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 static void bch2_trans_mutex_lock_norelock(struct btree_trans *trans,
  39                                            struct mutex *lock)
  40 {
  41         if (!mutex_trylock(lock)) {
  42                 bch2_trans_unlock(trans);
  43                 mutex_lock(lock);
  44         }
  45 }
  46
  47 const char * const bch2_watermarks[] = {
  48 #define x(t) #t,
  49         BCH_WATERMARKS()
  50 #undef x
  51         NULL
  52 };
  53
  54 /*
  55  * Open buckets represent a bucket that's currently being allocated from.  They
  56  * serve two purposes:
  57  *
  58  *  - They track buckets that have been partially allocated, allowing for
  59  *    sub-bucket sized allocations - they're used by the sector allocator below
  60  *
  61  *  - They provide a reference to the buckets they own that mark and sweep GC
  62  *    can find, until the new allocation has a pointer to it inserted into the
  63  *    btree
  64  *
  65  * When allocating some space with the sector allocator, the allocation comes
  66  * with a reference to an open bucket - the caller is required to put that
  67  * reference _after_ doing the index update that makes its allocation reachable.
  68  */
  69
  70 void bch2_reset_alloc_cursors(struct bch_fs *c)
  71 {
  72         struct bch_dev *ca;
  73         unsigned i;
  74
  75         rcu_read_lock();
  76         for_each_member_device_rcu(ca, c, i, NULL)
  77                 ca->alloc_cursor = 0;
  78         rcu_read_unlock();
  79 }
  80
  81 static void bch2_open_bucket_hash_add(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         ob->hash = *slot;
  87         *slot = idx;
  88 }
  89
  90 static void bch2_open_bucket_hash_remove(struct bch_fs *c, struct open_bucket *ob)
  91 {
  92         open_bucket_idx_t idx = ob - c->open_buckets;
  93         open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
  94
  95         while (*slot != idx) {
  96                 BUG_ON(!*slot);
  97                 slot = &c->open_buckets[*slot].hash;
  98         }
  99
 100         *slot = ob->hash;
 101         ob->hash = 0;
 102 }
 103
 104 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
 105 {
 106         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 107
 108         if (ob->ec) {
 109                 ec_stripe_new_put(c, ob->ec, STRIPE_REF_io);
 110                 return;
 111         }
 112
 113         percpu_down_read(&c->mark_lock);
 114         spin_lock(&ob->lock);
 115
 116         ob->valid = false;
 117         ob->data_type = 0;
 118
 119         spin_unlock(&ob->lock);
 120         percpu_up_read(&c->mark_lock);
 121
 122         spin_lock(&c->freelist_lock);
 123         bch2_open_bucket_hash_remove(c, ob);
 124
 125         ob->freelist = c->open_buckets_freelist;
 126         c->open_buckets_freelist = ob - c->open_buckets;
 127
 128         c->open_buckets_nr_free++;
 129         ca->nr_open_buckets--;
 130         spin_unlock(&c->freelist_lock);
 131
 132         closure_wake_up(&c->open_buckets_wait);
 133 }
 134
 135 void bch2_open_bucket_write_error(struct bch_fs *c,
 136                                   struct open_buckets *obs,
 137                                   unsigned dev)
 138 {
 139         struct open_bucket *ob;
 140         unsigned i;
 141
 142         open_bucket_for_each(c, obs, ob, i)
 143                 if (ob->dev == dev && ob->ec)
 144                         bch2_ec_bucket_cancel(c, ob);
 145 }
 146
 147 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
 148 {
 149         struct open_bucket *ob;
 150
 151         BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
 152
 153         ob = c->open_buckets + c->open_buckets_freelist;
 154         c->open_buckets_freelist = ob->freelist;
 155         atomic_set(&ob->pin, 1);
 156         ob->data_type = 0;
 157
 158         c->open_buckets_nr_free--;
 159         return ob;
 160 }
 161
 162 static void open_bucket_free_unused(struct bch_fs *c, struct open_bucket *ob)
 163 {
 164         BUG_ON(c->open_buckets_partial_nr >=
 165                ARRAY_SIZE(c->open_buckets_partial));
 166
 167         spin_lock(&c->freelist_lock);
 168         ob->on_partial_list = true;
 169         c->open_buckets_partial[c->open_buckets_partial_nr++] =
 170                 ob - c->open_buckets;
 171         spin_unlock(&c->freelist_lock);
 172
 173         closure_wake_up(&c->open_buckets_wait);
 174         closure_wake_up(&c->freelist_wait);
 175 }
 176
 177 /* _only_ for allocating the journal on a new device: */
 178 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
 179 {
 180         while (ca->new_fs_bucket_idx < ca->mi.nbuckets) {
 181                 u64 b = ca->new_fs_bucket_idx++;
 182
 183                 if (!is_superblock_bucket(ca, b) &&
 184                     (!ca->buckets_nouse || !test_bit(b, ca->buckets_nouse)))
 185                         return b;
 186         }
 187
 188         return -1;
 189 }
 190
 191 static inline unsigned open_buckets_reserved(enum bch_watermark watermark)
 192 {
 193         switch (watermark) {
 194         case BCH_WATERMARK_reclaim:
 195                 return 0;
 196         case BCH_WATERMARK_btree:
 197         case BCH_WATERMARK_btree_copygc:
 198                 return OPEN_BUCKETS_COUNT / 4;
 199         case BCH_WATERMARK_copygc:
 200                 return OPEN_BUCKETS_COUNT / 3;
 201         default:
 202                 return OPEN_BUCKETS_COUNT / 2;
 203         }
 204 }
 205
 206 static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 207                                               u64 bucket,
 208                                               enum bch_watermark watermark,
 209                                               const struct bch_alloc_v4 *a,
 210                                               struct bucket_alloc_state *s,
 211                                               struct closure *cl)
 212 {
 213         struct open_bucket *ob;
 214
 215         if (unlikely(ca->buckets_nouse && test_bit(bucket, ca->buckets_nouse))) {
 216                 s->skipped_nouse++;
 217                 return NULL;
 218         }
 219
 220         if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
 221                 s->skipped_open++;
 222                 return NULL;
 223         }
 224
 225         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 226                         c->journal.flushed_seq_ondisk, ca->dev_idx, bucket)) {
 227                 s->skipped_need_journal_commit++;
 228                 return NULL;
 229         }
 230
 231         if (bch2_bucket_nocow_is_locked(&c->nocow_locks, POS(ca->dev_idx, bucket))) {
 232                 s->skipped_nocow++;
 233                 return NULL;
 234         }
 235
 236         spin_lock(&c->freelist_lock);
 237
 238         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(watermark))) {
 239                 if (cl)
 240                         closure_wait(&c->open_buckets_wait, cl);
 241
 242                 if (!c->blocked_allocate_open_bucket)
 243                         c->blocked_allocate_open_bucket = local_clock();
 244
 245                 spin_unlock(&c->freelist_lock);
 246                 return ERR_PTR(-BCH_ERR_open_buckets_empty);
 247         }
 248
 249         /* Recheck under lock: */
