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         return ob;
 272 }
 273
 274 static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bch_dev *ca,
 275                                             enum alloc_reserve reserve, u64 free_entry,
 276                                             u64 *skipped_open,
 277                                             u64 *skipped_need_journal_commit,
 278                                             u64 *skipped_nouse,
 279                                             struct bkey_s_c freespace_k,
 280                                             struct closure *cl)
 281 {
 282         struct bch_fs *c = trans->c;
 283         struct btree_iter iter = { NULL };
 284         struct bkey_s_c k;
 285         struct open_bucket *ob;
 286         struct bch_alloc_v4 a;
 287         u64 b = free_entry & ~(~0ULL << 56);
 288         unsigned genbits = free_entry >> 56;
 289         struct printbuf buf = PRINTBUF;
 290         int ret;
 291
 292         if (b < ca->mi.first_bucket || b >= ca->mi.nbuckets) {
 293                 prt_printf(&buf, "freespace btree has bucket outside allowed range %u-%llu\n"
 294                        "  freespace key ",
 295                         ca->mi.first_bucket, ca->mi.nbuckets);
 296                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 297                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 298                 ob = ERR_PTR(-EIO);
 299                 goto err;
 300         }
 301
 302         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS(ca->dev_idx, b), BTREE_ITER_CACHED);
 303         k = bch2_btree_iter_peek_slot(&iter);
 304         ret = bkey_err(k);
 305         if (ret) {
 306                 ob = ERR_PTR(ret);
 307                 goto err;
 308         }
 309
 310         bch2_alloc_to_v4(k, &a);
 311
 312         if (genbits != (alloc_freespace_genbits(a) >> 56)) {
 313                 prt_printf(&buf, "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
 314                        "  freespace key ",
 315                        genbits, alloc_freespace_genbits(a) >> 56);
 316                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 317                 prt_printf(&buf, "\n  ");
 318                 bch2_bkey_val_to_text(&buf, c, k);
 319                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 320                 ob = ERR_PTR(-EIO);
 321                 goto err;
 322
 323         }
 324
 325         if (a.data_type != BCH_DATA_free) {
 326                 prt_printf(&buf, "non free bucket in freespace btree\n"
 327                        "  freespace key ");
 328                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 329                 prt_printf(&buf, "\n  ");
 330                 bch2_bkey_val_to_text(&buf, c, k);
 331                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 332                 ob = ERR_PTR(-EIO);
 333                 goto err;
 334         }
 335
 336         ob = __try_alloc_bucket(c, ca, b, reserve, &a,
 337                                 skipped_open,
 338                                 skipped_need_journal_commit,
 339                                 skipped_nouse,
 340                                 cl);
 341         if (!ob)
 342                 iter.path->preserve = false;
 343 err:
 344         set_btree_iter_dontneed(&iter);
 345         bch2_trans_iter_exit(trans, &iter);
 346         printbuf_exit(&buf);
 347         return ob;
 348 }
 349
 350 static struct open_bucket *try_alloc_partial_bucket(struct bch_fs *c, struct bch_dev *ca,
 351                                                     enum alloc_reserve reserve)
 352 {
 353         struct open_bucket *ob;
 354         int i;
 355
 356         spin_lock(&c->freelist_lock);
 357
 358         for (i = ca->open_buckets_partial_nr - 1; i >= 0; --i) {
 359                 ob = c->open_buckets + ca->open_buckets_partial[i];
 360
 361                 if (reserve <= ob->alloc_reserve) {
 362                         array_remove_item(ca->open_buckets_partial,
 363                                           ca->open_buckets_partial_nr,
 364                                           i);
 365                         ob->on_partial_list = false;
 366                         ob->alloc_reserve = reserve;
 367                         spin_unlock(&c->freelist_lock);
 368                         return ob;
 369                 }
 370         }
 371
 372         spin_unlock(&c->freelist_lock);
 373         return NULL;
 374 }
 375
 376 /*
 377  * This path is for before the freespace btree is initialized:
 378  *
 379  * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
 380  * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
 381  */
 382 static noinline struct open_bucket *
 383 bch2_bucket_alloc_early(struct btree_trans *trans,
 384                         struct bch_dev *ca,
 385                         enum alloc_reserve reserve,
 386                         u64 *cur_bucket,
 387                         u64 *buckets_seen,
 388                         u64 *skipped_open,
 389                         u64 *skipped_need_journal_commit,
 390                         u64 *skipped_nouse,
 391                         struct closure *cl)
 392 {
 393         struct btree_iter iter;
 394         struct bkey_s_c k;
 395         struct open_bucket *ob = NULL;
 396         int ret;
 397
 398         *cur_bucket = max_t(u64, *cur_bucket, ca->mi.first_bucket);
 399         *cur_bucket = max_t(u64, *cur_bucket, ca->new_fs_bucket_idx);
 400
 401         for_each_btree_key_norestart(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, *cur_bucket),
 402                            BTREE_ITER_SLOTS, k, ret) {
