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                                               u64 *skipped_open,
 196                                               u64 *skipped_need_journal_commit,
 197                                               u64 *skipped_nouse,
 198                                               struct closure *cl)
 199 {
 200         struct open_bucket *ob;
 201
 202         if (unlikely(ca->buckets_nouse && test_bit(a.bucket, ca->buckets_nouse))) {
 203                 (*skipped_nouse)++;
 204                 return NULL;
 205         }
 206
 207         if (bch2_bucket_is_open(c, ca->dev_idx, a.bucket)) {
 208                 (*skipped_open)++;
 209                 return NULL;
 210         }
 211
 212         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 213                         c->journal.flushed_seq_ondisk, ca->dev_idx, a.bucket)) {
 214                 (*skipped_need_journal_commit)++;
 215                 return NULL;
 216         }
 217
 218         spin_lock(&c->freelist_lock);
 219
 220         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
 221                 if (cl)
 222                         closure_wait(&c->open_buckets_wait, cl);
 223
 224                 if (!c->blocked_allocate_open_bucket)
 225                         c->blocked_allocate_open_bucket = local_clock();
 226
 227                 spin_unlock(&c->freelist_lock);
 228
 229                 trace_open_bucket_alloc_fail(ca, bch2_alloc_reserves[reserve]);
 230                 return ERR_PTR(-OPEN_BUCKETS_EMPTY);
 231         }
 232
 233         /* Recheck under lock: */
 234         if (bch2_bucket_is_open(c, ca->dev_idx, a.bucket)) {
 235                 spin_unlock(&c->freelist_lock);
 236                 (*skipped_open)++;
 237                 return NULL;
 238         }
 239
 240         ob = bch2_open_bucket_alloc(c);
 241
 242         spin_lock(&ob->lock);
 243
 244         ob->valid       = true;
 245         ob->sectors_free = ca->mi.bucket_size;
 246         ob->alloc_reserve = reserve;
 247         ob->dev         = ca->dev_idx;
 248         ob->gen         = a.gen;
 249         ob->bucket      = a.bucket;
 250         spin_unlock(&ob->lock);
 251
 252         ca->nr_open_buckets++;
 253         bch2_open_bucket_hash_add(c, ob);
 254
 255         if (c->blocked_allocate_open_bucket) {
 256                 bch2_time_stats_update(
 257                         &c->times[BCH_TIME_blocked_allocate_open_bucket],
 258                         c->blocked_allocate_open_bucket);
 259                 c->blocked_allocate_open_bucket = 0;
 260         }
 261
 262         if (c->blocked_allocate) {
 263                 bch2_time_stats_update(
 264                         &c->times[BCH_TIME_blocked_allocate],
 265                         c->blocked_allocate);
 266                 c->blocked_allocate = 0;
 267         }
 268
 269         spin_unlock(&c->freelist_lock);
 270
 271         trace_bucket_alloc(ca, bch2_alloc_reserves[reserve]);
 272         return ob;
 273 }
 274
 275 static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bch_dev *ca,
 276                                             enum alloc_reserve reserve, u64 free_entry,
 277                                             u64 *skipped_open,
 278                                             u64 *skipped_need_journal_commit,
 279                                             u64 *skipped_nouse,
 280                                             struct closure *cl)
 281 {
 282         struct bch_fs *c = trans->c;
 283         struct btree_iter iter;
 284         struct bkey_s_c k;
 285         struct open_bucket *ob;
 286         struct bkey_alloc_unpacked a;
 287         u64 b = free_entry & ~(~0ULL << 56);
 288         unsigned genbits = free_entry >> 56;
 289         struct printbuf buf = PRINTBUF;
 290         int ret;
 291
 292         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS(ca->dev_idx, b), BTREE_ITER_CACHED);
 293         k = bch2_btree_iter_peek_slot(&iter);
 294         ret = bkey_err(k);
 295         if (ret) {
 296                 ob = ERR_PTR(ret);
 297                 goto err;
 298         }
 299
 300         a = bch2_alloc_unpack(k);
 301
 302         if (bch2_fs_inconsistent_on(bucket_state(a) != BUCKET_free, c,
 303                         "non free bucket in freespace btree (state %s)\n"
 304                         "  %s\n"
 305                         "  at %llu (genbits %u)",
 306                         bch2_bucket_states[bucket_state(a)],
 307                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
 308                         free_entry, genbits)) {
 309                 ob = ERR_PTR(-EIO);
 310                 goto err;
 311         }
 312
 313         if (bch2_fs_inconsistent_on(genbits != (alloc_freespace_genbits(a) >> 56), c,
 314                         "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
 315                         "  %s",
 316                         genbits, alloc_freespace_genbits(a) >> 56,
 317                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 318                 ob = ERR_PTR(-EIO);
 319                 goto err;
 320         }
 321
 322         if (bch2_fs_inconsistent_on(b < ca->mi.first_bucket || b >= ca->mi.nbuckets, c,
