mutex_unlock(&wp->lock);
 698 }
 699
 700 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 701                                                  unsigned long write_point)
 702 {
 703         unsigned hash =
 704                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
 705
 706         return &c->write_points_hash[hash];
 707 }
 708
 709 static struct write_point *__writepoint_find(struct hlist_head *head,
 710                                              unsigned long write_point)
 711 {
 712         struct write_point *wp;
 713
 714         hlist_for_each_entry_rcu(wp, head, node)
 715                 if (wp->write_point == write_point)
 716                         return wp;
 717
 718         return NULL;
 719 }
 720
 721 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
 722 {
 723         u64 stranded    = c->write_points_nr * c->bucket_size_max;
 724         u64 free        = bch2_fs_sectors_free(c);
 725
 726         return stranded * factor > free;
 727 }
 728
 729 static bool try_increase_writepoints(struct bch_fs *c)
 730 {
 731         struct write_point *wp;
 732
 733         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
 734             too_many_writepoints(c, 32))
 735                 return false;
 736
 737         wp = c->write_points + c->write_points_nr++;
 738         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
 739         return true;
 740 }
 741
 742 static bool try_decrease_writepoints(struct bch_fs *c,
 743                                      unsigned old_nr)
 744 {
 745         struct write_point *wp;
 746
 747         mutex_lock(&c->write_points_hash_lock);
 748         if (c->write_points_nr < old_nr) {
 749                 mutex_unlock(&c->write_points_hash_lock);
 750                 return true;
 751         }
 752
 753         if (c->write_points_nr == 1 ||
 754             !too_many_writepoints(c, 8)) {
 755                 mutex_unlock(&c->write_points_hash_lock);
 756                 return false;
 757         }
 758
 759         wp = c->write_points + --c->write_points_nr;
 760
 761         hlist_del_rcu(&wp->node);
 762         mutex_unlock(&c->write_points_hash_lock);
 763
 764         bch2_writepoint_stop(c, NULL, wp);
 765         return true;
 766 }
 767
 768 static struct write_point *writepoint_find(struct bch_fs *c,
 769                                            unsigned long write_point)
 770 {
 771         struct write_point *wp, *oldest;
 772         struct hlist_head *head;
 773
 774         if (!(write_point & 1UL)) {
 775                 wp = (struct write_point *) write_point;
 776                 mutex_lock(&wp->lock);
 777                 return wp;
 778         }
 779
 780         head = writepoint_hash(c, write_point);
 781 restart_find:
 782         wp = __writepoint_find(head, write_point);
 783         if (wp) {
 784 lock_wp:
 785                 mutex_lock(&wp->lock);
 786                 if (wp->write_point == write_point)
 787                         goto out;
 788                 mutex_unlock(&wp->lock);
 789                 goto restart_find;
 790         }
 791 restart_find_oldest:
 792         oldest = NULL;
 793         for (wp = c->write_points;
 794              wp < c->write_points + c->write_points_nr; wp++)
 795                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
 796                         oldest = wp;
 797
 798         mutex_lock(&oldest->lock);
 799         mutex_lock(&c->write_points_hash_lock);
 800         if (oldest >= c->write_points + c->write_points_nr ||
 801             try_increase_writepoints(c)) {
 802                 mutex_unlock(&c->write_points_hash_lock);
 803                 mutex_unlock(&oldest->lock);
 804                 goto restart_find_oldest;
 805         }
 806
 807         wp = __writepoint_find(head, write_point);
 808         if (wp && wp != oldest) {
 809                 mutex_unlock(&c->write_points_hash_lock);
 810                 mutex_unlock(&oldest->lock);
 811                 goto lock_wp;
 812         }
 813
 814         wp = oldest;
 815         hlist_del_rcu(&wp->node);
 816         wp->write_point = write_point;
 817         hlist_add_head_rcu(&wp->node, head);
 818         mutex_unlock(&c->write_points_hash_lock);
 819 out:
 820         wp->last_used = sched_clock();
 821         return wp;
 822 }
 823
 824 /*
 825  * Get us an open_bucket we can allocate from, return with it locked:
 826  */
 827 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
 828                                 unsigned target,
 829                                 unsigned erasure_code,
 830                                 struct write_point_specifier write_point,
 831                                 struct bch_devs_list *devs_have,
 832                                 unsigned nr_replicas,
 833                                 unsigned nr_replicas_required,
 834                                 enum alloc_reserve reserve,
 835                                 unsigned flags,
 836                                 struct closure *cl)
 837 {
 838         struct write_point *wp;
 839         struct open_bucket *ob;
 840         unsigned nr_effective = 0;
 841         struct open_buckets ptrs = { .nr = 0 };
 842         bool have_cache = false;
 843         unsigned write_points_nr;
 844         int ret = 0, i;
 845
 846         BUG_ON(!nr_replicas || !nr_replicas_required);
 847 retry:
 848         write_points_nr = c->write_points_nr;
 849
 850         wp = writepoint_find(c, write_point.v);
 851
 852         /* metadata may not allocate on cache devices: */
 853         if (wp->type != BCH_DATA_USER)
