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
 249                 if (!c->blocked_allocate_open_bucket)
 250                         c->blocked_allocate_open_bucket = local_clock();
 251
 252                 spin_unlock(&c->freelist_lock);
 253                 trace_open_bucket_alloc_fail(ca, reserve);
 254                 return ERR_PTR(-OPEN_BUCKETS_EMPTY);
 255         }
 256
 257         if (likely(fifo_pop(&ca->free[RESERVE_NONE], bucket)))
 258                 goto out;
 259
 260         switch (reserve) {
 261         case RESERVE_ALLOC:
 262                 if (fifo_pop(&ca->free[RESERVE_BTREE], bucket))
 263                         goto out;
 264                 break;
 265         case RESERVE_BTREE:
 266                 if (fifo_used(&ca->free[RESERVE_BTREE]) * 2 >=
 267                     ca->free[RESERVE_BTREE].size &&
 268                     fifo_pop(&ca->free[RESERVE_BTREE], bucket))
 269                         goto out;
 270                 break;
 271         case RESERVE_MOVINGGC:
 272                 if (fifo_pop(&ca->free[RESERVE_MOVINGGC], bucket))
 273                         goto out;
 274                 break;
 275         default:
 276                 break;
 277         }
 278
 279         if (cl)
 280                 closure_wait(&c->freelist_wait, cl);
 281
 282         if (!c->blocked_allocate)
 283                 c->blocked_allocate = local_clock();
 284
 285         spin_unlock(&c->freelist_lock);
 286
 287         trace_bucket_alloc_fail(ca, reserve);
 288         return ERR_PTR(-FREELIST_EMPTY);
 289 out:
 290         verify_not_on_freelist(c, ca, bucket);
 291
 292         ob = bch2_open_bucket_alloc(c);
 293
 294         spin_lock(&ob->lock);
 295         buckets = bucket_array(ca);
 296
 297         ob->valid       = true;
 298         ob->sectors_free = ca->mi.bucket_size;
 299         ob->ptr         = (struct bch_extent_ptr) {
 300                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
 301                 .gen    = buckets->b[bucket].mark.gen,
 302                 .offset = bucket_to_sector(ca, bucket),
 303                 .dev    = ca->dev_idx,
 304         };
 305
 306         bucket_io_clock_reset(c, ca, bucket, READ);
 307         bucket_io_clock_reset(c, ca, bucket, WRITE);
 308         spin_unlock(&ob->lock);
 309
 310         if (c->blocked_allocate_open_bucket) {
 311                 bch2_time_stats_update(
 312                         &c->times[BCH_TIME_blocked_allocate_open_bucket],
 313                         c->blocked_allocate_open_bucket);
 314                 c->blocked_allocate_open_bucket = 0;
 315         }
 316
 317         if (c->blocked_allocate) {
 318                 bch2_time_stats_update(
 319                         &c->times[BCH_TIME_blocked_allocate],
 320                         c->blocked_allocate);
 321                 c->blocked_allocate = 0;
 322         }
 323
 324         spin_unlock(&c->freelist_lock);
 325
 326         bch2_wake_allocator(ca);
 327
 328         trace_bucket_alloc(ca, reserve);
 329         return ob;
 330 }
 331
 332 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 333                             unsigned l, unsigned r)
 334 {
 335         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 336                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 337 }
 338
 339 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 340
 341 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 342                                           struct dev_stripe_state *stripe,
 343                                           struct bch_devs_mask *devs)
 344 {
 345         struct dev_alloc_list ret = { .nr = 0 };
 346         struct bch_dev *ca;
 347         unsigned i;
 348
 349         for_each_member_device_rcu(ca, c, i, devs)
 350                 ret.devs[ret.nr++] = i;
 351
 352         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 353         return ret;
 354 }
 355
 356 void bch2_dev_stripe_increment(struct bch_fs *c, struct bch_dev *ca,
 357                                struct dev_stripe_state *stripe)
 358 {
 359         u64 *v = stripe->next_alloc + ca->dev_idx;
 360         u64 free_space = dev_buckets_free(c, ca);
 361         u64 free_space_inv = free_space
 362                 ? div64_u64(1ULL << 48, free_space)
 363                 : 1ULL << 48;
 364         u64 scale = *v / 4;
 365
 366         if (*v + free_space_inv >= *v)
 367                 *v += free_space_inv;
 368         else
 369                 *v = U64_MAX;
 370
 371         for (v = stripe->next_alloc;
 372              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 373                 *v = *v < scale ? 0 : *v - scale;
 374 }
 375
 376 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 377 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 378
 379 static int bch2_bucket_alloc_set(struct bch_fs *c,
 380                                  struct open_buckets *ptrs,
 381                                  struct dev_stripe_state *stripe,