 323                         "freespace btree has bucket outside allowed range (got %llu, valid %u-%llu)",
 324                         b, ca->mi.first_bucket, ca->mi.nbuckets)) {
 325                 ob = ERR_PTR(-EIO);
 326                 goto err;
 327         }
 328
 329         ob = __try_alloc_bucket(c, ca, reserve, a,
 330                                 skipped_open,
 331                                 skipped_need_journal_commit,
 332                                 skipped_nouse,
 333                                 cl);
 334 err:
 335         bch2_trans_iter_exit(trans, &iter);
 336         printbuf_exit(&buf);
 337         return ob;
 338 }
 339
 340 static struct open_bucket *try_alloc_partial_bucket(struct bch_fs *c, struct bch_dev *ca,
 341                                                     enum alloc_reserve reserve)
 342 {
 343         struct open_bucket *ob;
 344         int i;
 345
 346         spin_lock(&c->freelist_lock);
 347
 348         for (i = ca->open_buckets_partial_nr - 1; i >= 0; --i) {
 349                 ob = c->open_buckets + ca->open_buckets_partial[i];
 350
 351                 if (reserve <= ob->alloc_reserve) {
 352                         array_remove_item(ca->open_buckets_partial,
 353                                           ca->open_buckets_partial_nr,
 354                                           i);
 355                         ob->on_partial_list = false;
 356                         ob->alloc_reserve = reserve;
 357                         spin_unlock(&c->freelist_lock);
 358                         return ob;
 359                 }
 360         }
 361
 362         spin_unlock(&c->freelist_lock);
 363         return NULL;
 364 }
 365
 366 /*
 367  * This path is for before the freespace btree is initialized:
 368  *
 369  * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
 370  * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
 371  */
 372 static noinline struct open_bucket *
 373 bch2_bucket_alloc_trans_early(struct btree_trans *trans,
 374                               struct bch_dev *ca,
 375                               enum alloc_reserve reserve,
 376                               u64 *cur_bucket,
 377                               u64 *buckets_seen,
 378                               u64 *skipped_open,
 379                               u64 *skipped_need_journal_commit,
 380                               u64 *skipped_nouse,
 381                               struct closure *cl)
 382 {
 383         struct btree_iter iter;
 384         struct bkey_s_c k;
 385         struct open_bucket *ob = NULL;
 386         int ret;
 387
 388         *cur_bucket = max_t(u64, *cur_bucket, ca->mi.first_bucket);
 389         *cur_bucket = max_t(u64, *cur_bucket, ca->new_fs_bucket_idx);
 390
 391         for_each_btree_key(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, *cur_bucket),
 392                            BTREE_ITER_SLOTS, k, ret) {
 393                 struct bkey_alloc_unpacked a;
 394
 395                 if (bkey_cmp(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)) >= 0)
 396                         break;
 397
 398                 if (ca->new_fs_bucket_idx &&
 399                     is_superblock_bucket(ca, k.k->p.offset))
 400                         continue;
 401
 402                 a = bch2_alloc_unpack(k);
 403
 404                 if (bucket_state(a) != BUCKET_free)
 405                         continue;
 406
 407                 (*buckets_seen)++;
 408
 409                 ob = __try_alloc_bucket(trans->c, ca, reserve, a,
 410                                         skipped_open,
 411                                         skipped_need_journal_commit,
 412                                         skipped_nouse,
 413                                         cl);
 414                 if (ob)
 415                         break;
 416         }
 417         bch2_trans_iter_exit(trans, &iter);
 418
 419         *cur_bucket = iter.pos.offset;
 420
 421         return ob ?: ERR_PTR(ret ?: -FREELIST_EMPTY);
 422 }
 423
 424 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
 425                                                    struct bch_dev *ca,
 426                                                    enum alloc_reserve reserve,
 427                                                    u64 *cur_bucket,
 428                                                    u64 *buckets_seen,
 429                                                    u64 *skipped_open,
 430                                                    u64 *skipped_need_journal_commit,
 431                                                    u64 *skipped_nouse,
 432                                                    struct closure *cl)
 433 {
 434         struct btree_iter iter;
 435         struct bkey_s_c k;
 436         struct open_bucket *ob = NULL;
 437         int ret;
 438
 439         if (unlikely(!ca->mi.freespace_initialized))
 440                 return bch2_bucket_alloc_trans_early(trans, ca, reserve,
 441                                                      cur_bucket,
 442                                                      buckets_seen,
 443                                                      skipped_open,
 444                                                      skipped_need_journal_commit,
