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 "backpointers.h"
  18 #include "btree_iter.h"
  19 #include "btree_update.h"
  20 #include "btree_gc.h"
  21 #include "buckets.h"
  22 #include "buckets_waiting_for_journal.h"
  23 #include "clock.h"
  24 #include "debug.h"
  25 #include "disk_groups.h"
  26 #include "ec.h"
  27 #include "error.h"
  28 #include "io.h"
  29 #include "journal.h"
  30 #include "movinggc.h"
  31 #include "nocow_locking.h"
  32
  33 #include <linux/math64.h>
  34 #include <linux/rculist.h>
  35 #include <linux/rcupdate.h>
  36 #include <trace/events/bcachefs.h>
  37
  38 const char * const bch2_alloc_reserves[] = {
  39 #define x(t) #t,
  40         BCH_ALLOC_RESERVES()
  41 #undef x
  42         NULL
  43 };
  44
  45 /*
  46  * Open buckets represent a bucket that's currently being allocated from.  They
  47  * serve two purposes:
  48  *
  49  *  - They track buckets that have been partially allocated, allowing for
  50  *    sub-bucket sized allocations - they're used by the sector allocator below
  51  *
  52  *  - They provide a reference to the buckets they own that mark and sweep GC
  53  *    can find, until the new allocation has a pointer to it inserted into the
  54  *    btree
  55  *
  56  * When allocating some space with the sector allocator, the allocation comes
  57  * with a reference to an open bucket - the caller is required to put that
  58  * reference _after_ doing the index update that makes its allocation reachable.
  59  */
  60
  61 void bch2_reset_alloc_cursors(struct bch_fs *c)
  62 {
  63         struct bch_dev *ca;
  64         unsigned i;
  65
  66         rcu_read_lock();
  67         for_each_member_device_rcu(ca, c, i, NULL)
  68                 ca->alloc_cursor = 0;
  69         rcu_read_unlock();
  70 }
  71
  72 static void bch2_open_bucket_hash_add(struct bch_fs *c, struct open_bucket *ob)
  73 {
  74         open_bucket_idx_t idx = ob - c->open_buckets;
  75         open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
  76
  77         ob->hash = *slot;
  78         *slot = idx;
  79 }
  80
  81 static void bch2_open_bucket_hash_remove(struct bch_fs *c, struct open_bucket *ob)
  82 {
  83         open_bucket_idx_t idx = ob - c->open_buckets;
  84         open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
  85
  86         while (*slot != idx) {
  87                 BUG_ON(!*slot);
  88                 slot = &c->open_buckets[*slot].hash;
  89         }
  90
  91         *slot = ob->hash;
  92         ob->hash = 0;
  93 }
  94
  95 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
  96 {
  97         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
  98
  99         if (ob->ec) {
 100                 bch2_ec_bucket_written(c, ob);
 101                 return;
 102         }
 103
 104         percpu_down_read(&c->mark_lock);
 105         spin_lock(&ob->lock);
 106
 107         ob->valid = false;
 108         ob->data_type = 0;
 109
 110         spin_unlock(&ob->lock);
 111         percpu_up_read(&c->mark_lock);
 112
 113         spin_lock(&c->freelist_lock);
 114         bch2_open_bucket_hash_remove(c, ob);
 115
 116         ob->freelist = c->open_buckets_freelist;
 117         c->open_buckets_freelist = ob - c->open_buckets;
 118
 119         c->open_buckets_nr_free++;
 120         ca->nr_open_buckets--;
 121         spin_unlock(&c->freelist_lock);
 122
 123         closure_wake_up(&c->open_buckets_wait);
 124 }
 125
 126 void bch2_open_bucket_write_error(struct bch_fs *c,
 127                                   struct open_buckets *obs,
 128                                   unsigned dev)
 129 {
 130         struct open_bucket *ob;
 131         unsigned i;
 132
 133         open_bucket_for_each(c, obs, ob, i)
 134                 if (ob->dev == dev && ob->ec)
 135                         bch2_ec_bucket_cancel(c, ob);
 136 }
 137
 138 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
 139 {
 140         struct open_bucket *ob;
 141
 142         BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
 143
 144         ob = c->open_buckets + c->open_buckets_freelist;
 145         c->open_buckets_freelist = ob->freelist;
 146         atomic_set(&ob->pin, 1);
 147         ob->data_type = 0;
 148
 149         c->open_buckets_nr_free--;
 150         return ob;
 151 }
 152
 153 static void open_bucket_free_unused(struct bch_fs *c,
 154                                     struct write_point *wp,
 155                                     struct open_bucket *ob)
 156 {
 157         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 158         bool may_realloc = wp->data_type == BCH_DATA_user;
 159
 160         BUG_ON(ca->open_buckets_partial_nr >
 161                ARRAY_SIZE(ca->open_buckets_partial));
 162
 163         if (ca->open_buckets_partial_nr <
 164             ARRAY_SIZE(ca->open_buckets_partial) &&
 165             may_realloc) {
 166                 spin_lock(&c->freelist_lock);
 167                 ob->on_partial_list = true;
 168                 ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
 169                         ob - c->open_buckets;
 170                 spin_unlock(&c->freelist_lock);
 171
 172                 closure_wake_up(&c->open_buckets_wait);
 173                 closure_wake_up(&c->freelist_wait);
 174         } else {
 175                 bch2_open_bucket_put(c, ob);
 176         }
 177 }
 178
 179 /* _only_ for allocating the journal on a new device: */
 180 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
 181 {
 182         while (ca->new_fs_bucket_idx < ca->mi.nbuckets) {
 183                 u64 b = ca->new_fs_bucket_idx++;
 184
 185                 if (!is_superblock_bucket(ca, b) &&
 186                     (!ca->buckets_nouse || !test_bit(b, ca->buckets_nouse)))
 187                         return b;
 188         }
 189
 190         return -1;
 191 }
 192
 193 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
 194 {
 195         switch (reserve) {
 196         case RESERVE_btree:
 197         case RESERVE_btree_movinggc:
 198                 return 0;
 199         case RESERVE_movinggc:
 200                 return OPEN_BUCKETS_COUNT / 4;
 201         default:
 202                 return OPEN_BUCKETS_COUNT / 2;
 203         }
 204 }
 205
 206 static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 207                                               u64 bucket,
 208                                               enum alloc_reserve reserve,
 209                                               const struct bch_alloc_v4 *a,
 210                                               struct bucket_alloc_state *s,
