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

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Primary bucket allocation code
   4  *
   5  * Copyright 2012 Google, Inc.
   6  *
   7  * Allocation in bcache is done in terms of buckets:
   8  *
   9  * Each bucket has associated an 8 bit gen; this gen corresponds to the gen in
  10  * btree pointers - they must match for the pointer to be considered valid.
  11  *
  12  * Thus (assuming a bucket has no dirty data or metadata in it) we can reuse a
  13  * bucket simply by incrementing its gen.
  14  *
  15  * The gens (along with the priorities; it's really the gens are important but
  16  * the code is named as if it's the priorities) are written in an arbitrary list
  17  * of buckets on disk, with a pointer to them in the journal header.
  18  *
  19  * When we invalidate a bucket, we have to write its new gen to disk and wait
  20  * for that write to complete before we use it - otherwise after a crash we
  21  * could have pointers that appeared to be good but pointed to data that had
  22  * been overwritten.
  23  *
  24  * Since the gens and priorities are all stored contiguously on disk, we can
  25  * batch this up: We fill up the free_inc list with freshly invalidated buckets,
  26  * call prio_write(), and when prio_write() finishes we pull buckets off the
  27  * free_inc list and optionally discard them.
  28  *
  29  * free_inc isn't the only freelist - if it was, we'd often have to sleep while
  30  * priorities and gens were being written before we could allocate. c->free is a
  31  * smaller freelist, and buckets on that list are always ready to be used.
  32  *
  33  * If we've got discards enabled, that happens when a bucket moves from the
  34  * free_inc list to the free list.
  35  *
  36  * It's important to ensure that gens don't wrap around - with respect to
  37  * either the oldest gen in the btree or the gen on disk. This is quite
  38  * difficult to do in practice, but we explicitly guard against it anyways - if
  39  * a bucket is in danger of wrapping around we simply skip invalidating it that
  40  * time around, and we garbage collect or rewrite the priorities sooner than we
  41  * would have otherwise.
  42  *
  43  * bch2_bucket_alloc() allocates a single bucket from a specific device.
  44  *
  45  * bch2_bucket_alloc_set() allocates one or more buckets from different devices
  46  * in a given filesystem.
  47  *
  48  * invalidate_buckets() drives all the processes described above. It's called
  49  * from bch2_bucket_alloc() and a few other places that need to make sure free
  50  * buckets are ready.
  51  *
  52  * invalidate_buckets_(lru|fifo)() find buckets that are available to be
  53  * invalidated, and then invalidate them and stick them on the free_inc list -
  54  * in either lru or fifo order.
  55  */
  56
  57 #include "bcachefs.h"
  58 #include "alloc_background.h"
  59 #include "alloc_foreground.h"
  60 #include "btree_gc.h"
  61 #include "buckets.h"
  62 #include "clock.h"
  63 #include "debug.h"
  64 #include "disk_groups.h"
  65 #include "ec.h"
  66 #include "io.h"
  67
  68 #include <linux/math64.h>
  69 #include <linux/rculist.h>
  70 #include <linux/rcupdate.h>
  71 #include <trace/events/bcachefs.h>
  72
  73 enum bucket_alloc_ret {
  74         ALLOC_SUCCESS,
  75         OPEN_BUCKETS_EMPTY,
  76         FREELIST_EMPTY,         /* Allocator thread not keeping up */
  77 };
  78
  79 /*
  80  * Open buckets represent a bucket that's currently being allocated from.  They
  81  * serve two purposes:
  82  *
  83  *  - They track buckets that have been partially allocated, allowing for
  84  *    sub-bucket sized allocations - they're used by the sector allocator below
  85  *
  86  *  - They provide a reference to the buckets they own that mark and sweep GC
  87  *    can find, until the new allocation has a pointer to it inserted into the
  88  *    btree
  89  *
  90  * When allocating some space with the sector allocator, the allocation comes
  91  * with a reference to an open bucket - the caller is required to put that
  92  * reference _after_ doing the index update that makes its allocation reachable.
