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                 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         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS(ca->dev_idx, b), BTREE_ITER_CACHED);
 307         k = bch2_btree_iter_peek_slot(&iter);
 308         ret = bkey_err(k);
 309         if (ret) {
 310                 ob = ERR_PTR(ret);
 311                 goto err;
 312         }
 313
 314         a = bch2_alloc_to_v4(k, &a_convert);
 315
 316         if (a->data_type != BCH_DATA_free) {
 317                 if (!test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
 318                         ob = NULL;
 319                         goto err;
 320                 }
 321
 322                 prt_printf(&buf, "non free bucket in freespace btree\n"
 323                        "  freespace key ");
 324                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 325                 prt_printf(&buf, "\n  ");
 326                 bch2_bkey_val_to_text(&buf, c, k);
 327                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 328                 ob = ERR_PTR(-EIO);
 329                 goto err;
 330         }
 331
 332         if (genbits != (alloc_freespace_genbits(*a) >> 56) &&
 333             test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
 334                 prt_printf(&buf, "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
 335                        "  freespace key ",
 336                        genbits, alloc_freespace_genbits(*a) >> 56);
 337                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 338                 prt_printf(&buf, "\n  ");
 339                 bch2_bkey_val_to_text(&buf, c, k);
 340                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 341                 ob = ERR_PTR(-EIO);
 342                 goto err;
 343
 344         }
 345
 346         if (!test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags)) {
 347                 struct bch_backpointer bp;
 348                 struct bpos bp_pos = POS_MIN;
 349
 350                 ret = bch2_get_next_backpointer(trans, POS(ca->dev_idx, b), -1,
 351                                                 &bp_pos, &bp,
 352                                                 BTREE_ITER_NOPRESERVE);
 353                 if (ret) {
 354                         ob = ERR_PTR(ret);
 355                         goto err;
 356                 }
 357
 358                 if (!bkey_eq(bp_pos, POS_MAX)) {
 359                         /*
 360                          * Bucket may have data in it - we don't call
 361                          * bc2h_trans_inconnsistent() because fsck hasn't
 362                          * finished yet
 363                          */
 364                         ob = NULL;
 365                         goto err;
 366                 }
 367         }
 368
 369         ob = __try_alloc_bucket(c, ca, b, reserve, a, s, cl);
 370         if (!ob)
 371                 iter.path->preserve = false;
 372 err:
 373         set_btree_iter_dontneed(&iter);
 374         bch2_trans_iter_exit(trans, &iter);
 375         printbuf_exit(&buf);
 376         return ob;
 377 }
 378
 379 /*
 380  * This path is for before the freespace btree is initialized:
 381  *
 382  * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
 383  * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
 384  */
 385 static noinline struct open_bucket *
 386 bch2_bucket_alloc_early(struct btree_trans *trans,
 387                         struct bch_dev *ca,
 388                         enum alloc_reserve reserve,
 389                         struct bucket_alloc_state *s,
 390                         struct closure *cl)
 391 {
 392         struct btree_iter iter;
 393         struct bkey_s_c k;
 394         struct open_bucket *ob = NULL;
 395         u64 alloc_start = max_t(u64, ca->mi.first_bucket, ca->new_fs_bucket_idx);
 396         u64 alloc_cursor = max(alloc_start, READ_ONCE(ca->alloc_cursor));
 397         int ret;
 398 again:
 399         for_each_btree_key_norestart(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, alloc_cursor),
 400                            BTREE_ITER_SLOTS, k, ret) {
 401                 struct bch_alloc_v4 a_convert;
 402                 const struct bch_alloc_v4 *a;
 403
 404                 if (bkey_ge(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)))
 405                         break;
 406
 407                 if (ca->new_fs_bucket_idx &&
 408                     is_superblock_bucket(ca, k.k->p.offset))
 409                         continue;
 410
 411                 a = bch2_alloc_to_v4(k, &a_convert);
 412
 413                 if (a->data_type != BCH_DATA_free)
 414                         continue;
 415
 416                 s->buckets_seen++;
 417
 418                 ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, reserve, a, s, cl);
 419                 if (ob)
 420                         break;
 421         }
 422         bch2_trans_iter_exit(trans, &iter);
 423
 424         ca->alloc_cursor = alloc_cursor;
 425
 426         if (!ob && ret)
 427                 ob = ERR_PTR(ret);
 428
 429         if (!ob && alloc_cursor > alloc_start) {
 430                 alloc_cursor = alloc_start;
 431                 goto again;
 432         }
 433
 434         return ob;
 435 }
 436
 437 static struct open_bucket *bch2_bucket_alloc_freelist(struct btree_trans *trans,
 438                                                    struct bch_dev *ca,
 439                                                    enum alloc_reserve reserve,
 440                                                    struct bucket_alloc_state *s,
 441                                                    struct closure *cl)
 442 {
 443         struct btree_iter iter;
 444         struct bkey_s_c k;
 445         struct open_bucket *ob = NULL;
 446         u64 alloc_start = max_t(u64, ca->mi.first_bucket, READ_ONCE(ca->alloc_cursor));
 447         u64 alloc_cursor = alloc_start;
 448         int ret;
 449
 450         BUG_ON(ca->new_fs_bucket_idx);
 451 again:
 452         for_each_btree_key_norestart(trans, iter, BTREE_ID_freespace,
 453                                      POS(ca->dev_idx, alloc_cursor), 0, k, ret) {
 454                 if (k.k->p.inode != ca->dev_idx)
 455                         break;
 456
 457                 for (alloc_cursor = max(alloc_cursor, bkey_start_offset(k.k));
 458                      alloc_cursor < k.k->p.offset;
 459                      alloc_cursor++) {
 460                         ret = btree_trans_too_many_iters(trans);
 461                         if (ret) {
 462                                 ob = ERR_PTR(ret);
 463                                 break;
 464                         }
 465
 466                         s->buckets_seen++;
 467
 468                         ob = try_alloc_bucket(trans, ca, reserve,
 469                                               alloc_cursor, s, k, cl);
 470                         if (ob) {
 471                                 iter.path->preserve = false;
 472                                 break;
 473                         }
 474                 }
 475
 476                 if (ob || ret)
 477                         break;
 478         }
 479         bch2_trans_iter_exit(trans, &iter);
 480
 481         ca->alloc_cursor = alloc_cursor;
 482
 483         if (!ob && ret)
 484                 ob = ERR_PTR(ret);
 485
 486         if (!ob && alloc_start > ca->mi.first_bucket) {
 487                 alloc_cursor = alloc_start = ca->mi.first_bucket;
 488                 goto again;
 489         }
 490
 491         return ob;
 492 }
 493
 494 /**
 495  * bch_bucket_alloc - allocate a single bucket from a specific device
 496  *
 497  * Returns index of bucket on success, 0 on failure
 498  */
 499 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
 500                                       struct bch_dev *ca,
 501                                       enum alloc_reserve reserve,
 502                                       struct closure *cl,
 503                                       struct bch_dev_usage *usage)
 504 {
 505         struct bch_fs *c = trans->c;
 506         struct open_bucket *ob = NULL;
 507         bool freespace = READ_ONCE(ca->mi.freespace_initialized);
 508         u64 avail;
 509         struct bucket_alloc_state s = { 0 };
 510         bool waiting = false;
 511 again:
 512         bch2_dev_usage_read_fast(ca, usage);
 513         avail = dev_buckets_free(ca, *usage, reserve);
 514
 515         if (usage->d[BCH_DATA_need_discard].buckets > avail)
 516                 bch2_do_discards(c);
 517
 518         if (usage->d[BCH_DATA_need_gc_gens].buckets > avail)
 519                 bch2_do_gc_gens(c);
 520
 521         if (should_invalidate_buckets(ca, *usage))
 522                 bch2_do_invalidates(c);
 523
 524         if (!avail) {
 525                 if (cl && !waiting) {
 526                         closure_wait(&c->freelist_wait, cl);
 527                         waiting = true;
 528                         goto again;
 529                 }
 530
 531                 if (!c->blocked_allocate)
 532                         c->blocked_allocate = local_clock();
 533
 534                 ob = ERR_PTR(-BCH_ERR_freelist_empty);
 535                 goto err;
 536         }
 537
 538         if (waiting)
 539                 closure_wake_up(&c->freelist_wait);
 540 alloc:
 541         ob = likely(freespace)
 542                 ? bch2_bucket_alloc_freelist(trans, ca, reserve, &s, cl)
 543                 : bch2_bucket_alloc_early(trans, ca, reserve, &s, cl);
 544
 545         if (s.skipped_need_journal_commit * 2 > avail)
 546                 bch2_journal_flush_async(&c->journal, NULL);
 547
 548         if (!ob && freespace && !test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
 549                 freespace = false;
 550                 goto alloc;
 551         }
 552 err:
 553         if (!ob)
 554                 ob = ERR_PTR(-BCH_ERR_no_buckets_found);
 555
 556         if (!IS_ERR(ob))
 557                 trace_and_count(c, bucket_alloc, ca,
 558                                 bch2_alloc_reserves[reserve],
 559                                 ob->bucket,
 560                                 usage->d[BCH_DATA_free].buckets,
 561                                 avail,
 562                                 bch2_copygc_wait_amount(c),
 563                                 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
 564                                 &s,
 565                                 cl == NULL,
 566                                 "");
 567         else if (!bch2_err_matches(PTR_ERR(ob), BCH_ERR_transaction_restart))
 568                 trace_and_count(c, bucket_alloc_fail, ca,
 569                                 bch2_alloc_reserves[reserve],
 570                                 0,
 571                                 usage->d[BCH_DATA_free].buckets,
 572                                 avail,
 573                                 bch2_copygc_wait_amount(c),
 574                                 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
 575                                 &s,
 576                                 cl == NULL,
 577                                 bch2_err_str(PTR_ERR(ob)));
 578
 579         return ob;
 580 }
 581
 582 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 583                                       enum alloc_reserve reserve,
 584                                       struct closure *cl)
 585 {
 586         struct bch_dev_usage usage;
 587         struct open_bucket *ob;
 588
 589         bch2_trans_do(c, NULL, NULL, 0,
 590                       PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, reserve,
 591                                                         cl, &usage)));
 592         return ob;
 593 }
 594
 595 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 596                             unsigned l, unsigned r)
 597 {
 598         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 599                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 600 }
 601
