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 #include "trace.h"
  33
  34 #include <linux/math64.h>
  35 #include <linux/rculist.h>
  36 #include <linux/rcupdate.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                 ec_stripe_new_put(c, ob->ec, STRIPE_REF_io);
 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         BUG_ON(c->open_buckets_partial_nr >=
 158                ARRAY_SIZE(c->open_buckets_partial));
 159
 160         spin_lock(&c->freelist_lock);
 161         ob->on_partial_list = true;
 162         c->open_buckets_partial[c->open_buckets_partial_nr++] =
 163                 ob - c->open_buckets;
 164         spin_unlock(&c->freelist_lock);
 165
 166         closure_wake_up(&c->open_buckets_wait);
 167         closure_wake_up(&c->freelist_wait);
 168 }
 169
 170 /* _only_ for allocating the journal on a new device: */
 171 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
 172 {
 173         while (ca->new_fs_bucket_idx < ca->mi.nbuckets) {
 174                 u64 b = ca->new_fs_bucket_idx++;
 175
 176                 if (!is_superblock_bucket(ca, b) &&
 177                     (!ca->buckets_nouse || !test_bit(b, ca->buckets_nouse)))
 178                         return b;
 179         }
 180
 181         return -1;
 182 }
 183
 184 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
 185 {
 186         switch (reserve) {
 187         case RESERVE_btree:
 188         case RESERVE_btree_movinggc:
 189                 return 0;
 190         case RESERVE_movinggc:
 191                 return OPEN_BUCKETS_COUNT / 4;
 192         default:
 193                 return OPEN_BUCKETS_COUNT / 2;
 194         }
 195 }
 196
 197 static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 198                                               u64 bucket,
 199                                               enum alloc_reserve reserve,
 200                                               const struct bch_alloc_v4 *a,
 201                                               struct bucket_alloc_state *s,
 202                                               struct closure *cl)
 203 {
 204         struct open_bucket *ob;
 205
 206         if (unlikely(ca->buckets_nouse && test_bit(bucket, ca->buckets_nouse))) {
 207                 s->skipped_nouse++;
 208                 return NULL;
 209         }
 210
 211         if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
 212                 s->skipped_open++;
 213                 return NULL;
 214         }
 215
 216         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 217                         c->journal.flushed_seq_ondisk, ca->dev_idx, bucket)) {
 218                 s->skipped_need_journal_commit++;
 219                 return NULL;
 220         }
 221
 222         if (bch2_bucket_nocow_is_locked(&c->nocow_locks, POS(ca->dev_idx, bucket))) {
 223                 s->skipped_nocow++;
 224                 return NULL;
 225         }
 226
 227         spin_lock(&c->freelist_lock);
 228
 229         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
 230                 if (cl)
 231                         closure_wait(&c->open_buckets_wait, cl);
 232
 233                 if (!c->blocked_allocate_open_bucket)
 234                         c->blocked_allocate_open_bucket = local_clock();
 235
 236                 spin_unlock(&c->freelist_lock);
 237                 return ERR_PTR(-BCH_ERR_open_buckets_empty);
 238         }
 239
 240         /* Recheck under lock: */
 241         if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
 242                 spin_unlock(&c->freelist_lock);
 243                 s->skipped_open++;
 244                 return NULL;
 245         }
 246
 247         ob = bch2_open_bucket_alloc(c);
 248
 249         spin_lock(&ob->lock);
 250
 251         ob->valid       = true;
 252         ob->sectors_free = ca->mi.bucket_size;
 253         ob->dev         = ca->dev_idx;
 254         ob->gen         = a->gen;
 255         ob->bucket      = bucket;
 256         spin_unlock(&ob->lock);
 257
 258         ca->nr_open_buckets++;
 259         bch2_open_bucket_hash_add(c, ob);
 260
 261         if (c->blocked_allocate_open_bucket) {
 262                 bch2_time_stats_update(
 263                         &c->times[BCH_TIME_blocked_allocate_open_bucket],
 264                         c->blocked_allocate_open_bucket);
 265                 c->blocked_allocate_open_bucket = 0;
 266         }
 267
 268         if (c->blocked_allocate) {
 269                 bch2_time_stats_update(
 270                         &c->times[BCH_TIME_blocked_allocate],
 271                         c->blocked_allocate);
 272                 c->blocked_allocate = 0;
 273         }
 274
 275         spin_unlock(&c->freelist_lock);
 276         return ob;
 277 }
 278
 279 static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bch_dev *ca,
 280                                             enum alloc_reserve reserve, u64 free_entry,
 281                                             struct bucket_alloc_state *s,
 282                                             struct bkey_s_c freespace_k,
 283                                             struct closure *cl)
 284 {
 285         struct bch_fs *c = trans->c;
 286         struct btree_iter iter = { NULL };
 287         struct bkey_s_c k;
 288         struct open_bucket *ob;
 289         struct bch_alloc_v4 a_convert;
 290         const struct bch_alloc_v4 *a;
 291         u64 b = free_entry & ~(~0ULL << 56);
 292         unsigned genbits = free_entry >> 56;
 293         struct printbuf buf = PRINTBUF;
 294         int ret;
 295
 296         if (b < ca->mi.first_bucket || b >= ca->mi.nbuckets) {
 297                 prt_printf(&buf, "freespace btree has bucket outside allowed range %u-%llu\n"
 298                        "  freespace key ",
 299                         ca->mi.first_bucket, ca->mi.nbuckets);
 300                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 301                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 302                 ob = ERR_PTR(-EIO);
 303                 goto err;
 304         }
 305
 306         k = bch2_bkey_get_iter(trans, &iter,
 307                                BTREE_ID_alloc, POS(ca->dev_idx, b),
 308                                BTREE_ITER_CACHED);
 309         ret = bkey_err(k);
 310         if (ret) {
 311                 ob = ERR_PTR(ret);
 312                 goto err;
 313         }
 314
 315         a = bch2_alloc_to_v4(k, &a_convert);
 316
 317         if (a->data_type != BCH_DATA_free) {
 318                 if (!test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
 319                         ob = NULL;
 320                         goto err;
 321                 }
 322
 323                 prt_printf(&buf, "non free bucket in freespace btree\n"
 324                        "  freespace key ");