 403                 struct bch_alloc_v4 a;
 404
 405                 if (bkey_cmp(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)) >= 0)
 406                         break;
 407
 408                 if (ca->new_fs_bucket_idx &&
 409                     is_superblock_bucket(ca, k.k->p.offset))
 410                         continue;
 411
 412                 bch2_alloc_to_v4(k, &a);
 413
 414                 if (a.data_type != BCH_DATA_free)
 415                         continue;
 416
 417                 (*buckets_seen)++;
 418
 419                 ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, reserve, &a,
 420                                         skipped_open,
 421                                         skipped_need_journal_commit,
 422                                         skipped_nouse,
 423                                         cl);
 424                 if (ob)
 425                         break;
 426         }
 427         bch2_trans_iter_exit(trans, &iter);
 428
 429         *cur_bucket = iter.pos.offset;
 430
 431         return ob ?: ERR_PTR(ret ?: -BCH_ERR_no_buckets_found);
 432 }
 433
 434 static struct open_bucket *bch2_bucket_alloc_freelist(struct btree_trans *trans,
 435                                                    struct bch_dev *ca,
 436                                                    enum alloc_reserve reserve,
 437                                                    u64 *cur_bucket,
 438                                                    u64 *buckets_seen,
 439                                                    u64 *skipped_open,
 440                                                    u64 *skipped_need_journal_commit,
 441                                                    u64 *skipped_nouse,
 442                                                    struct closure *cl)
 443 {
 444         struct btree_iter iter;
 445         struct bkey_s_c k;
 446         struct open_bucket *ob = NULL;
 447         int ret;
 448
 449         BUG_ON(ca->new_fs_bucket_idx);
 450
 451         /*
 452          * XXX:
 453          * On transaction restart, we'd like to restart from the bucket we were
 454          * at previously
 455          */
 456         for_each_btree_key_norestart(trans, iter, BTREE_ID_freespace,
 457                                      POS(ca->dev_idx, *cur_bucket), 0, k, ret) {
 458                 if (k.k->p.inode != ca->dev_idx)
 459                         break;
 460
 461                 for (*cur_bucket = max(*cur_bucket, bkey_start_offset(k.k));
 462                      *cur_bucket < k.k->p.offset;
 463                      (*cur_bucket)++) {
 464                         ret = btree_trans_too_many_iters(trans);
 465                         if (ret)
 466                                 break;
 467
 468                         (*buckets_seen)++;
 469
 470                         ob = try_alloc_bucket(trans, ca, reserve,
 471                                               *cur_bucket,
 472                                               skipped_open,
 473                                               skipped_need_journal_commit,
 474                                               skipped_nouse,
 475                                               k, cl);
 476                         if (ob)
 477                                 break;
 478                 }
 479
 480                 if (ob || ret)
 481                         break;
 482         }
 483         bch2_trans_iter_exit(trans, &iter);
 484
 485         return ob ?: ERR_PTR(ret);
 486 }
 487
 488 /**
 489  * bch_bucket_alloc - allocate a single bucket from a specific device
 490  *
 491  * Returns index of bucket on success, 0 on failure
 492  */
 493 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
 494                                       struct bch_dev *ca,
 495                                       enum alloc_reserve reserve,
 496                                       bool may_alloc_partial,
 497                                       struct closure *cl,
 498                                       struct bch_dev_usage *usage)
 499 {
 500         struct bch_fs *c = trans->c;
 501         struct open_bucket *ob = NULL;
 502         bool freespace_initialized = READ_ONCE(ca->mi.freespace_initialized);
 503         u64 start = freespace_initialized ? 0 : ca->bucket_alloc_trans_early_cursor;
 504         u64 avail;
 505         u64 cur_bucket = start;
 506         u64 buckets_seen = 0;
 507         u64 skipped_open = 0;
 508         u64 skipped_need_journal_commit = 0;
 509         u64 skipped_nouse = 0;
 510         bool waiting = false;
 511 again:
 512         bch2_dev_usage_read_fast(ca, usage);
 513         avail = dev_buckets_free(ca, *usage, reserve);
 514
 515         if (usage->d[BCH_DATA_need_discard].buckets > avail)
 516                 bch2_do_discards(c);
 517
 518         if (usage->d[BCH_DATA_need_gc_gens].buckets > avail)
 519                 bch2_do_gc_gens(c);
 520
 521         if (should_invalidate_buckets(ca, *usage))
 522                 bch2_do_invalidates(c);
 523
 524         if (!avail) {
 525                 if (cl && !waiting) {
 526                         closure_wait(&c->freelist_wait, cl);
 527                         waiting = true;
 528                         goto again;
 529                 }
 530
 531                 if (!c->blocked_allocate)
 532                         c->blocked_allocate = local_clock();
 533
 534                 ob = ERR_PTR(-BCH_ERR_freelist_empty);
 535                 goto err;
 536         }
 537
 538         if (waiting)
 539                 closure_wake_up(&c->freelist_wait);
 540
 541         if (may_alloc_partial) {
 542                 ob = try_alloc_partial_bucket(c, ca, reserve);