 602 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 603
 604 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 605                                           struct dev_stripe_state *stripe,
 606                                           struct bch_devs_mask *devs)
 607 {
 608         struct dev_alloc_list ret = { .nr = 0 };
 609         unsigned i;
 610
 611         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 612                 ret.devs[ret.nr++] = i;
 613
 614         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 615         return ret;
 616 }
 617
 618 static inline void bch2_dev_stripe_increment_inlined(struct bch_dev *ca,
 619                                struct dev_stripe_state *stripe,
 620                                struct bch_dev_usage *usage)
 621 {
 622         u64 *v = stripe->next_alloc + ca->dev_idx;
 623         u64 free_space = dev_buckets_available(ca, RESERVE_none);
 624         u64 free_space_inv = free_space
 625                 ? div64_u64(1ULL << 48, free_space)
 626                 : 1ULL << 48;
 627         u64 scale = *v / 4;
 628
 629         if (*v + free_space_inv >= *v)
 630                 *v += free_space_inv;
 631         else
 632                 *v = U64_MAX;
 633
 634         for (v = stripe->next_alloc;
 635              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 636                 *v = *v < scale ? 0 : *v - scale;
 637 }
 638
 639 void bch2_dev_stripe_increment(struct bch_dev *ca,
 640                                struct dev_stripe_state *stripe)
 641 {
 642         struct bch_dev_usage usage;
 643
 644         bch2_dev_usage_read_fast(ca, &usage);
 645         bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
 646 }
 647
 648 static int add_new_bucket(struct bch_fs *c,
 649                            struct open_buckets *ptrs,
 650                            struct bch_devs_mask *devs_may_alloc,
 651                            unsigned nr_replicas,
 652                            unsigned *nr_effective,
 653                            bool *have_cache,
 654                            unsigned flags,
 655                            struct open_bucket *ob)
 656 {
 657         unsigned durability =
 658                 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
 659
 660         BUG_ON(*nr_effective >= nr_replicas);
 661         BUG_ON(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS);
 662
 663         __clear_bit(ob->dev, devs_may_alloc->d);
 664         *nr_effective   += (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)
 665                 ? durability : 1;
 666         *have_cache     |= !durability;
 667
 668         ob_push(c, ptrs, ob);
 669
 670         if (*nr_effective >= nr_replicas)
 671                 return 1;
 672         if (ob->ec)
 673                 return 1;
 674         return 0;
 675 }
 676
 677 int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
 678                       struct open_buckets *ptrs,
 679                       struct dev_stripe_state *stripe,
 680                       struct bch_devs_mask *devs_may_alloc,
 681                       unsigned nr_replicas,
 682                       unsigned *nr_effective,
 683                       bool *have_cache,
 684                       unsigned flags,
 685                       enum bch_data_type data_type,
 686                       enum alloc_reserve reserve,
 687                       struct closure *cl)
 688 {
 689         struct bch_fs *c = trans->c;
 690         struct dev_alloc_list devs_sorted =
 691                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 692         unsigned dev;
 693         struct bch_dev *ca;
 694         int ret = -BCH_ERR_insufficient_devices;
 695         unsigned i;
 696
 697         BUG_ON(*nr_effective >= nr_replicas);
 698
 699         for (i = 0; i < devs_sorted.nr; i++) {
 700                 struct bch_dev_usage usage;
 701                 struct open_bucket *ob;
 702
 703                 dev = devs_sorted.devs[i];
 704
 705                 rcu_read_lock();
 706                 ca = rcu_dereference(c->devs[dev]);
 707                 if (ca)
 708                         percpu_ref_get(&ca->ref);
 709                 rcu_read_unlock();
 710
 711                 if (!ca)
 712                         continue;
 713
 714                 if (!ca->mi.durability && *have_cache) {
 715                         percpu_ref_put(&ca->ref);
 716                         continue;
 717                 }
 718
 719                 ob = bch2_bucket_alloc_trans(trans, ca, reserve, cl, &usage);
 720                 if (!IS_ERR(ob))
 721                         bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
 722                 percpu_ref_put(&ca->ref);
 723
 724                 if (IS_ERR(ob)) {
 725                         ret = PTR_ERR(ob);
 726                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) || cl)
 727                                 break;
 728                         continue;
 729                 }
 730
 731                 ob->data_type = data_type;
 732
 733                 if (add_new_bucket(c, ptrs, devs_may_alloc,
 734                                    nr_replicas, nr_effective,
 735                                    have_cache, flags, ob)) {
 736                         ret = 0;
 737                         break;
 738                 }
 739         }
 740
 741         return ret;
 742 }
 743
 744 /* Allocate from stripes: */
 745
 746 /*
 747  * if we can't allocate a new stripe because there are already too many
 748  * partially filled stripes, force allocating from an existing stripe even when
 749  * it's to a device we don't want:
 750  */
 751
 752 static int bucket_alloc_from_stripe(struct btree_trans *trans,
 753                          struct open_buckets *ptrs,
 754                          struct write_point *wp,
 755                          struct bch_devs_mask *devs_may_alloc,
 756                          u16 target,
 757                          unsigned nr_replicas,
 758                          unsigned *nr_effective,
 759                          bool *have_cache,