 211                                               struct closure *cl)
 212 {
 213         struct open_bucket *ob;
 214
 215         if (unlikely(ca->buckets_nouse && test_bit(bucket, ca->buckets_nouse))) {
 216                 s->skipped_nouse++;
 217                 return NULL;
 218         }
 219
 220         if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
 221                 s->skipped_open++;
 222                 return NULL;
 223         }
 224
 225         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 226                         c->journal.flushed_seq_ondisk, ca->dev_idx, bucket)) {
 227                 s->skipped_need_journal_commit++;
 228                 return NULL;
 229         }
 230
 231         if (bch2_bucket_nocow_is_locked(&c->nocow_locks, POS(ca->dev_idx, bucket))) {
 232                 s->skipped_nocow++;
 233                 return NULL;
 234         }
 235
 236         spin_lock(&c->freelist_lock);
 237
 238         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
 239                 if (cl)
 240                         closure_wait(&c->open_buckets_wait, cl);
 241
 242                 if (!c->blocked_allocate_open_bucket)
 243                         c->blocked_allocate_open_bucket = local_clock();
 244
 245                 spin_unlock(&c->freelist_lock);
 246                 return ERR_PTR(-BCH_ERR_open_buckets_empty);
 247         }
 248
 249         /* Recheck under lock: */
 250         if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
 251                 spin_unlock(&c->freelist_lock);
 252                 s->skipped_open++;
 253                 return NULL;
 254         }
 255
 256         ob = bch2_open_bucket_alloc(c);
 257
 258         spin_lock(&ob->lock);
 259
 260         ob->valid       = true;
 261         ob->sectors_free = ca->mi.bucket_size;
 262         ob->alloc_reserve = reserve;
 263         ob->dev         = ca->dev_idx;
 264         ob->gen         = a->gen;
 265         ob->bucket      = bucket;
 266         spin_unlock(&ob->lock);
 267
 268         ca->nr_open_buckets++;
 269         bch2_open_bucket_hash_add(c, ob);
 270
 271         if (c->blocked_allocate_open_bucket) {
 272                 bch2_time_stats_update(
 273                         &c->times[BCH_TIME_blocked_allocate_open_bucket],
 274                         c->blocked_allocate_open_bucket);
 275                 c->blocked_allocate_open_bucket = 0;
 276         }
 277
 278         if (c->blocked_allocate) {
 279                 bch2_time_stats_update(
 280                         &c->times[BCH_TIME_blocked_allocate],
 281                         c->blocked_allocate);
 282                 c->blocked_allocate = 0;
 283         }
 284
 285         spin_unlock(&c->freelist_lock);
 286         return ob;
 287 }
 288
 289 static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bch_dev *ca,
 290                                             enum alloc_reserve reserve, u64 free_entry,
 291                                             struct bucket_alloc_state *s,
 292                                             struct bkey_s_c freespace_k,
 293                                             struct closure *cl)
 294 {
 295         struct bch_fs *c = trans->c;
 296         struct btree_iter iter = { NULL };
 297         struct bkey_s_c k;
 298         struct open_bucket *ob;
 299         struct bch_alloc_v4 a_convert;
 300         const struct bch_alloc_v4 *a;
 301         u64 b = free_entry & ~(~0ULL << 56);
 302         unsigned genbits = free_entry >> 56;
 303         struct printbuf buf = PRINTBUF;
 304         int ret;
 305
 306         if (b < ca->mi.first_bucket || b >= ca->mi.nbuckets) {
 307                 prt_printf(&buf, "freespace btree has bucket outside allowed range %u-%llu\n"
 308                        "  freespace key ",
 309                         ca->mi.first_bucket, ca->mi.nbuckets);
 310                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 311                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 312                 ob = ERR_PTR(-EIO);
 313                 goto err;
 314         }
 315
 316         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS(ca->dev_idx, b), BTREE_ITER_CACHED);
 317         k = bch2_btree_iter_peek_slot(&iter);
 318         ret = bkey_err(k);
 319         if (ret) {
 320                 ob = ERR_PTR(ret);
 321                 goto err;
 322         }
 323
 324         a = bch2_alloc_to_v4(k, &a_convert);
 325
 326         if (a->data_type != BCH_DATA_free) {
 327                 if (!test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
 328                         ob = NULL;
 329                         goto err;
 330                 }
 331
 332                 prt_printf(&buf, "non free bucket in freespace btree\n"
 333                        "  freespace key ");
 334                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 335                 prt_printf(&buf, "\n  ");
 336                 bch2_bkey_val_to_text(&buf, c, k);
 337                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 338                 ob = ERR_PTR(-EIO);
 339                 goto err;
 340         }
 341
 342         if (genbits != (alloc_freespace_genbits(*a) >> 56) &&
 343             test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
 344                 prt_printf(&buf, "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
 345                        "  freespace key ",
 346                        genbits, alloc_freespace_genbits(*a) >> 56);
 347                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 348                 prt_printf(&buf, "\n  ");
 349                 bch2_bkey_val_to_text(&buf, c, k);
 350                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 351                 ob = ERR_PTR(-EIO);
 352                 goto err;
 353
 354         }
 355
 356         if (!test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags)) {
 357                 struct bch_backpointer bp;
 358                 u64 bp_offset = 0;
 359
 360                 ret = bch2_get_next_backpointer(trans, POS(ca->dev_idx, b), -1,
 361                                                 &bp_offset, &bp,
 362                                                 BTREE_ITER_NOPRESERVE);
 363                 if (ret) {
 364                         ob = ERR_PTR(ret);
 365                         goto err;
 366                 }
 367
 368                 if (bp_offset != U64_MAX) {
 369                         /*
 370                          * Bucket may have data in it - we don't call
 371                          * bc2h_trans_inconnsistent() because fsck hasn't