  93  */
  94
  95 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
  96 {
  97         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
  98
  99         if (ob->ec) {
 100                 bch2_ec_bucket_written(c, ob);
 101                 return;
 102         }
 103
 104         percpu_down_read(&c->mark_lock);
 105         spin_lock(&ob->lock);
 106
 107         bch2_mark_alloc_bucket(c, ca, PTR_BUCKET_NR(ca, &ob->ptr),
 108                                false, gc_pos_alloc(c, ob), 0);
 109         ob->valid = false;
 110         ob->type = 0;
 111
 112         spin_unlock(&ob->lock);
 113         percpu_up_read(&c->mark_lock);
 114
 115         spin_lock(&c->freelist_lock);
 116         ob->freelist = c->open_buckets_freelist;
 117         c->open_buckets_freelist = ob - c->open_buckets;
 118         c->open_buckets_nr_free++;
 119         spin_unlock(&c->freelist_lock);
 120
 121         closure_wake_up(&c->open_buckets_wait);
 122 }
 123
 124 void bch2_open_bucket_write_error(struct bch_fs *c,
 125                                   struct open_buckets *obs,
 126                                   unsigned dev)
 127 {
 128         struct open_bucket *ob;
 129         unsigned i;
 130
 131         open_bucket_for_each(c, obs, ob, i)
 132                 if (ob->ptr.dev == dev &&
 133                     ob->ec)
 134                         bch2_ec_bucket_cancel(c, ob);
 135 }
 136
 137 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
 138 {
 139         struct open_bucket *ob;
 140
 141         BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
 142
 143         ob = c->open_buckets + c->open_buckets_freelist;
 144         c->open_buckets_freelist = ob->freelist;
 145         atomic_set(&ob->pin, 1);
 146         ob->type = 0;
 147
 148         c->open_buckets_nr_free--;
 149         return ob;
 150 }
 151
 152 static void open_bucket_free_unused(struct bch_fs *c,
 153                                     struct open_bucket *ob,
 154                                     bool may_realloc)
 155 {
 156         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 157
 158         BUG_ON(ca->open_buckets_partial_nr >=
 159                ARRAY_SIZE(ca->open_buckets_partial));
 160
 161         if (ca->open_buckets_partial_nr <
 162             ARRAY_SIZE(ca->open_buckets_partial) &&
 163             may_realloc) {
 164                 spin_lock(&c->freelist_lock);
 165                 ob->on_partial_list = true;
 166                 ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
 167                         ob - c->open_buckets;
 168                 spin_unlock(&c->freelist_lock);
 169
 170                 closure_wake_up(&c->open_buckets_wait);
 171                 closure_wake_up(&c->freelist_wait);
 172         } else {
 173                 bch2_open_bucket_put(c, ob);
 174         }
 175 }
 176
 177 static void verify_not_stale(struct bch_fs *c, const struct open_buckets *obs)
 178 {
 179 #ifdef CONFIG_BCACHEFS_DEBUG
 180         struct open_bucket *ob;
 181         unsigned i;
 182
 183         open_bucket_for_each(c, obs, ob, i) {
 184                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 185
 186                 BUG_ON(ptr_stale(ca, &ob->ptr));
 187         }
 188 #endif
 189 }
 190
 191 /* _only_ for allocating the journal on a new device: */
 192 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
 193 {
 194         struct bucket_array *buckets;
 195         ssize_t b;
 196
 197         rcu_read_lock();
 198         buckets = bucket_array(ca);
 199
 200         for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++)
 201                 if (is_available_bucket(buckets->b[b].mark))
 202                         goto success;
 203         b = -1;
 204 success:
 205         rcu_read_unlock();
 206         return b;
 207 }
 208
 209 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
 210 {
 211         switch (reserve) {
 212         case RESERVE_ALLOC:
 213                 return 0;
 214         case RESERVE_BTREE:
 215                 return OPEN_BUCKETS_COUNT / 4;
 216         default:
 217                 return OPEN_BUCKETS_COUNT / 2;
 218         }
 219 }
 220
 221 /**
 222  * bch_bucket_alloc - allocate a single bucket from a specific device
 223  *
 224  * Returns index of bucket on success, 0 on failure
 225  * */
 226 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 227                                       enum alloc_reserve reserve,
 228                                       bool may_alloc_partial,
 229                                       struct closure *cl)
 230 {
 231         struct bucket_array *buckets;
 232         struct open_bucket *ob;
 233         long bucket = 0;
 234
 235         spin_lock(&c->freelist_lock);
 236
 237         if (may_alloc_partial &&
 238             ca->open_buckets_partial_nr) {
 239                 ob = c->open_buckets +
 240                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
 241                 ob->on_partial_list = false;
 242                 spin_unlock(&c->freelist_lock);
 243                 return ob;
 244         }
 245
