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

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright 2012 Google, Inc.
   4  *
   5  * Foreground allocator code: allocate buckets from freelist, and allocate in
   6  * sector granularity from writepoints.
   7  *
   8  * bch2_bucket_alloc() allocates a single bucket from a specific device.
   9  *
  10  * bch2_bucket_alloc_set() allocates one or more buckets from different devices
  11  * in a given filesystem.
  12  */
  13
  14 #include "bcachefs.h"
  15 #include "alloc_background.h"
  16 #include "alloc_foreground.h"
  17 #include "btree_iter.h"
  18 #include "btree_update.h"
  19 #include "btree_gc.h"
  20 #include "buckets.h"
  21 #include "buckets_waiting_for_journal.h"
  22 #include "clock.h"
  23 #include "debug.h"
  24 #include "disk_groups.h"
  25 #include "ec.h"
  26 #include "error.h"
  27 #include "io.h"
  28 #include "journal.h"
  29
  30 #include <linux/math64.h>
  31 #include <linux/rculist.h>
  32 #include <linux/rcupdate.h>
  33 #include <trace/events/bcachefs.h>
  34
  35 const char * const bch2_alloc_reserves[] = {
  36 #define x(t) #t,
  37         BCH_ALLOC_RESERVES()
  38 #undef x
  39         NULL
  40 };
  41
  42 /*
  43  * Open buckets represent a bucket that's currently being allocated from.  They
  44  * serve two purposes:
  45  *
  46  *  - They track buckets that have been partially allocated, allowing for
  47  *    sub-bucket sized allocations - they're used by the sector allocator below
  48  *
  49  *  - They provide a reference to the buckets they own that mark and sweep GC
  50  *    can find, until the new allocation has a pointer to it inserted into the
  51  *    btree
  52  *
  53  * When allocating some space with the sector allocator, the allocation comes
  54  * with a reference to an open bucket - the caller is required to put that
  55  * reference _after_ doing the index update that makes its allocation reachable.
  56  */
  57
  58 static void bch2_open_bucket_hash_add(struct bch_fs *c, struct open_bucket *ob)
  59 {
  60         open_bucket_idx_t idx = ob - c->open_buckets;
  61         open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
  62
  63         ob->hash = *slot;
  64         *slot = idx;
  65 }
  66
  67 static void bch2_open_bucket_hash_remove(struct bch_fs *c, struct open_bucket *ob)
  68 {
  69         open_bucket_idx_t idx = ob - c->open_buckets;
  70         open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
  71
  72         while (*slot != idx) {
  73                 BUG_ON(!*slot);
  74                 slot = &c->open_buckets[*slot].hash;
  75         }
  76
  77         *slot = ob->hash;
  78         ob->hash = 0;
  79 }
  80
  81 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
  82 {
  83         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
  84
  85         if (ob->ec) {
  86                 bch2_ec_bucket_written(c, ob);
  87                 return;
  88         }
  89
  90         percpu_down_read(&c->mark_lock);
  91         spin_lock(&ob->lock);
  92
  93         ob->valid = false;
  94         ob->data_type = 0;
  95
  96         spin_unlock(&ob->lock);
  97         percpu_up_read(&c->mark_lock);
  98
  99         spin_lock(&c->freelist_lock);
 100         bch2_open_bucket_hash_remove(c, ob);
 101
 102         ob->freelist = c->open_buckets_freelist;
 103         c->open_buckets_freelist = ob - c->open_buckets;
 104
 105         c->open_buckets_nr_free++;
 106         ca->nr_open_buckets--;
 107         spin_unlock(&c->freelist_lock);
 108
 109         closure_wake_up(&c->open_buckets_wait);
 110 }
 111
 112 void bch2_open_bucket_write_error(struct bch_fs *c,
 113                                   struct open_buckets *obs,
 114                                   unsigned dev)
 115 {
 116         struct open_bucket *ob;
 117         unsigned i;
 118
 119         open_bucket_for_each(c, obs, ob, i)
 120                 if (ob->dev == dev && ob->ec)
 121                         bch2_ec_bucket_cancel(c, ob);
 122 }
 123
 124 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
 125 {
 126         struct open_bucket *ob;
 127
 128         BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
 129
 130         ob = c->open_buckets + c->open_buckets_freelist;
 131         c->open_buckets_freelist = ob->freelist;
 132         atomic_set(&ob->pin, 1);
 133         ob->data_type = 0;
 134
 135         c->open_buckets_nr_free--;
 136         return ob;
 137 }
 138
 139 static void open_bucket_free_unused(struct bch_fs *c,
 140                                     struct write_point *wp,
 141                                     struct open_bucket *ob)
 142 {
 143         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 144         bool may_realloc = wp->data_type == BCH_DATA_user;
 145
 146         BUG_ON(ca->open_buckets_partial_nr >
 147                ARRAY_SIZE(ca->open_buckets_partial));
 148
 149         if (ca->open_buckets_partial_nr <
 150             ARRAY_SIZE(ca->open_buckets_partial) &&
 151             may_realloc) {
 152                 spin_lock(&c->freelist_lock);
 153                 ob->on_partial_list = true;
 154                 ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
 155                         ob - c->open_buckets;
 156                 spin_unlock(&c->freelist_lock);
 157
 158                 closure_wake_up(&c->open_buckets_wait);
 159                 closure_wake_up(&c->freelist_wait);
 160         } else {
 161                 bch2_open_bucket_put(c, ob);
 162         }
 163 }
 164
 165 /* _only_ for allocating the journal on a new device: */
 166 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
 167 {
 168         while (ca->new_fs_bucket_idx < ca->mi.nbuckets) {
 169                 u64 b = ca->new_fs_bucket_idx++;