 445                                                      skipped_nouse,
 446                                                      cl);
 447
 448         BUG_ON(ca->new_fs_bucket_idx);
 449
 450         for_each_btree_key(trans, iter, BTREE_ID_freespace,
 451                            POS(ca->dev_idx, *cur_bucket), 0, k, ret) {
 452                 if (k.k->p.inode != ca->dev_idx)
 453                         break;
 454
 455                 for (*cur_bucket = max(*cur_bucket, bkey_start_offset(k.k));
 456                      *cur_bucket != k.k->p.offset && !ob;
 457                      (*cur_bucket)++) {
 458                         if (btree_trans_too_many_iters(trans)) {
 459                                 ob = ERR_PTR(-EINTR);
 460                                 break;
 461                         }
 462
 463                         (*buckets_seen)++;
 464
 465                         ob = try_alloc_bucket(trans, ca, reserve,
 466                                               *cur_bucket,
 467                                               skipped_open,
 468                                               skipped_need_journal_commit,
 469                                               skipped_nouse,
 470                                               cl);
 471                 }
 472                 if (ob)
 473                         break;
 474         }
 475         bch2_trans_iter_exit(trans, &iter);
 476
 477         return ob ?: ERR_PTR(ret);
 478 }
 479
 480 /**
 481  * bch_bucket_alloc - allocate a single bucket from a specific device
 482  *
 483  * Returns index of bucket on success, 0 on failure
 484  * */
 485 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 486                                       enum alloc_reserve reserve,
 487                                       bool may_alloc_partial,
 488                                       struct closure *cl)
 489 {
 490         struct open_bucket *ob = NULL;
 491         u64 avail = dev_buckets_available(ca, reserve);
 492         u64 cur_bucket = 0;
 493         u64 buckets_seen = 0;
 494         u64 skipped_open = 0;
 495         u64 skipped_need_journal_commit = 0;
 496         u64 skipped_nouse = 0;
 497         int ret;
 498
 499         if (may_alloc_partial) {
 500                 ob = try_alloc_partial_bucket(c, ca, reserve);
 501                 if (ob)
 502                         return ob;
 503         }
 504 again:
 505         if (!avail) {
 506                 if (cl) {
 507                         closure_wait(&c->freelist_wait, cl);
 508                         /* recheck after putting ourself on waitlist */
 509                         avail = dev_buckets_available(ca, reserve);
 510                         if (avail) {
 511                                 closure_wake_up(&c->freelist_wait);
 512                                 goto again;
 513                         }
 514                 }
 515
 516                 if (!c->blocked_allocate)
 517                         c->blocked_allocate = local_clock();
 518
 519                 ob = ERR_PTR(-FREELIST_EMPTY);
 520                 goto err;
 521         }
 522
 523         ret = bch2_trans_do(c, NULL, NULL, 0,
 524                         PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, reserve,
 525                                                         &cur_bucket,
 526                                                         &buckets_seen,
 527                                                         &skipped_open,
 528                                                         &skipped_need_journal_commit,
 529                                                         &skipped_nouse,
 530                                                         cl)));
 531
 532         if (skipped_need_journal_commit * 2 > avail)
 533                 bch2_journal_flush_async(&c->journal, NULL);
 534 err:
 535         if (!ob)
 536                 ob = ERR_PTR(ret ?: -FREELIST_EMPTY);
 537
 538         if (IS_ERR(ob)) {
 539                 trace_bucket_alloc_fail(ca, bch2_alloc_reserves[reserve], avail,
 540                                         buckets_seen,
 541                                         skipped_open,
 542                                         skipped_need_journal_commit,
 543                                         skipped_nouse,
 544                                         cl == NULL, PTR_ERR(ob));
 545                 atomic_long_inc(&c->bucket_alloc_fail);
 546         }
 547
 548         return ob;
 549 }
 550
 551 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 552                             unsigned l, unsigned r)
 553 {
 554         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 555                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 556 }
 557
 558 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 559
 560 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 561                                           struct dev_stripe_state *stripe,
 562                                           struct bch_devs_mask *devs)
 563 {
 564         struct dev_alloc_list ret = { .nr = 0 };
 565         unsigned i;
 566
 567         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 568                 ret.devs[ret.nr++] = i;
 569