 250         if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
 251                 spin_unlock(&c->freelist_lock);
 252                 s->skipped_open++;
 253                 return NULL;
 254         }
 255
 256         ob = bch2_open_bucket_alloc(c);
 257
 258         spin_lock(&ob->lock);
 259
 260         ob->valid       = true;
 261         ob->sectors_free = ca->mi.bucket_size;
 262         ob->dev         = ca->dev_idx;
 263         ob->gen         = a->gen;
 264         ob->bucket      = bucket;
 265         spin_unlock(&ob->lock);
 266
 267         ca->nr_open_buckets++;
 268         bch2_open_bucket_hash_add(c, ob);
 269
 270         if (c->blocked_allocate_open_bucket) {
 271                 bch2_time_stats_update(
 272                         &c->times[BCH_TIME_blocked_allocate_open_bucket],
 273                         c->blocked_allocate_open_bucket);
 274                 c->blocked_allocate_open_bucket = 0;
 275         }
 276
 277         if (c->blocked_allocate) {
 278                 bch2_time_stats_update(
 279                         &c->times[BCH_TIME_blocked_allocate],
 280                         c->blocked_allocate);
 281                 c->blocked_allocate = 0;
 282         }
 283
 284         spin_unlock(&c->freelist_lock);
 285         return ob;
 286 }
 287
 288 static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bch_dev *ca,
 289                                             enum bch_watermark watermark, u64 free_entry,
 290                                             struct bucket_alloc_state *s,
 291                                             struct bkey_s_c freespace_k,
 292                                             struct closure *cl)
 293 {
 294         struct bch_fs *c = trans->c;
 295         struct btree_iter iter = { NULL };
 296         struct bkey_s_c k;
 297         struct open_bucket *ob;
 298         struct bch_alloc_v4 a_convert;
 299         const struct bch_alloc_v4 *a;
 300         u64 b = free_entry & ~(~0ULL << 56);
 301         unsigned genbits = free_entry >> 56;
 302         struct printbuf buf = PRINTBUF;
 303         int ret;
 304
 305         if (b < ca->mi.first_bucket || b >= ca->mi.nbuckets) {
 306                 prt_printf(&buf, "freespace btree has bucket outside allowed range %u-%llu\n"
 307                        "  freespace key ",
 308                         ca->mi.first_bucket, ca->mi.nbuckets);
 309                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 310                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 311                 ob = ERR_PTR(-EIO);
 312                 goto err;
 313         }
 314
 315         k = bch2_bkey_get_iter(trans, &iter,
 316                                BTREE_ID_alloc, POS(ca->dev_idx, b),
 317                                BTREE_ITER_CACHED);
 318         ret = bkey_err(k);
 319         if (ret) {
 320                 ob = ERR_PTR(ret);
 321                 goto err;
 322         }
 323
 324         a = bch2_alloc_to_v4(k, &a_convert);
 325
 326         if (a->data_type != BCH_DATA_free) {
 327                 if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_alloc_info) {
 328                         ob = NULL;
 329                         goto err;
 330                 }
 331
 332                 prt_printf(&buf, "non free bucket in freespace btree\n"
 333                        "  freespace key ");
 334                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 335                 prt_printf(&buf, "\n  ");
 336                 bch2_bkey_val_to_text(&buf, c, k);
 337                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 338                 ob = ERR_PTR(-EIO);
 339                 goto err;
 340         }
 341
 342         if (genbits != (alloc_freespace_genbits(*a) >> 56) &&
 343             c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
 344                 prt_printf(&buf, "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
 345                        "  freespace key ",
 346                        genbits, alloc_freespace_genbits(*a) >> 56);
 347                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 348                 prt_printf(&buf, "\n  ");
 349                 bch2_bkey_val_to_text(&buf, c, k);
 350                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 351                 ob = ERR_PTR(-EIO);
 352                 goto err;
 353         }
 354
 355         if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_extents_to_backpointers) {
 356                 struct bch_backpointer bp;
 357                 struct bpos bp_pos = POS_MIN;
 358
 359                 ret = bch2_get_next_backpointer(trans, POS(ca->dev_idx, b), -1,
 360                                                 &bp_pos, &bp,
 361                                                 BTREE_ITER_NOPRESERVE);
 362                 if (ret) {
 363                         ob = ERR_PTR(ret);
 364                         goto err;
 365                 }
 366
 367                 if (!bkey_eq(bp_pos, POS_MAX)) {
 368                         /*
 369                          * Bucket may have data in it - we don't call
 370                          * bc2h_trans_inconnsistent() because fsck hasn't
 371                          * finished yet
 372                          */
 373                         ob = NULL;
 374                         goto err;
 375                 }
 376         }
 377
 378         ob = __try_alloc_bucket(c, ca, b, watermark, a, s, cl);
 379         if (!ob)
 380                 iter.path->preserve = false;
 381 err:
 382         if (iter.trans && iter.path)
 383                 set_btree_iter_dontneed(&iter);
 384         bch2_trans_iter_exit(trans, &iter);
 385         printbuf_exit(&buf);
 386         return ob;
 387 }
 388
 389 /*
 390  * This path is for before the freespace btree is initialized:
 391  *
 392  * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
 393  * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
 394  */
 395 static noinline struct open_bucket *
 396 bch2_bucket_alloc_early(struct btree_trans *trans,
 397                         struct bch_dev *ca,
 398                         enum bch_watermark watermark,
 399                         struct bucket_alloc_state *s,
 400                         struct closure *cl)
 401 {
 402         struct btree_iter iter;
 403         struct bkey_s_c k;
 404         struct open_bucket *ob = NULL;
 405         u64 alloc_start = max_t(u64, ca->mi.first_bucket, ca->new_fs_bucket_idx);
 406         u64 alloc_cursor = max(alloc_start, READ_ONCE(ca->alloc_cursor));
 407         int ret;
 408 again:
 409         for_each_btree_key_norestart(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, alloc_cursor),
 410                            BTREE_ITER_SLOTS, k, ret) {
 411                 struct bch_alloc_v4 a_convert;
 412                 const struct bch_alloc_v4 *a;
 413
 414                 if (bkey_ge(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)))
 415                         break;
 416
 417                 if (ca->new_fs_bucket_idx &&
 418                     is_superblock_bucket(ca, k.k->p.offset))
 419                         continue;
 420
 421                 a = bch2_alloc_to_v4(k, &a_convert);
 422
 423                 if (a->data_type != BCH_DATA_free)
 424                         continue;
 425