 170
 171                 if (!is_superblock_bucket(ca, b) &&
 172                     (!ca->buckets_nouse || !test_bit(b, ca->buckets_nouse)))
 173                         return b;
 174         }
 175
 176         return -1;
 177 }
 178
 179 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
 180 {
 181         switch (reserve) {
 182         case RESERVE_btree:
 183         case RESERVE_btree_movinggc:
 184                 return 0;
 185         case RESERVE_movinggc:
 186                 return OPEN_BUCKETS_COUNT / 4;
 187         default:
 188                 return OPEN_BUCKETS_COUNT / 2;
 189         }
 190 }
 191
 192 static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 193                                               enum alloc_reserve reserve,
 194                                               struct bkey_alloc_unpacked a,
 195                                               size_t *need_journal_commit,
 196                                               struct closure *cl)
 197 {
 198         struct open_bucket *ob;
 199
 200         if (unlikely(ca->buckets_nouse && test_bit(a.bucket, ca->buckets_nouse)))
 201                 return NULL;
 202
 203         if (bch2_bucket_is_open(c, ca->dev_idx, a.bucket))
 204                 return NULL;
 205
 206         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 207                         c->journal.flushed_seq_ondisk, ca->dev_idx, a.bucket)) {
 208                 (*need_journal_commit)++;
 209                 return NULL;
 210         }
 211
 212         spin_lock(&c->freelist_lock);
 213
 214         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
 215                 if (cl)
 216                         closure_wait(&c->open_buckets_wait, cl);
 217
 218                 if (!c->blocked_allocate_open_bucket)
 219                         c->blocked_allocate_open_bucket = local_clock();
 220
 221                 spin_unlock(&c->freelist_lock);
 222
 223                 trace_open_bucket_alloc_fail(ca, bch2_alloc_reserves[reserve]);
 224                 return ERR_PTR(-OPEN_BUCKETS_EMPTY);
 225         }
 226
 227         /* Recheck under lock: */
 228         if (bch2_bucket_is_open(c, ca->dev_idx, a.bucket)) {
 229                 spin_unlock(&c->freelist_lock);
 230                 return NULL;
 231         }
 232
 233         ob = bch2_open_bucket_alloc(c);
 234
 235         spin_lock(&ob->lock);
 236
 237         ob->valid       = true;
 238         ob->sectors_free = ca->mi.bucket_size;
 239         ob->alloc_reserve = reserve;
 240         ob->dev         = ca->dev_idx;
 241         ob->gen         = a.gen;
 242         ob->bucket      = a.bucket;
 243         spin_unlock(&ob->lock);
 244
 245         ca->nr_open_buckets++;
 246         bch2_open_bucket_hash_add(c, ob);
 247
 248         if (c->blocked_allocate_open_bucket) {
 249                 bch2_time_stats_update(
 250                         &c->times[BCH_TIME_blocked_allocate_open_bucket],
 251                         c->blocked_allocate_open_bucket);
 252                 c->blocked_allocate_open_bucket = 0;
 253         }
 254
 255         if (c->blocked_allocate) {
 256                 bch2_time_stats_update(
 257                         &c->times[BCH_TIME_blocked_allocate],
 258                         c->blocked_allocate);
 259                 c->blocked_allocate = 0;
 260         }
 261
 262         spin_unlock(&c->freelist_lock);
 263
 264         trace_bucket_alloc(ca, bch2_alloc_reserves[reserve]);
 265         return ob;
 266 }
 267
 268 static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bch_dev *ca,
 269                                             enum alloc_reserve reserve, u64 free_entry,
 270                                             size_t *need_journal_commit,
 271                                             struct closure *cl)
 272 {
 273         struct bch_fs *c = trans->c;
 274         struct btree_iter iter;
 275         struct bkey_s_c k;
 276         struct open_bucket *ob;
 277         struct bkey_alloc_unpacked a;
 278         u64 b = free_entry & ~(~0ULL << 56);
 279         unsigned genbits = free_entry >> 56;
 280         struct printbuf buf = PRINTBUF;
 281         int ret;
 282
 283         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS(ca->dev_idx, b), BTREE_ITER_CACHED);
 284         k = bch2_btree_iter_peek_slot(&iter);
 285         ret = bkey_err(k);
 286         if (ret) {
 287                 ob = ERR_PTR(ret);
 288                 goto err;
 289         }
 290
 291         a = bch2_alloc_unpack(k);
 292
 293         if (bch2_fs_inconsistent_on(bucket_state(a) != BUCKET_free, c,
 294                         "non free bucket in freespace btree (state %s)\n"
 295                         "  %s\n"
 296                         "  at %llu (genbits %u)",
 297                         bch2_bucket_states[bucket_state(a)],
 298                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
 299                         free_entry, genbits)) {
 300                 ob = ERR_PTR(-EIO);
 301                 goto err;
 302         }
 303
 304         if (bch2_fs_inconsistent_on(genbits != (alloc_freespace_genbits(a) >> 56), c,
 305                         "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
 306                         "  %s",
 307                         genbits, alloc_freespace_genbits(a) >> 56,
 308                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 309                 ob = ERR_PTR(-EIO);
 310                 goto err;
 311         }
 312
 313         if (bch2_fs_inconsistent_on(b < ca->mi.first_bucket || b >= ca->mi.nbuckets, c,
 314                         "freespace btree has bucket outside allowed range (got %llu, valid %u-%llu)",
 315                         b, ca->mi.first_bucket, ca->mi.nbuckets)) {
