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 #include "movinggc.h"
  30
  31 #include <linux/math64.h>
  32 #include <linux/rculist.h>
  33 #include <linux/rcupdate.h>
  34 #include <trace/events/bcachefs.h>
  35
  36 const char * const bch2_alloc_reserves[] = {
  37 #define x(t) #t,
  38         BCH_ALLOC_RESERVES()
  39 #undef x
  40         NULL
  41 };
  42
  43 /*
  44  * Open buckets represent a bucket that's currently being allocated from.  They
  45  * serve two purposes:
  46  *
  47  *  - They track buckets that have been partially allocated, allowing for
  48  *    sub-bucket sized allocations - they're used by the sector allocator below
  49  *
  50  *  - They provide a reference to the buckets they own that mark and sweep GC
  51  *    can find, until the new allocation has a pointer to it inserted into the
  52  *    btree
  53  *
  54  * When allocating some space with the sector allocator, the allocation comes
  55  * with a reference to an open bucket - the caller is required to put that
  56  * reference _after_ doing the index update that makes its allocation reachable.
  57  */
  58
  59 static void bch2_open_bucket_hash_add(struct bch_fs *c, struct open_bucket *ob)
  60 {
  61         open_bucket_idx_t idx = ob - c->open_buckets;
  62         open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
  63
  64         ob->hash = *slot;
  65         *slot = idx;
  66 }
  67
  68 static void bch2_open_bucket_hash_remove(struct bch_fs *c, struct open_bucket *ob)
  69 {
  70         open_bucket_idx_t idx = ob - c->open_buckets;
  71         open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
  72
  73         while (*slot != idx) {
  74                 BUG_ON(!*slot);
  75                 slot = &c->open_buckets[*slot].hash;
  76         }
  77
  78         *slot = ob->hash;
  79         ob->hash = 0;
  80 }
  81
  82 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
  83 {
  84         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
  85
  86         if (ob->ec) {
  87                 bch2_ec_bucket_written(c, ob);
  88                 return;
  89         }
  90
  91         percpu_down_read(&c->mark_lock);
  92         spin_lock(&ob->lock);
  93
  94         ob->valid = false;
  95         ob->data_type = 0;
  96
  97         spin_unlock(&ob->lock);
  98         percpu_up_read(&c->mark_lock);
  99
 100         spin_lock(&c->freelist_lock);
 101         bch2_open_bucket_hash_remove(c, ob);
 102
 103         ob->freelist = c->open_buckets_freelist;
 104         c->open_buckets_freelist = ob - c->open_buckets;
 105
 106         c->open_buckets_nr_free++;
 107         ca->nr_open_buckets--;
 108         spin_unlock(&c->freelist_lock);
 109
 110         closure_wake_up(&c->open_buckets_wait);
 111 }
 112
 113 void bch2_open_bucket_write_error(struct bch_fs *c,
 114                                   struct open_buckets *obs,
 115                                   unsigned dev)
 116 {
 117         struct open_bucket *ob;
 118         unsigned i;
 119
 120         open_bucket_for_each(c, obs, ob, i)
 121                 if (ob->dev == dev && ob->ec)
 122                         bch2_ec_bucket_cancel(c, ob);
 123 }
 124
 125 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
 126 {
 127         struct open_bucket *ob;
 128
 129         BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
 130
 131         ob = c->open_buckets + c->open_buckets_freelist;
 132         c->open_buckets_freelist = ob->freelist;
 133         atomic_set(&ob->pin, 1);
 134         ob->data_type = 0;
 135
 136         c->open_buckets_nr_free--;
 137         return ob;
 138 }
 139
 140 static void open_bucket_free_unused(struct bch_fs *c,
 141                                     struct write_point *wp,
 142                                     struct open_bucket *ob)
 143 {
 144         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 145         bool may_realloc = wp->data_type == BCH_DATA_user;
 146
 147         BUG_ON(ca->open_buckets_partial_nr >
 148                ARRAY_SIZE(ca->open_buckets_partial));
 149
 150         if (ca->open_buckets_partial_nr <
 151             ARRAY_SIZE(ca->open_buckets_partial) &&
 152             may_realloc) {
 153                 spin_lock(&c->freelist_lock);
 154                 ob->on_partial_list = true;
 155                 ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
 156                         ob - c->open_buckets;
 157                 spin_unlock(&c->freelist_lock);
 158
 159                 closure_wake_up(&c->open_buckets_wait);
 160                 closure_wake_up(&c->freelist_wait);
 161         } else {
 162                 bch2_open_bucket_put(c, ob);
 163         }
 164 }
 165
 166 /* _only_ for allocating the journal on a new device: */
 167 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
 168 {
 169         while (ca->new_fs_bucket_idx < ca->mi.nbuckets) {
 170                 u64 b = ca->new_fs_bucket_idx++;
 171
 172                 if (!is_superblock_bucket(ca, b) &&
 173                     (!ca->buckets_nouse || !test_bit(b, ca->buckets_nouse)))
 174                         return b;
 175         }
 176
 177         return -1;
 178 }
 179
 180 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
 181 {
 182         switch (reserve) {
 183         case RESERVE_btree:
 184         case RESERVE_btree_movinggc:
 185                 return 0;
 186         case RESERVE_movinggc:
 187                 return OPEN_BUCKETS_COUNT / 4;
 188         default:
 189                 return OPEN_BUCKETS_COUNT / 2;
 190         }
 191 }
 192
 193 static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 194                                               u64 bucket,
 195                                               enum alloc_reserve reserve,
 196                                               struct bch_alloc_v4 *a,
 197                                               u64 *skipped_open,
 198                                               u64 *skipped_need_journal_commit,
 199                                               u64 *skipped_nouse,
 200                                               struct closure *cl)
 201 {
 202         struct open_bucket *ob;
 203
 204         if (unlikely(ca->buckets_nouse && test_bit(bucket, ca->buckets_nouse))) {
 205                 (*skipped_nouse)++;
 206                 return NULL;
 207         }
 208
 209         if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
 210                 (*skipped_open)++;