 382                                  struct bch_devs_mask *devs_may_alloc,
 383                                  unsigned nr_replicas,
 384                                  unsigned *nr_effective,
 385                                  bool *have_cache,
 386                                  enum alloc_reserve reserve,
 387                                  unsigned flags,
 388                                  struct closure *cl)
 389 {
 390         struct dev_alloc_list devs_sorted =
 391                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 392         struct bch_dev *ca;
 393         bool alloc_failure = false;
 394         unsigned i, durability;
 395
 396         BUG_ON(*nr_effective >= nr_replicas);
 397
 398         for (i = 0; i < devs_sorted.nr; i++) {
 399                 struct open_bucket *ob;
 400
 401                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
 402                 if (!ca)
 403                         continue;
 404
 405                 if (!ca->mi.durability && *have_cache)
 406                         continue;
 407
 408                 ob = bch2_bucket_alloc(c, ca, reserve,
 409                                 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
 410                 if (IS_ERR(ob)) {
 411                         enum bucket_alloc_ret ret = -PTR_ERR(ob);
 412
 413                         WARN_ON(reserve == RESERVE_MOVINGGC &&
 414                                 ret != OPEN_BUCKETS_EMPTY);
 415
 416                         if (cl)
 417                                 return -EAGAIN;
 418                         if (ret == OPEN_BUCKETS_EMPTY)
 419                                 return -ENOSPC;
 420                         alloc_failure = true;
 421                         continue;
 422                 }
 423
 424                 durability = (flags & BUCKET_ALLOC_USE_DURABILITY)
 425                         ? ca->mi.durability : 1;
 426
 427                 __clear_bit(ca->dev_idx, devs_may_alloc->d);
 428                 *nr_effective   += durability;
 429                 *have_cache     |= !durability;
 430
 431                 ob_push(c, ptrs, ob);
 432
 433                 bch2_dev_stripe_increment(c, ca, stripe);
 434
 435                 if (*nr_effective >= nr_replicas)
 436                         return 0;
 437         }
 438
 439         return alloc_failure ? -ENOSPC : -EROFS;
 440 }
 441
 442 /* Allocate from stripes: */
 443
 444 /*
 445  * XXX: use a higher watermark for allocating open buckets here:
 446  */
 447 static int ec_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
 448 {
 449         struct bch_devs_mask devs;
 450         struct open_bucket *ob;
 451         unsigned i, nr_have = 0, nr_data =
 452                 min_t(unsigned, h->nr_active_devs,
 453                       EC_STRIPE_MAX) - h->redundancy;
 454         bool have_cache = true;
 455         int ret = 0;
 456
 457         BUG_ON(h->blocks.nr > nr_data);
 458         BUG_ON(h->parity.nr > h->redundancy);
 459
 460         devs = h->devs;
 461
 462         open_bucket_for_each(c, &h->parity, ob, i)
 463                 __clear_bit(ob->ptr.dev, devs.d);
 464         open_bucket_for_each(c, &h->blocks, ob, i)
 465                 __clear_bit(ob->ptr.dev, devs.d);
 466
 467         percpu_down_read_preempt_disable(&c->mark_lock);
 468         rcu_read_lock();
 469
 470         if (h->parity.nr < h->redundancy) {
 471                 nr_have = h->parity.nr;
 472
 473                 ret = bch2_bucket_alloc_set(c, &h->parity,
 474                                             &h->parity_stripe,
 475                                             &devs,
 476                                             h->redundancy,
 477                                             &nr_have,
 478                                             &have_cache,
 479                                             RESERVE_NONE,
 480                                             0,
 481                                             NULL);
 482                 if (ret)
 483                         goto err;
 484         }
 485
 486         if (h->blocks.nr < nr_data) {
 487                 nr_have = h->blocks.nr;
 488
 489                 ret = bch2_bucket_alloc_set(c, &h->blocks,
 490                                             &h->block_stripe,
 491                                             &devs,
 492                                             nr_data,
 493                                             &nr_have,
 494                                             &have_cache,
 495                                             RESERVE_NONE,
 496                                             0,
 497                                             NULL);
 498                 if (ret)
 499                         goto err;
 500         }
 501
 502         rcu_read_unlock();
 503         percpu_up_read_preempt_enable(&c->mark_lock);
 504
 505         return bch2_ec_stripe_new_alloc(c, h);
 506 err:
 507         rcu_read_unlock();
 508         percpu_up_read_preempt_enable(&c->mark_lock);
 509         return -1;
 510 }
 511
 512 /*
 513  * if we can't allocate a new stripe because there are already too many
 514  * partially filled stripes, force allocating from an existing stripe even when
 515  * it's to a device we don't want:
 516  */
 517
 518 static void bucket_alloc_from_stripe(struct bch_fs *c,
 519                                      struct open_buckets *ptrs,
 520                                      struct write_point *wp,
 521                                      struct bch_devs_mask *devs_may_alloc,
 522                                      u16 target,
 523                                      unsigned erasure_code,
 524                                      unsigned nr_replicas,
 525                                      unsigned *nr_effective,
 526                                      bool *have_cache)
 527 {
 528         struct dev_alloc_list devs_sorted;
 529         struct ec_stripe_head *h;
 530         struct open_bucket *ob;
 531         struct bch_dev *ca;
 532         unsigned i, ec_idx;
 533
 534         if (!erasure_code)
 535                 return;
 536
 537         if (nr_replicas < 2)
 538                 return;
 539
 540         if (ec_open_bucket(c, ptrs))
 541                 return;
 542
 543         h = bch2_ec_stripe_head_get(c, target, erasure_code, nr_replicas - 1);
 544         if (!h)
 545                 return;
 546
 547         if (!h->s && ec_stripe_alloc(c, h))
 548                 goto out_put_head;
 549
 550         rcu_read_lock();
 551         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 552         rcu_read_unlock();
 553
 554         for (i = 0; i < devs_sorted.nr; i++)
 555                 open_bucket_for_each(c, &h->s->blocks, ob, ec_idx)
 556                         if (ob->ptr.dev == devs_sorted.devs[i] &&
 557                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 558                                 goto got_bucket;
 559         goto out_put_head;
 560 got_bucket:
 561         ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 562
 563         ob->ec_idx      = ec_idx;
 564         ob->ec          = h->s;
 565
 566         __clear_bit(ob->ptr.dev, devs_may_alloc->d);
 567         *nr_effective   += ca->mi.durability;
 568         *have_cache     |= !ca->mi.durability;
 569
 570         ob_push(c, ptrs, ob);
 571         atomic_inc(&h->s->pin);
 572 out_put_head:
 573         bch2_ec_stripe_head_put(h);
 574 }
 575
 576 /* Sector allocator */
 577
 578 static void get_buckets_from_writepoint(struct bch_fs *c,
 579                                         struct open_buckets *ptrs,
 580                                         struct write_point *wp,
 581                                         struct bch_devs_mask *devs_may_alloc,
 582                                         unsigned nr_replicas,
 583                                         unsigned *nr_effective,
 584                                         bool *have_cache,
 585                                         bool need_ec)
 586 {
 587         struct open_buckets ptrs_skip = { .nr = 0 };
 588         struct open_bucket *ob;
 589         unsigned i;
 590
 591         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 592                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 593
 594                 if (*nr_effective < nr_replicas &&
 595                     test_bit(ob->ptr.dev, devs_may_alloc->d) &&
 596                     (ca->mi.durability ||
 597                      (wp->type == BCH_DATA_USER && !*have_cache)) &&
 598                     (ob->ec || !need_ec)) {
 599                         __clear_bit(ob->ptr.dev, devs_may_alloc->d);
 600                         *nr_effective   += ca->mi.durability;
 601                         *have_cache     |= !ca->mi.durability;
 602
 603                         ob_push(c, ptrs, ob);
 604                 } else {
 605                         ob_push(c, &ptrs_skip, ob);
 606                 }
 607         }
 608         wp->ptrs = ptrs_skip;
 609 }
 610
 611 static int open_bucket_add_buckets(struct bch_fs *c,
 612                                    struct open_buckets *ptrs,
 613                                    struct write_point *wp,
 614                                    struct bch_devs_list *devs_have,
 615                                    u16 target,
 616                                    unsigned erasure_code,
 617                                    unsigned nr_replicas,
 618                                    unsigned *nr_effective,
 619                                    bool *have_cache,
 620                                    enum alloc_reserve reserve,
 621                                    struct closure *_cl)
 622 {
 623         struct bch_devs_mask devs;
 624         struct open_bucket *ob;
 625         struct closure *cl = NULL;
 626         unsigned i, flags = BUCKET_ALLOC_USE_DURABILITY;
 627         int ret;