 316                 ob = ERR_PTR(-EIO);
 317                 goto err;
 318         }
 319
 320         ob = __try_alloc_bucket(c, ca, reserve, a, need_journal_commit, cl);
 321 err:
 322         bch2_trans_iter_exit(trans, &iter);
 323         printbuf_exit(&buf);
 324         return ob;
 325 }
 326
 327 static struct open_bucket *try_alloc_partial_bucket(struct bch_fs *c, struct bch_dev *ca,
 328                                                     enum alloc_reserve reserve)
 329 {
 330         struct open_bucket *ob;
 331         int i;
 332
 333         spin_lock(&c->freelist_lock);
 334
 335         for (i = ca->open_buckets_partial_nr - 1; i >= 0; --i) {
 336                 ob = c->open_buckets + ca->open_buckets_partial[i];
 337
 338                 if (reserve <= ob->alloc_reserve) {
 339                         array_remove_item(ca->open_buckets_partial,
 340                                           ca->open_buckets_partial_nr,
 341                                           i);
 342                         ob->on_partial_list = false;
 343                         ob->alloc_reserve = reserve;
 344                         spin_unlock(&c->freelist_lock);
 345                         return ob;
 346                 }
 347         }
 348
 349         spin_unlock(&c->freelist_lock);
 350         return NULL;
 351 }
 352
 353 /*
 354  * This path is for before the freespace btree is initialized:
 355  *
 356  * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
 357  * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
 358  */
 359 static noinline struct open_bucket *
 360 bch2_bucket_alloc_trans_early(struct btree_trans *trans,
 361                               struct bch_dev *ca,
 362                               enum alloc_reserve reserve,
 363                               u64 *b,
 364                               size_t *need_journal_commit,
 365                               struct closure *cl)
 366 {
 367         struct btree_iter iter;
 368         struct bkey_s_c k;
 369         struct open_bucket *ob = NULL;
 370         int ret;
 371
 372         *b = max_t(u64, *b, ca->mi.first_bucket);
 373         *b = max_t(u64, *b, ca->new_fs_bucket_idx);
 374
 375         for_each_btree_key(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, *b),
 376                            BTREE_ITER_SLOTS, k, ret) {
 377                 struct bkey_alloc_unpacked a;
 378
 379                 if (bkey_cmp(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)) >= 0)
 380                         break;
 381
 382                 if (ca->new_fs_bucket_idx &&
 383                     is_superblock_bucket(ca, k.k->p.offset))
 384                         continue;
 385
 386                 a = bch2_alloc_unpack(k);
 387
 388                 if (bucket_state(a) != BUCKET_free)
 389                         continue;
 390
 391                 ob = __try_alloc_bucket(trans->c, ca, reserve, a,
 392                                         need_journal_commit, cl);
 393                 if (ob)
 394                         break;
 395         }
 396         bch2_trans_iter_exit(trans, &iter);
 397
 398         *b = iter.pos.offset;
 399
 400         return ob ?: ERR_PTR(ret ?: -FREELIST_EMPTY);
 401 }
 402
 403 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
 404                                                    struct bch_dev *ca,
 405                                                    enum alloc_reserve reserve,
 406                                                    u64 *b,
 407                                                    size_t *need_journal_commit,
 408                                                    struct closure *cl)
 409 {
 410         struct btree_iter iter;
 411         struct bkey_s_c k;
 412         struct open_bucket *ob = NULL;
 413         int ret;
 414
 415         if (unlikely(!ca->mi.freespace_initialized))
 416                 return bch2_bucket_alloc_trans_early(trans, ca, reserve, b,
 417                                                      need_journal_commit, cl);
 418
 419         BUG_ON(ca->new_fs_bucket_idx);
 420
 421         for_each_btree_key(trans, iter, BTREE_ID_freespace,
 422                            POS(ca->dev_idx, *b), 0, k, ret) {
 423                 if (k.k->p.inode != ca->dev_idx)
 424                         break;
 425
 426                 for (*b = max(*b, bkey_start_offset(k.k));
 427                      *b != k.k->p.offset && !ob;
 428                      (*b)++) {
 429                         if (btree_trans_too_many_iters(trans)) {
 430                                 ob = ERR_PTR(-EINTR);
 431                                 break;
 432                         }
 433
 434                         ob = try_alloc_bucket(trans, ca, reserve, *b,
 435                                               need_journal_commit, cl);
 436                 }
 437                 if (ob)
 438                         break;
 439         }
 440         bch2_trans_iter_exit(trans, &iter);
 441
 442         return ob ?: ERR_PTR(ret);
 443 }
 444
 445 /**
 446  * bch_bucket_alloc - allocate a single bucket from a specific device
 447  *
 448  * Returns index of bucket on success, 0 on failure
 449  * */
 450 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 451                                       enum alloc_reserve reserve,
 452                                       bool may_alloc_partial,
 453                                       struct closure *cl)
 454 {
 455         struct open_bucket *ob = NULL;
 456         size_t need_journal_commit = 0;
 457         u64 avail = dev_buckets_available(ca, reserve);
 458         u64 b = 0;
 459         int ret;
 460
 461         if (may_alloc_partial) {
 462                 ob = try_alloc_partial_bucket(c, ca, reserve);