 854                 have_cache = true;
 855
 856         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
 857                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 858                                               target, erasure_code,
 859                                               nr_replicas, &nr_effective,
 860                                               &have_cache, reserve, cl);
 861         } else {
 862                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 863                                               target, erasure_code,
 864                                               nr_replicas, &nr_effective,
 865                                               &have_cache, reserve, NULL);
 866                 if (!ret)
 867                         goto alloc_done;
 868
 869                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 870                                               0, erasure_code,
 871                                               nr_replicas, &nr_effective,
 872                                               &have_cache, reserve, cl);
 873         }
 874 alloc_done:
 875         BUG_ON(!ret && nr_effective < nr_replicas);
 876
 877         if (erasure_code && !ec_open_bucket(c, &ptrs))
 878                 pr_debug("failed to get ec bucket: ret %u", ret);
 879
 880         if (ret == -EROFS &&
 881             nr_effective >= nr_replicas_required)
 882                 ret = 0;
 883
 884         if (ret)
 885                 goto err;
 886
 887         /* Free buckets we didn't use: */
 888         open_bucket_for_each(c, &wp->ptrs, ob, i)
 889                 open_bucket_free_unused(c, ob, wp->type == BCH_DATA_USER);
 890
 891         wp->ptrs = ptrs;
 892
 893         wp->sectors_free = UINT_MAX;
 894
 895         open_bucket_for_each(c, &wp->ptrs, ob, i)
 896                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
 897
 898         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
 899
 900         verify_not_stale(c, &wp->ptrs);
 901
 902         return wp;
 903 err:
 904         open_bucket_for_each(c, &wp->ptrs, ob, i)
 905                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
 906                         ob_push(c, &ptrs, ob);
 907                 else
 908                         open_bucket_free_unused(c, ob,
 909                                         wp->type == BCH_DATA_USER);
 910         wp->ptrs = ptrs;
 911
 912         mutex_unlock(&wp->lock);
 913
 914         if (ret == -ENOSPC &&
 915             try_decrease_writepoints(c, write_points_nr))
 916                 goto retry;
 917
 918         return ERR_PTR(ret);
 919 }
 920
 921 /*
 922  * Append pointers to the space we just allocated to @k, and mark @sectors space
 923  * as allocated out of @ob
 924  */
 925 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
 926                                     struct bkey_i *k, unsigned sectors)
 927
 928 {
 929         struct open_bucket *ob;
 930         unsigned i;
 931
 932         BUG_ON(sectors > wp->sectors_free);
 933         wp->sectors_free -= sectors;
 934
 935         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 936                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 937                 struct bch_extent_ptr tmp = ob->ptr;
 938
 939                 tmp.cached = !ca->mi.durability &&
 940                         wp->type == BCH_DATA_USER;
 941
 942                 tmp.offset += ca->mi.bucket_size - ob->sectors_free;
 943                 bch2_bkey_append_ptr(k, tmp);
 944
 945                 BUG_ON(sectors > ob->sectors_free);
 946                 ob->sectors_free -= sectors;
 947         }
 948 }
 949
 950 /*
 951  * Append pointers to the space we just allocated to @k, and mark @sectors space
 952  * as allocated out of @ob
 953  */
 954 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
 955 {
 956         struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
 957         struct open_bucket *ob;
 958         unsigned i;
 959
 960         open_bucket_for_each(c, &wp->ptrs, ob, i)
 961                 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
 962         wp->ptrs = keep;
 963
 964         mutex_unlock(&wp->lock);
 965
 966         bch2_open_buckets_put(c, &ptrs);
 967 }
 968
 969 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
 970 {
 971         struct open_bucket *ob;
 972         struct write_point *wp;
 973
 974         mutex_init(&c->write_points_hash_lock);
 975         c->write_points_nr = ARRAY_SIZE(c->write_points);
 976
 977         /* open bucket 0 is a sentinal NULL: */
 978         spin_lock_init(&c->open_buckets[0].lock);
 979
 980         for (ob = c->open_buckets + 1;
 981              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
 982                 spin_lock_init(&ob->lock);
 983                 c->open_buckets_nr_free++;
 984
 985                 ob->freelist = c->open_buckets_freelist;
 986                 c->open_buckets_freelist = ob - c->open_buckets;
 987         }
 988
 989         writepoint_init(&c->btree_write_point, BCH_DATA_BTREE);
 990         writepoint_init(&c->rebalance_write_point, BCH_DATA_USER);
 991
 992         for (wp = c->write_points;
 993              wp < c->write_points + c->write_points_nr; wp++) {
 994                 writepoint_init(wp, BCH_DATA_USER);
 995
 996                 wp->last_used   = sched_clock();
 997                 wp->write_point = (unsigned long) wp;
 998                 hlist_add_head_rcu(&wp->node,
 999                                    writepoint_hash(c, wp->write_point));
1000         }
1001 }