 325                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 326                 prt_printf(&buf, "\n  ");
 327                 bch2_bkey_val_to_text(&buf, c, k);
 328                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 329                 ob = ERR_PTR(-EIO);
 330                 goto err;
 331         }
 332
 333         if (genbits != (alloc_freespace_genbits(*a) >> 56) &&
 334             test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
 335                 prt_printf(&buf, "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
 336                        "  freespace key ",
 337                        genbits, alloc_freespace_genbits(*a) >> 56);
 338                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 339                 prt_printf(&buf, "\n  ");
 340                 bch2_bkey_val_to_text(&buf, c, k);
 341                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 342                 ob = ERR_PTR(-EIO);
 343                 goto err;
 344
 345         }
 346
 347         if (!test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags)) {
 348                 struct bch_backpointer bp;
 349                 struct bpos bp_pos = POS_MIN;
 350
 351                 ret = bch2_get_next_backpointer(trans, POS(ca->dev_idx, b), -1,
 352                                                 &bp_pos, &bp,
 353                                                 BTREE_ITER_NOPRESERVE);
 354                 if (ret) {
 355                         ob = ERR_PTR(ret);
 356                         goto err;
 357                 }
 358
 359                 if (!bkey_eq(bp_pos, POS_MAX)) {
 360                         /*
 361                          * Bucket may have data in it - we don't call
 362                          * bc2h_trans_inconnsistent() because fsck hasn't
 363                          * finished yet
 364                          */
 365                         ob = NULL;
 366                         goto err;
 367                 }
 368         }
 369
 370         ob = __try_alloc_bucket(c, ca, b, reserve, a, s, cl);
 371         if (!ob)
 372                 iter.path->preserve = false;
 373 err:
 374         set_btree_iter_dontneed(&iter);
 375         bch2_trans_iter_exit(trans, &iter);
 376         printbuf_exit(&buf);
 377         return ob;
 378 }
 379
 380 /*
 381  * This path is for before the freespace btree is initialized:
 382  *
 383  * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
 384  * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
 385  */
 386 static noinline struct open_bucket *
 387 bch2_bucket_alloc_early(struct btree_trans *trans,
 388                         struct bch_dev *ca,
 389                         enum alloc_reserve reserve,
 390                         struct bucket_alloc_state *s,
 391                         struct closure *cl)
 392 {
 393         struct btree_iter iter;
 394         struct bkey_s_c k;
 395         struct open_bucket *ob = NULL;
 396         u64 alloc_start = max_t(u64, ca->mi.first_bucket, ca->new_fs_bucket_idx);
 397         u64 alloc_cursor = max(alloc_start, READ_ONCE(ca->alloc_cursor));
 398         int ret;
 399 again:
 400         for_each_btree_key_norestart(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, alloc_cursor),
 401                            BTREE_ITER_SLOTS, k, ret) {
 402                 struct bch_alloc_v4 a_convert;
 403                 const struct bch_alloc_v4 *a;
 404
 405                 if (bkey_ge(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)))
 406                         break;
 407
 408                 if (ca->new_fs_bucket_idx &&
 409                     is_superblock_bucket(ca, k.k->p.offset))
 410                         continue;
 411
 412                 a = bch2_alloc_to_v4(k, &a_convert);
 413
 414                 if (a->data_type != BCH_DATA_free)
 415                         continue;
 416
 417                 s->buckets_seen++;
 418
 419                 ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, reserve, a, s, cl);
 420                 if (ob)
 421                         break;
 422         }
 423         bch2_trans_iter_exit(trans, &iter);
 424
 425         ca->alloc_cursor = alloc_cursor;
 426
 427         if (!ob && ret)
 428                 ob = ERR_PTR(ret);
 429
 430         if (!ob && alloc_cursor > alloc_start) {
 431                 alloc_cursor = alloc_start;
 432                 goto again;
 433         }
 434
 435         return ob;
 436 }
 437
 438 static struct open_bucket *bch2_bucket_alloc_freelist(struct btree_trans *trans,
 439                                                    struct bch_dev *ca,
 440                                                    enum alloc_reserve reserve,
 441                                                    struct bucket_alloc_state *s,
 442                                                    struct closure *cl)
 443 {
 444         struct btree_iter iter;
 445         struct bkey_s_c k;
 446         struct open_bucket *ob = NULL;
 447         u64 alloc_start = max_t(u64, ca->mi.first_bucket, READ_ONCE(ca->alloc_cursor));
 448         u64 alloc_cursor = alloc_start;
 449         int ret;
 450
 451         BUG_ON(ca->new_fs_bucket_idx);
 452 again:
 453         for_each_btree_key_norestart(trans, iter, BTREE_ID_freespace,
 454                                      POS(ca->dev_idx, alloc_cursor), 0, k, ret) {
 455                 if (k.k->p.inode != ca->dev_idx)
 456                         break;
 457
 458                 for (alloc_cursor = max(alloc_cursor, bkey_start_offset(k.k));
 459                      alloc_cursor < k.k->p.offset;
 460                      alloc_cursor++) {
 461                         ret = btree_trans_too_many_iters(trans);
 462                         if (ret) {
 463                                 ob = ERR_PTR(ret);
 464                                 break;
 465                         }
 466
 467                         s->buckets_seen++;
 468
 469                         ob = try_alloc_bucket(trans, ca, reserve,
 470                                               alloc_cursor, s, k, cl);
 471                         if (ob) {
 472                                 iter.path->preserve = false;
 473                                 break;
 474                         }
 475                 }
 476
 477                 if (ob || ret)
 478                         break;
 479         }
 480         bch2_trans_iter_exit(trans, &iter);
 481
 482         ca->alloc_cursor = alloc_cursor;
 483
 484         if (!ob && ret)
 485                 ob = ERR_PTR(ret);
 486
 487         if (!ob && alloc_start > ca->mi.first_bucket) {
 488                 alloc_cursor = alloc_start = ca->mi.first_bucket;
 489                 goto again;
 490         }
 491
 492         return ob;
 493 }
 494