 426                 s->buckets_seen++;
 427
 428                 ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, watermark, a, s, cl);
 429                 if (ob)
 430                         break;
 431         }
 432         bch2_trans_iter_exit(trans, &iter);
 433
 434         ca->alloc_cursor = alloc_cursor;
 435
 436         if (!ob && ret)
 437                 ob = ERR_PTR(ret);
 438
 439         if (!ob && alloc_cursor > alloc_start) {
 440                 alloc_cursor = alloc_start;
 441                 goto again;
 442         }
 443
 444         return ob;
 445 }
 446
 447 static struct open_bucket *bch2_bucket_alloc_freelist(struct btree_trans *trans,
 448                                                    struct bch_dev *ca,
 449                                                    enum bch_watermark watermark,
 450                                                    struct bucket_alloc_state *s,
 451                                                    struct closure *cl)
 452 {
 453         struct btree_iter iter;
 454         struct bkey_s_c k;
 455         struct open_bucket *ob = NULL;
 456         u64 alloc_start = max_t(u64, ca->mi.first_bucket, READ_ONCE(ca->alloc_cursor));
 457         u64 alloc_cursor = alloc_start;
 458         int ret;
 459
 460         BUG_ON(ca->new_fs_bucket_idx);
 461 again:
 462         for_each_btree_key_norestart(trans, iter, BTREE_ID_freespace,
 463                                      POS(ca->dev_idx, alloc_cursor), 0, k, ret) {
 464                 if (k.k->p.inode != ca->dev_idx)
 465                         break;
 466
 467                 for (alloc_cursor = max(alloc_cursor, bkey_start_offset(k.k));
 468                      alloc_cursor < k.k->p.offset;
 469                      alloc_cursor++) {
 470                         ret = btree_trans_too_many_iters(trans);
 471                         if (ret) {
 472                                 ob = ERR_PTR(ret);
 473                                 break;
 474                         }
 475
 476                         s->buckets_seen++;
 477
 478                         ob = try_alloc_bucket(trans, ca, watermark,
 479                                               alloc_cursor, s, k, cl);
 480                         if (ob) {
 481                                 iter.path->preserve = false;
 482                                 break;
 483                         }
 484                 }
 485
 486                 if (ob || ret)
 487                         break;
 488         }
 489         bch2_trans_iter_exit(trans, &iter);
 490
 491         ca->alloc_cursor = alloc_cursor;
 492
 493         if (!ob && ret)
 494                 ob = ERR_PTR(ret);
 495
 496         if (!ob && alloc_start > ca->mi.first_bucket) {
 497                 alloc_cursor = alloc_start = ca->mi.first_bucket;
 498                 goto again;
 499         }
 500
 501         return ob;
 502 }
 503
 504 /**
 505  * bch_bucket_alloc - allocate a single bucket from a specific device
 506  *
 507  * Returns index of bucket on success, 0 on failure
 508  */
 509 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
 510                                       struct bch_dev *ca,
 511                                       enum bch_watermark watermark,
 512                                       struct closure *cl,
 513                                       struct bch_dev_usage *usage)
 514 {
 515         struct bch_fs *c = trans->c;
 516         struct open_bucket *ob = NULL;
 517         bool freespace = READ_ONCE(ca->mi.freespace_initialized);
 518         u64 avail;
 519         struct bucket_alloc_state s = { 0 };
 520         bool waiting = false;
 521 again:
 522         bch2_dev_usage_read_fast(ca, usage);
 523         avail = dev_buckets_free(ca, *usage, watermark);
 524
 525         if (usage->d[BCH_DATA_need_discard].buckets > avail)
 526                 bch2_do_discards(c);
 527
 528         if (usage->d[BCH_DATA_need_gc_gens].buckets > avail)
 529                 bch2_do_gc_gens(c);
 530
 531         if (should_invalidate_buckets(ca, *usage))
 532                 bch2_do_invalidates(c);
 533
 534         if (!avail) {
 535                 if (cl && !waiting) {
 536                         closure_wait(&c->freelist_wait, cl);
 537                         waiting = true;
 538                         goto again;
 539                 }
 540
 541                 if (!c->blocked_allocate)
 542                         c->blocked_allocate = local_clock();
 543
 544                 ob = ERR_PTR(-BCH_ERR_freelist_empty);
 545                 goto err;
 546         }
 547
 548         if (waiting)
 549                 closure_wake_up(&c->freelist_wait);
 550 alloc:
 551         ob = likely(freespace)
 552                 ? bch2_bucket_alloc_freelist(trans, ca, watermark, &s, cl)
 553                 : bch2_bucket_alloc_early(trans, ca, watermark, &s, cl);
 554
 555         if (s.skipped_need_journal_commit * 2 > avail)
 556                 bch2_journal_flush_async(&c->journal, NULL);
 557
 558         if (!ob && freespace && c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_alloc_info) {
 559                 freespace = false;
 560                 goto alloc;
 561         }
 562 err:
 563         if (!ob)
 564                 ob = ERR_PTR(-BCH_ERR_no_buckets_found);
 565
 566         if (!IS_ERR(ob))
 567                 trace_and_count(c, bucket_alloc, ca,
 568                                 bch2_watermarks[watermark],
 569                                 ob->bucket,
 570                                 usage->d[BCH_DATA_free].buckets,
 571                                 avail,
 572                                 bch2_copygc_wait_amount(c),
 573                                 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
 574                                 &s,
 575                                 cl == NULL,
 576                                 "");
 577         else if (!bch2_err_matches(PTR_ERR(ob), BCH_ERR_transaction_restart))
 578                 trace_and_count(c, bucket_alloc_fail, ca,
 579                                 bch2_watermarks[watermark],
 580                                 0,
 581                                 usage->d[BCH_DATA_free].buckets,
 582                                 avail,
 583                                 bch2_copygc_wait_amount(c),
 584                                 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
 585                                 &s,
 586                                 cl == NULL,
 587                                 bch2_err_str(PTR_ERR(ob)));
 588
 589         return ob;
 590 }
 591
 592 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 593                                       enum bch_watermark watermark,
 594                                       struct closure *cl)
 595 {
 596         struct bch_dev_usage usage;
 597         struct open_bucket *ob;
 598
 599         bch2_trans_do(c, NULL, NULL, 0,
 600                       PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, watermark,
 601                                                         cl, &usage)));
 602         return ob;
 603 }
 604
 605 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 606                             unsigned l, unsigned r)
 607 {
 608         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 609                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 610 }