 246         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
 247                 if (cl)
 248                         closure_wait(&c->open_buckets_wait, cl);
 249
 250                 if (!c->blocked_allocate_open_bucket)
 251                         c->blocked_allocate_open_bucket = local_clock();
 252
 253                 spin_unlock(&c->freelist_lock);
 254                 trace_open_bucket_alloc_fail(ca, reserve);
 255                 return ERR_PTR(-OPEN_BUCKETS_EMPTY);
 256         }
 257
 258         if (likely(fifo_pop(&ca->free[RESERVE_NONE], bucket)))
 259                 goto out;
 260
 261         switch (reserve) {
 262         case RESERVE_ALLOC:
 263                 if (fifo_pop(&ca->free[RESERVE_BTREE], bucket))
 264                         goto out;
 265                 break;
 266         case RESERVE_BTREE:
 267                 if (fifo_used(&ca->free[RESERVE_BTREE]) * 2 >=
 268                     ca->free[RESERVE_BTREE].size &&
 269                     fifo_pop(&ca->free[RESERVE_BTREE], bucket))
 270                         goto out;
 271                 break;
 272         case RESERVE_MOVINGGC:
 273                 if (fifo_pop(&ca->free[RESERVE_MOVINGGC], bucket))
 274                         goto out;
 275                 break;
 276         default:
 277                 break;
 278         }
 279
 280         if (cl)
 281                 closure_wait(&c->freelist_wait, cl);
 282
 283         if (!c->blocked_allocate)
 284                 c->blocked_allocate = local_clock();
 285
 286         spin_unlock(&c->freelist_lock);
 287
 288         trace_bucket_alloc_fail(ca, reserve);
 289         return ERR_PTR(-FREELIST_EMPTY);
 290 out:
 291         verify_not_on_freelist(c, ca, bucket);
 292
 293         ob = bch2_open_bucket_alloc(c);
 294
 295         spin_lock(&ob->lock);
 296         buckets = bucket_array(ca);
 297
 298         ob->valid       = true;
 299         ob->sectors_free = ca->mi.bucket_size;
 300         ob->ptr         = (struct bch_extent_ptr) {
 301                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
 302                 .gen    = buckets->b[bucket].mark.gen,
 303                 .offset = bucket_to_sector(ca, bucket),
 304                 .dev    = ca->dev_idx,
 305         };
 306
 307         bucket_io_clock_reset(c, ca, bucket, READ);
 308         bucket_io_clock_reset(c, ca, bucket, WRITE);
 309         spin_unlock(&ob->lock);
 310
 311         if (c->blocked_allocate_open_bucket) {
 312                 bch2_time_stats_update(
 313                         &c->times[BCH_TIME_blocked_allocate_open_bucket],
 314                         c->blocked_allocate_open_bucket);
 315                 c->blocked_allocate_open_bucket = 0;
 316         }
 317
 318         if (c->blocked_allocate) {
 319                 bch2_time_stats_update(
 320                         &c->times[BCH_TIME_blocked_allocate],
 321                         c->blocked_allocate);
 322                 c->blocked_allocate = 0;
 323         }
 324
 325         spin_unlock(&c->freelist_lock);
 326
 327         bch2_wake_allocator(ca);
 328
 329         trace_bucket_alloc(ca, reserve);
 330         return ob;
 331 }
 332
 333 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 334                             unsigned l, unsigned r)
 335 {
 336         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 337                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 338 }
 339
 340 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 341
 342 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 343                                           struct dev_stripe_state *stripe,
 344                                           struct bch_devs_mask *devs)
 345 {
 346         struct dev_alloc_list ret = { .nr = 0 };
 347         struct bch_dev *ca;
 348         unsigned i;
 349
 350         for_each_member_device_rcu(ca, c, i, devs)
 351                 ret.devs[ret.nr++] = i;
 352
 353         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 354         return ret;
 355 }
 356
 357 void bch2_dev_stripe_increment(struct bch_fs *c, struct bch_dev *ca,
 358                                struct dev_stripe_state *stripe)
 359 {
 360         u64 *v = stripe->next_alloc + ca->dev_idx;
 361         u64 free_space = dev_buckets_free(c, ca);
 362         u64 free_space_inv = free_space
 363                 ? div64_u64(1ULL << 48, free_space)
 364                 : 1ULL << 48;
 365         u64 scale = *v / 4;
 366
 367         if (*v + free_space_inv >= *v)
 368                 *v += free_space_inv;
 369         else
 370                 *v = U64_MAX;
 371
 372         for (v = stripe->next_alloc;
 373              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 374                 *v = *v < scale ? 0 : *v - scale;
 375 }
 376
 377 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 378 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 379
 380 static void add_new_bucket(struct bch_fs *c,
 381                            struct open_buckets *ptrs,
 382                            struct bch_devs_mask *devs_may_alloc,
 383                            unsigned *nr_effective,
