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