 372                          * finished yet
 373                          */
 374                         ob = NULL;
 375                         goto err;
 376                 }
 377         }
 378
 379         ob = __try_alloc_bucket(c, ca, b, reserve, a, s, cl);
 380         if (!ob)
 381                 iter.path->preserve = false;
 382 err:
 383         set_btree_iter_dontneed(&iter);
 384         bch2_trans_iter_exit(trans, &iter);
 385         printbuf_exit(&buf);
 386         return ob;
 387 }
 388
 389 static struct open_bucket *try_alloc_partial_bucket(struct bch_fs *c, struct bch_dev *ca,
 390                                                     enum alloc_reserve reserve)
 391 {
 392         struct open_bucket *ob;
 393         int i;
 394
 395         spin_lock(&c->freelist_lock);
 396
 397         for (i = ca->open_buckets_partial_nr - 1; i >= 0; --i) {
 398                 ob = c->open_buckets + ca->open_buckets_partial[i];
 399
 400                 if (reserve <= ob->alloc_reserve) {
 401                         array_remove_item(ca->open_buckets_partial,
 402                                           ca->open_buckets_partial_nr,
 403                                           i);
 404                         ob->on_partial_list = false;
 405                         ob->alloc_reserve = reserve;
 406                         spin_unlock(&c->freelist_lock);
 407                         return ob;
 408                 }
 409         }
 410
 411         spin_unlock(&c->freelist_lock);
 412         return NULL;
 413 }
 414
 415 /*
 416  * This path is for before the freespace btree is initialized:
 417  *
 418  * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
 419  * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
 420  */
 421 static noinline struct open_bucket *
 422 bch2_bucket_alloc_early(struct btree_trans *trans,
 423                         struct bch_dev *ca,
 424                         enum alloc_reserve reserve,
 425                         struct bucket_alloc_state *s,
 426                         struct closure *cl)
 427 {
 428         struct btree_iter iter;
 429         struct bkey_s_c k;
 430         struct open_bucket *ob = NULL;
 431         u64 alloc_start = max_t(u64, ca->mi.first_bucket, ca->new_fs_bucket_idx);
 432         u64 alloc_cursor = max(alloc_start, READ_ONCE(ca->alloc_cursor));
 433         int ret;
 434 again:
 435         for_each_btree_key_norestart(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, alloc_cursor),
 436                            BTREE_ITER_SLOTS, k, ret) {
 437                 struct bch_alloc_v4 a_convert;
 438                 const struct bch_alloc_v4 *a;
 439
 440                 if (bkey_ge(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)))
 441                         break;
 442
 443                 if (ca->new_fs_bucket_idx &&
 444                     is_superblock_bucket(ca, k.k->p.offset))
 445                         continue;
 446
 447                 a = bch2_alloc_to_v4(k, &a_convert);
 448
 449                 if (a->data_type != BCH_DATA_free)
 450                         continue;
 451
 452                 s->buckets_seen++;
 453
 454                 ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, reserve, a, s, cl);
 455                 if (ob)
 456                         break;
 457         }
 458         bch2_trans_iter_exit(trans, &iter);
 459
 460         ca->alloc_cursor = alloc_cursor;
 461
 462         if (!ob && ret)
 463                 ob = ERR_PTR(ret);
 464
 465         if (!ob && alloc_cursor > alloc_start) {
 466                 alloc_cursor = alloc_start;
 467                 goto again;
 468         }
 469
 470         return ob;
 471 }
 472
 473 static struct open_bucket *bch2_bucket_alloc_freelist(struct btree_trans *trans,
 474                                                    struct bch_dev *ca,
 475                                                    enum alloc_reserve reserve,
 476                                                    struct bucket_alloc_state *s,
 477                                                    struct closure *cl)
 478 {
 479         struct btree_iter iter;
 480         struct bkey_s_c k;
 481         struct open_bucket *ob = NULL;
 482         u64 alloc_start = max_t(u64, ca->mi.first_bucket, READ_ONCE(ca->alloc_cursor));
 483         u64 alloc_cursor = alloc_start;
 484         int ret;
 485
 486         BUG_ON(ca->new_fs_bucket_idx);
 487 again:
 488         for_each_btree_key_norestart(trans, iter, BTREE_ID_freespace,
 489                                      POS(ca->dev_idx, alloc_cursor), 0, k, ret) {
 490                 if (k.k->p.inode != ca->dev_idx)
 491                         break;
 492
 493                 for (alloc_cursor = max(alloc_cursor, bkey_start_offset(k.k));
 494                      alloc_cursor < k.k->p.offset;
 495                      alloc_cursor++) {
 496                         ret = btree_trans_too_many_iters(trans);
 497                         if (ret) {
 498                                 ob = ERR_PTR(ret);
 499                                 break;
 500                         }
 501
 502                         s->buckets_seen++;
 503
 504                         ob = try_alloc_bucket(trans, ca, reserve,
 505                                               alloc_cursor, s, k, cl);
 506                         if (ob) {
 507                                 iter.path->preserve = false;
 508                                 break;
 509                         }
 510                 }
 511
 512                 if (ob || ret)
 513                         break;
 514         }
 515         bch2_trans_iter_exit(trans, &iter);
 516
 517         ca->alloc_cursor = alloc_cursor;
 518
 519         if (!ob && ret)
 520                 ob = ERR_PTR(ret);
 521
 522         if (!ob && alloc_start > ca->mi.first_bucket) {
 523                 alloc_cursor = alloc_start = ca->mi.first_bucket;
 524                 goto again;
 525         }
 526
 527         return ob;
 528 }
 529
 530 /**
 531  * bch_bucket_alloc - allocate a single bucket from a specific device
 532  *
 533  * Returns index of bucket on success, 0 on failure
 534  */
 535 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
 536                                       struct bch_dev *ca,
 537                                       enum alloc_reserve reserve,
 538                                       bool may_alloc_partial,
 539                                       struct closure *cl,
 540                                       struct bch_dev_usage *usage)
 541 {