 211                 return NULL;
 212         }
 213
 214         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 215                         c->journal.flushed_seq_ondisk, ca->dev_idx, bucket)) {
 216                 (*skipped_need_journal_commit)++;
 217                 return NULL;
 218         }
 219
 220         spin_lock(&c->freelist_lock);
 221
 222         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
 223                 if (cl)
 224                         closure_wait(&c->open_buckets_wait, cl);
 225
 226                 if (!c->blocked_allocate_open_bucket)
 227                         c->blocked_allocate_open_bucket = local_clock();
 228
 229                 spin_unlock(&c->freelist_lock);
 230                 return ERR_PTR(-BCH_ERR_open_buckets_empty);
 231         }
 232
 233         /* Recheck under lock: */
 234         if (bch2_bucket_is_open(c, ca->dev_idx, 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      = 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 bkey_s_c freespace_k,
 281                                             struct closure *cl)
 282 {
 283         struct bch_fs *c = trans->c;
 284         struct btree_iter iter = { NULL };
 285         struct bkey_s_c k;
 286         struct open_bucket *ob;
 287         struct bch_alloc_v4 a;
 288         u64 b = free_entry & ~(~0ULL << 56);
 289         unsigned genbits = free_entry >> 56;
 290         struct printbuf buf = PRINTBUF;
 291         int ret;
 292
 293         if (b < ca->mi.first_bucket || b >= ca->mi.nbuckets) {
 294                 prt_printf(&buf, "freespace btree has bucket outside allowed range %u-%llu\n"
 295                        "  freespace key ",
 296                         ca->mi.first_bucket, ca->mi.nbuckets);
 297                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 298                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 299                 ob = ERR_PTR(-EIO);
 300                 goto err;
 301         }
 302
 303         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS(ca->dev_idx, b), BTREE_ITER_CACHED);
 304         k = bch2_btree_iter_peek_slot(&iter);
 305         ret = bkey_err(k);
 306         if (ret) {
 307                 ob = ERR_PTR(ret);
 308                 goto err;
 309         }
 310
 311         bch2_alloc_to_v4(k, &a);
 312
 313         if (genbits != (alloc_freespace_genbits(a) >> 56)) {
 314                 prt_printf(&buf, "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
 315                        "  freespace key ",
 316                        genbits, alloc_freespace_genbits(a) >> 56);
 317                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 318                 prt_printf(&buf, "\n  ");
 319                 bch2_bkey_val_to_text(&buf, c, k);
 320                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 321                 ob = ERR_PTR(-EIO);
 322                 goto err;
 323
 324         }
 325
 326         if (a.data_type != BCH_DATA_free) {
 327                 prt_printf(&buf, "non free bucket in freespace btree\n"
 328                        "  freespace key ");
 329                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 330                 prt_printf(&buf, "\n  ");
 331                 bch2_bkey_val_to_text(&buf, c, k);
 332                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 333                 ob = ERR_PTR(-EIO);
 334                 goto err;
 335         }
 336
 337         ob = __try_alloc_bucket(c, ca, b, reserve, &a,
 338                                 skipped_open,
 339                                 skipped_need_journal_commit,
 340                                 skipped_nouse,
 341                                 cl);
 342         if (!ob)
 343                 iter.path->preserve = false;
 344 err:
 345         set_btree_iter_dontneed(&iter);
 346         bch2_trans_iter_exit(trans, &iter);
 347         printbuf_exit(&buf);
 348         return ob;
 349 }
 350
 351 static struct open_bucket *try_alloc_partial_bucket(struct bch_fs *c, struct bch_dev *ca,
 352                                                     enum alloc_reserve reserve)
 353 {
 354         struct open_bucket *ob;
 355         int i;
 356
 357         spin_lock(&c->freelist_lock);
 358
 359         for (i = ca->open_buckets_partial_nr - 1; i >= 0; --i) {
 360                 ob = c->open_buckets + ca->open_buckets_partial[i];
 361
 362                 if (reserve <= ob->alloc_reserve) {
 363                         array_remove_item(ca->open_buckets_partial,
 364                                           ca->open_buckets_partial_nr,
 365                                           i);
 366                         ob->on_partial_list = false;
 367                         ob->alloc_reserve = reserve;
 368                         spin_unlock(&c->freelist_lock);
 369                         return ob;
 370                 }
 371         }
 372
 373         spin_unlock(&c->freelist_lock);
 374         return NULL;
 375 }
 376
 377 /*
 378  * This path is for before the freespace btree is initialized:
 379  *
 380  * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
 381  * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
 382  */
 383 static noinline struct open_bucket *
 384 bch2_bucket_alloc_early(struct btree_trans *trans,
 385                         struct bch_dev *ca,
 386                         enum alloc_reserve reserve,
 387                         u64 *cur_bucket,
 388                         u64 *buckets_seen,
 389                         u64 *skipped_open,
 390                         u64 *skipped_need_journal_commit,
 391                         u64 *skipped_nouse,
 392                         struct closure *cl)
 393 {
 394         struct btree_iter iter;
 395         struct bkey_s_c k;
 396         struct open_bucket *ob = NULL;
 397         int ret;
 398
 399         *cur_bucket = max_t(u64, *cur_bucket, ca->mi.first_bucket);
 400         *cur_bucket = max_t(u64, *cur_bucket, ca->new_fs_bucket_idx);
 401
 402         for_each_btree_key_norestart(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, *cur_bucket),
 403                            BTREE_ITER_SLOTS, k, ret) {
 404                 struct bch_alloc_v4 a;
 405
 406                 if (bkey_cmp(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)) >= 0)
 407                         break;
 408
 409                 if (ca->new_fs_bucket_idx &&
 410                     is_superblock_bucket(ca, k.k->p.offset))