 463                 if (ob)
 464                         return ob;
 465         }
 466 again:
 467         if (!avail) {
 468                 if (cl) {
 469                         closure_wait(&c->freelist_wait, cl);
 470                         /* recheck after putting ourself on waitlist */
 471                         avail = dev_buckets_available(ca, reserve);
 472                         if (avail) {
 473                                 closure_wake_up(&c->freelist_wait);
 474                                 goto again;
 475                         }
 476                 }
 477
 478                 if (!c->blocked_allocate)
 479                         c->blocked_allocate = local_clock();
 480
 481                 ob = ERR_PTR(-FREELIST_EMPTY);
 482                 goto err;
 483         }
 484
 485         ret = bch2_trans_do(c, NULL, NULL, 0,
 486                         PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans,
 487                                                         ca, reserve, &b,
 488                                                         &need_journal_commit, cl)));
 489
 490         if (need_journal_commit * 2 > avail)
 491                 bch2_journal_flush_async(&c->journal, NULL);
 492 err:
 493         if (!ob)
 494                 ob = ERR_PTR(ret ?: -FREELIST_EMPTY);
 495
 496         if (ob == ERR_PTR(-FREELIST_EMPTY)) {
 497                 trace_bucket_alloc_fail(ca, bch2_alloc_reserves[reserve], avail,
 498                                         need_journal_commit, cl == NULL);
 499                 atomic_long_inc(&c->bucket_alloc_fail);
 500         }
 501
 502         return ob;
 503 }
 504
 505 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 506                             unsigned l, unsigned r)
 507 {
 508         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 509                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 510 }
 511
 512 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 513
 514 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 515                                           struct dev_stripe_state *stripe,
 516                                           struct bch_devs_mask *devs)
 517 {
 518         struct dev_alloc_list ret = { .nr = 0 };
 519         unsigned i;
 520
 521         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 522                 ret.devs[ret.nr++] = i;
 523
 524         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 525         return ret;
 526 }
 527
 528 void bch2_dev_stripe_increment(struct bch_dev *ca,
 529                                struct dev_stripe_state *stripe)
 530 {
 531         u64 *v = stripe->next_alloc + ca->dev_idx;
 532         u64 free_space = dev_buckets_available(ca, RESERVE_none);
 533         u64 free_space_inv = free_space
 534                 ? div64_u64(1ULL << 48, free_space)
 535                 : 1ULL << 48;
 536         u64 scale = *v / 4;
 537
 538         if (*v + free_space_inv >= *v)
 539                 *v += free_space_inv;
 540         else
 541                 *v = U64_MAX;
 542
 543         for (v = stripe->next_alloc;
 544              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 545                 *v = *v < scale ? 0 : *v - scale;
 546 }
 547
 548 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 549 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 550
 551 static void add_new_bucket(struct bch_fs *c,
 552                            struct open_buckets *ptrs,
 553                            struct bch_devs_mask *devs_may_alloc,
 554                            unsigned *nr_effective,
 555                            bool *have_cache,
 556                            unsigned flags,
 557                            struct open_bucket *ob)
 558 {
 559         unsigned durability =
 560                 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
 561
 562         __clear_bit(ob->dev, devs_may_alloc->d);
 563         *nr_effective   += (flags & BUCKET_ALLOC_USE_DURABILITY)
 564                 ? durability : 1;
 565         *have_cache     |= !durability;
 566
 567         ob_push(c, ptrs, ob);
 568 }
 569
 570 int bch2_bucket_alloc_set(struct bch_fs *c,
 571                       struct open_buckets *ptrs,
 572                       struct dev_stripe_state *stripe,
 573                       struct bch_devs_mask *devs_may_alloc,
 574                       unsigned nr_replicas,
 575                       unsigned *nr_effective,
 576                       bool *have_cache,
 577                       enum alloc_reserve reserve,
 578                       unsigned flags,
 579                       struct closure *cl)
 580 {
 581         struct dev_alloc_list devs_sorted =
 582                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 583         unsigned dev;
 584         struct bch_dev *ca;
 585         int ret = -INSUFFICIENT_DEVICES;
 586         unsigned i;
 587
 588         BUG_ON(*nr_effective >= nr_replicas);
 589
 590         for (i = 0; i < devs_sorted.nr; i++) {
 591                 struct open_bucket *ob;
 592
 593                 dev = devs_sorted.devs[i];
 594
 595                 rcu_read_lock();
 596                 ca = rcu_dereference(c->devs[dev]);
 597                 if (ca)
 598                         percpu_ref_get(&ca->ref);
 599                 rcu_read_unlock();
 600
 601                 if (!ca)
 602                         continue;
 603
 604                 if (!ca->mi.durability && *have_cache) {
 605                         percpu_ref_put(&ca->ref);
 606                         continue;
 607                 }
 608
 609                 ob = bch2_bucket_alloc(c, ca, reserve,
 610                                 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