 542         struct bch_fs *c = trans->c;
 543         struct open_bucket *ob = NULL;
 544         bool freespace = READ_ONCE(ca->mi.freespace_initialized);
 545         u64 avail;
 546         struct bucket_alloc_state s = { 0 };
 547         bool waiting = false;
 548 again:
 549         bch2_dev_usage_read_fast(ca, usage);
 550         avail = dev_buckets_free(ca, *usage, reserve);
 551
 552         if (usage->d[BCH_DATA_need_discard].buckets > avail)
 553                 bch2_do_discards(c);
 554
 555         if (usage->d[BCH_DATA_need_gc_gens].buckets > avail)
 556                 bch2_do_gc_gens(c);
 557
 558         if (should_invalidate_buckets(ca, *usage))
 559                 bch2_do_invalidates(c);
 560
 561         if (!avail) {
 562                 if (cl && !waiting) {
 563                         closure_wait(&c->freelist_wait, cl);
 564                         waiting = true;
 565                         goto again;
 566                 }
 567
 568                 if (!c->blocked_allocate)
 569                         c->blocked_allocate = local_clock();
 570
 571                 ob = ERR_PTR(-BCH_ERR_freelist_empty);
 572                 goto err;
 573         }
 574
 575         if (waiting)
 576                 closure_wake_up(&c->freelist_wait);
 577
 578         if (may_alloc_partial) {
 579                 ob = try_alloc_partial_bucket(c, ca, reserve);
 580                 if (ob)
 581                         return ob;
 582         }
 583 alloc:
 584         ob = likely(freespace)
 585                 ? bch2_bucket_alloc_freelist(trans, ca, reserve, &s, cl)
 586                 : bch2_bucket_alloc_early(trans, ca, reserve, &s, cl);
 587
 588         if (s.skipped_need_journal_commit * 2 > avail)
 589                 bch2_journal_flush_async(&c->journal, NULL);
 590
 591         if (!ob && freespace && !test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
 592                 freespace = false;
 593                 goto alloc;
 594         }
 595 err:
 596         if (!ob)
 597                 ob = ERR_PTR(-BCH_ERR_no_buckets_found);
 598
 599         if (!IS_ERR(ob))
 600                 trace_and_count(c, bucket_alloc, ca,
 601                                 bch2_alloc_reserves[reserve],
 602                                 may_alloc_partial,
 603                                 ob->bucket,
 604                                 usage->d[BCH_DATA_free].buckets,
 605                                 avail,
 606                                 bch2_copygc_wait_amount(c),
 607                                 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
 608                                 &s,
 609                                 cl == NULL,
 610                                 "");
 611         else if (!bch2_err_matches(PTR_ERR(ob), BCH_ERR_transaction_restart))
 612                 trace_and_count(c, bucket_alloc_fail, ca,
 613                                 bch2_alloc_reserves[reserve],
 614                                 may_alloc_partial,
 615                                 0,
 616                                 usage->d[BCH_DATA_free].buckets,
 617                                 avail,
 618                                 bch2_copygc_wait_amount(c),
 619                                 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
 620                                 &s,
 621                                 cl == NULL,
 622                                 bch2_err_str(PTR_ERR(ob)));
 623
 624         return ob;
 625 }
 626
 627 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 628                                       enum alloc_reserve reserve,
 629                                       bool may_alloc_partial,
 630                                       struct closure *cl)
 631 {
 632         struct bch_dev_usage usage;
 633         struct open_bucket *ob;
 634
 635         bch2_trans_do(c, NULL, NULL, 0,
 636                       PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, reserve,
 637                                                         may_alloc_partial, cl, &usage)));
 638         return ob;
 639 }
 640
 641 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 642                             unsigned l, unsigned r)
 643 {
 644         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 645                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 646 }
 647
 648 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 649
 650 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 651                                           struct dev_stripe_state *stripe,
 652                                           struct bch_devs_mask *devs)
 653 {
 654         struct dev_alloc_list ret = { .nr = 0 };
 655         unsigned i;
 656
 657         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 658                 ret.devs[ret.nr++] = i;
 659
 660         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 661         return ret;
 662 }
 663
 664 static inline void bch2_dev_stripe_increment_inlined(struct bch_dev *ca,
 665                                struct dev_stripe_state *stripe,
 666                                struct bch_dev_usage *usage)
 667 {
 668         u64 *v = stripe->next_alloc + ca->dev_idx;
 669         u64 free_space = dev_buckets_available(ca, RESERVE_none);
 670         u64 free_space_inv = free_space
 671                 ? div64_u64(1ULL << 48, free_space)
 672                 : 1ULL << 48;
 673         u64 scale = *v / 4;
 674
 675         if (*v + free_space_inv >= *v)
 676                 *v += free_space_inv;
 677         else
 678                 *v = U64_MAX;
 679
 680         for (v = stripe->next_alloc;
 681              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 682                 *v = *v < scale ? 0 : *v - scale;
 683 }
 684
 685 void bch2_dev_stripe_increment(struct bch_dev *ca,
 686                                struct dev_stripe_state *stripe)
 687 {
 688         struct bch_dev_usage usage;
 689
 690         bch2_dev_usage_read_fast(ca, &usage);
 691         bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
 692 }
 693
 694 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 695 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 696
 697 static void add_new_bucket(struct bch_fs *c,
 698                            struct open_buckets *ptrs,
 699                            struct bch_devs_mask *devs_may_alloc,
 700                            unsigned *nr_effective,
 701                            bool *have_cache,
 702                            unsigned flags,
 703                            struct open_bucket *ob)