 384                            bool *have_cache,
 385                            unsigned flags,
 386                            struct open_bucket *ob)
 387 {
 388         unsigned durability =
 389                 bch_dev_bkey_exists(c, ob->ptr.dev)->mi.durability;
 390
 391         __clear_bit(ob->ptr.dev, devs_may_alloc->d);
 392         *nr_effective   += (flags & BUCKET_ALLOC_USE_DURABILITY)
 393                 ? durability : 1;
 394         *have_cache     |= !durability;
 395
 396         ob_push(c, ptrs, ob);
 397 }
 398
 399 static int bch2_bucket_alloc_set(struct bch_fs *c,
 400                                  struct open_buckets *ptrs,
 401                                  struct dev_stripe_state *stripe,
 402                                  struct bch_devs_mask *devs_may_alloc,
 403                                  unsigned nr_replicas,
 404                                  unsigned *nr_effective,
 405                                  bool *have_cache,
 406                                  enum alloc_reserve reserve,
 407                                  unsigned flags,
 408                                  struct closure *cl)
 409 {
 410         struct dev_alloc_list devs_sorted =
 411                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 412         struct bch_dev *ca;
 413         bool alloc_failure = false;
 414         unsigned i;
 415
 416         BUG_ON(*nr_effective >= nr_replicas);
 417
 418         for (i = 0; i < devs_sorted.nr; i++) {
 419                 struct open_bucket *ob;
 420
 421                 ca = rcu_dereference(c->devs[devs_sorted.devs[i]]);
 422                 if (!ca)
 423                         continue;
 424
 425                 if (!ca->mi.durability && *have_cache)
 426                         continue;
 427
 428                 ob = bch2_bucket_alloc(c, ca, reserve,
 429                                 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
 430                 if (IS_ERR(ob)) {
 431                         enum bucket_alloc_ret ret = -PTR_ERR(ob);
 432
 433                         WARN_ON(reserve == RESERVE_MOVINGGC &&
 434                                 ret != OPEN_BUCKETS_EMPTY);
 435
 436                         if (cl)
 437                                 return -EAGAIN;
 438                         if (ret == OPEN_BUCKETS_EMPTY)
 439                                 return -ENOSPC;
 440                         alloc_failure = true;
 441                         continue;
 442                 }
 443
 444                 add_new_bucket(c, ptrs, devs_may_alloc,
 445                                nr_effective, have_cache, flags, ob);
 446
 447                 bch2_dev_stripe_increment(c, ca, stripe);
 448
 449                 if (*nr_effective >= nr_replicas)
 450                         return 0;
 451         }
 452
 453         return alloc_failure ? -ENOSPC : -EROFS;
 454 }
 455
 456 /* Allocate from stripes: */
 457
 458 /*
 459  * XXX: use a higher watermark for allocating open buckets here:
 460  */
 461 static int ec_stripe_alloc(struct bch_fs *c, struct ec_stripe_head *h)
 462 {
 463         struct bch_devs_mask devs;
 464         struct open_bucket *ob;
 465         unsigned i, nr_have = 0, nr_data =
 466                 min_t(unsigned, h->nr_active_devs,
 467                       EC_STRIPE_MAX) - h->redundancy;
 468         bool have_cache = true;
 469         int ret = 0;
 470
 471         BUG_ON(h->blocks.nr > nr_data);
 472         BUG_ON(h->parity.nr > h->redundancy);
 473
 474         devs = h->devs;
 475
 476         open_bucket_for_each(c, &h->parity, ob, i)
 477                 __clear_bit(ob->ptr.dev, devs.d);
 478         open_bucket_for_each(c, &h->blocks, ob, i)
 479                 __clear_bit(ob->ptr.dev, devs.d);
 480
 481         percpu_down_read(&c->mark_lock);
 482         rcu_read_lock();
 483
 484         if (h->parity.nr < h->redundancy) {
 485                 nr_have = h->parity.nr;
 486
 487                 ret = bch2_bucket_alloc_set(c, &h->parity,
 488                                             &h->parity_stripe,
 489                                             &devs,
 490                                             h->redundancy,
 491                                             &nr_have,
 492                                             &have_cache,
 493                                             RESERVE_NONE,
 494                                             0,
 495                                             NULL);
 496                 if (ret)
 497                         goto err;
 498         }
 499
 500         if (h->blocks.nr < nr_data) {
 501                 nr_have = h->blocks.nr;
 502
 503                 ret = bch2_bucket_alloc_set(c, &h->blocks,
 504                                             &h->block_stripe,
 505                                             &devs,
 506                                             nr_data,
 507                                             &nr_have,
 508                                             &have_cache,
 509                                             RESERVE_NONE,
 510                                             0,