 611                 if (!IS_ERR(ob))
 612                         bch2_dev_stripe_increment(ca, stripe);
 613                 percpu_ref_put(&ca->ref);
 614
 615                 if (IS_ERR(ob)) {
 616                         ret = PTR_ERR(ob);
 617
 618                         if (cl)
 619                                 break;
 620                         continue;
 621                 }
 622
 623                 add_new_bucket(c, ptrs, devs_may_alloc,
 624                                nr_effective, have_cache, flags, ob);
 625
 626                 if (*nr_effective >= nr_replicas) {
 627                         ret = 0;
 628                         break;
 629                 }
 630         }
 631
 632         return ret;
 633 }
 634
 635 /* Allocate from stripes: */
 636
 637 /*
 638  * if we can't allocate a new stripe because there are already too many
 639  * partially filled stripes, force allocating from an existing stripe even when
 640  * it's to a device we don't want:
 641  */
 642
 643 static int bucket_alloc_from_stripe(struct bch_fs *c,
 644                          struct open_buckets *ptrs,
 645                          struct write_point *wp,
 646                          struct bch_devs_mask *devs_may_alloc,
 647                          u16 target,
 648                          unsigned erasure_code,
 649                          unsigned nr_replicas,
 650                          unsigned *nr_effective,
 651                          bool *have_cache,
 652                          unsigned flags,
 653                          struct closure *cl)
 654 {
 655         struct dev_alloc_list devs_sorted;
 656         struct ec_stripe_head *h;
 657         struct open_bucket *ob;
 658         struct bch_dev *ca;
 659         unsigned i, ec_idx;
 660
 661         if (!erasure_code)
 662                 return 0;
 663
 664         if (nr_replicas < 2)
 665                 return 0;
 666
 667         if (ec_open_bucket(c, ptrs))
 668                 return 0;
 669
 670         h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1,
 671                                     wp == &c->copygc_write_point,
 672                                     cl);
 673         if (IS_ERR(h))
 674                 return -PTR_ERR(h);
 675         if (!h)
 676                 return 0;
 677
 678         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 679
 680         for (i = 0; i < devs_sorted.nr; i++)
 681                 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 682                         if (!h->s->blocks[ec_idx])
 683                                 continue;
 684
 685                         ob = c->open_buckets + h->s->blocks[ec_idx];
 686                         if (ob->dev == devs_sorted.devs[i] &&
 687                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 688                                 goto got_bucket;
 689                 }
 690         goto out_put_head;
 691 got_bucket:
 692         ca = bch_dev_bkey_exists(c, ob->dev);
 693
 694         ob->ec_idx      = ec_idx;
 695         ob->ec          = h->s;
 696
 697         add_new_bucket(c, ptrs, devs_may_alloc,
 698                        nr_effective, have_cache, flags, ob);
 699         atomic_inc(&h->s->pin);
 700 out_put_head:
 701         bch2_ec_stripe_head_put(c, h);
 702         return 0;
 703 }
 704
 705 /* Sector allocator */
 706
 707 static void get_buckets_from_writepoint(struct bch_fs *c,
 708                                         struct open_buckets *ptrs,
 709                                         struct write_point *wp,
 710                                         struct bch_devs_mask *devs_may_alloc,
 711                                         unsigned nr_replicas,
 712                                         unsigned *nr_effective,
 713                                         bool *have_cache,
 714                                         unsigned flags,
 715                                         bool need_ec)
 716 {
 717         struct open_buckets ptrs_skip = { .nr = 0 };
 718         struct open_bucket *ob;
 719         unsigned i;
 720
 721         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 722                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 723
 724                 if (*nr_effective < nr_replicas &&
 725                     test_bit(ob->dev, devs_may_alloc->d) &&
 726                     (ca->mi.durability ||
 727                      (wp->data_type == BCH_DATA_user && !*have_cache)) &&
 728                     (ob->ec || !need_ec)) {
 729                         add_new_bucket(c, ptrs, devs_may_alloc,
 730                                        nr_effective, have_cache,
 731                                        flags, ob);
 732                 } else {
 733                         ob_push(c, &ptrs_skip, ob);
 734                 }
 735         }
 736         wp->ptrs = ptrs_skip;
 737 }
 738
 739 static int open_bucket_add_buckets(struct bch_fs *c,
 740                         struct open_buckets *ptrs,
 741                         struct write_point *wp,
 742                         struct bch_devs_list *devs_have,
 743                         u16 target,
 744                         unsigned erasure_code,
 745                         unsigned nr_replicas,
 746                         unsigned *nr_effective,
 747                         bool *have_cache,
 748                         enum alloc_reserve reserve,
 749                         unsigned flags,
 750                         struct closure *_cl)
 751 {
 752         struct bch_devs_mask devs;
 753         struct open_bucket *ob;
 754         struct closure *cl = NULL;
 755         int ret;
 756         unsigned i;
 757
 758         rcu_read_lock();
 759         devs = target_rw_devs(c, wp->data_type, target);