 611
 612 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 613
 614 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 615                                           struct dev_stripe_state *stripe,
 616                                           struct bch_devs_mask *devs)
 617 {
 618         struct dev_alloc_list ret = { .nr = 0 };
 619         unsigned i;
 620
 621         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 622                 ret.devs[ret.nr++] = i;
 623
 624         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 625         return ret;
 626 }
 627
 628 static inline void bch2_dev_stripe_increment_inlined(struct bch_dev *ca,
 629                                struct dev_stripe_state *stripe,
 630                                struct bch_dev_usage *usage)
 631 {
 632         u64 *v = stripe->next_alloc + ca->dev_idx;
 633         u64 free_space = dev_buckets_available(ca, BCH_WATERMARK_normal);
 634         u64 free_space_inv = free_space
 635                 ? div64_u64(1ULL << 48, free_space)
 636                 : 1ULL << 48;
 637         u64 scale = *v / 4;
 638
 639         if (*v + free_space_inv >= *v)
 640                 *v += free_space_inv;
 641         else
 642                 *v = U64_MAX;
 643
 644         for (v = stripe->next_alloc;
 645              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 646                 *v = *v < scale ? 0 : *v - scale;
 647 }
 648
 649 void bch2_dev_stripe_increment(struct bch_dev *ca,
 650                                struct dev_stripe_state *stripe)
 651 {
 652         struct bch_dev_usage usage;
 653
 654         bch2_dev_usage_read_fast(ca, &usage);
 655         bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
 656 }
 657
 658 static int add_new_bucket(struct bch_fs *c,
 659                            struct open_buckets *ptrs,
 660                            struct bch_devs_mask *devs_may_alloc,
 661                            unsigned nr_replicas,
 662                            unsigned *nr_effective,
 663                            bool *have_cache,
 664                            unsigned flags,
 665                            struct open_bucket *ob)
 666 {
 667         unsigned durability =
 668                 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
 669
 670         BUG_ON(*nr_effective >= nr_replicas);
 671         BUG_ON(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS);
 672
 673         __clear_bit(ob->dev, devs_may_alloc->d);
 674         *nr_effective   += (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)
 675                 ? durability : 1;
 676         *have_cache     |= !durability;
 677
 678         ob_push(c, ptrs, ob);
 679
 680         if (*nr_effective >= nr_replicas)
 681                 return 1;
 682         if (ob->ec)
 683                 return 1;
 684         return 0;
 685 }
 686
 687 int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
 688                       struct open_buckets *ptrs,
 689                       struct dev_stripe_state *stripe,
 690                       struct bch_devs_mask *devs_may_alloc,
 691                       unsigned nr_replicas,
 692                       unsigned *nr_effective,
 693                       bool *have_cache,
 694                       unsigned flags,
 695                       enum bch_data_type data_type,
 696                       enum bch_watermark watermark,
 697                       struct closure *cl)
 698 {
 699         struct bch_fs *c = trans->c;
 700         struct dev_alloc_list devs_sorted =
 701                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 702         unsigned dev;
 703         struct bch_dev *ca;
 704         int ret = -BCH_ERR_insufficient_devices;
 705         unsigned i;
 706
 707         BUG_ON(*nr_effective >= nr_replicas);
 708
 709         for (i = 0; i < devs_sorted.nr; i++) {
 710                 struct bch_dev_usage usage;
 711                 struct open_bucket *ob;
 712
 713                 dev = devs_sorted.devs[i];
 714
 715                 rcu_read_lock();
 716                 ca = rcu_dereference(c->devs[dev]);
 717                 if (ca)
 718                         percpu_ref_get(&ca->ref);
 719                 rcu_read_unlock();
 720
 721                 if (!ca)
 722                         continue;
 723
 724                 if (!ca->mi.durability && *have_cache) {
 725                         percpu_ref_put(&ca->ref);
 726                         continue;
 727                 }
 728
 729                 ob = bch2_bucket_alloc_trans(trans, ca, watermark, cl, &usage);
 730                 if (!IS_ERR(ob))
 731                         bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
 732                 percpu_ref_put(&ca->ref);
 733
 734                 if (IS_ERR(ob)) {
 735                         ret = PTR_ERR(ob);
 736                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) || cl)
 737                                 break;
 738                         continue;
 739                 }
 740
 741                 ob->data_type = data_type;
 742
 743                 if (add_new_bucket(c, ptrs, devs_may_alloc,
 744                                    nr_replicas, nr_effective,
 745                                    have_cache, flags, ob)) {
 746                         ret = 0;
 747                         break;
 748                 }
 749         }
 750
 751         return ret;
 752 }
 753
 754 /* Allocate from stripes: */
 755
 756 /*
 757  * if we can't allocate a new stripe because there are already too many
 758  * partially filled stripes, force allocating from an existing stripe even when
 759  * it's to a device we don't want:
 760  */
 761
 762 static int bucket_alloc_from_stripe(struct btree_trans *trans,
 763                          struct open_buckets *ptrs,
 764                          struct write_point *wp,
 765                          struct bch_devs_mask *devs_may_alloc,
 766                          u16 target,
 767                          unsigned nr_replicas,
 768                          unsigned *nr_effective,
 769                          bool *have_cache,
 770                          enum bch_watermark watermark,
 771                          unsigned flags,
 772                          struct closure *cl)
 773 {
 774         struct bch_fs *c = trans->c;
 775         struct dev_alloc_list devs_sorted;
 776         struct ec_stripe_head *h;
 777         struct open_bucket *ob;
 778         struct bch_dev *ca;
 779         unsigned i, ec_idx;
 780         int ret = 0;
 781
 782         if (nr_replicas < 2)
 783                 return 0;
 784
 785         if (ec_open_bucket(c, ptrs))
 786                 return 0;
 787
 788         h = bch2_ec_stripe_head_get(trans, target, 0, nr_replicas - 1, watermark, cl);
 789         if (IS_ERR(h))
 790                 return PTR_ERR(h);
 791         if (!h)
 792                 return 0;
 793
 794         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 795
 796         for (i = 0; i < devs_sorted.nr; i++)
 797                 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 798                         if (!h->s->blocks[ec_idx])
 799                                 continue;
 800
 801                         ob = c->open_buckets + h->s->blocks[ec_idx];