 411                         continue;
 412
 413                 bch2_alloc_to_v4(k, &a);
 414
 415                 if (a.data_type != BCH_DATA_free)
 416                         continue;
 417
 418                 (*buckets_seen)++;
 419
 420                 ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, reserve, &a,
 421                                         skipped_open,
 422                                         skipped_need_journal_commit,
 423                                         skipped_nouse,
 424                                         cl);
 425                 if (ob)
 426                         break;
 427         }
 428         bch2_trans_iter_exit(trans, &iter);
 429
 430         *cur_bucket = iter.pos.offset;
 431
 432         return ob ?: ERR_PTR(ret ?: -BCH_ERR_no_buckets_found);
 433 }
 434
 435 static struct open_bucket *bch2_bucket_alloc_freelist(struct btree_trans *trans,
 436                                                    struct bch_dev *ca,
 437                                                    enum alloc_reserve reserve,
 438                                                    u64 *cur_bucket,
 439                                                    u64 *buckets_seen,
 440                                                    u64 *skipped_open,
 441                                                    u64 *skipped_need_journal_commit,
 442                                                    u64 *skipped_nouse,
 443                                                    struct closure *cl)
 444 {
 445         struct btree_iter iter;
 446         struct bkey_s_c k;
 447         struct open_bucket *ob = NULL;
 448         int ret;
 449
 450         BUG_ON(ca->new_fs_bucket_idx);
 451
 452         /*
 453          * XXX:
 454          * On transaction restart, we'd like to restart from the bucket we were
 455          * at previously
 456          */
 457         for_each_btree_key_norestart(trans, iter, BTREE_ID_freespace,
 458                                      POS(ca->dev_idx, *cur_bucket), 0, k, ret) {
 459                 if (k.k->p.inode != ca->dev_idx)
 460                         break;
 461
 462                 for (*cur_bucket = max(*cur_bucket, bkey_start_offset(k.k));
 463                      *cur_bucket < k.k->p.offset;
 464                      (*cur_bucket)++) {
 465                         ret = btree_trans_too_many_iters(trans);
 466                         if (ret)
 467                                 break;
 468
 469                         (*buckets_seen)++;
 470
 471                         ob = try_alloc_bucket(trans, ca, reserve,
 472                                               *cur_bucket,
 473                                               skipped_open,
 474                                               skipped_need_journal_commit,
 475                                               skipped_nouse,
 476                                               k, cl);
 477                         if (ob)
 478                                 break;
 479                 }
 480
 481                 if (ob || ret)
 482                         break;
 483         }
 484         bch2_trans_iter_exit(trans, &iter);
 485
 486         return ob ?: ERR_PTR(ret);
 487 }
 488
 489 /**
 490  * bch_bucket_alloc - allocate a single bucket from a specific device
 491  *
 492  * Returns index of bucket on success, 0 on failure
 493  * */
 494 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
 495                                       struct bch_dev *ca,
 496                                       enum alloc_reserve reserve,
 497                                       bool may_alloc_partial,
 498                                       struct closure *cl)
 499 {
 500         struct bch_fs *c = trans->c;
 501         struct open_bucket *ob = NULL;
 502         struct bch_dev_usage usage;
 503         bool freespace_initialized = READ_ONCE(ca->mi.freespace_initialized);
 504         u64 start = freespace_initialized ? 0 : ca->bucket_alloc_trans_early_cursor;
 505         u64 avail;
 506         u64 cur_bucket = start;
 507         u64 buckets_seen = 0;
 508         u64 skipped_open = 0;
 509         u64 skipped_need_journal_commit = 0;
 510         u64 skipped_nouse = 0;
 511         bool waiting = false;
 512 again:
 513         usage = bch2_dev_usage_read(ca);
 514         avail = dev_buckets_free(ca, usage, reserve);
 515
 516         if (usage.d[BCH_DATA_need_discard].buckets > avail)
 517                 bch2_do_discards(c);
 518
 519         if (usage.d[BCH_DATA_need_gc_gens].buckets > avail)
 520                 bch2_do_gc_gens(c);
 521
 522         if (should_invalidate_buckets(ca, usage))
 523                 bch2_do_invalidates(c);
 524
 525         if (!avail) {
 526                 if (cl && !waiting) {
 527                         closure_wait(&c->freelist_wait, cl);
 528                         waiting = true;
 529                         goto again;
 530                 }
 531
 532                 if (!c->blocked_allocate)
 533                         c->blocked_allocate = local_clock();
 534
 535                 ob = ERR_PTR(-BCH_ERR_freelist_empty);
 536                 goto err;
 537         }
 538
 539         if (waiting)
 540                 closure_wake_up(&c->freelist_wait);
 541
 542         if (may_alloc_partial) {
 543                 ob = try_alloc_partial_bucket(c, ca, reserve);
 544                 if (ob)
 545                         return ob;
 546         }
 547
 548         ob = likely(ca->mi.freespace_initialized)
 549                 ? bch2_bucket_alloc_freelist(trans, ca, reserve,
 550                                         &cur_bucket,
 551                                         &buckets_seen,
 552                                         &skipped_open,
 553                                         &skipped_need_journal_commit,
 554                                         &skipped_nouse,
 555                                         cl)
 556                 : bch2_bucket_alloc_early(trans, ca, reserve,
 557                                         &cur_bucket,
 558                                         &buckets_seen,
 559                                         &skipped_open,
 560                                         &skipped_need_journal_commit,
 561                                         &skipped_nouse,
 562                                         cl);
 563
 564         if (skipped_need_journal_commit * 2 > avail)
 565                 bch2_journal_flush_async(&c->journal, NULL);