 760                          enum alloc_reserve reserve,
 761                          unsigned flags,
 762                          struct closure *cl)
 763 {
 764         struct bch_fs *c = trans->c;
 765         struct dev_alloc_list devs_sorted;
 766         struct ec_stripe_head *h;
 767         struct open_bucket *ob;
 768         struct bch_dev *ca;
 769         unsigned i, ec_idx;
 770         int ret = 0;
 771
 772         if (nr_replicas < 2)
 773                 return 0;
 774
 775         if (ec_open_bucket(c, ptrs))
 776                 return 0;
 777
 778         h = bch2_ec_stripe_head_get(trans, target, 0, nr_replicas - 1, reserve, cl);
 779         if (IS_ERR(h))
 780                 return PTR_ERR(h);
 781         if (!h)
 782                 return 0;
 783
 784         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 785
 786         for (i = 0; i < devs_sorted.nr; i++)
 787                 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 788                         if (!h->s->blocks[ec_idx])
 789                                 continue;
 790
 791                         ob = c->open_buckets + h->s->blocks[ec_idx];
 792                         if (ob->dev == devs_sorted.devs[i] &&
 793                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 794                                 goto got_bucket;
 795                 }
 796         goto out_put_head;
 797 got_bucket:
 798         ca = bch_dev_bkey_exists(c, ob->dev);
 799
 800         ob->ec_idx      = ec_idx;
 801         ob->ec          = h->s;
 802         ec_stripe_new_get(h->s, STRIPE_REF_io);
 803
 804         ret = add_new_bucket(c, ptrs, devs_may_alloc,
 805                              nr_replicas, nr_effective,
 806                              have_cache, flags, ob);
 807 out_put_head:
 808         bch2_ec_stripe_head_put(c, h);
 809         return ret;
 810 }
 811
 812 /* Sector allocator */
 813
 814 static bool want_bucket(struct bch_fs *c,
 815                         struct write_point *wp,
 816                         struct bch_devs_mask *devs_may_alloc,
 817                         bool *have_cache, bool ec,
 818                         struct open_bucket *ob)
 819 {
 820         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 821
 822         if (!test_bit(ob->dev, devs_may_alloc->d))
 823                 return false;
 824
 825         if (ob->data_type != wp->data_type)
 826                 return false;
 827
 828         if (!ca->mi.durability &&
 829             (wp->data_type == BCH_DATA_btree || ec || *have_cache))
 830                 return false;
 831
 832         if (ec != (ob->ec != NULL))
 833                 return false;
 834
 835         return true;
 836 }
 837
 838 static int bucket_alloc_set_writepoint(struct bch_fs *c,
 839                                        struct open_buckets *ptrs,
 840                                        struct write_point *wp,
 841                                        struct bch_devs_mask *devs_may_alloc,
 842                                        unsigned nr_replicas,
 843                                        unsigned *nr_effective,
 844                                        bool *have_cache,
 845                                        bool ec, unsigned flags)
 846 {
 847         struct open_buckets ptrs_skip = { .nr = 0 };
 848         struct open_bucket *ob;
 849         unsigned i;
 850         int ret = 0;
 851
 852         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 853                 if (!ret && want_bucket(c, wp, devs_may_alloc,
 854                                         have_cache, ec, ob))
 855                         ret = add_new_bucket(c, ptrs, devs_may_alloc,
 856                                        nr_replicas, nr_effective,
 857                                        have_cache, flags, ob);
 858                 else
 859                         ob_push(c, &ptrs_skip, ob);
 860         }
 861         wp->ptrs = ptrs_skip;
 862
 863         return ret;
 864 }
 865
 866 static int bucket_alloc_set_partial(struct bch_fs *c,
 867                                     struct open_buckets *ptrs,
 868                                     struct write_point *wp,
 869                                     struct bch_devs_mask *devs_may_alloc,
 870                                     unsigned nr_replicas,
 871                                     unsigned *nr_effective,
 872                                     bool *have_cache, bool ec,
 873                                     enum alloc_reserve reserve,
 874                                     unsigned flags)
 875 {
 876         int i, ret = 0;
 877
 878         if (!c->open_buckets_partial_nr)
 879                 return 0;
 880
 881         spin_lock(&c->freelist_lock);
 882
 883         if (!c->open_buckets_partial_nr)
 884                 goto unlock;
 885
 886         for (i = c->open_buckets_partial_nr - 1; i >= 0; --i) {
 887                 struct open_bucket *ob = c->open_buckets + c->open_buckets_partial[i];
 888
 889                 if (want_bucket(c, wp, devs_may_alloc, have_cache, ec, ob)) {
 890                         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 891                         struct bch_dev_usage usage;
 892                         u64 avail;
 893
 894                         bch2_dev_usage_read_fast(ca, &usage);
 895                         avail = dev_buckets_free(ca, usage, reserve);
 896                         if (!avail)
 897                                 continue;
 898
 899                         array_remove_item(c->open_buckets_partial,
 900                                           c->open_buckets_partial_nr,
 901                                           i);
 902                         ob->on_partial_list = false;
 903
 904                         ret = add_new_bucket(c, ptrs, devs_may_alloc,
 905                                              nr_replicas, nr_effective,
 906                                              have_cache, flags, ob);
 907                         if (ret)
 908                                 break;