 543                 if (ob)
 544                         return ob;
 545         }
 546
 547         ob = likely(ca->mi.freespace_initialized)
 548                 ? bch2_bucket_alloc_freelist(trans, ca, reserve,
 549                                         &cur_bucket,
 550                                         &buckets_seen,
 551                                         &skipped_open,
 552                                         &skipped_need_journal_commit,
 553                                         &skipped_nouse,
 554                                         cl)
 555                 : bch2_bucket_alloc_early(trans, ca, reserve,
 556                                         &cur_bucket,
 557                                         &buckets_seen,
 558                                         &skipped_open,
 559                                         &skipped_need_journal_commit,
 560                                         &skipped_nouse,
 561                                         cl);
 562
 563         if (skipped_need_journal_commit * 2 > avail)
 564                 bch2_journal_flush_async(&c->journal, NULL);
 565
 566         if (!ob && !freespace_initialized && start) {
 567                 start = cur_bucket = 0;
 568                 goto again;
 569         }
 570
 571         if (!freespace_initialized)
 572                 ca->bucket_alloc_trans_early_cursor = cur_bucket;
 573 err:
 574         if (!ob)
 575                 ob = ERR_PTR(-BCH_ERR_no_buckets_found);
 576
 577         if (!IS_ERR(ob))
 578                 trace_and_count(c, bucket_alloc, ca, bch2_alloc_reserves[reserve],
 579                                 may_alloc_partial, ob->bucket);
 580         else if (!bch2_err_matches(PTR_ERR(ob), BCH_ERR_transaction_restart))
 581                 trace_and_count(c, bucket_alloc_fail,
 582                                 ca, bch2_alloc_reserves[reserve],
 583                                 usage->d[BCH_DATA_free].buckets,
 584                                 avail,
 585                                 bch2_copygc_wait_amount(c),
 586                                 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
 587                                 buckets_seen,
 588                                 skipped_open,
 589                                 skipped_need_journal_commit,
 590                                 skipped_nouse,
 591                                 cl == NULL,
 592                                 bch2_err_str(PTR_ERR(ob)));
 593
 594         return ob;
 595 }
 596
 597 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 598                                       enum alloc_reserve reserve,
 599                                       bool may_alloc_partial,
 600                                       struct closure *cl)
 601 {
 602         struct bch_dev_usage usage;
 603         struct open_bucket *ob;
 604
 605         bch2_trans_do(c, NULL, NULL, 0,
 606                       PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, reserve,
 607                                                         may_alloc_partial, cl, &usage)));
 608         return ob;
 609 }
 610
 611 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 612                             unsigned l, unsigned r)
 613 {
 614         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 615                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 616 }
 617
 618 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 619
 620 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 621                                           struct dev_stripe_state *stripe,
 622                                           struct bch_devs_mask *devs)
 623 {
 624         struct dev_alloc_list ret = { .nr = 0 };
 625         unsigned i;
 626
 627         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 628                 ret.devs[ret.nr++] = i;
 629
 630         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 631         return ret;
 632 }
 633
 634 static inline void bch2_dev_stripe_increment_inlined(struct bch_dev *ca,
 635                                struct dev_stripe_state *stripe,
 636                                struct bch_dev_usage *usage)
 637 {
 638         u64 *v = stripe->next_alloc + ca->dev_idx;
 639         u64 free_space = dev_buckets_available(ca, RESERVE_none);
 640         u64 free_space_inv = free_space
 641                 ? div64_u64(1ULL << 48, free_space)
 642                 : 1ULL << 48;
 643         u64 scale = *v / 4;
 644
 645         if (*v + free_space_inv >= *v)
 646                 *v += free_space_inv;
 647         else
 648                 *v = U64_MAX;
 649
 650         for (v = stripe->next_alloc;
 651              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 652                 *v = *v < scale ? 0 : *v - scale;
 653 }
 654
 655 void bch2_dev_stripe_increment(struct bch_dev *ca,
 656                                struct dev_stripe_state *stripe)
 657 {
 658         struct bch_dev_usage usage;
 659
 660         bch2_dev_usage_read_fast(ca, &usage);
 661         bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
 662 }
 663
 664 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 665 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 666
 667 static void add_new_bucket(struct bch_fs *c,
 668                            struct open_buckets *ptrs,
 669                            struct bch_devs_mask *devs_may_alloc,
 670                            unsigned *nr_effective,
 671                            bool *have_cache,
 672                            unsigned flags,
 673                            struct open_bucket *ob)
 674 {
 675         unsigned durability =