 495 /**
 496  * bch_bucket_alloc - allocate a single bucket from a specific device
 497  *
 498  * Returns index of bucket on success, 0 on failure
 499  */
 500 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
 501                                       struct bch_dev *ca,
 502                                       enum alloc_reserve reserve,
 503                                       struct closure *cl,
 504                                       struct bch_dev_usage *usage)
 505 {
 506         struct bch_fs *c = trans->c;
 507         struct open_bucket *ob = NULL;
 508         bool freespace = READ_ONCE(ca->mi.freespace_initialized);
 509         u64 avail;
 510         struct bucket_alloc_state s = { 0 };
 511         bool waiting = false;
 512 again:
 513         bch2_dev_usage_read_fast(ca, usage);
 514         avail = dev_buckets_free(ca, *usage, reserve);
 515
 516         if (usage->d[BCH_DATA_need_discard].buckets > avail)
 517                 bch2_do_discards(c);
 518
 519         if (usage->d[BCH_DATA_need_gc_gens].buckets > avail)
 520                 bch2_do_gc_gens(c);
 521
 522         if (should_invalidate_buckets(ca, *usage))
 523                 bch2_do_invalidates(c);
 524
 525         if (!avail) {
 526                 if (cl && !waiting) {
 527                         closure_wait(&c->freelist_wait, cl);
 528                         waiting = true;
 529                         goto again;
 530                 }
 531
 532                 if (!c->blocked_allocate)
 533                         c->blocked_allocate = local_clock();
 534
 535                 ob = ERR_PTR(-BCH_ERR_freelist_empty);
 536                 goto err;
 537         }
 538
 539         if (waiting)
 540                 closure_wake_up(&c->freelist_wait);
 541 alloc:
 542         ob = likely(freespace)
 543                 ? bch2_bucket_alloc_freelist(trans, ca, reserve, &s, cl)
 544                 : bch2_bucket_alloc_early(trans, ca, reserve, &s, cl);
 545
 546         if (s.skipped_need_journal_commit * 2 > avail)
 547                 bch2_journal_flush_async(&c->journal, NULL);
 548
 549         if (!ob && freespace && !test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
 550                 freespace = false;
 551                 goto alloc;
 552         }
 553 err:
 554         if (!ob)
 555                 ob = ERR_PTR(-BCH_ERR_no_buckets_found);
 556
 557         if (!IS_ERR(ob))
 558                 trace_and_count(c, bucket_alloc, ca,
 559                                 bch2_alloc_reserves[reserve],
 560                                 ob->bucket,
 561                                 usage->d[BCH_DATA_free].buckets,
 562                                 avail,
 563                                 bch2_copygc_wait_amount(c),
 564                                 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
 565                                 &s,
 566                                 cl == NULL,
 567                                 "");
 568         else if (!bch2_err_matches(PTR_ERR(ob), BCH_ERR_transaction_restart))
 569                 trace_and_count(c, bucket_alloc_fail, ca,
 570                                 bch2_alloc_reserves[reserve],
 571                                 0,
 572                                 usage->d[BCH_DATA_free].buckets,
 573                                 avail,
 574                                 bch2_copygc_wait_amount(c),
 575                                 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
 576                                 &s,
 577                                 cl == NULL,
 578                                 bch2_err_str(PTR_ERR(ob)));
 579
 580         return ob;
 581 }
 582
 583 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 584                                       enum alloc_reserve reserve,
 585                                       struct closure *cl)
 586 {
 587         struct bch_dev_usage usage;
 588         struct open_bucket *ob;
 589
 590         bch2_trans_do(c, NULL, NULL, 0,
 591                       PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, reserve,
 592                                                         cl, &usage)));
 593         return ob;
 594 }
 595
 596 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 597                             unsigned l, unsigned r)
 598 {
 599         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 600                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 601 }
 602
 603 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 604
 605 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 606                                           struct dev_stripe_state *stripe,
 607                                           struct bch_devs_mask *devs)
 608 {
 609         struct dev_alloc_list ret = { .nr = 0 };
 610         unsigned i;
 611
 612         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 613                 ret.devs[ret.nr++] = i;
 614
 615         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 616         return ret;
 617 }
 618
 619 static inline void bch2_dev_stripe_increment_inlined(struct bch_dev *ca,
 620                                struct dev_stripe_state *stripe,
 621                                struct bch_dev_usage *usage)
 622 {
 623         u64 *v = stripe->next_alloc + ca->dev_idx;
 624         u64 free_space = dev_buckets_available(ca, RESERVE_none);
 625         u64 free_space_inv = free_space
 626                 ? div64_u64(1ULL << 48, free_space)
 627                 : 1ULL << 48;
 628         u64 scale = *v / 4;
 629
 630         if (*v + free_space_inv >= *v)
 631                 *v += free_space_inv;
 632         else
 633                 *v = U64_MAX;
 634
 635         for (v = stripe->next_alloc;
 636              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 637                 *v = *v < scale ? 0 : *v - scale;
 638 }
 639
 640 void bch2_dev_stripe_increment(struct bch_dev *ca,
 641                                struct dev_stripe_state *stripe)
 642 {
 643         struct bch_dev_usage usage;
 644
 645         bch2_dev_usage_read_fast(ca, &usage);
 646         bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
 647 }
 648
 649 static int add_new_bucket(struct bch_fs *c,
 650                            struct open_buckets *ptrs,
 651                            struct bch_devs_mask *devs_may_alloc,
 652                            unsigned nr_replicas,
 653                            unsigned *nr_effective,
 654                            bool *have_cache,
 655                            unsigned flags,
 656                            struct open_bucket *ob)
 657 {
 658         unsigned durability =
 659                 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
 660
 661         BUG_ON(*nr_effective >= nr_replicas);