 566
 567         if (!ob && !freespace_initialized && start) {
 568                 start = cur_bucket = 0;
 569                 goto again;
 570         }
 571
 572         if (!freespace_initialized)
 573                 ca->bucket_alloc_trans_early_cursor = cur_bucket;
 574 err:
 575         if (!ob)
 576                 ob = ERR_PTR(-BCH_ERR_no_buckets_found);
 577
 578         if (IS_ERR(ob)) {
 579                 trace_bucket_alloc_fail(ca, bch2_alloc_reserves[reserve],
 580                                         usage.d[BCH_DATA_free].buckets,
 581                                         avail,
 582                                         bch2_copygc_wait_amount(c),
 583                                         c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
 584                                         buckets_seen,
 585                                         skipped_open,
 586                                         skipped_need_journal_commit,
 587                                         skipped_nouse,
 588                                         cl == NULL,
 589                                         bch2_err_str(PTR_ERR(ob)));
 590                 atomic_long_inc(&c->bucket_alloc_fail);
 591         }
 592
 593         return ob;
 594 }
 595
 596 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 597                                       enum alloc_reserve reserve,
 598                                       bool may_alloc_partial,
 599                                       struct closure *cl)
 600 {
 601         struct open_bucket *ob;
 602
 603         bch2_trans_do(c, NULL, NULL, 0,
 604                       PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, reserve,
 605                                                                    may_alloc_partial, cl)));
 606         return ob;
 607 }
 608
 609 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 610                             unsigned l, unsigned r)
 611 {
 612         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 613                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 614 }
 615
 616 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 617
 618 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 619                                           struct dev_stripe_state *stripe,
 620                                           struct bch_devs_mask *devs)
 621 {
 622         struct dev_alloc_list ret = { .nr = 0 };
 623         unsigned i;
 624
 625         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 626                 ret.devs[ret.nr++] = i;
 627
 628         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 629         return ret;
 630 }
 631
 632 void bch2_dev_stripe_increment(struct bch_dev *ca,
 633                                struct dev_stripe_state *stripe)
 634 {
 635         u64 *v = stripe->next_alloc + ca->dev_idx;
 636         u64 free_space = dev_buckets_available(ca, RESERVE_none);
 637         u64 free_space_inv = free_space
 638                 ? div64_u64(1ULL << 48, free_space)
 639                 : 1ULL << 48;
 640         u64 scale = *v / 4;
 641
 642         if (*v + free_space_inv >= *v)
 643                 *v += free_space_inv;
 644         else
 645                 *v = U64_MAX;
 646
 647         for (v = stripe->next_alloc;
 648              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 649                 *v = *v < scale ? 0 : *v - scale;
 650 }
 651
 652 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 653 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 654
 655 static void add_new_bucket(struct bch_fs *c,
 656                            struct open_buckets *ptrs,
 657                            struct bch_devs_mask *devs_may_alloc,
 658                            unsigned *nr_effective,
 659                            bool *have_cache,
 660                            unsigned flags,
 661                            struct open_bucket *ob)
 662 {
 663         unsigned durability =
 664                 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
 665
 666         __clear_bit(ob->dev, devs_may_alloc->d);
 667         *nr_effective   += (flags & BUCKET_ALLOC_USE_DURABILITY)
 668                 ? durability : 1;
 669         *have_cache     |= !durability;
 670
 671         ob_push(c, ptrs, ob);
 672 }
 673
 674 static int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
 675                       struct open_buckets *ptrs,
 676                       struct dev_stripe_state *stripe,
 677                       struct bch_devs_mask *devs_may_alloc,
 678                       unsigned nr_replicas,
 679                       unsigned *nr_effective,
 680                       bool *have_cache,
 681                       enum alloc_reserve reserve,
 682                       unsigned flags,
 683                       struct closure *cl)
 684 {
 685         struct bch_fs *c = trans->c;
 686         struct dev_alloc_list devs_sorted =
 687                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 688         unsigned dev;
 689         struct bch_dev *ca;
 690         int ret = 0;
 691         unsigned i;
 692
 693         BUG_ON(*nr_effective >= nr_replicas);
 694
 695         for (i = 0; i < devs_sorted.nr; i++) {
 696                 struct open_bucket *ob;
 697
 698                 dev = devs_sorted.devs[i];
 699
 700                 rcu_read_lock();
 701                 ca = rcu_dereference(c->devs[dev]);
 702                 if (ca)
 703                         percpu_ref_get(&ca->ref);
 704                 rcu_read_unlock();
 705
 706                 if (!ca)
 707                         continue;
 708
 709                 if (!ca->mi.durability && *have_cache) {
 710                         percpu_ref_put(&ca->ref);
 711                         continue;
 712                 }
 713
 714                 ob = bch2_bucket_alloc_trans(trans, ca, reserve,
 715                                 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
 716                 if (!IS_ERR(ob))
 717                         bch2_dev_stripe_increment(ca, stripe);
 718                 percpu_ref_put(&ca->ref);
 719
 720                 ret = PTR_ERR_OR_ZERO(ob);
 721                 if (ret) {
 722                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) || cl)
 723                                 break;
 724                         continue;
 725                 }
 726
 727                 add_new_bucket(c, ptrs, devs_may_alloc,