 628
 629         if (wp->type == BCH_DATA_USER)
 630                 flags |= BUCKET_MAY_ALLOC_PARTIAL;
 631
 632         rcu_read_lock();
 633         devs = target_rw_devs(c, wp->type, target);
 634         rcu_read_unlock();
 635
 636         /* Don't allocate from devices we already have pointers to: */
 637         for (i = 0; i < devs_have->nr; i++)
 638                 __clear_bit(devs_have->devs[i], devs.d);
 639
 640         open_bucket_for_each(c, ptrs, ob, i)
 641                 __clear_bit(ob->ptr.dev, devs.d);
 642
 643         if (erasure_code) {
 644                 get_buckets_from_writepoint(c, ptrs, wp, &devs,
 645                                             nr_replicas, nr_effective,
 646                                             have_cache, true);
 647                 if (*nr_effective >= nr_replicas)
 648                         return 0;
 649
 650                 bucket_alloc_from_stripe(c, ptrs, wp, &devs,
 651                                          target, erasure_code,
 652                                          nr_replicas, nr_effective,
 653                                          have_cache);
 654                 if (*nr_effective >= nr_replicas)
 655                         return 0;
 656         }
 657
 658         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 659                                     nr_replicas, nr_effective,
 660                                     have_cache, false);
 661         if (*nr_effective >= nr_replicas)
 662                 return 0;
 663
 664         percpu_down_read_preempt_disable(&c->mark_lock);
 665         rcu_read_lock();
 666
 667 retry_blocking:
 668         /*
 669          * Try nonblocking first, so that if one device is full we'll try from
 670          * other devices:
 671          */
 672         ret = bch2_bucket_alloc_set(c, ptrs, &wp->stripe, &devs,
 673                                 nr_replicas, nr_effective, have_cache,
 674                                 reserve, flags, cl);
 675         if (ret && ret != -EROFS && !cl && _cl) {
 676                 cl = _cl;
 677                 goto retry_blocking;
 678         }
 679
 680         rcu_read_unlock();
 681         percpu_up_read_preempt_enable(&c->mark_lock);
 682
 683         return ret;
 684 }
 685
 686 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 687                                 struct open_buckets *obs,
 688                                 enum bch_data_type data_type)
 689 {
 690         struct open_buckets ptrs = { .nr = 0 };
 691         struct open_bucket *ob, *ob2;
 692         unsigned i, j;
 693
 694         open_bucket_for_each(c, obs, ob, i) {
 695                 bool drop = !ca || ob->ptr.dev == ca->dev_idx;
 696
 697                 if (!drop && ob->ec) {
 698                         mutex_lock(&ob->ec->lock);
 699                         open_bucket_for_each(c, &ob->ec->blocks, ob2, j)
 700                                 drop |= ob2->ptr.dev == ca->dev_idx;
 701                         open_bucket_for_each(c, &ob->ec->parity, ob2, j)
 702                                 drop |= ob2->ptr.dev == ca->dev_idx;
 703                         mutex_unlock(&ob->ec->lock);
 704                 }
 705
 706                 if (drop)
 707                         bch2_open_bucket_put(c, ob);
 708                 else
 709                         ob_push(c, &ptrs, ob);
 710         }
 711
 712         *obs = ptrs;
 713 }
 714
 715 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 716                           struct write_point *wp)
 717 {
 718         mutex_lock(&wp->lock);
 719         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs, wp->type);
 720         mutex_unlock(&wp->lock);
 721 }
 722
 723 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 724                                                  unsigned long write_point)
 725 {
 726         unsigned hash =
 727                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
 728
 729         return &c->write_points_hash[hash];
 730 }
 731
 732 static struct write_point *__writepoint_find(struct hlist_head *head,
 733                                              unsigned long write_point)
 734 {
 735         struct write_point *wp;
 736
 737         hlist_for_each_entry_rcu(wp, head, node)
 738                 if (wp->write_point == write_point)
 739                         return wp;
 740
 741         return NULL;
 742 }
 743
 744 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
 745 {
 746         u64 stranded    = c->write_points_nr * c->bucket_size_max;
 747         u64 free        = bch2_fs_usage_read_short(c).free;
 748
 749         return stranded * factor > free;
 750 }
 751
 752 static bool try_increase_writepoints(struct bch_fs *c)
 753 {
 754         struct write_point *wp;
 755
 756         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
 757             too_many_writepoints(c, 32))
 758                 return false;
 759
 760         wp = c->write_points + c->write_points_nr++;