 570         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 571         return ret;
 572 }
 573
 574 void bch2_dev_stripe_increment(struct bch_dev *ca,
 575                                struct dev_stripe_state *stripe)
 576 {
 577         u64 *v = stripe->next_alloc + ca->dev_idx;
 578         u64 free_space = dev_buckets_available(ca, RESERVE_none);
 579         u64 free_space_inv = free_space
 580                 ? div64_u64(1ULL << 48, free_space)
 581                 : 1ULL << 48;
 582         u64 scale = *v / 4;
 583
 584         if (*v + free_space_inv >= *v)
 585                 *v += free_space_inv;
 586         else
 587                 *v = U64_MAX;
 588
 589         for (v = stripe->next_alloc;
 590              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 591                 *v = *v < scale ? 0 : *v - scale;
 592 }
 593
 594 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 595 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 596
 597 static void add_new_bucket(struct bch_fs *c,
 598                            struct open_buckets *ptrs,
 599                            struct bch_devs_mask *devs_may_alloc,
 600                            unsigned *nr_effective,
 601                            bool *have_cache,
 602                            unsigned flags,
 603                            struct open_bucket *ob)
 604 {
 605         unsigned durability =
 606                 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
 607
 608         __clear_bit(ob->dev, devs_may_alloc->d);
 609         *nr_effective   += (flags & BUCKET_ALLOC_USE_DURABILITY)
 610                 ? durability : 1;
 611         *have_cache     |= !durability;
 612
 613         ob_push(c, ptrs, ob);
 614 }
 615
 616 int bch2_bucket_alloc_set(struct bch_fs *c,
 617                       struct open_buckets *ptrs,
 618                       struct dev_stripe_state *stripe,
 619                       struct bch_devs_mask *devs_may_alloc,
 620                       unsigned nr_replicas,
 621                       unsigned *nr_effective,
 622                       bool *have_cache,
 623                       enum alloc_reserve reserve,
 624                       unsigned flags,
 625                       struct closure *cl)
 626 {
 627         struct dev_alloc_list devs_sorted =
 628                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 629         unsigned dev;
 630         struct bch_dev *ca;
 631         int ret = -INSUFFICIENT_DEVICES;
 632         unsigned i;
 633
 634         BUG_ON(*nr_effective >= nr_replicas);
 635
 636         for (i = 0; i < devs_sorted.nr; i++) {
 637                 struct open_bucket *ob;
 638
 639                 dev = devs_sorted.devs[i];
 640
 641                 rcu_read_lock();
 642                 ca = rcu_dereference(c->devs[dev]);
 643                 if (ca)
 644                         percpu_ref_get(&ca->ref);
 645                 rcu_read_unlock();
 646
 647                 if (!ca)
 648                         continue;
 649
 650                 if (!ca->mi.durability && *have_cache) {
 651                         percpu_ref_put(&ca->ref);
 652                         continue;
 653                 }
 654
 655                 ob = bch2_bucket_alloc(c, ca, reserve,
 656                                 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
 657                 if (!IS_ERR(ob))
 658                         bch2_dev_stripe_increment(ca, stripe);
 659                 percpu_ref_put(&ca->ref);
 660
 661                 if (IS_ERR(ob)) {
 662                         ret = PTR_ERR(ob);
 663
 664                         if (cl)
 665                                 break;
 666                         continue;
 667                 }
 668
 669                 add_new_bucket(c, ptrs, devs_may_alloc,
 670                                nr_effective, have_cache, flags, ob);
 671
 672                 if (*nr_effective >= nr_replicas) {
 673                         ret = 0;
 674                         break;
 675                 }
 676         }
 677
 678         return ret;
 679 }
 680
 681 /* Allocate from stripes: */
 682
 683 /*
 684  * if we can't allocate a new stripe because there are already too many
 685  * partially filled stripes, force allocating from an existing stripe even when
 686  * it's to a device we don't want:
 687  */
 688
 689 static int bucket_alloc_from_stripe(struct bch_fs *c,
 690                          struct open_buckets *ptrs,
 691                          struct write_point *wp,
 692                          struct bch_devs_mask *devs_may_alloc,
 693                          u16 target,
 694                          unsigned erasure_code,
 695                          unsigned nr_replicas,
 696                          unsigned *nr_effective,
 697                          bool *have_cache,
 698                          unsigned flags,
 699                          struct closure *cl)
 700 {
 701         struct dev_alloc_list devs_sorted;
 702         struct ec_stripe_head *h;
 703         struct open_bucket *ob;
 704         struct bch_dev *ca;
 705         unsigned i, ec_idx;
 706
 707         if (!erasure_code)
 708                 return 0;
 709
 710         if (nr_replicas < 2)
 711                 return 0;
 712
 713         if (ec_open_bucket(c, ptrs))