 728                                nr_effective, have_cache, flags, ob);
 729
 730                 if (*nr_effective >= nr_replicas)
 731                         break;
 732         }
 733
 734         if (*nr_effective >= nr_replicas)
 735                 ret = 0;
 736         else if (!ret)
 737                 ret = -BCH_ERR_insufficient_devices;
 738
 739         return ret;
 740 }
 741
 742 int bch2_bucket_alloc_set(struct bch_fs *c,
 743                       struct open_buckets *ptrs,
 744                       struct dev_stripe_state *stripe,
 745                       struct bch_devs_mask *devs_may_alloc,
 746                       unsigned nr_replicas,
 747                       unsigned *nr_effective,
 748                       bool *have_cache,
 749                       enum alloc_reserve reserve,
 750                       unsigned flags,
 751                       struct closure *cl)
 752 {
 753         return bch2_trans_do(c, NULL, NULL, 0,
 754                       bch2_bucket_alloc_set_trans(&trans, ptrs, stripe,
 755                                               devs_may_alloc, nr_replicas,
 756                                               nr_effective, have_cache, reserve,
 757                                               flags, cl));
 758 }
 759
 760 /* Allocate from stripes: */
 761
 762 /*
 763  * if we can't allocate a new stripe because there are already too many
 764  * partially filled stripes, force allocating from an existing stripe even when
 765  * it's to a device we don't want:
 766  */
 767
 768 static int bucket_alloc_from_stripe(struct bch_fs *c,
 769                          struct open_buckets *ptrs,
 770                          struct write_point *wp,
 771                          struct bch_devs_mask *devs_may_alloc,
 772                          u16 target,
 773                          unsigned erasure_code,
 774                          unsigned nr_replicas,
 775                          unsigned *nr_effective,
 776                          bool *have_cache,
 777                          unsigned flags,
 778                          struct closure *cl)
 779 {
 780         struct dev_alloc_list devs_sorted;
 781         struct ec_stripe_head *h;
 782         struct open_bucket *ob;
 783         struct bch_dev *ca;
 784         unsigned i, ec_idx;
 785
 786         if (!erasure_code)
 787                 return 0;
 788
 789         if (nr_replicas < 2)
 790                 return 0;
 791
 792         if (ec_open_bucket(c, ptrs))
 793                 return 0;
 794
 795         h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1,
 796                                     wp == &c->copygc_write_point,
 797                                     cl);
 798         if (IS_ERR(h))
 799                 return -PTR_ERR(h);
 800         if (!h)
 801                 return 0;
 802
 803         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 804
 805         for (i = 0; i < devs_sorted.nr; i++)
 806                 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 807                         if (!h->s->blocks[ec_idx])
 808                                 continue;
 809
 810                         ob = c->open_buckets + h->s->blocks[ec_idx];
 811                         if (ob->dev == devs_sorted.devs[i] &&
 812                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 813                                 goto got_bucket;
 814                 }
 815         goto out_put_head;
 816 got_bucket:
 817         ca = bch_dev_bkey_exists(c, ob->dev);
 818
 819         ob->ec_idx      = ec_idx;
 820         ob->ec          = h->s;
 821
 822         add_new_bucket(c, ptrs, devs_may_alloc,
 823                        nr_effective, have_cache, flags, ob);
 824         atomic_inc(&h->s->pin);
 825 out_put_head:
 826         bch2_ec_stripe_head_put(c, h);
 827         return 0;
 828 }
 829
 830 /* Sector allocator */
 831
 832 static void get_buckets_from_writepoint(struct bch_fs *c,
 833                                         struct open_buckets *ptrs,
 834                                         struct write_point *wp,
 835                                         struct bch_devs_mask *devs_may_alloc,
 836                                         unsigned nr_replicas,
 837                                         unsigned *nr_effective,
 838                                         bool *have_cache,
 839                                         unsigned flags,
 840                                         bool need_ec)
 841 {
 842         struct open_buckets ptrs_skip = { .nr = 0 };
 843         struct open_bucket *ob;
 844         unsigned i;
 845
 846         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 847                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 848
 849                 if (*nr_effective < nr_replicas &&
 850                     test_bit(ob->dev, devs_may_alloc->d) &&
 851                     (ca->mi.durability ||
 852                      (wp->data_type == BCH_DATA_user && !*have_cache)) &&
 853                     (ob->ec || !need_ec)) {
 854                         add_new_bucket(c, ptrs, devs_may_alloc,
 855                                        nr_effective, have_cache,
 856                                        flags, ob);
 857                 } else {
 858                         ob_push(c, &ptrs_skip, ob);
 859                 }
 860         }
 861         wp->ptrs = ptrs_skip;
 862 }
 863
 864 static int open_bucket_add_buckets(struct btree_trans *trans,
 865                         struct open_buckets *ptrs,
 866                         struct write_point *wp,
 867                         struct bch_devs_list *devs_have,
 868                         u16 target,
 869                         unsigned erasure_code,
 870                         unsigned nr_replicas,
 871                         unsigned *nr_effective,
 872                         bool *have_cache,
 873                         enum alloc_reserve reserve,
 874                         unsigned flags,
 875                         struct closure *_cl)
 876 {
 877         struct bch_fs *c = trans->c;
 878         struct bch_devs_mask devs;
 879         struct open_bucket *ob;
 880         struct closure *cl = NULL;
 881         int ret;
 882         unsigned i;
 883
 884         rcu_read_lock();
 885         devs = target_rw_devs(c, wp->data_type, target);
 886         rcu_read_unlock();
 887
 888         /* Don't allocate from devices we already have pointers to: */