 909                 }
 910         }
 911 unlock:
 912         spin_unlock(&c->freelist_lock);
 913         return ret;
 914 }
 915
 916 static int __open_bucket_add_buckets(struct btree_trans *trans,
 917                         struct open_buckets *ptrs,
 918                         struct write_point *wp,
 919                         struct bch_devs_list *devs_have,
 920                         u16 target,
 921                         bool erasure_code,
 922                         unsigned nr_replicas,
 923                         unsigned *nr_effective,
 924                         bool *have_cache,
 925                         enum alloc_reserve reserve,
 926                         unsigned flags,
 927                         struct closure *_cl)
 928 {
 929         struct bch_fs *c = trans->c;
 930         struct bch_devs_mask devs;
 931         struct open_bucket *ob;
 932         struct closure *cl = NULL;
 933         unsigned i;
 934         int ret;
 935
 936         rcu_read_lock();
 937         devs = target_rw_devs(c, wp->data_type, target);
 938         rcu_read_unlock();
 939
 940         /* Don't allocate from devices we already have pointers to: */
 941         for (i = 0; i < devs_have->nr; i++)
 942                 __clear_bit(devs_have->devs[i], devs.d);
 943
 944         open_bucket_for_each(c, ptrs, ob, i)
 945                 __clear_bit(ob->dev, devs.d);
 946
 947         if (erasure_code && ec_open_bucket(c, ptrs))
 948                 return 0;
 949
 950         ret = bucket_alloc_set_writepoint(c, ptrs, wp, &devs,
 951                                  nr_replicas, nr_effective,
 952                                  have_cache, erasure_code, flags);
 953         if (ret)
 954                 return ret;
 955
 956         ret = bucket_alloc_set_partial(c, ptrs, wp, &devs,
 957                                  nr_replicas, nr_effective,
 958                                  have_cache, erasure_code, reserve, flags);
 959         if (ret)
 960                 return ret;
 961
 962         if (erasure_code) {
 963                 ret = bucket_alloc_from_stripe(trans, ptrs, wp, &devs,
 964                                          target,
 965                                          nr_replicas, nr_effective,
 966                                          have_cache,
 967                                          reserve, flags, _cl);
 968         } else {
 969 retry_blocking:
 970                 /*
 971                  * Try nonblocking first, so that if one device is full we'll try from
 972                  * other devices:
 973                  */
 974                 ret = bch2_bucket_alloc_set_trans(trans, ptrs, &wp->stripe, &devs,
 975                                         nr_replicas, nr_effective, have_cache,
 976                                         flags, wp->data_type, reserve, cl);
 977                 if (ret &&
 978                     !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
 979                     !bch2_err_matches(ret, BCH_ERR_insufficient_devices) &&
 980                     !cl && _cl) {
 981                         cl = _cl;
 982                         goto retry_blocking;
 983                 }
 984
 985         }
 986
 987         return ret;
 988 }
 989
 990 static int open_bucket_add_buckets(struct btree_trans *trans,
 991                         struct open_buckets *ptrs,
 992                         struct write_point *wp,
 993                         struct bch_devs_list *devs_have,
 994                         u16 target,
 995                         unsigned erasure_code,
 996                         unsigned nr_replicas,
 997                         unsigned *nr_effective,
 998                         bool *have_cache,
 999                         enum alloc_reserve reserve,
1000                         unsigned flags,
1001                         struct closure *cl)
1002 {
1003         int ret;
1004
1005         if (erasure_code) {
1006                 ret = __open_bucket_add_buckets(trans, ptrs, wp,
1007                                 devs_have, target, erasure_code,
1008                                 nr_replicas, nr_effective, have_cache,
1009                                 reserve, flags, cl);
1010                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1011                     bch2_err_matches(ret, BCH_ERR_operation_blocked) ||
1012                     bch2_err_matches(ret, BCH_ERR_freelist_empty) ||
1013                     bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
1014                         return ret;
1015                 if (*nr_effective >= nr_replicas)
1016                         return 0;
1017         }
1018
1019         ret = __open_bucket_add_buckets(trans, ptrs, wp,
1020                         devs_have, target, false,
1021                         nr_replicas, nr_effective, have_cache,
1022                         reserve, flags, cl);
1023         return ret < 0 ? ret : 0;
1024 }
1025
1026 static bool should_drop_bucket(struct open_bucket *ob, struct bch_fs *c,
1027                                struct bch_dev *ca, bool ec)
1028 {
1029         if (ec) {
1030                 return ob->ec != NULL;
1031         } else if (ca) {
1032                 bool drop = ob->dev == ca->dev_idx;
1033                 struct open_bucket *ob2;
1034                 unsigned i;
1035
1036                 if (!drop && ob->ec) {
1037                         mutex_lock(&ob->ec->lock);
1038                         for (i = 0; i < ob->ec->new_stripe.key.v.nr_blocks; i++) {
1039                                 if (!ob->ec->blocks[i])
1040                                         continue;
1041
1042                                 ob2 = c->open_buckets + ob->ec->blocks[i];
1043                                 drop |= ob2->dev == ca->dev_idx;
1044                         }
1045                         mutex_unlock(&ob->ec->lock);
1046                 }
1047
1048                 return drop;
1049         } else {
1050                 return true;
1051         }
1052 }
1053