 760         rcu_read_unlock();
 761
 762         /* Don't allocate from devices we already have pointers to: */
 763         for (i = 0; i < devs_have->nr; i++)
 764                 __clear_bit(devs_have->devs[i], devs.d);
 765
 766         open_bucket_for_each(c, ptrs, ob, i)
 767                 __clear_bit(ob->dev, devs.d);
 768
 769         if (erasure_code) {
 770                 if (!ec_open_bucket(c, ptrs)) {
 771                         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 772                                                     nr_replicas, nr_effective,
 773                                                     have_cache, flags, true);
 774                         if (*nr_effective >= nr_replicas)
 775                                 return 0;
 776                 }
 777
 778                 if (!ec_open_bucket(c, ptrs)) {
 779                         ret = bucket_alloc_from_stripe(c, ptrs, wp, &devs,
 780                                                  target, erasure_code,
 781                                                  nr_replicas, nr_effective,
 782                                                  have_cache, flags, _cl);
 783                         if (ret == -FREELIST_EMPTY ||
 784                             ret == -OPEN_BUCKETS_EMPTY)
 785                                 return ret;
 786                         if (*nr_effective >= nr_replicas)
 787                                 return 0;
 788                 }
 789         }
 790
 791         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 792                                     nr_replicas, nr_effective,
 793                                     have_cache, flags, false);
 794         if (*nr_effective >= nr_replicas)
 795                 return 0;
 796
 797 retry_blocking:
 798         /*
 799          * Try nonblocking first, so that if one device is full we'll try from
 800          * other devices:
 801          */
 802         ret = bch2_bucket_alloc_set(c, ptrs, &wp->stripe, &devs,
 803                                 nr_replicas, nr_effective, have_cache,
 804                                 reserve, flags, cl);
 805         if (ret && ret != -INSUFFICIENT_DEVICES && !cl && _cl) {
 806                 cl = _cl;
 807                 goto retry_blocking;
 808         }
 809
 810         return ret;
 811 }
 812
 813 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 814                                 struct open_buckets *obs)
 815 {
 816         struct open_buckets ptrs = { .nr = 0 };
 817         struct open_bucket *ob, *ob2;
 818         unsigned i, j;
 819
 820         open_bucket_for_each(c, obs, ob, i) {
 821                 bool drop = !ca || ob->dev == ca->dev_idx;
 822
 823                 if (!drop && ob->ec) {
 824                         mutex_lock(&ob->ec->lock);
 825                         for (j = 0; j < ob->ec->new_stripe.key.v.nr_blocks; j++) {
 826                                 if (!ob->ec->blocks[j])
 827                                         continue;
 828
 829                                 ob2 = c->open_buckets + ob->ec->blocks[j];
 830                                 drop |= ob2->dev == ca->dev_idx;
 831                         }
 832                         mutex_unlock(&ob->ec->lock);
 833                 }
 834
 835                 if (drop)
 836                         bch2_open_bucket_put(c, ob);
 837                 else
 838                         ob_push(c, &ptrs, ob);
 839         }
 840
 841         *obs = ptrs;
 842 }
 843
 844 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 845                           struct write_point *wp)
 846 {
 847         mutex_lock(&wp->lock);
 848         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
 849         mutex_unlock(&wp->lock);
 850 }
 851
 852 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 853                                                  unsigned long write_point)
 854 {
 855         unsigned hash =
 856                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
 857
 858         return &c->write_points_hash[hash];
 859 }
 860
 861 static struct write_point *__writepoint_find(struct hlist_head *head,
 862                                              unsigned long write_point)
 863 {
 864         struct write_point *wp;
 865
 866         rcu_read_lock();
 867         hlist_for_each_entry_rcu(wp, head, node)
 868                 if (wp->write_point == write_point)
 869                         goto out;
 870         wp = NULL;
 871 out:
 872         rcu_read_unlock();
 873         return wp;
 874 }
 875
 876 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
 877 {
 878         u64 stranded    = c->write_points_nr * c->bucket_size_max;
 879         u64 free        = bch2_fs_usage_read_short(c).free;
 880
 881         return stranded * factor > free;
 882 }
 883
 884 static bool try_increase_writepoints(struct bch_fs *c)
 885 {
 886         struct write_point *wp;
 887
 888         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
 889             too_many_writepoints(c, 32))
 890                 return false;
 891
 892         wp = c->write_points + c->write_points_nr++;
 893         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
 894         return true;
 895 }
 896
 897 static bool try_decrease_writepoints(struct bch_fs *c,
 898                                      unsigned old_nr)
 899 {
 900         struct write_point *wp;
 901
 902         mutex_lock(&c->write_points_hash_lock);
 903         if (c->write_points_nr < old_nr) {
 904                 mutex_unlock(&c->write_points_hash_lock);
 905                 return true;
 906         }
 907
 908         if (c->write_points_nr == 1 ||
 909             !too_many_writepoints(c, 8)) {
 910                 mutex_unlock(&c->write_points_hash_lock);