 704 {
 705         unsigned durability =
 706                 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
 707
 708         __clear_bit(ob->dev, devs_may_alloc->d);
 709         *nr_effective   += (flags & BUCKET_ALLOC_USE_DURABILITY)
 710                 ? durability : 1;
 711         *have_cache     |= !durability;
 712
 713         ob_push(c, ptrs, ob);
 714 }
 715
 716 int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
 717                       struct open_buckets *ptrs,
 718                       struct dev_stripe_state *stripe,
 719                       struct bch_devs_mask *devs_may_alloc,
 720                       unsigned nr_replicas,
 721                       unsigned *nr_effective,
 722                       bool *have_cache,
 723                       enum alloc_reserve reserve,
 724                       unsigned flags,
 725                       struct closure *cl)
 726 {
 727         struct bch_fs *c = trans->c;
 728         struct dev_alloc_list devs_sorted =
 729                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 730         unsigned dev;
 731         struct bch_dev *ca;
 732         int ret = -BCH_ERR_insufficient_devices;
 733         unsigned i;
 734
 735         BUG_ON(*nr_effective >= nr_replicas);
 736
 737         for (i = 0; i < devs_sorted.nr; i++) {
 738                 struct bch_dev_usage usage;
 739                 struct open_bucket *ob;
 740
 741                 dev = devs_sorted.devs[i];
 742
 743                 rcu_read_lock();
 744                 ca = rcu_dereference(c->devs[dev]);
 745                 if (ca)
 746                         percpu_ref_get(&ca->ref);
 747                 rcu_read_unlock();
 748
 749                 if (!ca)
 750                         continue;
 751
 752                 if (!ca->mi.durability && *have_cache) {
 753                         percpu_ref_put(&ca->ref);
 754                         continue;
 755                 }
 756
 757                 ob = bch2_bucket_alloc_trans(trans, ca, reserve,
 758                                 flags & BUCKET_MAY_ALLOC_PARTIAL, cl, &usage);
 759                 if (!IS_ERR(ob))
 760                         bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
 761                 percpu_ref_put(&ca->ref);
 762
 763                 if (IS_ERR(ob)) {
 764                         ret = PTR_ERR(ob);
 765                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) || cl)
 766                                 break;
 767                         continue;
 768                 }
 769
 770                 add_new_bucket(c, ptrs, devs_may_alloc,
 771                                nr_effective, have_cache, flags, ob);
 772
 773                 if (*nr_effective >= nr_replicas) {
 774                         ret = 0;
 775                         break;
 776                 }
 777         }
 778
 779         return ret;
 780 }
 781
 782 /* Allocate from stripes: */
 783
 784 /*
 785  * if we can't allocate a new stripe because there are already too many
 786  * partially filled stripes, force allocating from an existing stripe even when
 787  * it's to a device we don't want:
 788  */
 789
 790 static int bucket_alloc_from_stripe(struct btree_trans *trans,
 791                          struct open_buckets *ptrs,
 792                          struct write_point *wp,
 793                          struct bch_devs_mask *devs_may_alloc,
 794                          u16 target,
 795                          unsigned erasure_code,
 796                          unsigned nr_replicas,
 797                          unsigned *nr_effective,
 798                          bool *have_cache,
 799                          unsigned flags,
 800                          struct closure *cl)
 801 {
 802         struct bch_fs *c = trans->c;
 803         struct dev_alloc_list devs_sorted;
 804         struct ec_stripe_head *h;
 805         struct open_bucket *ob;
 806         struct bch_dev *ca;
 807         unsigned i, ec_idx;
 808
 809         if (!erasure_code)
 810                 return 0;
 811
 812         if (nr_replicas < 2)
 813                 return 0;
 814
 815         if (ec_open_bucket(c, ptrs))
 816                 return 0;
 817
 818         h = bch2_ec_stripe_head_get(trans, target, 0, nr_replicas - 1,
 819                                     wp == &c->copygc_write_point,
 820                                     cl);
 821         if (IS_ERR(h))
 822                 return PTR_ERR(h);
 823         if (!h)
 824                 return 0;
 825
 826         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 827
 828         for (i = 0; i < devs_sorted.nr; i++)
 829                 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 830                         if (!h->s->blocks[ec_idx])
 831                                 continue;
 832
 833                         ob = c->open_buckets + h->s->blocks[ec_idx];
 834                         if (ob->dev == devs_sorted.devs[i] &&
 835                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 836                                 goto got_bucket;
 837                 }
 838         goto out_put_head;
 839 got_bucket:
 840         ca = bch_dev_bkey_exists(c, ob->dev);
 841
 842         ob->ec_idx      = ec_idx;
 843         ob->ec          = h->s;
 844
 845         add_new_bucket(c, ptrs, devs_may_alloc,
 846                        nr_effective, have_cache, flags, ob);
 847         atomic_inc(&h->s->pin);
 848 out_put_head:
 849         bch2_ec_stripe_head_put(c, h);
 850         return 0;
 851 }
 852
 853 /* Sector allocator */
 854
 855 static void get_buckets_from_writepoint(struct bch_fs *c,
 856                                         struct open_buckets *ptrs,
 857                                         struct write_point *wp,
 858                                         struct bch_devs_mask *devs_may_alloc,
 859                                         unsigned nr_replicas,
 860                                         unsigned *nr_effective,
 861                                         bool *have_cache,
 862                                         unsigned flags,
 863                                         bool need_ec)
 864 {
 865         struct open_buckets ptrs_skip = { .nr = 0 };
 866         struct open_bucket *ob;
 867         unsigned i;
 868
 869         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 870                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 871
 872                 if (*nr_effective < nr_replicas &&
 873                     test_bit(ob->dev, devs_may_alloc->d) &&
 874                     (ca->mi.durability ||