 889         for (i = 0; i < devs_have->nr; i++)
 890                 __clear_bit(devs_have->devs[i], devs.d);
 891
 892         open_bucket_for_each(c, ptrs, ob, i)
 893                 __clear_bit(ob->dev, devs.d);
 894
 895         if (erasure_code) {
 896                 if (!ec_open_bucket(c, ptrs)) {
 897                         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 898                                                     nr_replicas, nr_effective,
 899                                                     have_cache, flags, true);
 900                         if (*nr_effective >= nr_replicas)
 901                                 return 0;
 902                 }
 903
 904                 if (!ec_open_bucket(c, ptrs)) {
 905                         ret = bucket_alloc_from_stripe(c, ptrs, wp, &devs,
 906                                                  target, erasure_code,
 907                                                  nr_replicas, nr_effective,
 908                                                  have_cache, flags, _cl);
 909                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
 910                             bch2_err_matches(ret, BCH_ERR_freelist_empty) ||
 911                             bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
 912                                 return ret;
 913                         if (*nr_effective >= nr_replicas)
 914                                 return 0;
 915                 }
 916         }
 917
 918         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 919                                     nr_replicas, nr_effective,
 920                                     have_cache, flags, false);
 921         if (*nr_effective >= nr_replicas)
 922                 return 0;
 923
 924 retry_blocking:
 925         /*
 926          * Try nonblocking first, so that if one device is full we'll try from
 927          * other devices:
 928          */
 929         ret = bch2_bucket_alloc_set_trans(trans, ptrs, &wp->stripe, &devs,
 930                                 nr_replicas, nr_effective, have_cache,
 931                                 reserve, flags, cl);
 932         if (ret &&
 933             !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
 934             !bch2_err_matches(ret, BCH_ERR_insufficient_devices) &&
 935             !cl && _cl) {
 936                 cl = _cl;
 937                 goto retry_blocking;
 938         }
 939
 940         return ret;
 941 }
 942
 943 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 944                                 struct open_buckets *obs)
 945 {
 946         struct open_buckets ptrs = { .nr = 0 };
 947         struct open_bucket *ob, *ob2;
 948         unsigned i, j;
 949
 950         open_bucket_for_each(c, obs, ob, i) {
 951                 bool drop = !ca || ob->dev == ca->dev_idx;
 952
 953                 if (!drop && ob->ec) {
 954                         mutex_lock(&ob->ec->lock);
 955                         for (j = 0; j < ob->ec->new_stripe.key.v.nr_blocks; j++) {
 956                                 if (!ob->ec->blocks[j])
 957                                         continue;
 958
 959                                 ob2 = c->open_buckets + ob->ec->blocks[j];
 960                                 drop |= ob2->dev == ca->dev_idx;
 961                         }
 962                         mutex_unlock(&ob->ec->lock);
 963                 }
 964
 965                 if (drop)
 966                         bch2_open_bucket_put(c, ob);
 967                 else
 968                         ob_push(c, &ptrs, ob);
 969         }
 970
 971         *obs = ptrs;
 972 }
 973
 974 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 975                           struct write_point *wp)
 976 {
 977         mutex_lock(&wp->lock);
 978         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
 979         mutex_unlock(&wp->lock);
 980 }
 981
 982 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 983                                                  unsigned long write_point)
 984 {
 985         unsigned hash =
 986                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
 987
 988         return &c->write_points_hash[hash];
 989 }
 990
 991 static struct write_point *__writepoint_find(struct hlist_head *head,
 992                                              unsigned long write_point)
 993 {
 994         struct write_point *wp;
 995
 996         rcu_read_lock();
 997         hlist_for_each_entry_rcu(wp, head, node)
 998                 if (wp->write_point == write_point)
 999                         goto out;
1000         wp = NULL;
1001 out:
1002         rcu_read_unlock();
1003         return wp;
1004 }
1005
1006 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
1007 {
1008         u64 stranded    = c->write_points_nr * c->bucket_size_max;
1009         u64 free        = bch2_fs_usage_read_short(c).free;
1010
1011         return stranded * factor > free;
1012 }
1013
1014 static bool try_increase_writepoints(struct bch_fs *c)
1015 {
1016         struct write_point *wp;
1017
1018         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
1019             too_many_writepoints(c, 32))
1020                 return false;
1021
1022         wp = c->write_points + c->write_points_nr++;
1023         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
1024         return true;
1025 }
1026
1027 static bool try_decrease_writepoints(struct bch_fs *c,
1028                                      unsigned old_nr)
1029 {
1030         struct write_point *wp;
1031
1032         mutex_lock(&c->write_points_hash_lock);
1033         if (c->write_points_nr < old_nr) {
1034                 mutex_unlock(&c->write_points_hash_lock);
1035                 return true;
1036         }
1037
1038         if (c->write_points_nr == 1 ||
1039             !too_many_writepoints(c, 8)) {
1040                 mutex_unlock(&c->write_points_hash_lock);
1041                 return false;
1042         }
1043
1044         wp = c->write_points + --c->write_points_nr;
1045
1046         hlist_del_rcu(&wp->node);
1047         mutex_unlock(&c->write_points_hash_lock);
1048
1049         bch2_writepoint_stop(c, NULL, wp);
1050         return true;
1051 }
1052
1053 static void bch2_trans_mutex_lock(struct btree_trans *trans,
1054                                   struct mutex *lock)
1055 {
1056         if (!mutex_trylock(lock)) {
1057                 bch2_trans_unlock(trans);
1058                 mutex_lock(lock);
1059         }
1060 }