 802                         if (ob->dev == devs_sorted.devs[i] &&
 803                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 804                                 goto got_bucket;
 805                 }
 806         goto out_put_head;
 807 got_bucket:
 808         ca = bch_dev_bkey_exists(c, ob->dev);
 809
 810         ob->ec_idx      = ec_idx;
 811         ob->ec          = h->s;
 812         ec_stripe_new_get(h->s, STRIPE_REF_io);
 813
 814         ret = add_new_bucket(c, ptrs, devs_may_alloc,
 815                              nr_replicas, nr_effective,
 816                              have_cache, flags, ob);
 817 out_put_head:
 818         bch2_ec_stripe_head_put(c, h);
 819         return ret;
 820 }
 821
 822 /* Sector allocator */
 823
 824 static bool want_bucket(struct bch_fs *c,
 825                         struct write_point *wp,
 826                         struct bch_devs_mask *devs_may_alloc,
 827                         bool *have_cache, bool ec,
 828                         struct open_bucket *ob)
 829 {
 830         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 831
 832         if (!test_bit(ob->dev, devs_may_alloc->d))
 833                 return false;
 834
 835         if (ob->data_type != wp->data_type)
 836                 return false;
 837
 838         if (!ca->mi.durability &&
 839             (wp->data_type == BCH_DATA_btree || ec || *have_cache))
 840                 return false;
 841
 842         if (ec != (ob->ec != NULL))
 843                 return false;
 844
 845         return true;
 846 }
 847
 848 static int bucket_alloc_set_writepoint(struct bch_fs *c,
 849                                        struct open_buckets *ptrs,
 850                                        struct write_point *wp,
 851                                        struct bch_devs_mask *devs_may_alloc,
 852                                        unsigned nr_replicas,
 853                                        unsigned *nr_effective,
 854                                        bool *have_cache,
 855                                        bool ec, unsigned flags)
 856 {
 857         struct open_buckets ptrs_skip = { .nr = 0 };
 858         struct open_bucket *ob;
 859         unsigned i;
 860         int ret = 0;
 861
 862         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 863                 if (!ret && want_bucket(c, wp, devs_may_alloc,
 864                                         have_cache, ec, ob))
 865                         ret = add_new_bucket(c, ptrs, devs_may_alloc,
 866                                        nr_replicas, nr_effective,
 867                                        have_cache, flags, ob);
 868                 else
 869                         ob_push(c, &ptrs_skip, ob);
 870         }
 871         wp->ptrs = ptrs_skip;
 872
 873         return ret;
 874 }
 875
 876 static int bucket_alloc_set_partial(struct bch_fs *c,
 877                                     struct open_buckets *ptrs,
 878                                     struct write_point *wp,
 879                                     struct bch_devs_mask *devs_may_alloc,
 880                                     unsigned nr_replicas,
 881                                     unsigned *nr_effective,
 882                                     bool *have_cache, bool ec,
 883                                     enum bch_watermark watermark,
 884                                     unsigned flags)
 885 {
 886         int i, ret = 0;
 887
 888         if (!c->open_buckets_partial_nr)
 889                 return 0;
 890
 891         spin_lock(&c->freelist_lock);
 892
 893         if (!c->open_buckets_partial_nr)
 894                 goto unlock;
 895
 896         for (i = c->open_buckets_partial_nr - 1; i >= 0; --i) {
 897                 struct open_bucket *ob = c->open_buckets + c->open_buckets_partial[i];
 898
 899                 if (want_bucket(c, wp, devs_may_alloc, have_cache, ec, ob)) {
 900                         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 901                         struct bch_dev_usage usage;
 902                         u64 avail;
 903
 904                         bch2_dev_usage_read_fast(ca, &usage);
 905                         avail = dev_buckets_free(ca, usage, watermark);
 906                         if (!avail)
 907                                 continue;
 908
 909                         array_remove_item(c->open_buckets_partial,
 910                                           c->open_buckets_partial_nr,
 911                                           i);
 912                         ob->on_partial_list = false;
 913
 914                         ret = add_new_bucket(c, ptrs, devs_may_alloc,
 915                                              nr_replicas, nr_effective,
 916                                              have_cache, flags, ob);
 917                         if (ret)
 918                                 break;
 919                 }
 920         }
 921 unlock:
 922         spin_unlock(&c->freelist_lock);
 923         return ret;
 924 }
 925
 926 static int __open_bucket_add_buckets(struct btree_trans *trans,
 927                         struct open_buckets *ptrs,
 928                         struct write_point *wp,
 929                         struct bch_devs_list *devs_have,
 930                         u16 target,
 931                         bool erasure_code,
 932                         unsigned nr_replicas,
 933                         unsigned *nr_effective,
 934                         bool *have_cache,
 935                         enum bch_watermark watermark,
 936                         unsigned flags,
 937                         struct closure *_cl)
 938 {
 939         struct bch_fs *c = trans->c;
 940         struct bch_devs_mask devs;
 941         struct open_bucket *ob;
 942         struct closure *cl = NULL;
 943         unsigned i;
 944         int ret;
 945
 946         devs = target_rw_devs(c, wp->data_type, target);
 947
 948         /* Don't allocate from devices we already have pointers to: */
 949         for (i = 0; i < devs_have->nr; i++)
 950                 __clear_bit(devs_have->devs[i], devs.d);
 951
 952         open_bucket_for_each(c, ptrs, ob, i)
 953                 __clear_bit(ob->dev, devs.d);
 954
 955         if (erasure_code && ec_open_bucket(c, ptrs))
 956                 return 0;
 957
 958         ret = bucket_alloc_set_writepoint(c, ptrs, wp, &devs,
 959                                  nr_replicas, nr_effective,
 960                                  have_cache, erasure_code, flags);
 961         if (ret)
 962                 return ret;
 963
 964         ret = bucket_alloc_set_partial(c, ptrs, wp, &devs,
 965                                  nr_replicas, nr_effective,
 966                                  have_cache, erasure_code, watermark, flags);
 967         if (ret)
 968                 return ret;
 969
 970         if (erasure_code) {
 971                 ret = bucket_alloc_from_stripe(trans, ptrs, wp, &devs,
 972                                          target,
 973                                          nr_replicas, nr_effective,
 974                                          have_cache,
 975                                          watermark, flags, _cl);
 976         } else {
 977 retry_blocking:
 978                 /*