 714                 return 0;
 715
 716         h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1,
 717                                     wp == &c->copygc_write_point,
 718                                     cl);
 719         if (IS_ERR(h))
 720                 return -PTR_ERR(h);
 721         if (!h)
 722                 return 0;
 723
 724         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 725
 726         for (i = 0; i < devs_sorted.nr; i++)
 727                 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 728                         if (!h->s->blocks[ec_idx])
 729                                 continue;
 730
 731                         ob = c->open_buckets + h->s->blocks[ec_idx];
 732                         if (ob->dev == devs_sorted.devs[i] &&
 733                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 734                                 goto got_bucket;
 735                 }
 736         goto out_put_head;
 737 got_bucket:
 738         ca = bch_dev_bkey_exists(c, ob->dev);
 739
 740         ob->ec_idx      = ec_idx;
 741         ob->ec          = h->s;
 742
 743         add_new_bucket(c, ptrs, devs_may_alloc,
 744                        nr_effective, have_cache, flags, ob);
 745         atomic_inc(&h->s->pin);
 746 out_put_head:
 747         bch2_ec_stripe_head_put(c, h);
 748         return 0;
 749 }
 750
 751 /* Sector allocator */
 752
 753 static void get_buckets_from_writepoint(struct bch_fs *c,
 754                                         struct open_buckets *ptrs,
 755                                         struct write_point *wp,
 756                                         struct bch_devs_mask *devs_may_alloc,
 757                                         unsigned nr_replicas,
 758                                         unsigned *nr_effective,
 759                                         bool *have_cache,
 760                                         unsigned flags,
 761                                         bool need_ec)
 762 {
 763         struct open_buckets ptrs_skip = { .nr = 0 };
 764         struct open_bucket *ob;
 765         unsigned i;
 766
 767         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 768                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 769
 770                 if (*nr_effective < nr_replicas &&
 771                     test_bit(ob->dev, devs_may_alloc->d) &&
 772                     (ca->mi.durability ||
 773                      (wp->data_type == BCH_DATA_user && !*have_cache)) &&
 774                     (ob->ec || !need_ec)) {
 775                         add_new_bucket(c, ptrs, devs_may_alloc,
 776                                        nr_effective, have_cache,
 777                                        flags, ob);
 778                 } else {
 779                         ob_push(c, &ptrs_skip, ob);
 780                 }
 781         }
 782         wp->ptrs = ptrs_skip;
 783 }
 784
 785 static int open_bucket_add_buckets(struct bch_fs *c,
 786                         struct open_buckets *ptrs,
 787                         struct write_point *wp,
 788                         struct bch_devs_list *devs_have,
 789                         u16 target,
 790                         unsigned erasure_code,
 791                         unsigned nr_replicas,
 792                         unsigned *nr_effective,
 793                         bool *have_cache,
 794                         enum alloc_reserve reserve,
 795                         unsigned flags,
 796                         struct closure *_cl)
 797 {
 798         struct bch_devs_mask devs;
 799         struct open_bucket *ob;
 800         struct closure *cl = NULL;
 801         int ret;
 802         unsigned i;
 803
 804         rcu_read_lock();
 805         devs = target_rw_devs(c, wp->data_type, target);
 806         rcu_read_unlock();
 807
 808         /* Don't allocate from devices we already have pointers to: */
 809         for (i = 0; i < devs_have->nr; i++)
 810                 __clear_bit(devs_have->devs[i], devs.d);
 811
 812         open_bucket_for_each(c, ptrs, ob, i)
 813                 __clear_bit(ob->dev, devs.d);
 814
 815         if (erasure_code) {
 816                 if (!ec_open_bucket(c, ptrs)) {
 817                         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 818                                                     nr_replicas, nr_effective,
 819                                                     have_cache, flags, true);
 820                         if (*nr_effective >= nr_replicas)
 821                                 return 0;
 822                 }
 823
 824                 if (!ec_open_bucket(c, ptrs)) {
 825                         ret = bucket_alloc_from_stripe(c, ptrs, wp, &devs,
 826                                                  target, erasure_code,
 827                                                  nr_replicas, nr_effective,
 828                                                  have_cache, flags, _cl);
 829                         if (ret == -FREELIST_EMPTY ||
 830                             ret == -OPEN_BUCKETS_EMPTY)
 831                                 return ret;
 832                         if (*nr_effective >= nr_replicas)
 833                                 return 0;
 834                 }
