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