 676                 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
 677
 678         __clear_bit(ob->dev, devs_may_alloc->d);
 679         *nr_effective   += (flags & BUCKET_ALLOC_USE_DURABILITY)
 680                 ? durability : 1;
 681         *have_cache     |= !durability;
 682
 683         ob_push(c, ptrs, ob);
 684 }
 685
 686 static int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
 687                       struct open_buckets *ptrs,
 688                       struct dev_stripe_state *stripe,
 689                       struct bch_devs_mask *devs_may_alloc,
 690                       unsigned nr_replicas,
 691                       unsigned *nr_effective,
 692                       bool *have_cache,
 693                       enum alloc_reserve reserve,
 694                       unsigned flags,
 695                       struct closure *cl)
 696 {
 697         struct bch_fs *c = trans->c;
 698         struct dev_alloc_list devs_sorted =
 699                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 700         unsigned dev;
 701         struct bch_dev *ca;
 702         int ret = -BCH_ERR_insufficient_devices;
 703         unsigned i;
 704
 705         BUG_ON(*nr_effective >= nr_replicas);
 706
 707         for (i = 0; i < devs_sorted.nr; i++) {
 708                 struct bch_dev_usage usage;
 709                 struct open_bucket *ob;
 710
 711                 dev = devs_sorted.devs[i];
 712
 713                 rcu_read_lock();
 714                 ca = rcu_dereference(c->devs[dev]);
 715                 if (ca)
 716                         percpu_ref_get(&ca->ref);
 717                 rcu_read_unlock();
 718
 719                 if (!ca)
 720                         continue;
 721
 722                 if (!ca->mi.durability && *have_cache) {
 723                         percpu_ref_put(&ca->ref);
 724                         continue;
 725                 }
 726
 727                 ob = bch2_bucket_alloc_trans(trans, ca, reserve,
 728                                 flags & BUCKET_MAY_ALLOC_PARTIAL, cl, &usage);
 729                 if (!IS_ERR(ob))
 730                         bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
 731                 percpu_ref_put(&ca->ref);
 732
 733                 if (IS_ERR(ob)) {
 734                         ret = PTR_ERR(ob);
 735                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) || cl)
 736                                 break;
 737                         continue;
 738                 }
 739
 740                 add_new_bucket(c, ptrs, devs_may_alloc,
 741                                nr_effective, have_cache, flags, ob);
 742
 743                 if (*nr_effective >= nr_replicas) {
 744                         ret = 0;
 745                         break;
 746                 }
 747         }
 748
 749         return ret;
 750 }
 751
 752 int bch2_bucket_alloc_set(struct bch_fs *c,
 753                       struct open_buckets *ptrs,
 754                       struct dev_stripe_state *stripe,
 755                       struct bch_devs_mask *devs_may_alloc,
 756                       unsigned nr_replicas,
 757                       unsigned *nr_effective,
 758                       bool *have_cache,
 759                       enum alloc_reserve reserve,
 760                       unsigned flags,
 761                       struct closure *cl)
 762 {
 763         return bch2_trans_do(c, NULL, NULL, 0,
 764                       bch2_bucket_alloc_set_trans(&trans, ptrs, stripe,
 765                                               devs_may_alloc, nr_replicas,
 766                                               nr_effective, have_cache, reserve,
 767                                               flags, cl));
 768 }
 769
 770 /* Allocate from stripes: */
 771
 772 /*
 773  * if we can't allocate a new stripe because there are already too many
 774  * partially filled stripes, force allocating from an existing stripe even when
 775  * it's to a device we don't want:
 776  */
 777
 778 static int bucket_alloc_from_stripe(struct bch_fs *c,
 779                          struct open_buckets *ptrs,
 780                          struct write_point *wp,
 781                          struct bch_devs_mask *devs_may_alloc,
 782                          u16 target,
 783                          unsigned erasure_code,
 784                          unsigned nr_replicas,
 785                          unsigned *nr_effective,
 786                          bool *have_cache,
 787                          unsigned flags,
 788                          struct closure *cl)
 789 {
 790         struct dev_alloc_list devs_sorted;
 791         struct ec_stripe_head *h;
 792         struct open_bucket *ob;
 793         struct bch_dev *ca;
 794         unsigned i, ec_idx;
 795
 796         if (!erasure_code)
 797                 return 0;
 798
 799         if (nr_replicas < 2)
 800                 return 0;
 801
 802         if (ec_open_bucket(c, ptrs))
 803                 return 0;
 804
 805         h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1,
 806                                     wp == &c->copygc_write_point,
 807                                     cl);
 808         if (IS_ERR(h))
 809                 return -PTR_ERR(h);
 810         if (!h)
 811                 return 0;
 812
 813         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 814
 815         for (i = 0; i < devs_sorted.nr; i++)
 816                 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 817                         if (!h->s->blocks[ec_idx])
 818                                 continue;
 819
 820                         ob = c->open_buckets + h->s->blocks[ec_idx];