 875                      (wp->data_type == BCH_DATA_user && !*have_cache)) &&
 876                     (ob->ec || !need_ec)) {
 877                         add_new_bucket(c, ptrs, devs_may_alloc,
 878                                        nr_effective, have_cache,
 879                                        flags, ob);
 880                 } else {
 881                         ob_push(c, &ptrs_skip, ob);
 882                 }
 883         }
 884         wp->ptrs = ptrs_skip;
 885 }
 886
 887 static int open_bucket_add_buckets(struct btree_trans *trans,
 888                         struct open_buckets *ptrs,
 889                         struct write_point *wp,
 890                         struct bch_devs_list *devs_have,
 891                         u16 target,
 892                         unsigned erasure_code,
 893                         unsigned nr_replicas,
 894                         unsigned *nr_effective,
 895                         bool *have_cache,
 896                         enum alloc_reserve reserve,
 897                         unsigned flags,
 898                         struct closure *_cl)
 899 {
 900         struct bch_fs *c = trans->c;
 901         struct bch_devs_mask devs;
 902         struct open_bucket *ob;
 903         struct closure *cl = NULL;
 904         int ret;
 905         unsigned i;
 906
 907         rcu_read_lock();
 908         devs = target_rw_devs(c, wp->data_type, target);
 909         rcu_read_unlock();
 910
 911         /* Don't allocate from devices we already have pointers to: */
 912         for (i = 0; i < devs_have->nr; i++)
 913                 __clear_bit(devs_have->devs[i], devs.d);
 914
 915         open_bucket_for_each(c, ptrs, ob, i)
 916                 __clear_bit(ob->dev, devs.d);
 917
 918         if (erasure_code) {
 919                 if (!ec_open_bucket(c, ptrs)) {
 920                         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 921                                                     nr_replicas, nr_effective,
 922                                                     have_cache, flags, true);
 923                         if (*nr_effective >= nr_replicas)
 924                                 return 0;
 925                 }
 926
 927                 if (!ec_open_bucket(c, ptrs)) {
 928                         ret = bucket_alloc_from_stripe(trans, ptrs, wp, &devs,
 929                                                  target, erasure_code,
 930                                                  nr_replicas, nr_effective,
 931                                                  have_cache, flags, _cl);
 932                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
 933                             bch2_err_matches(ret, BCH_ERR_freelist_empty) ||
 934                             bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
 935                                 return ret;
 936                         if (*nr_effective >= nr_replicas)
 937                                 return 0;
 938                 }
 939         }
 940
 941         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 942                                     nr_replicas, nr_effective,
 943                                     have_cache, flags, false);
 944         if (*nr_effective >= nr_replicas)
 945                 return 0;
 946
 947 retry_blocking:
 948         /*
 949          * Try nonblocking first, so that if one device is full we'll try from
 950          * other devices:
 951          */
 952         ret = bch2_bucket_alloc_set_trans(trans, ptrs, &wp->stripe, &devs,
 953                                 nr_replicas, nr_effective, have_cache,
 954                                 reserve, flags, cl);
 955         if (ret &&
 956             !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
 957             !bch2_err_matches(ret, BCH_ERR_insufficient_devices) &&
 958             !cl && _cl) {
 959                 cl = _cl;
 960                 goto retry_blocking;
 961         }
 962
 963         return ret;
 964 }
 965
 966 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 967                                 struct open_buckets *obs)
 968 {
 969         struct open_buckets ptrs = { .nr = 0 };
 970         struct open_bucket *ob, *ob2;
 971         unsigned i, j;
 972
 973         open_bucket_for_each(c, obs, ob, i) {
 974                 bool drop = !ca || ob->dev == ca->dev_idx;
 975
 976                 if (!drop && ob->ec) {
 977                         mutex_lock(&ob->ec->lock);
 978                         for (j = 0; j < ob->ec->new_stripe.key.v.nr_blocks; j++) {
 979                                 if (!ob->ec->blocks[j])
 980                                         continue;
 981
 982                                 ob2 = c->open_buckets + ob->ec->blocks[j];
 983                                 drop |= ob2->dev == ca->dev_idx;
 984                         }
 985                         mutex_unlock(&ob->ec->lock);
 986                 }
 987
 988                 if (drop)
 989                         bch2_open_bucket_put(c, ob);
 990                 else
 991                         ob_push(c, &ptrs, ob);
 992         }
 993
 994         *obs = ptrs;
 995 }
 996
 997 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 998                           struct write_point *wp)
 999 {
1000         mutex_lock(&wp->lock);
1001         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
1002         mutex_unlock(&wp->lock);
1003 }
1004
1005 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
1006                                                  unsigned long write_point)
1007 {
1008         unsigned hash =
1009                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
1010
1011         return &c->write_points_hash[hash];
1012 }
1013
1014 static struct write_point *__writepoint_find(struct hlist_head *head,
1015                                              unsigned long write_point)
1016 {
1017         struct write_point *wp;
1018
1019         rcu_read_lock();
1020         hlist_for_each_entry_rcu(wp, head, node)
1021                 if (wp->write_point == write_point)
1022                         goto out;
1023         wp = NULL;
1024 out:
1025         rcu_read_unlock();
1026         return wp;
1027 }
1028
1029 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
1030 {
1031         u64 stranded    = c->write_points_nr * c->bucket_size_max;
1032         u64 free        = bch2_fs_usage_read_short(c).free;
1033
1034         return stranded * factor > free;
1035 }
1036
1037 static bool try_increase_writepoints(struct bch_fs *c)
1038 {