 511                                             NULL);
 512                 if (ret)
 513                         goto err;
 514         }
 515
 516         rcu_read_unlock();
 517         percpu_up_read(&c->mark_lock);
 518
 519         return bch2_ec_stripe_new_alloc(c, h);
 520 err:
 521         rcu_read_unlock();
 522         percpu_up_read(&c->mark_lock);
 523         return -1;
 524 }
 525
 526 /*
 527  * if we can't allocate a new stripe because there are already too many
 528  * partially filled stripes, force allocating from an existing stripe even when
 529  * it's to a device we don't want:
 530  */
 531
 532 static void bucket_alloc_from_stripe(struct bch_fs *c,
 533                                      struct open_buckets *ptrs,
 534                                      struct write_point *wp,
 535                                      struct bch_devs_mask *devs_may_alloc,
 536                                      u16 target,
 537                                      unsigned erasure_code,
 538                                      unsigned nr_replicas,
 539                                      unsigned *nr_effective,
 540                                      bool *have_cache,
 541                                      unsigned flags)
 542 {
 543         struct dev_alloc_list devs_sorted;
 544         struct ec_stripe_head *h;
 545         struct open_bucket *ob;
 546         struct bch_dev *ca;
 547         unsigned i, ec_idx;
 548
 549         if (!erasure_code)
 550                 return;
 551
 552         if (nr_replicas < 2)
 553                 return;
 554
 555         if (ec_open_bucket(c, ptrs))
 556                 return;
 557
 558         h = bch2_ec_stripe_head_get(c, target, erasure_code, nr_replicas - 1);
 559         if (!h)
 560                 return;
 561
 562         if (!h->s && ec_stripe_alloc(c, h))
 563                 goto out_put_head;
 564
 565         rcu_read_lock();
 566         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 567         rcu_read_unlock();
 568
 569         for (i = 0; i < devs_sorted.nr; i++)
 570                 open_bucket_for_each(c, &h->s->blocks, ob, ec_idx)
 571                         if (ob->ptr.dev == devs_sorted.devs[i] &&
 572                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 573                                 goto got_bucket;
 574         goto out_put_head;
 575 got_bucket:
 576         ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 577
 578         ob->ec_idx      = ec_idx;
 579         ob->ec          = h->s;
 580
 581         add_new_bucket(c, ptrs, devs_may_alloc,
 582                        nr_effective, have_cache, flags, ob);
 583         atomic_inc(&h->s->pin);
 584 out_put_head:
 585         bch2_ec_stripe_head_put(h);
 586 }
 587
 588 /* Sector allocator */
 589
 590 static void get_buckets_from_writepoint(struct bch_fs *c,
 591                                         struct open_buckets *ptrs,
 592                                         struct write_point *wp,
 593                                         struct bch_devs_mask *devs_may_alloc,
 594                                         unsigned nr_replicas,
 595                                         unsigned *nr_effective,
 596                                         bool *have_cache,
 597                                         unsigned flags,
 598                                         bool need_ec)
 599 {
 600         struct open_buckets ptrs_skip = { .nr = 0 };
 601         struct open_bucket *ob;
 602         unsigned i;
 603
 604         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 605                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 606
 607                 if (*nr_effective < nr_replicas &&
 608                     test_bit(ob->ptr.dev, devs_may_alloc->d) &&
 609                     (ca->mi.durability ||
 610                      (wp->type == BCH_DATA_USER && !*have_cache)) &&
 611                     (ob->ec || !need_ec)) {
 612                         add_new_bucket(c, ptrs, devs_may_alloc,
 613                                        nr_effective, have_cache,
 614                                        flags, ob);
 615                 } else {
 616                         ob_push(c, &ptrs_skip, ob);
 617                 }
 618         }
 619         wp->ptrs = ptrs_skip;
 620 }
 621
 622 static int open_bucket_add_buckets(struct bch_fs *c,
 623                                    struct open_buckets *ptrs,
 624                                    struct write_point *wp,
 625                                    struct bch_devs_list *devs_have,
 626                                    u16 target,
 627                                    unsigned erasure_code,
 628                                    unsigned nr_replicas,
 629                                    unsigned *nr_effective,
 630                                    bool *have_cache,
 631                                    enum alloc_reserve reserve,
 632                                    unsigned flags,
 633                                    struct closure *_cl)
 634 {
 635         struct bch_devs_mask devs;