1054 static void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
1055                                  bool ec, struct write_point *wp)
1056 {
1057         struct open_buckets ptrs = { .nr = 0 };
1058         struct open_bucket *ob;
1059         unsigned i;
1060
1061         mutex_lock(&wp->lock);
1062         open_bucket_for_each(c, &wp->ptrs, ob, i)
1063                 if (should_drop_bucket(ob, c, ca, ec))
1064                         bch2_open_bucket_put(c, ob);
1065                 else
1066                         ob_push(c, &ptrs, ob);
1067         wp->ptrs = ptrs;
1068         mutex_unlock(&wp->lock);
1069 }
1070
1071 void bch2_open_buckets_stop(struct bch_fs *c, struct bch_dev *ca,
1072                             bool ec)
1073 {
1074         unsigned i;
1075
1076         /* Next, close write points that point to this device... */
1077         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1078                 bch2_writepoint_stop(c, ca, ec, &c->write_points[i]);
1079
1080         bch2_writepoint_stop(c, ca, ec, &c->copygc_write_point);
1081         bch2_writepoint_stop(c, ca, ec, &c->rebalance_write_point);
1082         bch2_writepoint_stop(c, ca, ec, &c->btree_write_point);
1083
1084         mutex_lock(&c->btree_reserve_cache_lock);
1085         while (c->btree_reserve_cache_nr) {
1086                 struct btree_alloc *a =
1087                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1088
1089                 bch2_open_buckets_put(c, &a->ob);
1090         }
1091         mutex_unlock(&c->btree_reserve_cache_lock);
1092
1093         spin_lock(&c->freelist_lock);
1094         i = 0;
1095         while (i < c->open_buckets_partial_nr) {
1096                 struct open_bucket *ob =
1097                         c->open_buckets + c->open_buckets_partial[i];
1098
1099                 if (should_drop_bucket(ob, c, ca, ec)) {
1100                         --c->open_buckets_partial_nr;
1101                         swap(c->open_buckets_partial[i],
1102                              c->open_buckets_partial[c->open_buckets_partial_nr]);
1103                         ob->on_partial_list = false;
1104                         spin_unlock(&c->freelist_lock);
1105                         bch2_open_bucket_put(c, ob);
1106                         spin_lock(&c->freelist_lock);
1107                 } else {
1108                         i++;
1109                 }
1110         }
1111         spin_unlock(&c->freelist_lock);
1112
1113         bch2_ec_stop_dev(c, ca);
1114 }
1115
1116 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
1117                                                  unsigned long write_point)
1118 {
1119         unsigned hash =
1120                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
1121
1122         return &c->write_points_hash[hash];
1123 }
1124
1125 static struct write_point *__writepoint_find(struct hlist_head *head,
1126                                              unsigned long write_point)
1127 {
1128         struct write_point *wp;
1129
1130         rcu_read_lock();
1131         hlist_for_each_entry_rcu(wp, head, node)
1132                 if (wp->write_point == write_point)
1133                         goto out;
1134         wp = NULL;
1135 out:
1136         rcu_read_unlock();
1137         return wp;
1138 }
1139
1140 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
1141 {
1142         u64 stranded    = c->write_points_nr * c->bucket_size_max;
1143         u64 free        = bch2_fs_usage_read_short(c).free;
1144
1145         return stranded * factor > free;
1146 }
1147
1148 static bool try_increase_writepoints(struct bch_fs *c)
1149 {
1150         struct write_point *wp;
1151
1152         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
1153             too_many_writepoints(c, 32))
1154                 return false;
1155
1156         wp = c->write_points + c->write_points_nr++;
1157         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
1158         return true;
1159 }
1160
1161 static bool try_decrease_writepoints(struct bch_fs *c, unsigned old_nr)
1162 {
1163         struct write_point *wp;
1164
1165         mutex_lock(&c->write_points_hash_lock);
1166         if (c->write_points_nr < old_nr) {
1167                 mutex_unlock(&c->write_points_hash_lock);
1168                 return true;
1169         }
1170
1171         if (c->write_points_nr == 1 ||
1172             !too_many_writepoints(c, 8)) {
1173                 mutex_unlock(&c->write_points_hash_lock);
1174                 return false;
1175         }
1176
1177         wp = c->write_points + --c->write_points_nr;
1178
1179         hlist_del_rcu(&wp->node);
1180         mutex_unlock(&c->write_points_hash_lock);
1181
1182         bch2_writepoint_stop(c, NULL, false, wp);
1183         return true;
1184 }
1185
1186 static void bch2_trans_mutex_lock_norelock(struct btree_trans *trans,
1187                                   struct mutex *lock)
1188 {
1189         if (!mutex_trylock(lock)) {
1190                 bch2_trans_unlock(trans);
1191                 mutex_lock(lock);
1192         }
1193 }
1194
1195 static struct write_point *writepoint_find(struct btree_trans *trans,
1196                                            unsigned long write_point)
1197 {
1198         struct bch_fs *c = trans->c;
1199         struct write_point *wp, *oldest;
1200         struct hlist_head *head;
1201
1202         if (!(write_point & 1UL)) {
1203                 wp = (struct write_point *) write_point;
1204                 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1205                 return wp;
1206         }
1207
1208         head = writepoint_hash(c, write_point);
1209 restart_find:
1210         wp = __writepoint_find(head, write_point);
1211         if (wp) {
1212 lock_wp:
1213                 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1214                 if (wp->write_point == write_point)
1215                         goto out;
1216                 mutex_unlock(&wp->lock);
1217                 goto restart_find;
1218         }
1219 restart_find_oldest:
1220         oldest = NULL;
1221         for (wp = c->write_points;