 979                  * Try nonblocking first, so that if one device is full we'll try from
 980                  * other devices:
 981                  */
 982                 ret = bch2_bucket_alloc_set_trans(trans, ptrs, &wp->stripe, &devs,
 983                                         nr_replicas, nr_effective, have_cache,
 984                                         flags, wp->data_type, watermark, cl);
 985                 if (ret &&
 986                     !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
 987                     !bch2_err_matches(ret, BCH_ERR_insufficient_devices) &&
 988                     !cl && _cl) {
 989                         cl = _cl;
 990                         goto retry_blocking;
 991                 }
 992
 993         }
 994
 995         return ret;
 996 }
 997
 998 static int open_bucket_add_buckets(struct btree_trans *trans,
 999                         struct open_buckets *ptrs,
1000                         struct write_point *wp,
1001                         struct bch_devs_list *devs_have,
1002                         u16 target,
1003                         unsigned erasure_code,
1004                         unsigned nr_replicas,
1005                         unsigned *nr_effective,
1006                         bool *have_cache,
1007                         enum bch_watermark watermark,
1008                         unsigned flags,
1009                         struct closure *cl)
1010 {
1011         int ret;
1012
1013         if (erasure_code) {
1014                 ret = __open_bucket_add_buckets(trans, ptrs, wp,
1015                                 devs_have, target, erasure_code,
1016                                 nr_replicas, nr_effective, have_cache,
1017                                 watermark, flags, cl);
1018                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1019                     bch2_err_matches(ret, BCH_ERR_operation_blocked) ||
1020                     bch2_err_matches(ret, BCH_ERR_freelist_empty) ||
1021                     bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
1022                         return ret;
1023                 if (*nr_effective >= nr_replicas)
1024                         return 0;
1025         }
1026
1027         ret = __open_bucket_add_buckets(trans, ptrs, wp,
1028                         devs_have, target, false,
1029                         nr_replicas, nr_effective, have_cache,
1030                         watermark, flags, cl);
1031         return ret < 0 ? ret : 0;
1032 }
1033
1034 static bool should_drop_bucket(struct open_bucket *ob, struct bch_fs *c,
1035                                struct bch_dev *ca, bool ec)
1036 {
1037         if (ec) {
1038                 return ob->ec != NULL;
1039         } else if (ca) {
1040                 bool drop = ob->dev == ca->dev_idx;
1041                 struct open_bucket *ob2;
1042                 unsigned i;
1043
1044                 if (!drop && ob->ec) {
1045                         unsigned nr_blocks;
1046
1047                         mutex_lock(&ob->ec->lock);
1048                         nr_blocks = bkey_i_to_stripe(&ob->ec->new_stripe.key)->v.nr_blocks;
1049
1050                         for (i = 0; i < nr_blocks; i++) {
1051                                 if (!ob->ec->blocks[i])
1052                                         continue;
1053
1054                                 ob2 = c->open_buckets + ob->ec->blocks[i];
1055                                 drop |= ob2->dev == ca->dev_idx;
1056                         }
1057                         mutex_unlock(&ob->ec->lock);
1058                 }
1059
1060                 return drop;
1061         } else {
1062                 return true;
1063         }
1064 }
1065
1066 static void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
1067                                  bool ec, struct write_point *wp)
1068 {
1069         struct open_buckets ptrs = { .nr = 0 };
1070         struct open_bucket *ob;
1071         unsigned i;
1072
1073         mutex_lock(&wp->lock);
1074         open_bucket_for_each(c, &wp->ptrs, ob, i)
1075                 if (should_drop_bucket(ob, c, ca, ec))
1076                         bch2_open_bucket_put(c, ob);
1077                 else
1078                         ob_push(c, &ptrs, ob);
1079         wp->ptrs = ptrs;
1080         mutex_unlock(&wp->lock);
1081 }
1082
1083 void bch2_open_buckets_stop(struct bch_fs *c, struct bch_dev *ca,
1084                             bool ec)
1085 {
1086         unsigned i;
1087
1088         /* Next, close write points that point to this device... */
1089         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1090                 bch2_writepoint_stop(c, ca, ec, &c->write_points[i]);
1091
1092         bch2_writepoint_stop(c, ca, ec, &c->copygc_write_point);
1093         bch2_writepoint_stop(c, ca, ec, &c->rebalance_write_point);
1094         bch2_writepoint_stop(c, ca, ec, &c->btree_write_point);
1095
1096         mutex_lock(&c->btree_reserve_cache_lock);
1097         while (c->btree_reserve_cache_nr) {
1098                 struct btree_alloc *a =
1099                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1100
1101                 bch2_open_buckets_put(c, &a->ob);
1102         }
1103         mutex_unlock(&c->btree_reserve_cache_lock);
1104
1105         spin_lock(&c->freelist_lock);
1106         i = 0;
1107         while (i < c->open_buckets_partial_nr) {
1108                 struct open_bucket *ob =
1109                         c->open_buckets + c->open_buckets_partial[i];
1110
1111                 if (should_drop_bucket(ob, c, ca, ec)) {
1112                         --c->open_buckets_partial_nr;
1113                         swap(c->open_buckets_partial[i],
1114                              c->open_buckets_partial[c->open_buckets_partial_nr]);
1115                         ob->on_partial_list = false;
1116                         spin_unlock(&c->freelist_lock);
1117                         bch2_open_bucket_put(c, ob);
1118                         spin_lock(&c->freelist_lock);
1119                 } else {
1120                         i++;
1121                 }
1122         }
1123         spin_unlock(&c->freelist_lock);
1124
1125         bch2_ec_stop_dev(c, ca);
1126 }
1127
1128 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
1129                                                  unsigned long write_point)
1130 {
1131         unsigned hash =
1132                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
1133
1134         return &c->write_points_hash[hash];
1135 }
1136
1137 static struct write_point *__writepoint_find(struct hlist_head *head,
1138                                              unsigned long write_point)
1139 {
1140         struct write_point *wp;
1141
1142         rcu_read_lock();
1143         hlist_for_each_entry_rcu(wp, head, node)
1144                 if (wp->write_point == write_point)
1145                         goto out;
1146         wp = NULL;
1147 out:
1148         rcu_read_unlock();
1149         return wp;
1150 }
1151
1152 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
1153 {
1154         u64 stranded    = c->write_points_nr * c->bucket_size_max;
1155         u64 free        = bch2_fs_usage_read_short(c).free;
1156
1157         return stranded * factor > free;
1158 }
1159
1160 static bool try_increase_writepoints(struct bch_fs *c)