1061
1062 static struct write_point *writepoint_find(struct btree_trans *trans,
1063                                            unsigned long write_point)
1064 {
1065         struct bch_fs *c = trans->c;
1066         struct write_point *wp, *oldest;
1067         struct hlist_head *head;
1068
1069         if (!(write_point & 1UL)) {
1070                 wp = (struct write_point *) write_point;
1071                 bch2_trans_mutex_lock(trans, &wp->lock);
1072                 return wp;
1073         }
1074
1075         head = writepoint_hash(c, write_point);
1076 restart_find:
1077         wp = __writepoint_find(head, write_point);
1078         if (wp) {
1079 lock_wp:
1080                 bch2_trans_mutex_lock(trans, &wp->lock);
1081                 if (wp->write_point == write_point)
1082                         goto out;
1083                 mutex_unlock(&wp->lock);
1084                 goto restart_find;
1085         }
1086 restart_find_oldest:
1087         oldest = NULL;
1088         for (wp = c->write_points;
1089              wp < c->write_points + c->write_points_nr; wp++)
1090                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1091                         oldest = wp;
1092
1093         bch2_trans_mutex_lock(trans, &oldest->lock);
1094         bch2_trans_mutex_lock(trans, &c->write_points_hash_lock);
1095         if (oldest >= c->write_points + c->write_points_nr ||
1096             try_increase_writepoints(c)) {
1097                 mutex_unlock(&c->write_points_hash_lock);
1098                 mutex_unlock(&oldest->lock);
1099                 goto restart_find_oldest;
1100         }
1101
1102         wp = __writepoint_find(head, write_point);
1103         if (wp && wp != oldest) {
1104                 mutex_unlock(&c->write_points_hash_lock);
1105                 mutex_unlock(&oldest->lock);
1106                 goto lock_wp;
1107         }
1108
1109         wp = oldest;
1110         hlist_del_rcu(&wp->node);
1111         wp->write_point = write_point;
1112         hlist_add_head_rcu(&wp->node, head);
1113         mutex_unlock(&c->write_points_hash_lock);
1114 out:
1115         wp->last_used = sched_clock();
1116         return wp;
1117 }
1118
1119 /*
1120  * Get us an open_bucket we can allocate from, return with it locked:
1121  */
1122 struct write_point *bch2_alloc_sectors_start_trans(struct btree_trans *trans,
1123                                 unsigned target,
1124                                 unsigned erasure_code,
1125                                 struct write_point_specifier write_point,
1126                                 struct bch_devs_list *devs_have,
1127                                 unsigned nr_replicas,
1128                                 unsigned nr_replicas_required,
1129                                 enum alloc_reserve reserve,
1130                                 unsigned flags,
1131                                 struct closure *cl)
1132 {
1133         struct bch_fs *c = trans->c;
1134         struct write_point *wp;
1135         struct open_bucket *ob;
1136         struct open_buckets ptrs;
1137         unsigned nr_effective, write_points_nr;
1138         unsigned ob_flags = 0;
1139         bool have_cache;
1140         int ret;
1141         int i;
1142
1143         if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
1144                 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
1145
1146         BUG_ON(!nr_replicas || !nr_replicas_required);
1147 retry:
1148         ptrs.nr         = 0;
1149         nr_effective    = 0;
1150         write_points_nr = c->write_points_nr;
1151         have_cache      = false;
1152
1153         wp = writepoint_find(trans, write_point.v);
1154
1155         if (wp->data_type == BCH_DATA_user)
1156                 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
1157
1158         /* metadata may not allocate on cache devices: */
1159         if (wp->data_type != BCH_DATA_user)
1160                 have_cache = true;
1161
1162         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1163                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1164                                               target, erasure_code,
1165                                               nr_replicas, &nr_effective,
1166                                               &have_cache, reserve,
1167                                               ob_flags, cl);
1168         } else {
1169                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1170                                               target, erasure_code,
1171                                               nr_replicas, &nr_effective,
1172                                               &have_cache, reserve,
1173                                               ob_flags, NULL);
1174                 if (!ret ||
1175                     bch2_err_matches(ret, BCH_ERR_transaction_restart))
1176                         goto alloc_done;
1177
1178                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1179                                               0, erasure_code,
1180                                               nr_replicas, &nr_effective,
1181                                               &have_cache, reserve,
1182                                               ob_flags, cl);
1183         }
1184 alloc_done:
1185         BUG_ON(!ret && nr_effective < nr_replicas);
1186
1187         if (erasure_code && !ec_open_bucket(c, &ptrs))
1188                 pr_debug("failed to get ec bucket: ret %u", ret);
1189
1190         if (ret == -BCH_ERR_insufficient_devices &&
1191             nr_effective >= nr_replicas_required)
1192                 ret = 0;
1193
1194         if (ret)
1195                 goto err;
1196
1197         /* Free buckets we didn't use: */
1198         open_bucket_for_each(c, &wp->ptrs, ob, i)
1199                 open_bucket_free_unused(c, wp, ob);
1200
1201         wp->ptrs = ptrs;
1202
1203         wp->sectors_free = UINT_MAX;
1204
1205         open_bucket_for_each(c, &wp->ptrs, ob, i)
1206                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1207
1208         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1209
1210         return wp;
1211 err:
1212         open_bucket_for_each(c, &wp->ptrs, ob, i)
1213                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1214                         ob_push(c, &ptrs, ob);
1215                 else
1216                         open_bucket_free_unused(c, wp, ob);
1217         wp->ptrs = ptrs;