 821                         if (ob->dev == devs_sorted.devs[i] &&
 822                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 823                                 goto got_bucket;
 824                 }
 825         goto out_put_head;
 826 got_bucket:
 827         ca = bch_dev_bkey_exists(c, ob->dev);
 828
 829         ob->ec_idx      = ec_idx;
 830         ob->ec          = h->s;
 831
 832         add_new_bucket(c, ptrs, devs_may_alloc,
 833                        nr_effective, have_cache, flags, ob);
 834         atomic_inc(&h->s->pin);
 835 out_put_head:
 836         bch2_ec_stripe_head_put(c, h);
 837         return 0;
 838 }
 839
 840 /* Sector allocator */
 841
 842 static void get_buckets_from_writepoint(struct bch_fs *c,
 843                                         struct open_buckets *ptrs,
 844                                         struct write_point *wp,
 845                                         struct bch_devs_mask *devs_may_alloc,
 846                                         unsigned nr_replicas,
 847                                         unsigned *nr_effective,
 848                                         bool *have_cache,
 849                                         unsigned flags,
 850                                         bool need_ec)
 851 {
 852         struct open_buckets ptrs_skip = { .nr = 0 };
 853         struct open_bucket *ob;
 854         unsigned i;
 855
 856         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 857                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 858
 859                 if (*nr_effective < nr_replicas &&
 860                     test_bit(ob->dev, devs_may_alloc->d) &&
 861                     (ca->mi.durability ||
 862                      (wp->data_type == BCH_DATA_user && !*have_cache)) &&
 863                     (ob->ec || !need_ec)) {
 864                         add_new_bucket(c, ptrs, devs_may_alloc,
 865                                        nr_effective, have_cache,
 866                                        flags, ob);
 867                 } else {
 868                         ob_push(c, &ptrs_skip, ob);
 869                 }
 870         }
 871         wp->ptrs = ptrs_skip;
 872 }
 873
 874 static int open_bucket_add_buckets(struct btree_trans *trans,
 875                         struct open_buckets *ptrs,
 876                         struct write_point *wp,
 877                         struct bch_devs_list *devs_have,
 878                         u16 target,
 879                         unsigned erasure_code,
 880                         unsigned nr_replicas,
 881                         unsigned *nr_effective,
 882                         bool *have_cache,
 883                         enum alloc_reserve reserve,
 884                         unsigned flags,
 885                         struct closure *_cl)
 886 {
 887         struct bch_fs *c = trans->c;
 888         struct bch_devs_mask devs;
 889         struct open_bucket *ob;
 890         struct closure *cl = NULL;
 891         int ret;
 892         unsigned i;
 893
 894         rcu_read_lock();
 895         devs = target_rw_devs(c, wp->data_type, target);
 896         rcu_read_unlock();
 897
 898         /* Don't allocate from devices we already have pointers to: */
 899         for (i = 0; i < devs_have->nr; i++)
 900                 __clear_bit(devs_have->devs[i], devs.d);
 901
 902         open_bucket_for_each(c, ptrs, ob, i)
 903                 __clear_bit(ob->dev, devs.d);
 904
 905         if (erasure_code) {
 906                 if (!ec_open_bucket(c, ptrs)) {
 907                         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 908                                                     nr_replicas, nr_effective,
 909                                                     have_cache, flags, true);
 910                         if (*nr_effective >= nr_replicas)
 911                                 return 0;
 912                 }
 913
 914                 if (!ec_open_bucket(c, ptrs)) {
 915                         ret = bucket_alloc_from_stripe(c, ptrs, wp, &devs,
 916                                                  target, erasure_code,
 917                                                  nr_replicas, nr_effective,
 918                                                  have_cache, flags, _cl);
 919                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
 920                             bch2_err_matches(ret, BCH_ERR_freelist_empty) ||
 921                             bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
 922                                 return ret;
 923                         if (*nr_effective >= nr_replicas)
 924                                 return 0;
 925                 }
 926         }
 927
 928         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 929                                     nr_replicas, nr_effective,
 930                                     have_cache, flags, false);
 931         if (*nr_effective >= nr_replicas)
 932                 return 0;
 933
 934 retry_blocking:
 935         /*
 936          * Try nonblocking first, so that if one device is full we'll try from
 937          * other devices:
 938          */
 939         ret = bch2_bucket_alloc_set_trans(trans, ptrs, &wp->stripe, &devs,
 940                                 nr_replicas, nr_effective, have_cache,
 941                                 reserve, flags, cl);
 942         if (ret &&
 943             !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
 944             !bch2_err_matches(ret, BCH_ERR_insufficient_devices) &&
 945             !cl && _cl) {
 946                 cl = _cl;
 947                 goto retry_blocking;
 948         }
 949
 950         return ret;