 911                 return false;
 912         }
 913
 914         wp = c->write_points + --c->write_points_nr;
 915
 916         hlist_del_rcu(&wp->node);
 917         mutex_unlock(&c->write_points_hash_lock);
 918
 919         bch2_writepoint_stop(c, NULL, wp);
 920         return true;
 921 }
 922
 923 static struct write_point *writepoint_find(struct bch_fs *c,
 924                                            unsigned long write_point)
 925 {
 926         struct write_point *wp, *oldest;
 927         struct hlist_head *head;
 928
 929         if (!(write_point & 1UL)) {
 930                 wp = (struct write_point *) write_point;
 931                 mutex_lock(&wp->lock);
 932                 return wp;
 933         }
 934
 935         head = writepoint_hash(c, write_point);
 936 restart_find:
 937         wp = __writepoint_find(head, write_point);
 938         if (wp) {
 939 lock_wp:
 940                 mutex_lock(&wp->lock);
 941                 if (wp->write_point == write_point)
 942                         goto out;
 943                 mutex_unlock(&wp->lock);
 944                 goto restart_find;
 945         }
 946 restart_find_oldest:
 947         oldest = NULL;
 948         for (wp = c->write_points;
 949              wp < c->write_points + c->write_points_nr; wp++)
 950                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
 951                         oldest = wp;
 952
 953         mutex_lock(&oldest->lock);
 954         mutex_lock(&c->write_points_hash_lock);
 955         if (oldest >= c->write_points + c->write_points_nr ||
 956             try_increase_writepoints(c)) {
 957                 mutex_unlock(&c->write_points_hash_lock);
 958                 mutex_unlock(&oldest->lock);
 959                 goto restart_find_oldest;
 960         }
 961
 962         wp = __writepoint_find(head, write_point);
 963         if (wp && wp != oldest) {
 964                 mutex_unlock(&c->write_points_hash_lock);
 965                 mutex_unlock(&oldest->lock);
 966                 goto lock_wp;
 967         }
 968
 969         wp = oldest;
 970         hlist_del_rcu(&wp->node);
 971         wp->write_point = write_point;
 972         hlist_add_head_rcu(&wp->node, head);
 973         mutex_unlock(&c->write_points_hash_lock);
 974 out:
 975         wp->last_used = sched_clock();
 976         return wp;
 977 }
 978
 979 /*
 980  * Get us an open_bucket we can allocate from, return with it locked:
 981  */
 982 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
 983                                 unsigned target,
 984                                 unsigned erasure_code,
 985                                 struct write_point_specifier write_point,
 986                                 struct bch_devs_list *devs_have,
 987                                 unsigned nr_replicas,
 988                                 unsigned nr_replicas_required,
 989                                 enum alloc_reserve reserve,
 990                                 unsigned flags,
 991                                 struct closure *cl)
 992 {
 993         struct write_point *wp;
 994         struct open_bucket *ob;
 995         struct open_buckets ptrs;
 996         unsigned nr_effective, write_points_nr;
 997         unsigned ob_flags = 0;
 998         bool have_cache;
 999         int ret;
1000         int i;
1001
1002         if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
1003                 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
1004
1005         BUG_ON(!nr_replicas || !nr_replicas_required);
1006 retry:
1007         ptrs.nr         = 0;
1008         nr_effective    = 0;
1009         write_points_nr = c->write_points_nr;
1010         have_cache      = false;
1011
1012         wp = writepoint_find(c, write_point.v);
1013
1014         if (wp->data_type == BCH_DATA_user)
1015                 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
1016
1017         /* metadata may not allocate on cache devices: */
1018         if (wp->data_type != BCH_DATA_user)
1019                 have_cache = true;
1020
1021         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1022                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
1023                                               target, erasure_code,
1024                                               nr_replicas, &nr_effective,
1025                                               &have_cache, reserve,
1026                                               ob_flags, cl);
1027         } else {
1028                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
1029                                               target, erasure_code,
1030                                               nr_replicas, &nr_effective,
1031                                               &have_cache, reserve,
1032                                               ob_flags, NULL);
1033                 if (!ret)
1034                         goto alloc_done;
1035
1036                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
1037                                               0, erasure_code,
1038                                               nr_replicas, &nr_effective,
1039                                               &have_cache, reserve,
1040                                               ob_flags, cl);
1041         }
1042 alloc_done:
1043         BUG_ON(!ret && nr_effective < nr_replicas);
1044
1045         if (erasure_code && !ec_open_bucket(c, &ptrs))
1046                 pr_debug("failed to get ec bucket: ret %u", ret);
1047
1048         if (ret == -INSUFFICIENT_DEVICES &&
1049             nr_effective >= nr_replicas_required)
1050                 ret = 0;
1051
1052         if (ret)
1053                 goto err;
1054