 636         struct open_bucket *ob;
 637         struct closure *cl = NULL;
 638         unsigned i;
 639         int ret;
 640
 641         rcu_read_lock();
 642         devs = target_rw_devs(c, wp->type, target);
 643         rcu_read_unlock();
 644
 645         /* Don't allocate from devices we already have pointers to: */
 646         for (i = 0; i < devs_have->nr; i++)
 647                 __clear_bit(devs_have->devs[i], devs.d);
 648
 649         open_bucket_for_each(c, ptrs, ob, i)
 650                 __clear_bit(ob->ptr.dev, devs.d);
 651
 652         if (erasure_code) {
 653                 get_buckets_from_writepoint(c, ptrs, wp, &devs,
 654                                             nr_replicas, nr_effective,
 655                                             have_cache, flags, true);
 656                 if (*nr_effective >= nr_replicas)
 657                         return 0;
 658
 659                 bucket_alloc_from_stripe(c, ptrs, wp, &devs,
 660                                          target, erasure_code,
 661                                          nr_replicas, nr_effective,
 662                                          have_cache, flags);
 663                 if (*nr_effective >= nr_replicas)
 664                         return 0;
 665         }
 666
 667         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 668                                     nr_replicas, nr_effective,
 669                                     have_cache, flags, false);
 670         if (*nr_effective >= nr_replicas)
 671                 return 0;
 672
 673         percpu_down_read(&c->mark_lock);
 674         rcu_read_lock();
 675
 676 retry_blocking:
 677         /*
 678          * Try nonblocking first, so that if one device is full we'll try from
 679          * other devices:
 680          */
 681         ret = bch2_bucket_alloc_set(c, ptrs, &wp->stripe, &devs,
 682                                 nr_replicas, nr_effective, have_cache,
 683                                 reserve, flags, cl);
 684         if (ret && ret != -EROFS && !cl && _cl) {
 685                 cl = _cl;
 686                 goto retry_blocking;
 687         }
 688
 689         rcu_read_unlock();
 690         percpu_up_read(&c->mark_lock);
 691
 692         return ret;
 693 }
 694
 695 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 696                                 struct open_buckets *obs)
 697 {
 698         struct open_buckets ptrs = { .nr = 0 };
 699         struct open_bucket *ob, *ob2;
 700         unsigned i, j;
 701
 702         open_bucket_for_each(c, obs, ob, i) {
 703                 bool drop = !ca || ob->ptr.dev == ca->dev_idx;
 704
 705                 if (!drop && ob->ec) {
 706                         mutex_lock(&ob->ec->lock);
 707                         open_bucket_for_each(c, &ob->ec->blocks, ob2, j)
 708                                 drop |= ob2->ptr.dev == ca->dev_idx;
 709                         open_bucket_for_each(c, &ob->ec->parity, ob2, j)
 710                                 drop |= ob2->ptr.dev == ca->dev_idx;
 711                         mutex_unlock(&ob->ec->lock);
 712                 }
 713
 714                 if (drop)
 715                         bch2_open_bucket_put(c, ob);
 716                 else
 717                         ob_push(c, &ptrs, ob);
 718         }
 719
 720         *obs = ptrs;
 721 }
 722
 723 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 724                           struct write_point *wp)
 725 {
 726         mutex_lock(&wp->lock);
 727         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
 728         mutex_unlock(&wp->lock);
 729 }
 730
 731 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 732                                                  unsigned long write_point)
 733 {
 734         unsigned hash =
 735                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
 736
 737         return &c->write_points_hash[hash];
 738 }
 739
 740 static struct write_point *__writepoint_find(struct hlist_head *head,
 741                                              unsigned long write_point)
 742 {
 743         struct write_point *wp;
 744
 745         hlist_for_each_entry_rcu(wp, head, node)
 746                 if (wp->write_point == write_point)
 747                         return wp;
 748
 749         return NULL;
 750 }
 751
 752 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
 753 {
 754         u64 stranded    = c->write_points_nr * c->bucket_size_max;
 755         u64 free        = bch2_fs_usage_read_short(c).free;
 756
 757         return stranded * factor > free;
 758 }
 759
 760 static bool try_increase_writepoints(struct bch_fs *c)
 761 {
 762         struct write_point *wp;
 763
 764         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
 765             too_many_writepoints(c, 32))
 766                 return false;
 767
 768         wp = c->write_points + c->write_points_nr++;
 769         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
 770         return true;
 771 }
 772
 773 static bool try_decrease_writepoints(struct bch_fs *c,