 835         }
 836
 837         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 838                                     nr_replicas, nr_effective,
 839                                     have_cache, flags, false);
 840         if (*nr_effective >= nr_replicas)
 841                 return 0;
 842
 843 retry_blocking:
 844         /*
 845          * Try nonblocking first, so that if one device is full we'll try from
 846          * other devices:
 847          */
 848         ret = bch2_bucket_alloc_set(c, ptrs, &wp->stripe, &devs,
 849                                 nr_replicas, nr_effective, have_cache,
 850                                 reserve, flags, cl);
 851         if (ret && ret != -INSUFFICIENT_DEVICES && !cl && _cl) {
 852                 cl = _cl;
 853                 goto retry_blocking;
 854         }
 855
 856         return ret;
 857 }
 858
 859 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 860                                 struct open_buckets *obs)
 861 {
 862         struct open_buckets ptrs = { .nr = 0 };
 863         struct open_bucket *ob, *ob2;
 864         unsigned i, j;
 865
 866         open_bucket_for_each(c, obs, ob, i) {
 867                 bool drop = !ca || ob->dev == ca->dev_idx;
 868
 869                 if (!drop && ob->ec) {
 870                         mutex_lock(&ob->ec->lock);
 871                         for (j = 0; j < ob->ec->new_stripe.key.v.nr_blocks; j++) {
 872                                 if (!ob->ec->blocks[j])
 873                                         continue;
 874
 875                                 ob2 = c->open_buckets + ob->ec->blocks[j];
 876                                 drop |= ob2->dev == ca->dev_idx;
 877                         }
 878                         mutex_unlock(&ob->ec->lock);
 879                 }
 880
 881                 if (drop)
 882                         bch2_open_bucket_put(c, ob);
 883                 else
 884                         ob_push(c, &ptrs, ob);
 885         }
 886
 887         *obs = ptrs;
 888 }
 889
 890 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 891                           struct write_point *wp)
 892 {
 893         mutex_lock(&wp->lock);
 894         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
 895         mutex_unlock(&wp->lock);
 896 }
 897
 898 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 899                                                  unsigned long write_point)
 900 {
 901         unsigned hash =
 902                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
 903
 904         return &c->write_points_hash[hash];
 905 }
 906
 907 static struct write_point *__writepoint_find(struct hlist_head *head,
 908                                              unsigned long write_point)
 909 {
 910         struct write_point *wp;
 911
 912         rcu_read_lock();
 913         hlist_for_each_entry_rcu(wp, head, node)
 914                 if (wp->write_point == write_point)
 915                         goto out;
 916         wp = NULL;
 917 out:
 918         rcu_read_unlock();
 919         return wp;
 920 }
 921
 922 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
 923 {
 924         u64 stranded    = c->write_points_nr * c->bucket_size_max;
 925         u64 free        = bch2_fs_usage_read_short(c).free;
 926
 927         return stranded * factor > free;
 928 }
 929
 930 static bool try_increase_writepoints(struct bch_fs *c)
 931 {
 932         struct write_point *wp;
 933
 934         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
 935             too_many_writepoints(c, 32))
 936                 return false;
 937
 938         wp = c->write_points + c->write_points_nr++;
 939         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
 940         return true;
 941 }
 942
 943 static bool try_decrease_writepoints(struct bch_fs *c,
 944                                      unsigned old_nr)
 945 {
 946         struct write_point *wp;
 947
 948         mutex_lock(&c->write_points_hash_lock);
 949         if (c->write_points_nr < old_nr) {
 950                 mutex_unlock(&c->write_points_hash_lock);
 951                 return true;
 952         }
 953
 954         if (c->write_points_nr == 1 ||
 955             !too_many_writepoints(c, 8)) {
 956                 mutex_unlock(&c->write_points_hash_lock);
 957                 return false;
 958         }
 959
 960         wp = c->write_points + --c->write_points_nr;
 961
 962         hlist_del_rcu(&wp->node);
 963         mutex_unlock(&c->write_points_hash_lock);
 964
 965         bch2_writepoint_stop(c, NULL, wp);
 966         return true;
 967 }
 968
 969 static struct write_point *writepoint_find(struct bch_fs *c,
 970                                            unsigned long write_point)
 971 {
 972         struct write_point *wp, *oldest;
 973         struct hlist_head *head;
 974
 975         if (!(write_point & 1UL)) {
 976                 wp = (struct write_point *) write_point;
 977                 mutex_lock(&wp->lock);
 978                 return wp;
 979         }
 980
 981         head = writepoint_hash(c, write_point);
 982 restart_find:
 983         wp = __writepoint_find(head, write_point);
 984         if (wp) {
 985 lock_wp:
 986                 mutex_lock(&wp->lock);