 761         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
 762         return true;
 763 }
 764
 765 static bool try_decrease_writepoints(struct bch_fs *c,
 766                                      unsigned old_nr)
 767 {
 768         struct write_point *wp;
 769
 770         mutex_lock(&c->write_points_hash_lock);
 771         if (c->write_points_nr < old_nr) {
 772                 mutex_unlock(&c->write_points_hash_lock);
 773                 return true;
 774         }
 775
 776         if (c->write_points_nr == 1 ||
 777             !too_many_writepoints(c, 8)) {
 778                 mutex_unlock(&c->write_points_hash_lock);
 779                 return false;
 780         }
 781
 782         wp = c->write_points + --c->write_points_nr;
 783
 784         hlist_del_rcu(&wp->node);
 785         mutex_unlock(&c->write_points_hash_lock);
 786
 787         bch2_writepoint_stop(c, NULL, wp);
 788         return true;
 789 }
 790
 791 static struct write_point *writepoint_find(struct bch_fs *c,
 792                                            unsigned long write_point)
 793 {
 794         struct write_point *wp, *oldest;
 795         struct hlist_head *head;
 796
 797         if (!(write_point & 1UL)) {
 798                 wp = (struct write_point *) write_point;
 799                 mutex_lock(&wp->lock);
 800                 return wp;
 801         }
 802
 803         head = writepoint_hash(c, write_point);
 804 restart_find:
 805         wp = __writepoint_find(head, write_point);
 806         if (wp) {
 807 lock_wp:
 808                 mutex_lock(&wp->lock);
 809                 if (wp->write_point == write_point)
 810                         goto out;
 811                 mutex_unlock(&wp->lock);
 812                 goto restart_find;
 813         }
 814 restart_find_oldest:
 815         oldest = NULL;
 816         for (wp = c->write_points;
 817              wp < c->write_points + c->write_points_nr; wp++)
 818                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
 819                         oldest = wp;
 820
 821         mutex_lock(&oldest->lock);
 822         mutex_lock(&c->write_points_hash_lock);
 823         if (oldest >= c->write_points + c->write_points_nr ||
 824             try_increase_writepoints(c)) {
 825                 mutex_unlock(&c->write_points_hash_lock);
 826                 mutex_unlock(&oldest->lock);
 827                 goto restart_find_oldest;
 828         }
 829
 830         wp = __writepoint_find(head, write_point);
 831         if (wp && wp != oldest) {
 832                 mutex_unlock(&c->write_points_hash_lock);
 833                 mutex_unlock(&oldest->lock);
 834                 goto lock_wp;
 835         }
 836
 837         wp = oldest;
 838         hlist_del_rcu(&wp->node);
 839         wp->write_point = write_point;
 840         hlist_add_head_rcu(&wp->node, head);
 841         mutex_unlock(&c->write_points_hash_lock);
 842 out:
 843         wp->last_used = sched_clock();
 844         return wp;
 845 }
 846
 847 /*
 848  * Get us an open_bucket we can allocate from, return with it locked:
 849  */
 850 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
 851                                 unsigned target,
 852                                 unsigned erasure_code,
 853                                 struct write_point_specifier write_point,
 854                                 struct bch_devs_list *devs_have,
 855                                 unsigned nr_replicas,
 856                                 unsigned nr_replicas_required,
 857                                 enum alloc_reserve reserve,
 858                                 unsigned flags,
 859                                 struct closure *cl)
 860 {
 861         struct write_point *wp;
 862         struct open_bucket *ob;
 863         struct open_buckets ptrs;
 864         unsigned nr_effective, write_points_nr;
 865         bool have_cache;
 866         int ret, i;
 867
 868         BUG_ON(!nr_replicas || !nr_replicas_required);
 869 retry:
 870         ptrs.nr         = 0;
 871         nr_effective    = 0;
 872         write_points_nr = c->write_points_nr;
 873         have_cache      = false;
 874
 875         wp = writepoint_find(c, write_point.v);
 876
 877         /* metadata may not allocate on cache devices: */
 878         if (wp->type != BCH_DATA_USER)
 879                 have_cache = true;
 880
 881         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
 882                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 883                                               target, erasure_code,
 884                                               nr_replicas, &nr_effective,
 885                                               &have_cache, reserve, cl);
 886         } else {
 887                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 888                                               target, erasure_code,
 889                                               nr_replicas, &nr_effective,