 662         BUG_ON(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS);
 663
 664         __clear_bit(ob->dev, devs_may_alloc->d);
 665         *nr_effective   += (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)
 666                 ? durability : 1;
 667         *have_cache     |= !durability;
 668
 669         ob_push(c, ptrs, ob);
 670
 671         if (*nr_effective >= nr_replicas)
 672                 return 1;
 673         if (ob->ec)
 674                 return 1;
 675         return 0;
 676 }
 677
 678 int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
 679                       struct open_buckets *ptrs,
 680                       struct dev_stripe_state *stripe,
 681                       struct bch_devs_mask *devs_may_alloc,
 682                       unsigned nr_replicas,
 683                       unsigned *nr_effective,
 684                       bool *have_cache,
 685                       unsigned flags,
 686                       enum bch_data_type data_type,
 687                       enum alloc_reserve reserve,
 688                       struct closure *cl)
 689 {
 690         struct bch_fs *c = trans->c;
 691         struct dev_alloc_list devs_sorted =
 692                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 693         unsigned dev;
 694         struct bch_dev *ca;
 695         int ret = -BCH_ERR_insufficient_devices;
 696         unsigned i;
 697
 698         BUG_ON(*nr_effective >= nr_replicas);
 699
 700         for (i = 0; i < devs_sorted.nr; i++) {
 701                 struct bch_dev_usage usage;
 702                 struct open_bucket *ob;
 703
 704                 dev = devs_sorted.devs[i];
 705
 706                 rcu_read_lock();
 707                 ca = rcu_dereference(c->devs[dev]);
 708                 if (ca)
 709                         percpu_ref_get(&ca->ref);
 710                 rcu_read_unlock();
 711
 712                 if (!ca)
 713                         continue;
 714
 715                 if (!ca->mi.durability && *have_cache) {
 716                         percpu_ref_put(&ca->ref);
 717                         continue;
 718                 }
 719
 720                 ob = bch2_bucket_alloc_trans(trans, ca, reserve, cl, &usage);
 721                 if (!IS_ERR(ob))
 722                         bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
 723                 percpu_ref_put(&ca->ref);
 724
 725                 if (IS_ERR(ob)) {
 726                         ret = PTR_ERR(ob);
 727                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) || cl)
 728                                 break;
 729                         continue;
 730                 }
 731
 732                 ob->data_type = data_type;
 733
 734                 if (add_new_bucket(c, ptrs, devs_may_alloc,
 735                                    nr_replicas, nr_effective,
 736                                    have_cache, flags, ob)) {
 737                         ret = 0;
 738                         break;
 739                 }
 740         }
 741
 742         return ret;
 743 }
 744
 745 /* Allocate from stripes: */
 746
 747 /*
 748  * if we can't allocate a new stripe because there are already too many
 749  * partially filled stripes, force allocating from an existing stripe even when
 750  * it's to a device we don't want:
 751  */
 752
 753 static int bucket_alloc_from_stripe(struct btree_trans *trans,
 754                          struct open_buckets *ptrs,
 755                          struct write_point *wp,
 756                          struct bch_devs_mask *devs_may_alloc,
 757                          u16 target,
 758                          unsigned nr_replicas,
 759                          unsigned *nr_effective,
 760                          bool *have_cache,
 761                          enum alloc_reserve reserve,
 762                          unsigned flags,
 763                          struct closure *cl)
 764 {
 765         struct bch_fs *c = trans->c;
 766         struct dev_alloc_list devs_sorted;
 767         struct ec_stripe_head *h;
 768         struct open_bucket *ob;
 769         struct bch_dev *ca;
 770         unsigned i, ec_idx;
 771         int ret = 0;
 772
 773         if (nr_replicas < 2)
 774                 return 0;
 775
 776         if (ec_open_bucket(c, ptrs))
 777                 return 0;
 778
 779         h = bch2_ec_stripe_head_get(trans, target, 0, nr_replicas - 1, reserve, cl);
 780         if (IS_ERR(h))
 781                 return PTR_ERR(h);
 782         if (!h)
 783                 return 0;
 784
 785         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 786
 787         for (i = 0; i < devs_sorted.nr; i++)
 788                 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 789                         if (!h->s->blocks[ec_idx])
 790                                 continue;
 791
 792                         ob = c->open_buckets + h->s->blocks[ec_idx];
 793                         if (ob->dev == devs_sorted.devs[i] &&
 794                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 795                                 goto got_bucket;
 796                 }
 797         goto out_put_head;
 798 got_bucket:
 799         ca = bch_dev_bkey_exists(c, ob->dev);
 800
 801         ob->ec_idx      = ec_idx;
 802         ob->ec          = h->s;
 803         ec_stripe_new_get(h->s, STRIPE_REF_io);
 804
 805         ret = add_new_bucket(c, ptrs, devs_may_alloc,
 806                              nr_replicas, nr_effective,
 807                              have_cache, flags, ob);
 808 out_put_head:
 809         bch2_ec_stripe_head_put(c, h);
 810         return ret;
 811 }
 812
 813 /* Sector allocator */
 814
 815 static bool want_bucket(struct bch_fs *c,
 816                         struct write_point *wp,
 817                         struct bch_devs_mask *devs_may_alloc,
 818                         bool *have_cache, bool ec,
 819                         struct open_bucket *ob)
 820 {
 821         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 822
 823         if (!test_bit(ob->dev, devs_may_alloc->d))
 824                 return false;
 825
 826         if (ob->data_type != wp->data_type)
 827                 return false;
 828
 829         if (!ca->mi.durability &&
 830             (wp->data_type == BCH_DATA_btree || ec || *have_cache))
 831                 return false;
 832
 833         if (ec != (ob->ec != NULL))
 834                 return false;
 835
 836         return true;
 837 }
 838
 839 static int bucket_alloc_set_writepoint(struct bch_fs *c,
 840                                        struct open_buckets *ptrs,
 841                                        struct write_point *wp,
 842                                        struct bch_devs_mask *devs_may_alloc,
 843                                        unsigned nr_replicas,