 951 }
 952
 953 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 954                                 struct open_buckets *obs)
 955 {
 956         struct open_buckets ptrs = { .nr = 0 };
 957         struct open_bucket *ob, *ob2;
 958         unsigned i, j;
 959
 960         open_bucket_for_each(c, obs, ob, i) {
 961                 bool drop = !ca || ob->dev == ca->dev_idx;
 962
 963                 if (!drop && ob->ec) {
 964                         mutex_lock(&ob->ec->lock);
 965                         for (j = 0; j < ob->ec->new_stripe.key.v.nr_blocks; j++) {
 966                                 if (!ob->ec->blocks[j])
 967                                         continue;
 968
 969                                 ob2 = c->open_buckets + ob->ec->blocks[j];
 970                                 drop |= ob2->dev == ca->dev_idx;
 971                         }
 972                         mutex_unlock(&ob->ec->lock);
 973                 }
 974
 975                 if (drop)
 976                         bch2_open_bucket_put(c, ob);
 977                 else
 978                         ob_push(c, &ptrs, ob);
 979         }
 980
 981         *obs = ptrs;
 982 }
 983
 984 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 985                           struct write_point *wp)
 986 {
 987         mutex_lock(&wp->lock);
 988         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
 989         mutex_unlock(&wp->lock);
 990 }
 991
 992 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 993                                                  unsigned long write_point)
 994 {
 995         unsigned hash =
 996                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
 997
 998         return &c->write_points_hash[hash];
 999 }
1000
1001 static struct write_point *__writepoint_find(struct hlist_head *head,
1002                                              unsigned long write_point)
1003 {
1004         struct write_point *wp;
1005
1006         rcu_read_lock();
1007         hlist_for_each_entry_rcu(wp, head, node)
1008                 if (wp->write_point == write_point)
1009                         goto out;
1010         wp = NULL;
1011 out:
1012         rcu_read_unlock();
1013         return wp;
1014 }
1015
1016 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
1017 {
1018         u64 stranded    = c->write_points_nr * c->bucket_size_max;
1019         u64 free        = bch2_fs_usage_read_short(c).free;
1020
1021         return stranded * factor > free;
1022 }
1023
1024 static bool try_increase_writepoints(struct bch_fs *c)
1025 {
1026         struct write_point *wp;
1027
1028         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
1029             too_many_writepoints(c, 32))
1030                 return false;
1031
1032         wp = c->write_points + c->write_points_nr++;
1033         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
1034         return true;
1035 }
1036
1037 static bool try_decrease_writepoints(struct bch_fs *c,
1038                                      unsigned old_nr)
1039 {
1040         struct write_point *wp;
1041
1042         mutex_lock(&c->write_points_hash_lock);
1043         if (c->write_points_nr < old_nr) {
1044                 mutex_unlock(&c->write_points_hash_lock);
1045                 return true;
1046         }
1047
1048         if (c->write_points_nr == 1 ||
1049             !too_many_writepoints(c, 8)) {
1050                 mutex_unlock(&c->write_points_hash_lock);
1051                 return false;
1052         }
1053
1054         wp = c->write_points + --c->write_points_nr;
1055
1056         hlist_del_rcu(&wp->node);
1057         mutex_unlock(&c->write_points_hash_lock);
1058
1059         bch2_writepoint_stop(c, NULL, wp);
1060         return true;
1061 }
1062
1063 static void bch2_trans_mutex_lock(struct btree_trans *trans,
1064                                   struct mutex *lock)
1065 {
1066         if (!mutex_trylock(lock)) {
1067                 bch2_trans_unlock(trans);
1068                 mutex_lock(lock);
1069         }
1070 }
1071
1072 static struct write_point *writepoint_find(struct btree_trans *trans,
1073                                            unsigned long write_point)
1074 {
1075         struct bch_fs *c = trans->c;
1076         struct write_point *wp, *oldest;
1077         struct hlist_head *head;
1078
1079         if (!(write_point & 1UL)) {
1080                 wp = (struct write_point *) write_point;
1081                 bch2_trans_mutex_lock(trans, &wp->lock);
1082                 return wp;
1083         }
1084
1085         head = writepoint_hash(c, write_point);
1086 restart_find:
1087         wp = __writepoint_find(head, write_point);
1088         if (wp) {
1089 lock_wp:
1090                 bch2_trans_mutex_lock(trans, &wp->lock);
1091                 if (wp->write_point == write_point)
1092                         goto out;
1093                 mutex_unlock(&wp->lock);
1094                 goto restart_find;
1095         }
1096 restart_find_oldest:
1097         oldest = NULL;
1098         for (wp = c->write_points;
1099              wp < c->write_points + c->write_points_nr; wp++)
1100                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1101                         oldest = wp;
1102
1103         bch2_trans_mutex_lock(trans, &oldest->lock);
1104         bch2_trans_mutex_lock(trans, &c->write_points_hash_lock);
1105         if (oldest >= c->write_points + c->write_points_nr ||
1106             try_increase_writepoints(c)) {
1107                 mutex_unlock(&c->write_points_hash_lock);
1108                 mutex_unlock(&oldest->lock);
1109                 goto restart_find_oldest;
1110         }
1111
1112         wp = __writepoint_find(head, write_point);