1222              wp < c->write_points + c->write_points_nr; wp++)
1223                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1224                         oldest = wp;
1225
1226         bch2_trans_mutex_lock_norelock(trans, &oldest->lock);
1227         bch2_trans_mutex_lock_norelock(trans, &c->write_points_hash_lock);
1228         if (oldest >= c->write_points + c->write_points_nr ||
1229             try_increase_writepoints(c)) {
1230                 mutex_unlock(&c->write_points_hash_lock);
1231                 mutex_unlock(&oldest->lock);
1232                 goto restart_find_oldest;
1233         }
1234
1235         wp = __writepoint_find(head, write_point);
1236         if (wp && wp != oldest) {
1237                 mutex_unlock(&c->write_points_hash_lock);
1238                 mutex_unlock(&oldest->lock);
1239                 goto lock_wp;
1240         }
1241
1242         wp = oldest;
1243         hlist_del_rcu(&wp->node);
1244         wp->write_point = write_point;
1245         hlist_add_head_rcu(&wp->node, head);
1246         mutex_unlock(&c->write_points_hash_lock);
1247 out:
1248         wp->last_used = local_clock();
1249         return wp;
1250 }
1251
1252 /*
1253  * Get us an open_bucket we can allocate from, return with it locked:
1254  */
1255 int bch2_alloc_sectors_start_trans(struct btree_trans *trans,
1256                              unsigned target,
1257                              unsigned erasure_code,
1258                              struct write_point_specifier write_point,
1259                              struct bch_devs_list *devs_have,
1260                              unsigned nr_replicas,
1261                              unsigned nr_replicas_required,
1262                              enum alloc_reserve reserve,
1263                              unsigned flags,
1264                              struct closure *cl,
1265                              struct write_point **wp_ret)
1266 {
1267         struct bch_fs *c = trans->c;
1268         struct write_point *wp;
1269         struct open_bucket *ob;
1270         struct open_buckets ptrs;
1271         unsigned nr_effective, write_points_nr;
1272         bool have_cache;
1273         int ret;
1274         int i;
1275
1276         BUG_ON(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS);
1277
1278         BUG_ON(!nr_replicas || !nr_replicas_required);
1279 retry:
1280         ptrs.nr         = 0;
1281         nr_effective    = 0;
1282         write_points_nr = c->write_points_nr;
1283         have_cache      = false;
1284
1285         *wp_ret = wp = writepoint_find(trans, write_point.v);
1286
1287         /* metadata may not allocate on cache devices: */
1288         if (wp->data_type != BCH_DATA_user)
1289                 have_cache = true;
1290
1291         if (target && !(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1292                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1293                                               target, erasure_code,
1294                                               nr_replicas, &nr_effective,
1295                                               &have_cache, reserve,
1296                                               flags, NULL);
1297                 if (!ret ||
1298                     bch2_err_matches(ret, BCH_ERR_transaction_restart))
1299                         goto alloc_done;
1300
1301                 /* Don't retry from all devices if we're out of open buckets: */
1302                 if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
1303                         goto allocate_blocking;
1304
1305                 /*
1306                  * Only try to allocate cache (durability = 0 devices) from the
1307                  * specified target:
1308                  */
1309                 have_cache = true;
1310
1311                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1312                                               0, erasure_code,
1313                                               nr_replicas, &nr_effective,
1314                                               &have_cache, reserve,
1315                                               flags, cl);
1316         } else {
1317 allocate_blocking:
1318                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1319                                               target, erasure_code,
1320                                               nr_replicas, &nr_effective,
1321                                               &have_cache, reserve,
1322                                               flags, cl);
1323         }
1324 alloc_done:
1325         BUG_ON(!ret && nr_effective < nr_replicas);
1326
1327         if (erasure_code && !ec_open_bucket(c, &ptrs))
1328                 pr_debug("failed to get ec bucket: ret %u", ret);
1329
1330         if (ret == -BCH_ERR_insufficient_devices &&
1331             nr_effective >= nr_replicas_required)
1332                 ret = 0;
1333
1334         if (ret)
1335                 goto err;
1336
1337         /* Free buckets we didn't use: */
1338         open_bucket_for_each(c, &wp->ptrs, ob, i)
1339                 open_bucket_free_unused(c, wp, ob);
1340
1341         wp->ptrs = ptrs;
1342
1343         wp->sectors_free = UINT_MAX;
1344
1345         open_bucket_for_each(c, &wp->ptrs, ob, i)
1346                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1347
1348         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1349
1350         return 0;
1351 err:
1352         open_bucket_for_each(c, &wp->ptrs, ob, i)
1353                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1354                         ob_push(c, &ptrs, ob);
1355                 else
1356                         open_bucket_free_unused(c, wp, ob);
1357         wp->ptrs = ptrs;
1358
1359         mutex_unlock(&wp->lock);
1360
1361         if (bch2_err_matches(ret, BCH_ERR_freelist_empty) &&
1362             try_decrease_writepoints(c, write_points_nr))
1363                 goto retry;
1364
1365         if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty) ||