 844                                        unsigned *nr_effective,
 845                                        bool *have_cache,
 846                                        bool ec, unsigned flags)
 847 {
 848         struct open_buckets ptrs_skip = { .nr = 0 };
 849         struct open_bucket *ob;
 850         unsigned i;
 851         int ret = 0;
 852
 853         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 854                 if (!ret && want_bucket(c, wp, devs_may_alloc,
 855                                         have_cache, ec, ob))
 856                         ret = add_new_bucket(c, ptrs, devs_may_alloc,
 857                                        nr_replicas, nr_effective,
 858                                        have_cache, flags, ob);
 859                 else
 860                         ob_push(c, &ptrs_skip, ob);
 861         }
 862         wp->ptrs = ptrs_skip;
 863
 864         return ret;
 865 }
 866
 867 static int bucket_alloc_set_partial(struct bch_fs *c,
 868                                     struct open_buckets *ptrs,
 869                                     struct write_point *wp,
 870                                     struct bch_devs_mask *devs_may_alloc,
 871                                     unsigned nr_replicas,
 872                                     unsigned *nr_effective,
 873                                     bool *have_cache, bool ec,
 874                                     enum alloc_reserve reserve,
 875                                     unsigned flags)
 876 {
 877         int i, ret = 0;
 878
 879         if (!c->open_buckets_partial_nr)
 880                 return 0;
 881
 882         spin_lock(&c->freelist_lock);
 883
 884         if (!c->open_buckets_partial_nr)
 885                 goto unlock;
 886
 887         for (i = c->open_buckets_partial_nr - 1; i >= 0; --i) {
 888                 struct open_bucket *ob = c->open_buckets + c->open_buckets_partial[i];
 889
 890                 if (want_bucket(c, wp, devs_may_alloc, have_cache, ec, ob)) {
 891                         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 892                         struct bch_dev_usage usage;
 893                         u64 avail;
 894
 895                         bch2_dev_usage_read_fast(ca, &usage);
 896                         avail = dev_buckets_free(ca, usage, reserve);
 897                         if (!avail)
 898                                 continue;
 899
 900                         array_remove_item(c->open_buckets_partial,
 901                                           c->open_buckets_partial_nr,
 902                                           i);
 903                         ob->on_partial_list = false;
 904
 905                         ret = add_new_bucket(c, ptrs, devs_may_alloc,
 906                                              nr_replicas, nr_effective,
 907                                              have_cache, flags, ob);
 908                         if (ret)
 909                                 break;
 910                 }
 911         }
 912 unlock:
 913         spin_unlock(&c->freelist_lock);
 914         return ret;
 915 }
 916
 917 static int __open_bucket_add_buckets(struct btree_trans *trans,
 918                         struct open_buckets *ptrs,
 919                         struct write_point *wp,
 920                         struct bch_devs_list *devs_have,
 921                         u16 target,
 922                         bool erasure_code,
 923                         unsigned nr_replicas,
 924                         unsigned *nr_effective,
 925                         bool *have_cache,
 926                         enum alloc_reserve reserve,
 927                         unsigned flags,
 928                         struct closure *_cl)
 929 {
 930         struct bch_fs *c = trans->c;
 931         struct bch_devs_mask devs;
 932         struct open_bucket *ob;
 933         struct closure *cl = NULL;
 934         unsigned i;
 935         int ret;
 936
 937         rcu_read_lock();
 938         devs = target_rw_devs(c, wp->data_type, target);
 939         rcu_read_unlock();
 940
 941         /* Don't allocate from devices we already have pointers to: */
 942         for (i = 0; i < devs_have->nr; i++)
 943                 __clear_bit(devs_have->devs[i], devs.d);
 944
 945         open_bucket_for_each(c, ptrs, ob, i)
 946                 __clear_bit(ob->dev, devs.d);
 947
 948         if (erasure_code && ec_open_bucket(c, ptrs))
 949                 return 0;
 950
 951         ret = bucket_alloc_set_writepoint(c, ptrs, wp, &devs,
 952                                  nr_replicas, nr_effective,
 953                                  have_cache, erasure_code, flags);
 954         if (ret)
 955                 return ret;
 956
 957         ret = bucket_alloc_set_partial(c, ptrs, wp, &devs,
 958                                  nr_replicas, nr_effective,
 959                                  have_cache, erasure_code, reserve, flags);
 960         if (ret)
 961                 return ret;
 962
 963         if (erasure_code) {
 964                 ret = bucket_alloc_from_stripe(trans, ptrs, wp, &devs,
 965                                          target,
 966                                          nr_replicas, nr_effective,
 967                                          have_cache,
 968                                          reserve, flags, _cl);
 969         } else {
 970 retry_blocking:
 971                 /*
 972                  * Try nonblocking first, so that if one device is full we'll try from
 973                  * other devices:
 974                  */
 975                 ret = bch2_bucket_alloc_set_trans(trans, ptrs, &wp->stripe, &devs,
 976                                         nr_replicas, nr_effective, have_cache,
 977                                         flags, wp->data_type, reserve, cl);
 978                 if (ret &&
 979                     !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
 980                     !bch2_err_matches(ret, BCH_ERR_insufficient_devices) &&
 981                     !cl && _cl) {
 982                         cl = _cl;
 983                         goto retry_blocking;
 984                 }
 985
 986         }
 987
 988         return ret;
 989 }
 990
 991 static int open_bucket_add_buckets(struct btree_trans *trans,
 992                         struct open_buckets *ptrs,
 993                         struct write_point *wp,
 994                         struct bch_devs_list *devs_have,
 995                         u16 target,
 996                         unsigned erasure_code,
 997                         unsigned nr_replicas,
 998                         unsigned *nr_effective,
 999                         bool *have_cache,
1000                         enum alloc_reserve reserve,