1161 {
1162         struct write_point *wp;
1163
1164         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
1165             too_many_writepoints(c, 32))
1166                 return false;
1167
1168         wp = c->write_points + c->write_points_nr++;
1169         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
1170         return true;
1171 }
1172
1173 static bool try_decrease_writepoints(struct btree_trans *trans, unsigned old_nr)
1174 {
1175         struct bch_fs *c = trans->c;
1176         struct write_point *wp;
1177         struct open_bucket *ob;
1178         unsigned i;
1179
1180         mutex_lock(&c->write_points_hash_lock);
1181         if (c->write_points_nr < old_nr) {
1182                 mutex_unlock(&c->write_points_hash_lock);
1183                 return true;
1184         }
1185
1186         if (c->write_points_nr == 1 ||
1187             !too_many_writepoints(c, 8)) {
1188                 mutex_unlock(&c->write_points_hash_lock);
1189                 return false;
1190         }
1191
1192         wp = c->write_points + --c->write_points_nr;
1193
1194         hlist_del_rcu(&wp->node);
1195         mutex_unlock(&c->write_points_hash_lock);
1196
1197         bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1198         open_bucket_for_each(c, &wp->ptrs, ob, i)
1199                 open_bucket_free_unused(c, ob);
1200         wp->ptrs.nr = 0;
1201         mutex_unlock(&wp->lock);
1202         return true;
1203 }
1204
1205 static struct write_point *writepoint_find(struct btree_trans *trans,
1206                                            unsigned long write_point)
1207 {
1208         struct bch_fs *c = trans->c;
1209         struct write_point *wp, *oldest;
1210         struct hlist_head *head;
1211
1212         if (!(write_point & 1UL)) {
1213                 wp = (struct write_point *) write_point;
1214                 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1215                 return wp;
1216         }
1217
1218         head = writepoint_hash(c, write_point);
1219 restart_find:
1220         wp = __writepoint_find(head, write_point);
1221         if (wp) {
1222 lock_wp:
1223                 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1224                 if (wp->write_point == write_point)
1225                         goto out;
1226                 mutex_unlock(&wp->lock);
1227                 goto restart_find;
1228         }
1229 restart_find_oldest:
1230         oldest = NULL;
1231         for (wp = c->write_points;
1232              wp < c->write_points + c->write_points_nr; wp++)
1233                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1234                         oldest = wp;
1235
1236         bch2_trans_mutex_lock_norelock(trans, &oldest->lock);
1237         bch2_trans_mutex_lock_norelock(trans, &c->write_points_hash_lock);
1238         if (oldest >= c->write_points + c->write_points_nr ||
1239             try_increase_writepoints(c)) {
1240                 mutex_unlock(&c->write_points_hash_lock);
1241                 mutex_unlock(&oldest->lock);
1242                 goto restart_find_oldest;
1243         }
1244
1245         wp = __writepoint_find(head, write_point);
1246         if (wp && wp != oldest) {
1247                 mutex_unlock(&c->write_points_hash_lock);
1248                 mutex_unlock(&oldest->lock);
1249                 goto lock_wp;
1250         }
1251
1252         wp = oldest;
1253         hlist_del_rcu(&wp->node);
1254         wp->write_point = write_point;
1255         hlist_add_head_rcu(&wp->node, head);
1256         mutex_unlock(&c->write_points_hash_lock);
1257 out:
1258         wp->last_used = local_clock();
1259         return wp;
1260 }
1261
1262 /*
1263  * Get us an open_bucket we can allocate from, return with it locked:
1264  */
1265 int bch2_alloc_sectors_start_trans(struct btree_trans *trans,
1266                              unsigned target,
1267                              unsigned erasure_code,
1268                              struct write_point_specifier write_point,
1269                              struct bch_devs_list *devs_have,
1270                              unsigned nr_replicas,
1271                              unsigned nr_replicas_required,
1272                              enum bch_watermark watermark,
1273                              unsigned flags,
1274                              struct closure *cl,
1275                              struct write_point **wp_ret)
1276 {
1277         struct bch_fs *c = trans->c;
1278         struct write_point *wp;
1279         struct open_bucket *ob;
1280         struct open_buckets ptrs;
1281         unsigned nr_effective, write_points_nr;
1282         bool have_cache;
1283         int ret;
1284         int i;
1285
1286         BUG_ON(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS);
1287
1288         BUG_ON(!nr_replicas || !nr_replicas_required);
1289 retry:
1290         ptrs.nr         = 0;
1291         nr_effective    = 0;
1292         write_points_nr = c->write_points_nr;
1293         have_cache      = false;
1294
1295         *wp_ret = wp = writepoint_find(trans, write_point.v);
1296
1297         /* metadata may not allocate on cache devices: */
1298         if (wp->data_type != BCH_DATA_user)
1299                 have_cache = true;
1300
1301         if (target && !(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1302                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1303                                               target, erasure_code,
1304                                               nr_replicas, &nr_effective,
1305                                               &have_cache, watermark,
1306                                               flags, NULL);
1307                 if (!ret ||
1308                     bch2_err_matches(ret, BCH_ERR_transaction_restart))
1309                         goto alloc_done;
1310
1311                 /* Don't retry from all devices if we're out of open buckets: */
1312                 if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
1313                         goto allocate_blocking;
1314
1315                 /*
1316                  * Only try to allocate cache (durability = 0 devices) from the
1317                  * specified target:
1318                  */
1319                 have_cache = true;
1320
1321                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1322                                               0, erasure_code,
1323                                               nr_replicas, &nr_effective,
1324                                               &have_cache, watermark,
1325                                               flags, cl);
1326         } else {
1327 allocate_blocking:
1328                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1329                                               target, erasure_code,
1330                                               nr_replicas, &nr_effective,
1331                                               &have_cache, watermark,
1332                                               flags, cl);
1333         }
1334 alloc_done:
1335         BUG_ON(!ret && nr_effective < nr_replicas);
1336
1337         if (erasure_code && !ec_open_bucket(c, &ptrs))
1338                 pr_debug("failed to get ec bucket: ret %u", ret);
1339