 987                 if (wp->write_point == write_point)
 988                         goto out;
 989                 mutex_unlock(&wp->lock);
 990                 goto restart_find;
 991         }
 992 restart_find_oldest:
 993         oldest = NULL;
 994         for (wp = c->write_points;
 995              wp < c->write_points + c->write_points_nr; wp++)
 996                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
 997                         oldest = wp;
 998
 999         mutex_lock(&oldest->lock);
1000         mutex_lock(&c->write_points_hash_lock);
1001         if (oldest >= c->write_points + c->write_points_nr ||
1002             try_increase_writepoints(c)) {
1003                 mutex_unlock(&c->write_points_hash_lock);
1004                 mutex_unlock(&oldest->lock);
1005                 goto restart_find_oldest;
1006         }
1007
1008         wp = __writepoint_find(head, write_point);
1009         if (wp && wp != oldest) {
1010                 mutex_unlock(&c->write_points_hash_lock);
1011                 mutex_unlock(&oldest->lock);
1012                 goto lock_wp;
1013         }
1014
1015         wp = oldest;
1016         hlist_del_rcu(&wp->node);
1017         wp->write_point = write_point;
1018         hlist_add_head_rcu(&wp->node, head);
1019         mutex_unlock(&c->write_points_hash_lock);
1020 out:
1021         wp->last_used = sched_clock();
1022         return wp;
1023 }
1024
1025 /*
1026  * Get us an open_bucket we can allocate from, return with it locked:
1027  */
1028 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
1029                                 unsigned target,
1030                                 unsigned erasure_code,
1031                                 struct write_point_specifier write_point,
1032                                 struct bch_devs_list *devs_have,
1033                                 unsigned nr_replicas,
1034                                 unsigned nr_replicas_required,
1035                                 enum alloc_reserve reserve,
1036                                 unsigned flags,
1037                                 struct closure *cl)
1038 {
1039         struct write_point *wp;
1040         struct open_bucket *ob;
1041         struct open_buckets ptrs;
1042         unsigned nr_effective, write_points_nr;
1043         unsigned ob_flags = 0;
1044         bool have_cache;
1045         int ret;
1046         int i;
1047
1048         if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
1049                 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
1050
1051         BUG_ON(!nr_replicas || !nr_replicas_required);
1052 retry:
1053         ptrs.nr         = 0;
1054         nr_effective    = 0;
1055         write_points_nr = c->write_points_nr;
1056         have_cache      = false;
1057
1058         wp = writepoint_find(c, write_point.v);
1059
1060         if (wp->data_type == BCH_DATA_user)
1061                 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
1062
1063         /* metadata may not allocate on cache devices: */
1064         if (wp->data_type != BCH_DATA_user)
1065                 have_cache = true;
1066
1067         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1068                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
1069                                               target, erasure_code,
1070                                               nr_replicas, &nr_effective,
1071                                               &have_cache, reserve,
1072                                               ob_flags, cl);
1073         } else {
1074                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
1075                                               target, erasure_code,
1076                                               nr_replicas, &nr_effective,
1077                                               &have_cache, reserve,
1078                                               ob_flags, NULL);
1079                 if (!ret)
1080                         goto alloc_done;
1081
1082                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
1083                                               0, erasure_code,
1084                                               nr_replicas, &nr_effective,
1085                                               &have_cache, reserve,
1086                                               ob_flags, cl);
1087         }
1088 alloc_done:
1089         BUG_ON(!ret && nr_effective < nr_replicas);
1090
1091         if (erasure_code && !ec_open_bucket(c, &ptrs))
1092                 pr_debug("failed to get ec bucket: ret %u", ret);
1093
1094         if (ret == -INSUFFICIENT_DEVICES &&
1095             nr_effective >= nr_replicas_required)
1096                 ret = 0;
1097
1098         if (ret)
1099                 goto err;
1100
1101         /* Free buckets we didn't use: */
1102         open_bucket_for_each(c, &wp->ptrs, ob, i)
1103                 open_bucket_free_unused(c, wp, ob);
1104
1105         wp->ptrs = ptrs;
1106
1107         wp->sectors_free = UINT_MAX;
1108
1109         open_bucket_for_each(c, &wp->ptrs, ob, i)
1110                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1111
1112         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1113
1114         return wp;
1115 err:
1116         open_bucket_for_each(c, &wp->ptrs, ob, i)
1117                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1118                         ob_push(c, &ptrs, ob);
1119                 else
1120                         open_bucket_free_unused(c, wp, ob);
1121         wp->ptrs = ptrs;
1122
1123         mutex_unlock(&wp->lock);
1124
1125         if (ret == -FREELIST_EMPTY &&
1126             try_decrease_writepoints(c, write_points_nr))
1127                 goto retry;
1128
1129         switch (ret) {
1130         case -OPEN_BUCKETS_EMPTY:
1131         case -FREELIST_EMPTY:
1132                 return cl ? ERR_PTR(-EAGAIN) : ERR_PTR(-ENOSPC);
1133         case -INSUFFICIENT_DEVICES:
1134                 return ERR_PTR(-EROFS);
1135         default:
1136                 return ERR_PTR(ret);
1137         }
1138 }
1139
1140 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1141 {
1142         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1143
1144         return (struct bch_extent_ptr) {
1145                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
1146                 .gen    = ob->gen,
1147                 .dev    = ob->dev,
1148                 .offset = bucket_to_sector(ca, ob->bucket) +
1149                         ca->mi.bucket_size -
1150                         ob->sectors_free,
1151         };
1152 }
1153
1154 /*
1155  * Append pointers to the space we just allocated to @k, and mark @sectors space
1156  * as allocated out of @ob
1157  */
1158 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1159                                     struct bkey_i *k, unsigned sectors,
1160                                     bool cached)
1161
1162 {
1163         struct open_bucket *ob;
1164         unsigned i;
1165
1166         BUG_ON(sectors > wp->sectors_free);
1167         wp->sectors_free -= sectors;
1168
1169         open_bucket_for_each(c, &wp->ptrs, ob, i) {
1170                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1171                 struct bch_extent_ptr ptr = bch2_ob_ptr(c, ob);
1172
1173                 ptr.cached = cached ||
1174                         (!ca->mi.durability &&
1175                          wp->data_type == BCH_DATA_user);
1176
1177                 bch2_bkey_append_ptr(k, ptr);
1178
1179                 BUG_ON(sectors > ob->sectors_free);
1180                 ob->sectors_free -= sectors;
1181         }
1182 }
1183
1184 /*
1185  * Append pointers to the space we just allocated to @k, and mark @sectors space
1186  * as allocated out of @ob
1187  */
1188 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1189 {
1190         struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
1191         struct open_bucket *ob;
1192         unsigned i;
1193
1194         open_bucket_for_each(c, &wp->ptrs, ob, i)
1195                 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
1196         wp->ptrs = keep;
1197
1198         mutex_unlock(&wp->lock);
1199
1200         bch2_open_buckets_put(c, &ptrs);
1201 }
1202
1203 static inline void writepoint_init(struct write_point *wp,
1204                                    enum bch_data_type type)
1205 {
1206         mutex_init(&wp->lock);
1207         wp->data_type = type;
1208 }
1209
1210 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1211 {
1212         struct open_bucket *ob;
1213         struct write_point *wp;
1214
1215         mutex_init(&c->write_points_hash_lock);
1216         c->write_points_nr = ARRAY_SIZE(c->write_points);
1217
1218         /* open bucket 0 is a sentinal NULL: */
1219         spin_lock_init(&c->open_buckets[0].lock);
1220
1221         for (ob = c->open_buckets + 1;
1222              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1223                 spin_lock_init(&ob->lock);
1224                 c->open_buckets_nr_free++;
1225
1226                 ob->freelist = c->open_buckets_freelist;
1227                 c->open_buckets_freelist = ob - c->open_buckets;
1228         }
1229
1230         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
1231         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
1232         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
1233
1234         for (wp = c->write_points;
1235              wp < c->write_points + c->write_points_nr; wp++) {
1236                 writepoint_init(wp, BCH_DATA_user);
1237
1238                 wp->last_used   = sched_clock();
1239                 wp->write_point = (unsigned long) wp;
1240                 hlist_add_head_rcu(&wp->node,
1241                                    writepoint_hash(c, wp->write_point));
1242         }
1243 }
1244
1245 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1246 {
1247         struct open_bucket *ob;
1248
1249         for (ob = c->open_buckets;
1250              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1251              ob++) {
1252                 spin_lock(&ob->lock);
1253                 if (ob->valid && !ob->on_partial_list) {
1254                         pr_buf(out, "%zu ref %u type %s\n",
1255                                ob - c->open_buckets,
1256                                atomic_read(&ob->pin),
1257                                bch2_data_types[ob->data_type]);
1258                 }
1259                 spin_unlock(&ob->lock);
1260         }
1261
1262 }