1001                         unsigned flags,
1002                         struct closure *cl)
1003 {
1004         int ret;
1005
1006         if (erasure_code) {
1007                 ret = __open_bucket_add_buckets(trans, ptrs, wp,
1008                                 devs_have, target, erasure_code,
1009                                 nr_replicas, nr_effective, have_cache,
1010                                 reserve, flags, cl);
1011                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1012                     bch2_err_matches(ret, BCH_ERR_operation_blocked) ||
1013                     bch2_err_matches(ret, BCH_ERR_freelist_empty) ||
1014                     bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
1015                         return ret;
1016                 if (*nr_effective >= nr_replicas)
1017                         return 0;
1018         }
1019
1020         ret = __open_bucket_add_buckets(trans, ptrs, wp,
1021                         devs_have, target, false,
1022                         nr_replicas, nr_effective, have_cache,
1023                         reserve, flags, cl);
1024         return ret < 0 ? ret : 0;
1025 }
1026
1027 static bool should_drop_bucket(struct open_bucket *ob, struct bch_fs *c,
1028                                struct bch_dev *ca, bool ec)
1029 {
1030         if (ec) {
1031                 return ob->ec != NULL;
1032         } else if (ca) {
1033                 bool drop = ob->dev == ca->dev_idx;
1034                 struct open_bucket *ob2;
1035                 unsigned i;
1036
1037                 if (!drop && ob->ec) {
1038                         mutex_lock(&ob->ec->lock);
1039                         for (i = 0; i < ob->ec->new_stripe.key.v.nr_blocks; i++) {
1040                                 if (!ob->ec->blocks[i])
1041                                         continue;
1042
1043                                 ob2 = c->open_buckets + ob->ec->blocks[i];
1044                                 drop |= ob2->dev == ca->dev_idx;
1045                         }
1046                         mutex_unlock(&ob->ec->lock);
1047                 }
1048
1049                 return drop;
1050         } else {
1051                 return true;
1052         }
1053 }
1054
1055 static void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
1056                                  bool ec, struct write_point *wp)
1057 {
1058         struct open_buckets ptrs = { .nr = 0 };
1059         struct open_bucket *ob;
1060         unsigned i;
1061
1062         mutex_lock(&wp->lock);
1063         open_bucket_for_each(c, &wp->ptrs, ob, i)
1064                 if (should_drop_bucket(ob, c, ca, ec))
1065                         bch2_open_bucket_put(c, ob);
1066                 else
1067                         ob_push(c, &ptrs, ob);
1068         wp->ptrs = ptrs;
1069         mutex_unlock(&wp->lock);
1070 }
1071
1072 void bch2_open_buckets_stop(struct bch_fs *c, struct bch_dev *ca,
1073                             bool ec)
1074 {
1075         unsigned i;
1076
1077         /* Next, close write points that point to this device... */
1078         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1079                 bch2_writepoint_stop(c, ca, ec, &c->write_points[i]);
1080
1081         bch2_writepoint_stop(c, ca, ec, &c->copygc_write_point);
1082         bch2_writepoint_stop(c, ca, ec, &c->rebalance_write_point);
1083         bch2_writepoint_stop(c, ca, ec, &c->btree_write_point);
1084
1085         mutex_lock(&c->btree_reserve_cache_lock);
1086         while (c->btree_reserve_cache_nr) {
1087                 struct btree_alloc *a =
1088                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1089
1090                 bch2_open_buckets_put(c, &a->ob);
1091         }
1092         mutex_unlock(&c->btree_reserve_cache_lock);
1093
1094         spin_lock(&c->freelist_lock);
1095         i = 0;
1096         while (i < c->open_buckets_partial_nr) {
1097                 struct open_bucket *ob =
1098                         c->open_buckets + c->open_buckets_partial[i];
1099
1100                 if (should_drop_bucket(ob, c, ca, ec)) {
1101                         --c->open_buckets_partial_nr;
1102                         swap(c->open_buckets_partial[i],
1103                              c->open_buckets_partial[c->open_buckets_partial_nr]);
1104                         ob->on_partial_list = false;
1105                         spin_unlock(&c->freelist_lock);
1106                         bch2_open_bucket_put(c, ob);
1107                         spin_lock(&c->freelist_lock);
1108                 } else {
1109                         i++;
1110                 }
1111         }
1112         spin_unlock(&c->freelist_lock);
1113
1114         bch2_ec_stop_dev(c, ca);
1115 }
1116
1117 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
1118                                                  unsigned long write_point)
1119 {
1120         unsigned hash =
1121                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
1122
1123         return &c->write_points_hash[hash];
1124 }
1125
1126 static struct write_point *__writepoint_find(struct hlist_head *head,
1127                                              unsigned long write_point)
1128 {
1129         struct write_point *wp;
1130
1131         rcu_read_lock();
1132         hlist_for_each_entry_rcu(wp, head, node)
1133                 if (wp->write_point == write_point)
1134                         goto out;
1135         wp = NULL;
1136 out:
1137         rcu_read_unlock();
1138         return wp;
1139 }
1140
1141 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
1142 {
1143         u64 stranded    = c->write_points_nr * c->bucket_size_max;
1144         u64 free        = bch2_fs_usage_read_short(c).free;
1145
1146         return stranded * factor > free;
1147 }
1148
1149 static bool try_increase_writepoints(struct bch_fs *c)
1150 {
1151         struct write_point *wp;
1152
1153         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
1154             too_many_writepoints(c, 32))
1155                 return false;
1156
1157         wp = c->write_points + c->write_points_nr++;
1158         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
1159         return true;
1160 }
1161
1162 static bool try_decrease_writepoints(struct bch_fs *c, unsigned old_nr)
1163 {
1164         struct write_point *wp;
1165
1166         mutex_lock(&c->write_points_hash_lock);
1167         if (c->write_points_nr < old_nr) {
1168                 mutex_unlock(&c->write_points_hash_lock);
1169                 return true;
1170         }
1171
1172         if (c->write_points_nr == 1 ||
1173             !too_many_writepoints(c, 8)) {
1174                 mutex_unlock(&c->write_points_hash_lock);
1175                 return false;
1176         }
1177
1178         wp = c->write_points + --c->write_points_nr;
1179