1113         if (wp && wp != oldest) {
1114                 mutex_unlock(&c->write_points_hash_lock);
1115                 mutex_unlock(&oldest->lock);
1116                 goto lock_wp;
1117         }
1118
1119         wp = oldest;
1120         hlist_del_rcu(&wp->node);
1121         wp->write_point = write_point;
1122         hlist_add_head_rcu(&wp->node, head);
1123         mutex_unlock(&c->write_points_hash_lock);
1124 out:
1125         wp->last_used = local_clock();
1126         return wp;
1127 }
1128
1129 /*
1130  * Get us an open_bucket we can allocate from, return with it locked:
1131  */
1132 struct write_point *bch2_alloc_sectors_start_trans(struct btree_trans *trans,
1133                                 unsigned target,
1134                                 unsigned erasure_code,
1135                                 struct write_point_specifier write_point,
1136                                 struct bch_devs_list *devs_have,
1137                                 unsigned nr_replicas,
1138                                 unsigned nr_replicas_required,
1139                                 enum alloc_reserve reserve,
1140                                 unsigned flags,
1141                                 struct closure *cl)
1142 {
1143         struct bch_fs *c = trans->c;
1144         struct write_point *wp;
1145         struct open_bucket *ob;
1146         struct open_buckets ptrs;
1147         unsigned nr_effective, write_points_nr;
1148         unsigned ob_flags = 0;
1149         bool have_cache;
1150         int ret;
1151         int i;
1152
1153         if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
1154                 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
1155
1156         BUG_ON(!nr_replicas || !nr_replicas_required);
1157 retry:
1158         ptrs.nr         = 0;
1159         nr_effective    = 0;
1160         write_points_nr = c->write_points_nr;
1161         have_cache      = false;
1162
1163         wp = writepoint_find(trans, write_point.v);
1164
1165         if (wp->data_type == BCH_DATA_user)
1166                 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
1167
1168         /* metadata may not allocate on cache devices: */
1169         if (wp->data_type != BCH_DATA_user)
1170                 have_cache = true;
1171
1172         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1173                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1174                                               target, erasure_code,
1175                                               nr_replicas, &nr_effective,
1176                                               &have_cache, reserve,
1177                                               ob_flags, cl);
1178         } else {
1179                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1180                                               target, erasure_code,
1181                                               nr_replicas, &nr_effective,
1182                                               &have_cache, reserve,
1183                                               ob_flags, NULL);
1184                 if (!ret ||
1185                     bch2_err_matches(ret, BCH_ERR_transaction_restart))
1186                         goto alloc_done;
1187
1188                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1189                                               0, erasure_code,
1190                                               nr_replicas, &nr_effective,
1191                                               &have_cache, reserve,
1192                                               ob_flags, cl);
1193         }
1194 alloc_done:
1195         BUG_ON(!ret && nr_effective < nr_replicas);
1196
1197         if (erasure_code && !ec_open_bucket(c, &ptrs))
1198                 pr_debug("failed to get ec bucket: ret %u", ret);
1199
1200         if (ret == -BCH_ERR_insufficient_devices &&
1201             nr_effective >= nr_replicas_required)
1202                 ret = 0;
1203
1204         if (ret)
1205                 goto err;
1206
1207         /* Free buckets we didn't use: */
1208         open_bucket_for_each(c, &wp->ptrs, ob, i)
1209                 open_bucket_free_unused(c, wp, ob);
1210
1211         wp->ptrs = ptrs;
1212
1213         wp->sectors_free = UINT_MAX;
1214
1215         open_bucket_for_each(c, &wp->ptrs, ob, i)
1216                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1217
1218         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1219
1220         return wp;
1221 err:
1222         open_bucket_for_each(c, &wp->ptrs, ob, i)
1223                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1224                         ob_push(c, &ptrs, ob);
1225                 else
1226                         open_bucket_free_unused(c, wp, ob);
1227         wp->ptrs = ptrs;
1228
1229         mutex_unlock(&wp->lock);
1230
1231         if (bch2_err_matches(ret, BCH_ERR_freelist_empty) &&
1232             try_decrease_writepoints(c, write_points_nr))
1233                 goto retry;
1234
1235         if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty) ||
1236             bch2_err_matches(ret, BCH_ERR_freelist_empty))
1237                 return cl
1238                         ? ERR_PTR(-EAGAIN)
1239                         : ERR_PTR(-BCH_ERR_ENOSPC_bucket_alloc);
1240
1241         if (bch2_err_matches(ret, BCH_ERR_insufficient_devices))
1242                 return ERR_PTR(-EROFS);
1243
1244         return ERR_PTR(ret);
1245 }
1246
1247 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
1248                                 unsigned target,
1249                                 unsigned erasure_code,