1340         if (ret == -BCH_ERR_insufficient_devices &&
1341             nr_effective >= nr_replicas_required)
1342                 ret = 0;
1343
1344         if (ret)
1345                 goto err;
1346
1347         /* Free buckets we didn't use: */
1348         open_bucket_for_each(c, &wp->ptrs, ob, i)
1349                 open_bucket_free_unused(c, ob);
1350
1351         wp->ptrs = ptrs;
1352
1353         wp->sectors_free = UINT_MAX;
1354
1355         open_bucket_for_each(c, &wp->ptrs, ob, i)
1356                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1357
1358         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1359
1360         return 0;
1361 err:
1362         open_bucket_for_each(c, &wp->ptrs, ob, i)
1363                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1364                         ob_push(c, &ptrs, ob);
1365                 else
1366                         open_bucket_free_unused(c, ob);
1367         wp->ptrs = ptrs;
1368
1369         mutex_unlock(&wp->lock);
1370
1371         if (bch2_err_matches(ret, BCH_ERR_freelist_empty) &&
1372             try_decrease_writepoints(trans, write_points_nr))
1373                 goto retry;
1374
1375         if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty) ||
1376             bch2_err_matches(ret, BCH_ERR_freelist_empty))
1377                 return cl
1378                         ? -BCH_ERR_bucket_alloc_blocked
1379                         : -BCH_ERR_ENOSPC_bucket_alloc;
1380
1381         return ret;
1382 }
1383
1384 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1385 {
1386         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1387
1388         return (struct bch_extent_ptr) {
1389                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
1390                 .gen    = ob->gen,
1391                 .dev    = ob->dev,
1392                 .offset = bucket_to_sector(ca, ob->bucket) +
1393                         ca->mi.bucket_size -
1394                         ob->sectors_free,
1395         };
1396 }
1397
1398 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1399                                     struct bkey_i *k, unsigned sectors,
1400                                     bool cached)
1401 {
1402         bch2_alloc_sectors_append_ptrs_inlined(c, wp, k, sectors, cached);
1403 }
1404
1405 /*
1406  * Append pointers to the space we just allocated to @k, and mark @sectors space
1407  * as allocated out of @ob
1408  */
1409 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1410 {
1411         bch2_alloc_sectors_done_inlined(c, wp);
1412 }
1413
1414 static inline void writepoint_init(struct write_point *wp,
1415                                    enum bch_data_type type)
1416 {
1417         mutex_init(&wp->lock);
1418         wp->data_type = type;
1419
1420         INIT_WORK(&wp->index_update_work, bch2_write_point_do_index_updates);
1421         INIT_LIST_HEAD(&wp->writes);
1422         spin_lock_init(&wp->writes_lock);
1423 }
1424
1425 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1426 {
1427         struct open_bucket *ob;
1428         struct write_point *wp;
1429
1430         mutex_init(&c->write_points_hash_lock);
1431         c->write_points_nr = ARRAY_SIZE(c->write_points);
1432
1433         /* open bucket 0 is a sentinal NULL: */
1434         spin_lock_init(&c->open_buckets[0].lock);
1435
1436         for (ob = c->open_buckets + 1;
1437              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1438                 spin_lock_init(&ob->lock);
1439                 c->open_buckets_nr_free++;
1440
1441                 ob->freelist = c->open_buckets_freelist;
1442                 c->open_buckets_freelist = ob - c->open_buckets;
1443         }
1444
1445         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
1446         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
1447         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
1448
1449         for (wp = c->write_points;
1450              wp < c->write_points + c->write_points_nr; wp++) {
1451                 writepoint_init(wp, BCH_DATA_user);
1452
1453                 wp->last_used   = local_clock();
1454                 wp->write_point = (unsigned long) wp;
1455                 hlist_add_head_rcu(&wp->node,
1456                                    writepoint_hash(c, wp->write_point));
1457         }
1458 }
1459
1460 static void bch2_open_bucket_to_text(struct printbuf *out, struct bch_fs *c, struct open_bucket *ob)
1461 {
1462         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1463         unsigned data_type = ob->data_type;
1464         barrier(); /* READ_ONCE() doesn't work on bitfields */
1465
1466         prt_printf(out, "%zu ref %u %s %u:%llu gen %u allocated %u/%u",
1467                    ob - c->open_buckets,
1468                    atomic_read(&ob->pin),
1469                    data_type < BCH_DATA_NR ? bch2_data_types[data_type] : "invalid data type",
1470                    ob->dev, ob->bucket, ob->gen,
1471                    ca->mi.bucket_size - ob->sectors_free, ca->mi.bucket_size);
1472         if (ob->ec)
1473                 prt_printf(out, " ec idx %llu", ob->ec->idx);
1474         if (ob->on_partial_list)
1475                 prt_str(out, " partial");
1476         prt_newline(out);
1477 }
1478
1479 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1480 {
1481         struct open_bucket *ob;
1482
1483         out->atomic++;
1484
1485         for (ob = c->open_buckets;
1486              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1487              ob++) {
1488                 spin_lock(&ob->lock);
1489                 if (ob->valid && !ob->on_partial_list)
1490                         bch2_open_bucket_to_text(out, c, ob);
1491                 spin_unlock(&ob->lock);
1492         }
1493
1494         --out->atomic;
1495 }
1496
1497 void bch2_open_buckets_partial_to_text(struct printbuf *out, struct bch_fs *c)
1498 {
1499         unsigned i;
1500
1501         out->atomic++;
1502         spin_lock(&c->freelist_lock);
1503
1504         for (i = 0; i < c->open_buckets_partial_nr; i++)
1505                 bch2_open_bucket_to_text(out, c,
1506                                 c->open_buckets + c->open_buckets_partial[i]);
1507
1508         spin_unlock(&c->freelist_lock);
1509         --out->atomic;
1510 }
1511
1512 static const char * const bch2_write_point_states[] = {
1513 #define x(n)    #n,
1514         WRITE_POINT_STATES()
1515 #undef x
1516         NULL
1517 };
1518
1519 void bch2_write_points_to_text(struct printbuf *out, struct bch_fs *c)
1520 {
1521         struct write_point *wp;
1522         unsigned i;
1523
1524         for (wp = c->write_points;
1525              wp < c->write_points + ARRAY_SIZE(c->write_points);
1526              wp++) {
1527                 prt_printf(out, "%lu: ", wp->write_point);
1528                 prt_human_readable_u64(out, wp->sectors_allocated);
1529
1530                 prt_printf(out, " last wrote: ");
1531                 bch2_pr_time_units(out, sched_clock() - wp->last_used);
1532
1533                 for (i = 0; i < WRITE_POINT_STATE_NR; i++) {
1534                         prt_printf(out, " %s: ", bch2_write_point_states[i]);
1535                         bch2_pr_time_units(out, wp->time[i]);
1536                 }
1537
1538                 prt_newline(out);
1539         }
1540 }