1180         hlist_del_rcu(&wp->node);
1181         mutex_unlock(&c->write_points_hash_lock);
1182
1183         bch2_writepoint_stop(c, NULL, false, wp);
1184         return true;
1185 }
1186
1187 static void bch2_trans_mutex_lock_norelock(struct btree_trans *trans,
1188                                   struct mutex *lock)
1189 {
1190         if (!mutex_trylock(lock)) {
1191                 bch2_trans_unlock(trans);
1192                 mutex_lock(lock);
1193         }
1194 }
1195
1196 static struct write_point *writepoint_find(struct btree_trans *trans,
1197                                            unsigned long write_point)
1198 {
1199         struct bch_fs *c = trans->c;
1200         struct write_point *wp, *oldest;
1201         struct hlist_head *head;
1202
1203         if (!(write_point & 1UL)) {
1204                 wp = (struct write_point *) write_point;
1205                 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1206                 return wp;
1207         }
1208
1209         head = writepoint_hash(c, write_point);
1210 restart_find:
1211         wp = __writepoint_find(head, write_point);
1212         if (wp) {
1213 lock_wp:
1214                 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1215                 if (wp->write_point == write_point)
1216                         goto out;
1217                 mutex_unlock(&wp->lock);
1218                 goto restart_find;
1219         }
1220 restart_find_oldest:
1221         oldest = NULL;
1222         for (wp = c->write_points;
1223              wp < c->write_points + c->write_points_nr; wp++)
1224                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1225                         oldest = wp;
1226
1227         bch2_trans_mutex_lock_norelock(trans, &oldest->lock);
1228         bch2_trans_mutex_lock_norelock(trans, &c->write_points_hash_lock);
1229         if (oldest >= c->write_points + c->write_points_nr ||
1230             try_increase_writepoints(c)) {
1231                 mutex_unlock(&c->write_points_hash_lock);
1232                 mutex_unlock(&oldest->lock);
1233                 goto restart_find_oldest;
1234         }
1235
1236         wp = __writepoint_find(head, write_point);
1237         if (wp && wp != oldest) {
1238                 mutex_unlock(&c->write_points_hash_lock);
1239                 mutex_unlock(&oldest->lock);
1240                 goto lock_wp;
1241         }
1242
1243         wp = oldest;
1244         hlist_del_rcu(&wp->node);
1245         wp->write_point = write_point;
1246         hlist_add_head_rcu(&wp->node, head);
1247         mutex_unlock(&c->write_points_hash_lock);
1248 out:
1249         wp->last_used = local_clock();
1250         return wp;
1251 }
1252
1253 /*
1254  * Get us an open_bucket we can allocate from, return with it locked:
1255  */
1256 int bch2_alloc_sectors_start_trans(struct btree_trans *trans,
1257                              unsigned target,
1258                              unsigned erasure_code,
1259                              struct write_point_specifier write_point,
1260                              struct bch_devs_list *devs_have,
1261                              unsigned nr_replicas,
1262                              unsigned nr_replicas_required,
1263                              enum alloc_reserve reserve,
1264                              unsigned flags,
1265                              struct closure *cl,
1266                              struct write_point **wp_ret)
1267 {
1268         struct bch_fs *c = trans->c;
1269         struct write_point *wp;
1270         struct open_bucket *ob;
1271         struct open_buckets ptrs;
1272         unsigned nr_effective, write_points_nr;
1273         bool have_cache;
1274         int ret;
1275         int i;
1276
1277         BUG_ON(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS);
1278
1279         BUG_ON(!nr_replicas || !nr_replicas_required);
1280 retry:
1281         ptrs.nr         = 0;
1282         nr_effective    = 0;
1283         write_points_nr = c->write_points_nr;
1284         have_cache      = false;
1285
1286         *wp_ret = wp = writepoint_find(trans, write_point.v);
1287
1288         /* metadata may not allocate on cache devices: */
1289         if (wp->data_type != BCH_DATA_user)
1290                 have_cache = true;
1291
1292         if (target && !(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1293                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1294                                               target, erasure_code,
1295                                               nr_replicas, &nr_effective,
1296                                               &have_cache, reserve,
1297                                               flags, NULL);
1298                 if (!ret ||
1299                     bch2_err_matches(ret, BCH_ERR_transaction_restart))
1300                         goto alloc_done;
1301
1302                 /* Don't retry from all devices if we're out of open buckets: */
1303                 if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
1304                         goto allocate_blocking;
1305
1306                 /*
1307                  * Only try to allocate cache (durability = 0 devices) from the
1308                  * specified target:
1309                  */
1310                 have_cache = true;
1311
1312                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1313                                               0, erasure_code,
1314                                               nr_replicas, &nr_effective,
1315                                               &have_cache, reserve,
1316                                               flags, cl);
1317         } else {
1318 allocate_blocking:
1319                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1320                                               target, erasure_code,
1321                                               nr_replicas, &nr_effective,
1322                                               &have_cache, reserve,
1323                                               flags, cl);
1324         }
1325 alloc_done:
1326         BUG_ON(!ret && nr_effective < nr_replicas);
1327
1328         if (erasure_code && !ec_open_bucket(c, &ptrs))
1329                 pr_debug("failed to get ec bucket: ret %u", ret);
1330
1331         if (ret == -BCH_ERR_insufficient_devices &&
1332             nr_effective >= nr_replicas_required)
1333                 ret = 0;
1334
1335         if (ret)
1336                 goto err;
1337
1338         /* Free buckets we didn't use: */
1339         open_bucket_for_each(c, &wp->ptrs, ob, i)
1340                 open_bucket_free_unused(c, wp, ob);
1341
1342         wp->ptrs = ptrs;
1343
1344         wp->sectors_free = UINT_MAX;
1345
1346         open_bucket_for_each(c, &wp->ptrs, ob, i)