 774                                      unsigned old_nr)
 775 {
 776         struct write_point *wp;
 777
 778         mutex_lock(&c->write_points_hash_lock);
 779         if (c->write_points_nr < old_nr) {
 780                 mutex_unlock(&c->write_points_hash_lock);
 781                 return true;
 782         }
 783
 784         if (c->write_points_nr == 1 ||
 785             !too_many_writepoints(c, 8)) {
 786                 mutex_unlock(&c->write_points_hash_lock);
 787                 return false;
 788         }
 789
 790         wp = c->write_points + --c->write_points_nr;
 791
 792         hlist_del_rcu(&wp->node);
 793         mutex_unlock(&c->write_points_hash_lock);
 794
 795         bch2_writepoint_stop(c, NULL, wp);
 796         return true;
 797 }
 798
 799 static struct write_point *writepoint_find(struct bch_fs *c,
 800                                            unsigned long write_point)
 801 {
 802         struct write_point *wp, *oldest;
 803         struct hlist_head *head;
 804
 805         if (!(write_point & 1UL)) {
 806                 wp = (struct write_point *) write_point;
 807                 mutex_lock(&wp->lock);
 808                 return wp;
 809         }
 810
 811         head = writepoint_hash(c, write_point);
 812 restart_find:
 813         wp = __writepoint_find(head, write_point);
 814         if (wp) {
 815 lock_wp:
 816                 mutex_lock(&wp->lock);
 817                 if (wp->write_point == write_point)
 818                         goto out;
 819                 mutex_unlock(&wp->lock);
 820                 goto restart_find;
 821         }
 822 restart_find_oldest:
 823         oldest = NULL;
 824         for (wp = c->write_points;
 825              wp < c->write_points + c->write_points_nr; wp++)
 826                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
 827                         oldest = wp;
 828
 829         mutex_lock(&oldest->lock);
 830         mutex_lock(&c->write_points_hash_lock);
 831         if (oldest >= c->write_points + c->write_points_nr ||
 832             try_increase_writepoints(c)) {
 833                 mutex_unlock(&c->write_points_hash_lock);
 834                 mutex_unlock(&oldest->lock);
 835                 goto restart_find_oldest;
 836         }
 837
 838         wp = __writepoint_find(head, write_point);
 839         if (wp && wp != oldest) {
 840                 mutex_unlock(&c->write_points_hash_lock);
 841                 mutex_unlock(&oldest->lock);
 842                 goto lock_wp;
 843         }
 844
 845         wp = oldest;
 846         hlist_del_rcu(&wp->node);
 847         wp->write_point = write_point;
 848         hlist_add_head_rcu(&wp->node, head);
 849         mutex_unlock(&c->write_points_hash_lock);
 850 out:
 851         wp->last_used = sched_clock();
 852         return wp;
 853 }
 854
 855 /*
 856  * Get us an open_bucket we can allocate from, return with it locked:
 857  */
 858 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
 859                                 unsigned target,
 860                                 unsigned erasure_code,
 861                                 struct write_point_specifier write_point,
 862                                 struct bch_devs_list *devs_have,
 863                                 unsigned nr_replicas,
 864                                 unsigned nr_replicas_required,
 865                                 enum alloc_reserve reserve,
 866                                 unsigned flags,
 867                                 struct closure *cl)
 868 {
 869         struct write_point *wp;
 870         struct open_bucket *ob;
 871         struct open_buckets ptrs;
 872         unsigned nr_effective, write_points_nr;
 873         unsigned ob_flags = 0;
 874         bool have_cache;
 875         int ret, i;
 876
 877         if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
 878                 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
 879
 880         BUG_ON(!nr_replicas || !nr_replicas_required);
 881 retry:
 882         ptrs.nr         = 0;
 883         nr_effective    = 0;
 884         write_points_nr = c->write_points_nr;
 885         have_cache      = false;
 886
 887         wp = writepoint_find(c, write_point.v);
 888
 889         if (wp->type == BCH_DATA_USER)
 890                 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
 891
 892         /* metadata may not allocate on cache devices: */
 893         if (wp->type != BCH_DATA_USER)
 894                 have_cache = true;
 895
 896         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
 897                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 898                                               target, erasure_code,
 899                                               nr_replicas, &nr_effective,
 900                                               &have_cache, reserve,
 901                                               ob_flags, cl);
 902         } else {
 903                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 904                                               target, erasure_code,