1366             bch2_err_matches(ret, BCH_ERR_freelist_empty))
1367                 return cl
1368                         ? -BCH_ERR_bucket_alloc_blocked
1369                         : -BCH_ERR_ENOSPC_bucket_alloc;
1370
1371         return ret;
1372 }
1373
1374 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1375 {
1376         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1377
1378         return (struct bch_extent_ptr) {
1379                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
1380                 .gen    = ob->gen,
1381                 .dev    = ob->dev,
1382                 .offset = bucket_to_sector(ca, ob->bucket) +
1383                         ca->mi.bucket_size -
1384                         ob->sectors_free,
1385         };
1386 }
1387
1388 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1389                                     struct bkey_i *k, unsigned sectors,
1390                                     bool cached)
1391 {
1392         bch2_alloc_sectors_append_ptrs_inlined(c, wp, k, sectors, cached);
1393 }
1394
1395 /*
1396  * Append pointers to the space we just allocated to @k, and mark @sectors space
1397  * as allocated out of @ob
1398  */
1399 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1400 {
1401         bch2_alloc_sectors_done_inlined(c, wp);
1402 }
1403
1404 static inline void writepoint_init(struct write_point *wp,
1405                                    enum bch_data_type type)
1406 {
1407         mutex_init(&wp->lock);
1408         wp->data_type = type;
1409
1410         INIT_WORK(&wp->index_update_work, bch2_write_point_do_index_updates);
1411         INIT_LIST_HEAD(&wp->writes);
1412         spin_lock_init(&wp->writes_lock);
1413 }
1414
1415 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1416 {
1417         struct open_bucket *ob;
1418         struct write_point *wp;
1419
1420         mutex_init(&c->write_points_hash_lock);
1421         c->write_points_nr = ARRAY_SIZE(c->write_points);
1422
1423         /* open bucket 0 is a sentinal NULL: */
1424         spin_lock_init(&c->open_buckets[0].lock);
1425
1426         for (ob = c->open_buckets + 1;
1427              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1428                 spin_lock_init(&ob->lock);
1429                 c->open_buckets_nr_free++;
1430
1431                 ob->freelist = c->open_buckets_freelist;
1432                 c->open_buckets_freelist = ob - c->open_buckets;
1433         }
1434
1435         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
1436         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
1437         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
1438
1439         for (wp = c->write_points;
1440              wp < c->write_points + c->write_points_nr; wp++) {
1441                 writepoint_init(wp, BCH_DATA_user);
1442
1443                 wp->last_used   = local_clock();
1444                 wp->write_point = (unsigned long) wp;
1445                 hlist_add_head_rcu(&wp->node,
1446                                    writepoint_hash(c, wp->write_point));
1447         }
1448 }
1449
1450 static void bch2_open_bucket_to_text(struct printbuf *out, struct bch_fs *c, struct open_bucket *ob)
1451 {
1452         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1453         unsigned data_type = ob->data_type;
1454         barrier(); /* READ_ONCE() doesn't work on bitfields */
1455
1456         prt_printf(out, "%zu ref %u %s %u:%llu gen %u allocated %u/%u",
1457                    ob - c->open_buckets,
1458                    atomic_read(&ob->pin),
1459                    data_type < BCH_DATA_NR ? bch2_data_types[data_type] : "invalid data type",
1460                    ob->dev, ob->bucket, ob->gen,
1461                    ca->mi.bucket_size - ob->sectors_free, ca->mi.bucket_size);
1462         if (ob->ec)
1463                 prt_printf(out, " ec idx %llu", ob->ec->idx);
1464         if (ob->on_partial_list)
1465                 prt_str(out, " partial");
1466         prt_newline(out);
1467 }
1468
1469 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1470 {
1471         struct open_bucket *ob;
1472
1473         out->atomic++;
1474
1475         for (ob = c->open_buckets;
1476              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1477              ob++) {
1478                 spin_lock(&ob->lock);
1479                 if (ob->valid && !ob->on_partial_list)
1480                         bch2_open_bucket_to_text(out, c, ob);
1481                 spin_unlock(&ob->lock);
1482         }
1483
1484         --out->atomic;
1485 }
1486
1487 void bch2_open_buckets_partial_to_text(struct printbuf *out, struct bch_fs *c)
1488 {
1489         unsigned i;
1490
1491         out->atomic++;
1492         spin_lock(&c->freelist_lock);
1493
1494         for (i = 0; i < c->open_buckets_partial_nr; i++)
1495                 bch2_open_bucket_to_text(out, c,
1496                                 c->open_buckets + c->open_buckets_partial[i]);
1497
1498         spin_unlock(&c->freelist_lock);
1499         --out->atomic;
1500 }
1501
1502 static const char * const bch2_write_point_states[] = {
1503 #define x(n)    #n,
1504         WRITE_POINT_STATES()
1505 #undef x
1506         NULL
1507 };
1508
1509 void bch2_write_points_to_text(struct printbuf *out, struct bch_fs *c)
1510 {
1511         struct write_point *wp;
1512         unsigned i;
1513
1514         for (wp = c->write_points;
1515              wp < c->write_points + ARRAY_SIZE(c->write_points);
1516              wp++) {
1517                 prt_printf(out, "%lu: ", wp->write_point);
1518                 prt_human_readable_u64(out, wp->sectors_allocated);
1519
1520                 prt_printf(out, " last wrote: ");
1521                 bch2_pr_time_units(out, sched_clock() - wp->last_used);
1522
1523                 for (i = 0; i < WRITE_POINT_STATE_NR; i++) {
1524                         prt_printf(out, " %s: ", bch2_write_point_states[i]);
1525                         bch2_pr_time_units(out, wp->time[i]);
1526                 }
1527
1528                 prt_newline(out);
1529         }
1530 }