1250                                 struct write_point_specifier write_point,
1251                                 struct bch_devs_list *devs_have,
1252                                 unsigned nr_replicas,
1253                                 unsigned nr_replicas_required,
1254                                 enum alloc_reserve reserve,
1255                                 unsigned flags,
1256                                 struct closure *cl)
1257 {
1258         struct write_point *wp;
1259
1260         bch2_trans_do(c, NULL, NULL, 0,
1261                       PTR_ERR_OR_ZERO(wp = bch2_alloc_sectors_start_trans(&trans, target,
1262                                                         erasure_code,
1263                                                         write_point,
1264                                                         devs_have,
1265                                                         nr_replicas,
1266                                                         nr_replicas_required,
1267                                                         reserve,
1268                                                         flags, cl)));
1269         return wp;
1270
1271 }
1272
1273 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1274 {
1275         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1276
1277         return (struct bch_extent_ptr) {
1278                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
1279                 .gen    = ob->gen,
1280                 .dev    = ob->dev,
1281                 .offset = bucket_to_sector(ca, ob->bucket) +
1282                         ca->mi.bucket_size -
1283                         ob->sectors_free,
1284         };
1285 }
1286
1287 /*
1288  * Append pointers to the space we just allocated to @k, and mark @sectors space
1289  * as allocated out of @ob
1290  */
1291 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1292                                     struct bkey_i *k, unsigned sectors,
1293                                     bool cached)
1294
1295 {
1296         struct open_bucket *ob;
1297         unsigned i;
1298
1299         BUG_ON(sectors > wp->sectors_free);
1300         wp->sectors_free -= sectors;
1301
1302         open_bucket_for_each(c, &wp->ptrs, ob, i) {
1303                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1304                 struct bch_extent_ptr ptr = bch2_ob_ptr(c, ob);
1305
1306                 ptr.cached = cached ||
1307                         (!ca->mi.durability &&
1308                          wp->data_type == BCH_DATA_user);
1309
1310                 bch2_bkey_append_ptr(k, ptr);
1311
1312                 BUG_ON(sectors > ob->sectors_free);
1313                 ob->sectors_free -= sectors;
1314         }
1315 }
1316
1317 /*
1318  * Append pointers to the space we just allocated to @k, and mark @sectors space
1319  * as allocated out of @ob
1320  */
1321 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1322 {
1323         struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
1324         struct open_bucket *ob;
1325         unsigned i;
1326
1327         open_bucket_for_each(c, &wp->ptrs, ob, i)
1328                 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
1329         wp->ptrs = keep;
1330
1331         mutex_unlock(&wp->lock);
1332
1333         bch2_open_buckets_put(c, &ptrs);
1334 }
1335
1336 static inline void writepoint_init(struct write_point *wp,
1337                                    enum bch_data_type type)
1338 {
1339         mutex_init(&wp->lock);
1340         wp->data_type = type;
1341 }
1342
1343 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1344 {
1345         struct open_bucket *ob;
1346         struct write_point *wp;
1347
1348         mutex_init(&c->write_points_hash_lock);
1349         c->write_points_nr = ARRAY_SIZE(c->write_points);
1350
1351         /* open bucket 0 is a sentinal NULL: */
1352         spin_lock_init(&c->open_buckets[0].lock);
1353
1354         for (ob = c->open_buckets + 1;
1355              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1356                 spin_lock_init(&ob->lock);
1357                 c->open_buckets_nr_free++;
1358
1359                 ob->freelist = c->open_buckets_freelist;
1360                 c->open_buckets_freelist = ob - c->open_buckets;
1361         }
1362
1363         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
1364         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
1365         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
1366
1367         for (wp = c->write_points;
1368              wp < c->write_points + c->write_points_nr; wp++) {
1369                 writepoint_init(wp, BCH_DATA_user);
1370
1371                 wp->last_used   = local_clock();
1372                 wp->write_point = (unsigned long) wp;
1373                 hlist_add_head_rcu(&wp->node,
1374                                    writepoint_hash(c, wp->write_point));
1375         }
1376 }
1377
1378 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1379 {
1380         struct open_bucket *ob;
1381
1382         for (ob = c->open_buckets;
1383              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1384              ob++) {
1385                 spin_lock(&ob->lock);
1386                 if (ob->valid && !ob->on_partial_list) {
1387                         prt_printf(out, "%zu ref %u type %s %u:%llu:%u\n",
1388                                ob - c->open_buckets,
1389                                atomic_read(&ob->pin),
1390                                bch2_data_types[ob->data_type],
1391                                ob->dev, ob->bucket, ob->gen);
1392                 }
1393                 spin_unlock(&ob->lock);
1394         }
1395 }