 905                                               nr_replicas, &nr_effective,
 906                                               &have_cache, reserve,
 907                                               ob_flags, NULL);
 908                 if (!ret)
 909                         goto alloc_done;
 910
 911                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
 912                                               0, erasure_code,
 913                                               nr_replicas, &nr_effective,
 914                                               &have_cache, reserve,
 915                                               ob_flags, cl);
 916         }
 917 alloc_done:
 918         BUG_ON(!ret && nr_effective < nr_replicas);
 919
 920         if (erasure_code && !ec_open_bucket(c, &ptrs))
 921                 pr_debug("failed to get ec bucket: ret %u", ret);
 922
 923         if (ret == -EROFS &&
 924             nr_effective >= nr_replicas_required)
 925                 ret = 0;
 926
 927         if (ret)
 928                 goto err;
 929
 930         /* Free buckets we didn't use: */
 931         open_bucket_for_each(c, &wp->ptrs, ob, i)
 932                 open_bucket_free_unused(c, ob, wp->type == BCH_DATA_USER);
 933
 934         wp->ptrs = ptrs;
 935
 936         wp->sectors_free = UINT_MAX;
 937
 938         open_bucket_for_each(c, &wp->ptrs, ob, i)
 939                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
 940
 941         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
 942
 943         verify_not_stale(c, &wp->ptrs);
 944
 945         return wp;
 946 err:
 947         open_bucket_for_each(c, &wp->ptrs, ob, i)
 948                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
 949                         ob_push(c, &ptrs, ob);
 950                 else
 951                         open_bucket_free_unused(c, ob,
 952                                         wp->type == BCH_DATA_USER);
 953         wp->ptrs = ptrs;
 954
 955         mutex_unlock(&wp->lock);
 956
 957         if (ret == -ENOSPC &&
 958             try_decrease_writepoints(c, write_points_nr))
 959                 goto retry;
 960
 961         return ERR_PTR(ret);
 962 }
 963
 964 /*
 965  * Append pointers to the space we just allocated to @k, and mark @sectors space
 966  * as allocated out of @ob
 967  */
 968 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
 969                                     struct bkey_i *k, unsigned sectors)
 970
 971 {
 972         struct open_bucket *ob;
 973         unsigned i;
 974
 975         BUG_ON(sectors > wp->sectors_free);
 976         wp->sectors_free -= sectors;
 977
 978         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 979                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->ptr.dev);
 980                 struct bch_extent_ptr tmp = ob->ptr;
 981
 982                 tmp.cached = !ca->mi.durability &&
 983                         wp->type == BCH_DATA_USER;
 984
 985                 tmp.offset += ca->mi.bucket_size - ob->sectors_free;
 986                 bch2_bkey_append_ptr(k, tmp);
 987
 988                 BUG_ON(sectors > ob->sectors_free);
 989                 ob->sectors_free -= sectors;
 990         }
 991 }
 992
 993 /*
 994  * Append pointers to the space we just allocated to @k, and mark @sectors space
 995  * as allocated out of @ob
 996  */
 997 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
 998 {
 999         struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
1000         struct open_bucket *ob;
1001         unsigned i;
1002
1003         open_bucket_for_each(c, &wp->ptrs, ob, i)
1004                 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
1005         wp->ptrs = keep;
1006
1007         mutex_unlock(&wp->lock);
1008
1009         bch2_open_buckets_put(c, &ptrs);
1010 }
1011
1012 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1013 {
1014         struct open_bucket *ob;
1015         struct write_point *wp;
1016
1017         mutex_init(&c->write_points_hash_lock);
1018         c->write_points_nr = ARRAY_SIZE(c->write_points);
1019
1020         /* open bucket 0 is a sentinal NULL: */
1021         spin_lock_init(&c->open_buckets[0].lock);
1022
1023         for (ob = c->open_buckets + 1;
1024              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1025                 spin_lock_init(&ob->lock);
1026                 c->open_buckets_nr_free++;
1027
1028                 ob->freelist = c->open_buckets_freelist;
1029                 c->open_buckets_freelist = ob - c->open_buckets;
1030         }
1031
1032         writepoint_init(&c->btree_write_point, BCH_DATA_BTREE);
1033         writepoint_init(&c->rebalance_write_point, BCH_DATA_USER);
1034
1035         for (wp = c->write_points;
1036              wp < c->write_points + c->write_points_nr; wp++) {
1037                 writepoint_init(wp, BCH_DATA_USER);
1038
1039                 wp->last_used   = sched_clock();
1040                 wp->write_point = (unsigned long) wp;
1041                 hlist_add_head_rcu(&wp->node,
1042                                    writepoint_hash(c, wp->write_point));
1043         }
1044 }