1218
1219         mutex_unlock(&wp->lock);
1220
1221         if (bch2_err_matches(ret, BCH_ERR_freelist_empty) &&
1222             try_decrease_writepoints(c, write_points_nr))
1223                 goto retry;
1224
1225         if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty) ||
1226             bch2_err_matches(ret, BCH_ERR_freelist_empty))
1227                 return cl ? ERR_PTR(-EAGAIN) : ERR_PTR(-ENOSPC);
1228
1229         if (bch2_err_matches(ret, BCH_ERR_insufficient_devices))
1230                 return ERR_PTR(-EROFS);
1231
1232         return ERR_PTR(ret);
1233 }
1234
1235 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
1236                                 unsigned target,
1237                                 unsigned erasure_code,
1238                                 struct write_point_specifier write_point,
1239                                 struct bch_devs_list *devs_have,
1240                                 unsigned nr_replicas,
1241                                 unsigned nr_replicas_required,
1242                                 enum alloc_reserve reserve,
1243                                 unsigned flags,
1244                                 struct closure *cl)
1245 {
1246         struct write_point *wp;
1247
1248         bch2_trans_do(c, NULL, NULL, 0,
1249                       PTR_ERR_OR_ZERO(wp = bch2_alloc_sectors_start_trans(&trans, target,
1250                                                         erasure_code,
1251                                                         write_point,
1252                                                         devs_have,
1253                                                         nr_replicas,
1254                                                         nr_replicas_required,
1255                                                         reserve,
1256                                                         flags, cl)));
1257         return wp;
1258
1259 }
1260
1261 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1262 {
1263         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1264
1265         return (struct bch_extent_ptr) {
1266                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
1267                 .gen    = ob->gen,
1268                 .dev    = ob->dev,
1269                 .offset = bucket_to_sector(ca, ob->bucket) +
1270                         ca->mi.bucket_size -
1271                         ob->sectors_free,
1272         };
1273 }
1274
1275 /*
1276  * Append pointers to the space we just allocated to @k, and mark @sectors space
1277  * as allocated out of @ob
1278  */
1279 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1280                                     struct bkey_i *k, unsigned sectors,
1281                                     bool cached)
1282
1283 {
1284         struct open_bucket *ob;
1285         unsigned i;
1286
1287         BUG_ON(sectors > wp->sectors_free);
1288         wp->sectors_free -= sectors;
1289
1290         open_bucket_for_each(c, &wp->ptrs, ob, i) {
1291                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1292                 struct bch_extent_ptr ptr = bch2_ob_ptr(c, ob);
1293
1294                 ptr.cached = cached ||
1295                         (!ca->mi.durability &&
1296                          wp->data_type == BCH_DATA_user);
1297
1298                 bch2_bkey_append_ptr(k, ptr);
1299
1300                 BUG_ON(sectors > ob->sectors_free);
1301                 ob->sectors_free -= sectors;
1302         }
1303 }
1304
1305 /*
1306  * Append pointers to the space we just allocated to @k, and mark @sectors space
1307  * as allocated out of @ob
1308  */
1309 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1310 {
1311         struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
1312         struct open_bucket *ob;
1313         unsigned i;
1314
1315         open_bucket_for_each(c, &wp->ptrs, ob, i)
1316                 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
1317         wp->ptrs = keep;
1318
1319         mutex_unlock(&wp->lock);
1320
1321         bch2_open_buckets_put(c, &ptrs);
1322 }
1323
1324 static inline void writepoint_init(struct write_point *wp,
1325                                    enum bch_data_type type)
1326 {
1327         mutex_init(&wp->lock);
1328         wp->data_type = type;
1329 }
1330
1331 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1332 {
1333         struct open_bucket *ob;
1334         struct write_point *wp;
1335
1336         mutex_init(&c->write_points_hash_lock);
1337         c->write_points_nr = ARRAY_SIZE(c->write_points);
1338
1339         /* open bucket 0 is a sentinal NULL: */
1340         spin_lock_init(&c->open_buckets[0].lock);
1341
1342         for (ob = c->open_buckets + 1;
1343              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1344                 spin_lock_init(&ob->lock);
1345                 c->open_buckets_nr_free++;
1346
1347                 ob->freelist = c->open_buckets_freelist;
1348                 c->open_buckets_freelist = ob - c->open_buckets;
1349         }
1350
1351         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
1352         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
1353         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
1354
1355         for (wp = c->write_points;
1356              wp < c->write_points + c->write_points_nr; wp++) {
1357                 writepoint_init(wp, BCH_DATA_user);
1358
1359                 wp->last_used   = sched_clock();
1360                 wp->write_point = (unsigned long) wp;
1361                 hlist_add_head_rcu(&wp->node,
1362                                    writepoint_hash(c, wp->write_point));
1363         }
1364 }
1365
1366 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1367 {
1368         struct open_bucket *ob;
1369
1370         for (ob = c->open_buckets;
1371              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1372              ob++) {
1373                 spin_lock(&ob->lock);
1374                 if (ob->valid && !ob->on_partial_list) {
1375                         prt_printf(out, "%zu ref %u type %s %u:%llu:%u\n",
1376                                ob - c->open_buckets,
1377                                atomic_read(&ob->pin),
1378                                bch2_data_types[ob->data_type],
1379                                ob->dev, ob->bucket, ob->gen);
1380                 }
1381                 spin_unlock(&ob->lock);
1382         }
1383 }