1039         struct write_point *wp;
1040
1041         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
1042             too_many_writepoints(c, 32))
1043                 return false;
1044
1045         wp = c->write_points + c->write_points_nr++;
1046         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
1047         return true;
1048 }
1049
1050 static bool try_decrease_writepoints(struct bch_fs *c,
1051                                      unsigned old_nr)
1052 {
1053         struct write_point *wp;
1054
1055         mutex_lock(&c->write_points_hash_lock);
1056         if (c->write_points_nr < old_nr) {
1057                 mutex_unlock(&c->write_points_hash_lock);
1058                 return true;
1059         }
1060
1061         if (c->write_points_nr == 1 ||
1062             !too_many_writepoints(c, 8)) {
1063                 mutex_unlock(&c->write_points_hash_lock);
1064                 return false;
1065         }
1066
1067         wp = c->write_points + --c->write_points_nr;
1068
1069         hlist_del_rcu(&wp->node);
1070         mutex_unlock(&c->write_points_hash_lock);
1071
1072         bch2_writepoint_stop(c, NULL, wp);
1073         return true;
1074 }
1075
1076 static void bch2_trans_mutex_lock_norelock(struct btree_trans *trans,
1077                                   struct mutex *lock)
1078 {
1079         if (!mutex_trylock(lock)) {
1080                 bch2_trans_unlock(trans);
1081                 mutex_lock(lock);
1082         }
1083 }
1084
1085 static struct write_point *writepoint_find(struct btree_trans *trans,
1086                                            unsigned long write_point)
1087 {
1088         struct bch_fs *c = trans->c;
1089         struct write_point *wp, *oldest;
1090         struct hlist_head *head;
1091
1092         if (!(write_point & 1UL)) {
1093                 wp = (struct write_point *) write_point;
1094                 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1095                 return wp;
1096         }
1097
1098         head = writepoint_hash(c, write_point);
1099 restart_find:
1100         wp = __writepoint_find(head, write_point);
1101         if (wp) {
1102 lock_wp:
1103                 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1104                 if (wp->write_point == write_point)
1105                         goto out;
1106                 mutex_unlock(&wp->lock);
1107                 goto restart_find;
1108         }
1109 restart_find_oldest:
1110         oldest = NULL;
1111         for (wp = c->write_points;
1112              wp < c->write_points + c->write_points_nr; wp++)
1113                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1114                         oldest = wp;
1115
1116         bch2_trans_mutex_lock_norelock(trans, &oldest->lock);
1117         bch2_trans_mutex_lock_norelock(trans, &c->write_points_hash_lock);
1118         if (oldest >= c->write_points + c->write_points_nr ||
1119             try_increase_writepoints(c)) {
1120                 mutex_unlock(&c->write_points_hash_lock);
1121                 mutex_unlock(&oldest->lock);
1122                 goto restart_find_oldest;
1123         }
1124
1125         wp = __writepoint_find(head, write_point);
1126         if (wp && wp != oldest) {
1127                 mutex_unlock(&c->write_points_hash_lock);
1128                 mutex_unlock(&oldest->lock);
1129                 goto lock_wp;
1130         }
1131
1132         wp = oldest;
1133         hlist_del_rcu(&wp->node);
1134         wp->write_point = write_point;
1135         hlist_add_head_rcu(&wp->node, head);
1136         mutex_unlock(&c->write_points_hash_lock);
1137 out:
1138         wp->last_used = local_clock();
1139         return wp;
1140 }
1141
1142 /*
1143  * Get us an open_bucket we can allocate from, return with it locked:
1144  */
1145 int bch2_alloc_sectors_start_trans(struct btree_trans *trans,
1146                              unsigned target,
1147                              unsigned erasure_code,
1148                              struct write_point_specifier write_point,
1149                              struct bch_devs_list *devs_have,
1150                              unsigned nr_replicas,
1151                              unsigned nr_replicas_required,
1152                              enum alloc_reserve reserve,
1153                              unsigned flags,
1154                              struct closure *cl,
1155                              struct write_point **wp_ret)
1156 {
1157         struct bch_fs *c = trans->c;
1158         struct write_point *wp;
1159         struct open_bucket *ob;
1160         struct open_buckets ptrs;
1161         unsigned nr_effective, write_points_nr;
1162         unsigned ob_flags = 0;
1163         bool have_cache;
1164         int ret;
1165         int i;
1166
1167         if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
1168                 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
1169
1170         BUG_ON(!nr_replicas || !nr_replicas_required);
1171 retry:
1172         ptrs.nr         = 0;
1173         nr_effective    = 0;
1174         write_points_nr = c->write_points_nr;
1175         have_cache      = false;
1176
1177         *wp_ret = wp = writepoint_find(trans, write_point.v);
1178
1179         if (wp->data_type == BCH_DATA_user)
1180                 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
1181
1182         /* metadata may not allocate on cache devices: */
1183         if (wp->data_type != BCH_DATA_user)
1184                 have_cache = true;
1185
1186         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1187                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1188                                               target, erasure_code,
1189                                               nr_replicas, &nr_effective,
1190                                               &have_cache, reserve,
1191                                               ob_flags, cl);
1192         } else {
1193                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1194                                               target, erasure_code,
1195                                               nr_replicas, &nr_effective,
1196                                               &have_cache, reserve,
1197                                               ob_flags, NULL);
1198                 if (!ret ||
1199                     bch2_err_matches(ret, BCH_ERR_transaction_restart))
1200                         goto alloc_done;
1201
1202                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1203                                               0, erasure_code,