 890                                               &have_cache, reserve, NULL);
 891                 if (!ret)
 892                         goto alloc_done;
 893
 894                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 895                                               0, erasure_code,
 896                                               nr_replicas, &nr_effective,
 897                                               &have_cache, reserve, cl);
 898         }
 899 alloc_done:
 900         BUG_ON(!ret && nr_effective < nr_replicas);
 901
 902         if (erasure_code && !ec_open_bucket(c, &ptrs))
 903                 pr_debug("failed to get ec bucket: ret %u", ret);
 904
 905         if (ret == -EROFS &&
 906             nr_effective >= nr_replicas_required)
 907                 ret = 0;
 908
 909         if (ret)
 910                 goto err;
 911
 912         /* Free buckets we didn't use: */
 913         open_bucket_for_each(c, &wp->ptrs, ob, i)
 914                 open_bucket_free_unused(c, ob, wp->type == BCH_DATA_USER);
 915
 916         wp->ptrs = ptrs;
 917
 918         wp->sectors_free = UINT_MAX;
 919
 920         open_bucket_for_each(c, &wp->ptrs, ob, i)
 921                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
 922
 923         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
 924
 925         verify_not_stale(c, &wp->ptrs);
 926
 927         return wp;
 928 err:
 929         open_bucket_for_each(c, &wp->ptrs, ob, i)
 930                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
 931                         ob_push(c, &ptrs, ob);
 932                 else
 933                         open_bucket_free_unused(c, ob,
 934                                         wp->type == BCH_DATA_USER);
 935         wp->ptrs = ptrs;
 936
 937         mutex_unlock(&wp->lock);
 938
 939         if (ret == -ENOSPC &&
 940             try_decrease_writepoints(c, write_points_nr))
 941                 goto retry;
 942
 943         return ERR_PTR(ret);
 944 }
 945
 946 /*
 947  * Append pointers to the space we just allocated to @k, and mark @sectors space
 948  * as allocated out of @ob
 949  */
 950 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
 951                                     struct bkey_i *k, unsigned sectors)
 952
 953 {
 954         struct open_bucket *ob;
 955         unsigned i;
 956
 957         BUG_ON(sectors > wp->sectors_free);
 958         wp->sectors_free -= sectors;
 959
 960         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 961                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 962                 struct bch_extent_ptr tmp = ob->ptr;
 963
 964                 tmp.cached = !ca->mi.durability &&
 965                         wp->type == BCH_DATA_USER;
 966
 967                 tmp.offset += ca->mi.bucket_size - ob->sectors_free;
 968                 bch2_bkey_append_ptr(k, tmp);
 969
 970                 BUG_ON(sectors > ob->sectors_free);
 971                 ob->sectors_free -= sectors;
 972         }
 973 }
 974
 975 /*
 976  * Append pointers to the space we just allocated to @k, and mark @sectors space
 977  * as allocated out of @ob
 978  */
 979 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
 980 {
 981         struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
 982         struct open_bucket *ob;
 983         unsigned i;
 984
 985         open_bucket_for_each(c, &wp->ptrs, ob, i)
 986                 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
 987         wp->ptrs = keep;
 988
 989         mutex_unlock(&wp->lock);
 990
 991         bch2_open_buckets_put(c, &ptrs);
 992 }
 993
 994 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
 995 {
 996         struct open_bucket *ob;
 997         struct write_point *wp;
 998
 999         mutex_init(&c->write_points_hash_lock);
1000         c->write_points_nr = ARRAY_SIZE(c->write_points);
1001
1002         /* open bucket 0 is a sentinal NULL: */
1003         spin_lock_init(&c->open_buckets[0].lock);
1004
1005         for (ob = c->open_buckets + 1;
1006              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1007                 spin_lock_init(&ob->lock);
1008                 c->open_buckets_nr_free++;
1009
1010                 ob->freelist = c->open_buckets_freelist;
1011                 c->open_buckets_freelist = ob - c->open_buckets;
1012         }
1013
1014         writepoint_init(&c->btree_write_point, BCH_DATA_BTREE);
1015         writepoint_init(&c->rebalance_write_point, BCH_DATA_USER);
1016
1017         for (wp = c->write_points;
1018              wp < c->write_points + c->write_points_nr; wp++) {
1019                 writepoint_init(wp, BCH_DATA_USER);
1020
1021                 wp->last_used   = sched_clock();
1022                 wp->write_point = (unsigned long) wp;
1023                 hlist_add_head_rcu(&wp->node,
1024                                    writepoint_hash(c, wp->write_point));
1025         }
1026 }