1055         /* Free buckets we didn't use: */
1056         open_bucket_for_each(c, &wp->ptrs, ob, i)
1057                 open_bucket_free_unused(c, wp, ob);
1058
1059         wp->ptrs = ptrs;
1060
1061         wp->sectors_free = UINT_MAX;
1062
1063         open_bucket_for_each(c, &wp->ptrs, ob, i)
1064                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1065
1066         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1067
1068         return wp;
1069 err:
1070         open_bucket_for_each(c, &wp->ptrs, ob, i)
1071                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1072                         ob_push(c, &ptrs, ob);
1073                 else
1074                         open_bucket_free_unused(c, wp, ob);
1075         wp->ptrs = ptrs;
1076
1077         mutex_unlock(&wp->lock);
1078
1079         if (ret == -FREELIST_EMPTY &&
1080             try_decrease_writepoints(c, write_points_nr))
1081                 goto retry;
1082
1083         switch (ret) {
1084         case -OPEN_BUCKETS_EMPTY:
1085         case -FREELIST_EMPTY:
1086                 return cl ? ERR_PTR(-EAGAIN) : ERR_PTR(-ENOSPC);
1087         case -INSUFFICIENT_DEVICES:
1088                 return ERR_PTR(-EROFS);
1089         default:
1090                 return ERR_PTR(ret);
1091         }
1092 }
1093
1094 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1095 {
1096         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1097
1098         return (struct bch_extent_ptr) {
1099                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
1100                 .gen    = ob->gen,
1101                 .dev    = ob->dev,
1102                 .offset = bucket_to_sector(ca, ob->bucket) +
1103                         ca->mi.bucket_size -
1104                         ob->sectors_free,
1105         };
1106 }
1107
1108 /*
1109  * Append pointers to the space we just allocated to @k, and mark @sectors space
1110  * as allocated out of @ob
1111  */
1112 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1113                                     struct bkey_i *k, unsigned sectors,
1114                                     bool cached)
1115
1116 {
1117         struct open_bucket *ob;
1118         unsigned i;
1119
1120         BUG_ON(sectors > wp->sectors_free);
1121         wp->sectors_free -= sectors;
1122
1123         open_bucket_for_each(c, &wp->ptrs, ob, i) {
1124                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1125                 struct bch_extent_ptr ptr = bch2_ob_ptr(c, ob);
1126
1127                 ptr.cached = cached ||
1128                         (!ca->mi.durability &&
1129                          wp->data_type == BCH_DATA_user);
1130
1131                 bch2_bkey_append_ptr(k, ptr);
1132
1133                 BUG_ON(sectors > ob->sectors_free);
1134                 ob->sectors_free -= sectors;
1135         }
1136 }
1137
1138 /*
1139  * Append pointers to the space we just allocated to @k, and mark @sectors space
1140  * as allocated out of @ob
1141  */
1142 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1143 {
1144         struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
1145         struct open_bucket *ob;
1146         unsigned i;
1147
1148         open_bucket_for_each(c, &wp->ptrs, ob, i)
1149                 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
1150         wp->ptrs = keep;
1151
1152         mutex_unlock(&wp->lock);
1153
1154         bch2_open_buckets_put(c, &ptrs);
1155 }
1156
1157 static inline void writepoint_init(struct write_point *wp,
1158                                    enum bch_data_type type)
1159 {
1160         mutex_init(&wp->lock);
1161         wp->data_type = type;
1162 }
1163
1164 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1165 {
1166         struct open_bucket *ob;
1167         struct write_point *wp;
1168
1169         mutex_init(&c->write_points_hash_lock);
1170         c->write_points_nr = ARRAY_SIZE(c->write_points);
1171
1172         /* open bucket 0 is a sentinal NULL: */
1173         spin_lock_init(&c->open_buckets[0].lock);
1174
1175         for (ob = c->open_buckets + 1;
1176              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1177                 spin_lock_init(&ob->lock);
1178                 c->open_buckets_nr_free++;
1179
1180                 ob->freelist = c->open_buckets_freelist;
1181                 c->open_buckets_freelist = ob - c->open_buckets;
1182         }
1183
1184         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
1185         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
1186         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
1187
1188         for (wp = c->write_points;
1189              wp < c->write_points + c->write_points_nr; wp++) {
1190                 writepoint_init(wp, BCH_DATA_user);
1191
1192                 wp->last_used   = sched_clock();
1193                 wp->write_point = (unsigned long) wp;
1194                 hlist_add_head_rcu(&wp->node,
1195                                    writepoint_hash(c, wp->write_point));
1196         }
1197 }
1198
1199 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1200 {
1201         struct open_bucket *ob;
1202
1203         for (ob = c->open_buckets;
1204              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1205              ob++) {
1206                 spin_lock(&ob->lock);
1207                 if (ob->valid && !ob->on_partial_list) {
1208                         pr_buf(out, "%zu ref %u type %s\n",
1209                                ob - c->open_buckets,
1210                                atomic_read(&ob->pin),
1211                                bch2_data_types[ob->data_type]);
1212                 }
1213                 spin_unlock(&ob->lock);
1214         }
1215
1216 }