1347                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1348
1349         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1350
1351         return 0;
1352 err:
1353         open_bucket_for_each(c, &wp->ptrs, ob, i)
1354                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1355                         ob_push(c, &ptrs, ob);
1356                 else
1357                         open_bucket_free_unused(c, wp, ob);
1358         wp->ptrs = ptrs;
1359
1360         mutex_unlock(&wp->lock);
1361
1362         if (bch2_err_matches(ret, BCH_ERR_freelist_empty) &&
1363             try_decrease_writepoints(c, write_points_nr))
1364                 goto retry;
1365
1366         if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty) ||
1367             bch2_err_matches(ret, BCH_ERR_freelist_empty))
1368                 return cl
1369                         ? -BCH_ERR_bucket_alloc_blocked
1370                         : -BCH_ERR_ENOSPC_bucket_alloc;
1371
1372         return ret;
1373 }
1374
1375 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1376 {
1377         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1378
1379         return (struct bch_extent_ptr) {
1380                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
1381                 .gen    = ob->gen,
1382                 .dev    = ob->dev,
1383                 .offset = bucket_to_sector(ca, ob->bucket) +
1384                         ca->mi.bucket_size -
1385                         ob->sectors_free,
1386         };
1387 }
1388
1389 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1390                                     struct bkey_i *k, unsigned sectors,
1391                                     bool cached)
1392 {
1393         bch2_alloc_sectors_append_ptrs_inlined(c, wp, k, sectors, cached);
1394 }
1395
1396 /*
1397  * Append pointers to the space we just allocated to @k, and mark @sectors space
1398  * as allocated out of @ob
1399  */
1400 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1401 {
1402         bch2_alloc_sectors_done_inlined(c, wp);
1403 }
1404
1405 static inline void writepoint_init(struct write_point *wp,
1406                                    enum bch_data_type type)
1407 {
1408         mutex_init(&wp->lock);
1409         wp->data_type = type;
1410
1411         INIT_WORK(&wp->index_update_work, bch2_write_point_do_index_updates);
1412         INIT_LIST_HEAD(&wp->writes);
1413         spin_lock_init(&wp->writes_lock);
1414 }
1415
1416 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1417 {
1418         struct open_bucket *ob;
1419         struct write_point *wp;
1420
1421         mutex_init(&c->write_points_hash_lock);
1422         c->write_points_nr = ARRAY_SIZE(c->write_points);
1423
1424         /* open bucket 0 is a sentinal NULL: */
1425         spin_lock_init(&c->open_buckets[0].lock);
1426
1427         for (ob = c->open_buckets + 1;
1428              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1429                 spin_lock_init(&ob->lock);
1430                 c->open_buckets_nr_free++;
1431
1432                 ob->freelist = c->open_buckets_freelist;
1433                 c->open_buckets_freelist = ob - c->open_buckets;
1434         }
1435
1436         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
1437         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
1438         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
1439
1440         for (wp = c->write_points;
1441              wp < c->write_points + c->write_points_nr; wp++) {
1442                 writepoint_init(wp, BCH_DATA_user);
1443
1444                 wp->last_used   = local_clock();
1445                 wp->write_point = (unsigned long) wp;
1446                 hlist_add_head_rcu(&wp->node,
1447                                    writepoint_hash(c, wp->write_point));
1448         }
1449 }
1450
1451 static void bch2_open_bucket_to_text(struct printbuf *out, struct bch_fs *c, struct open_bucket *ob)
1452 {
1453         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1454         unsigned data_type = ob->data_type;
1455         barrier(); /* READ_ONCE() doesn't work on bitfields */
1456
1457         prt_printf(out, "%zu ref %u %s %u:%llu gen %u allocated %u/%u",
1458                    ob - c->open_buckets,
1459                    atomic_read(&ob->pin),
1460                    data_type < BCH_DATA_NR ? bch2_data_types[data_type] : "invalid data type",
1461                    ob->dev, ob->bucket, ob->gen,
1462                    ca->mi.bucket_size - ob->sectors_free, ca->mi.bucket_size);
1463         if (ob->ec)
1464                 prt_printf(out, " ec idx %llu", ob->ec->idx);
1465         if (ob->on_partial_list)
1466                 prt_str(out, " partial");
1467         prt_newline(out);
1468 }
1469
1470 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1471 {
1472         struct open_bucket *ob;
1473
1474         out->atomic++;
1475
1476         for (ob = c->open_buckets;
1477              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1478              ob++) {
1479                 spin_lock(&ob->lock);
1480                 if (ob->valid && !ob->on_partial_list)
1481                         bch2_open_bucket_to_text(out, c, ob);
1482                 spin_unlock(&ob->lock);
1483         }
1484
1485         --out->atomic;
1486 }
1487
1488 void bch2_open_buckets_partial_to_text(struct printbuf *out, struct bch_fs *c)
1489 {
1490         unsigned i;
1491
1492         out->atomic++;
1493         spin_lock(&c->freelist_lock);
1494
1495         for (i = 0; i < c->open_buckets_partial_nr; i++)
1496                 bch2_open_bucket_to_text(out, c,
1497                                 c->open_buckets + c->open_buckets_partial[i]);
1498
1499         spin_unlock(&c->freelist_lock);
1500         --out->atomic;
1501 }
1502
1503 static const char * const bch2_write_point_states[] = {
1504 #define x(n)    #n,
1505         WRITE_POINT_STATES()
1506 #undef x
1507         NULL
1508 };
1509
1510 void bch2_write_points_to_text(struct printbuf *out, struct bch_fs *c)
1511 {
1512         struct write_point *wp;
1513         unsigned i;
1514
1515         for (wp = c->write_points;
1516              wp < c->write_points + ARRAY_SIZE(c->write_points);
1517              wp++) {
1518                 prt_printf(out, "%lu: ", wp->write_point);
1519                 prt_human_readable_u64(out, wp->sectors_allocated);
1520
1521                 prt_printf(out, " last wrote: ");
1522                 bch2_pr_time_units(out, sched_clock() - wp->last_used);
1523
1524                 for (i = 0; i < WRITE_POINT_STATE_NR; i++) {
1525                         prt_printf(out, " %s: ", bch2_write_point_states[i]);
1526                         bch2_pr_time_units(out, wp->time[i]);
1527                 }
1528
1529                 prt_newline(out);
1530         }
1531 }