1204                                               nr_replicas, &nr_effective,
1205                                               &have_cache, reserve,
1206                                               ob_flags, cl);
1207         }
1208 alloc_done:
1209         BUG_ON(!ret && nr_effective < nr_replicas);
1210
1211         if (erasure_code && !ec_open_bucket(c, &ptrs))
1212                 pr_debug("failed to get ec bucket: ret %u", ret);
1213
1214         if (ret == -BCH_ERR_insufficient_devices &&
1215             nr_effective >= nr_replicas_required)
1216                 ret = 0;
1217
1218         if (ret)
1219                 goto err;
1220
1221         /* Free buckets we didn't use: */
1222         open_bucket_for_each(c, &wp->ptrs, ob, i)
1223                 open_bucket_free_unused(c, wp, ob);
1224
1225         wp->ptrs = ptrs;
1226
1227         wp->sectors_free = UINT_MAX;
1228
1229         open_bucket_for_each(c, &wp->ptrs, ob, i)
1230                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1231
1232         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1233
1234         return 0;
1235 err:
1236         open_bucket_for_each(c, &wp->ptrs, ob, i)
1237                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1238                         ob_push(c, &ptrs, ob);
1239                 else
1240                         open_bucket_free_unused(c, wp, ob);
1241         wp->ptrs = ptrs;
1242
1243         mutex_unlock(&wp->lock);
1244
1245         if (bch2_err_matches(ret, BCH_ERR_freelist_empty) &&
1246             try_decrease_writepoints(c, write_points_nr))
1247                 goto retry;
1248
1249         if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty) ||
1250             bch2_err_matches(ret, BCH_ERR_freelist_empty))
1251                 return cl
1252                         ? -BCH_ERR_bucket_alloc_blocked
1253                         : -BCH_ERR_ENOSPC_bucket_alloc;
1254
1255         return ret;
1256 }
1257
1258 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1259 {
1260         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1261
1262         return (struct bch_extent_ptr) {
1263                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
1264                 .gen    = ob->gen,
1265                 .dev    = ob->dev,
1266                 .offset = bucket_to_sector(ca, ob->bucket) +
1267                         ca->mi.bucket_size -
1268                         ob->sectors_free,
1269         };
1270 }
1271
1272 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1273                                     struct bkey_i *k, unsigned sectors,
1274                                     bool cached)
1275 {
1276         bch2_alloc_sectors_append_ptrs_inlined(c, wp, k, sectors, cached);
1277 }
1278
1279 /*
1280  * Append pointers to the space we just allocated to @k, and mark @sectors space
1281  * as allocated out of @ob
1282  */
1283 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1284 {
1285         bch2_alloc_sectors_done_inlined(c, wp);
1286 }
1287
1288 static inline void writepoint_init(struct write_point *wp,
1289                                    enum bch_data_type type)
1290 {
1291         mutex_init(&wp->lock);
1292         wp->data_type = type;
1293
1294         INIT_WORK(&wp->index_update_work, bch2_write_point_do_index_updates);
1295         INIT_LIST_HEAD(&wp->writes);
1296         spin_lock_init(&wp->writes_lock);
1297 }
1298
1299 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1300 {
1301         struct open_bucket *ob;
1302         struct write_point *wp;
1303
1304         mutex_init(&c->write_points_hash_lock);
1305         c->write_points_nr = ARRAY_SIZE(c->write_points);
1306
1307         /* open bucket 0 is a sentinal NULL: */
1308         spin_lock_init(&c->open_buckets[0].lock);
1309
1310         for (ob = c->open_buckets + 1;
1311              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1312                 spin_lock_init(&ob->lock);
1313                 c->open_buckets_nr_free++;
1314
1315                 ob->freelist = c->open_buckets_freelist;
1316                 c->open_buckets_freelist = ob - c->open_buckets;
1317         }
1318
1319         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
1320         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
1321         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
1322
1323         for (wp = c->write_points;
1324              wp < c->write_points + c->write_points_nr; wp++) {
1325                 writepoint_init(wp, BCH_DATA_user);
1326
1327                 wp->last_used   = local_clock();
1328                 wp->write_point = (unsigned long) wp;
1329                 hlist_add_head_rcu(&wp->node,
1330                                    writepoint_hash(c, wp->write_point));
1331         }
1332 }
1333
1334 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1335 {
1336         struct open_bucket *ob;
1337
1338         for (ob = c->open_buckets;
1339              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1340              ob++) {
1341                 spin_lock(&ob->lock);
1342                 if (ob->valid && !ob->on_partial_list) {
1343                         prt_printf(out, "%zu ref %u type %s %u:%llu:%u\n",
1344                                ob - c->open_buckets,
1345                                atomic_read(&ob->pin),
1346                                bch2_data_types[ob->data_type],
1347                                ob->dev, ob->bucket, ob->gen);
1348                 }
1349                 spin_unlock(&ob->lock);
1350         }
1351 }
1352
1353 static const char * const bch2_write_point_states[] = {
1354 #define x(n)    #n,
1355         WRITE_POINT_STATES()
1356 #undef x
1357         NULL
1358 };
1359
1360 void bch2_write_points_to_text(struct printbuf *out, struct bch_fs *c)
1361 {
1362         struct write_point *wp;
1363         unsigned i;
1364
1365         for (wp = c->write_points;
1366              wp < c->write_points + ARRAY_SIZE(c->write_points);
1367              wp++) {
1368                 prt_printf(out, "%lu: ", wp->write_point);
1369                 prt_human_readable_u64(out, wp->sectors_allocated);
1370
1371                 prt_printf(out, " last wrote: ");
1372                 bch2_pr_time_units(out, sched_clock() - wp->last_used);
1373
1374                 for (i = 0; i < WRITE_POINT_STATE_NR; i++) {
1375                         prt_printf(out, " %s: ", bch2_write_point_states[i]);
1376                         bch2_pr_time_units(out, wp->time[i]);
1377                 }
1378
1379                 prt_newline(out);
1380         }
1381 }