git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/alloc_foreground.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright 2012 Google, Inc.
   4  *
   5  * Foreground allocator code: allocate buckets from freelist, and allocate in
   6  * sector granularity from writepoints.
   7  *
   8  * bch2_bucket_alloc() allocates a single bucket from a specific device.
   9  *
  10  * bch2_bucket_alloc_set() allocates one or more buckets from different devices
  11  * in a given filesystem.
  12  */
  13
  14 #include "bcachefs.h"
  15 #include "alloc_background.h"
  16 #include "alloc_foreground.h"
  17 #include "btree_iter.h"
  18 #include "btree_update.h"
  19 #include "btree_gc.h"
  20 #include "buckets.h"
  21 #include "buckets_waiting_for_journal.h"
  22 #include "clock.h"
  23 #include "debug.h"
  24 #include "disk_groups.h"
  25 #include "ec.h"
  26 #include "error.h"
  27 #include "io.h"
  28 #include "journal.h"
  29
  30 #include <linux/math64.h>
  31 #include <linux/rculist.h>
  32 #include <linux/rcupdate.h>
  33 #include <trace/events/bcachefs.h>
  34
  35 const char * const bch2_alloc_reserves[] = {
  36 #define x(t) #t,
  37         BCH_ALLOC_RESERVES()
  38 #undef x
  39         NULL
  40 };
  41
  42 /*
  43  * Open buckets represent a bucket that's currently being allocated from.  They
  44  * serve two purposes:
  45  *
  46  *  - They track buckets that have been partially allocated, allowing for
  47  *    sub-bucket sized allocations - they're used by the sector allocator below
  48  *
  49  *  - They provide a reference to the buckets they own that mark and sweep GC
  50  *    can find, until the new allocation has a pointer to it inserted into the
  51  *    btree
  52  *
  53  * When allocating some space with the sector allocator, the allocation comes
  54  * with a reference to an open bucket - the caller is required to put that
  55  * reference _after_ doing the index update that makes its allocation reachable.
  56  */
  57
  58 static void bch2_open_bucket_hash_add(struct bch_fs *c, struct open_bucket *ob)
  59 {
  60         open_bucket_idx_t idx = ob - c->open_buckets;
  61         open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
  62
  63         ob->hash = *slot;
  64         *slot = idx;
  65 }
  66
  67 static void bch2_open_bucket_hash_remove(struct bch_fs *c, struct open_bucket *ob)
  68 {
  69         open_bucket_idx_t idx = ob - c->open_buckets;
  70         open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
  71
  72         while (*slot != idx) {
  73                 BUG_ON(!*slot);
  74                 slot = &c->open_buckets[*slot].hash;
  75         }
  76
  77         *slot = ob->hash;
  78         ob->hash = 0;
  79 }
  80
  81 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
  82 {
  83         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
  84
  85         if (ob->ec) {
  86                 bch2_ec_bucket_written(c, ob);
  87                 return;
  88         }
  89
  90         percpu_down_read(&c->mark_lock);
  91         spin_lock(&ob->lock);
  92
  93         ob->valid = false;
  94         ob->data_type = 0;
  95
  96         spin_unlock(&ob->lock);
  97         percpu_up_read(&c->mark_lock);
  98
  99         spin_lock(&c->freelist_lock);
 100         bch2_open_bucket_hash_remove(c, ob);
 101
 102         ob->freelist = c->open_buckets_freelist;
 103         c->open_buckets_freelist = ob - c->open_buckets;
 104
 105         c->open_buckets_nr_free++;
 106         ca->nr_open_buckets--;
 107         spin_unlock(&c->freelist_lock);
 108
 109         closure_wake_up(&c->open_buckets_wait);
 110 }
 111
 112 void bch2_open_bucket_write_error(struct bch_fs *c,
 113                                   struct open_buckets *obs,
 114                                   unsigned dev)
 115 {
 116         struct open_bucket *ob;
 117         unsigned i;
 118
 119         open_bucket_for_each(c, obs, ob, i)
 120                 if (ob->dev == dev && ob->ec)
 121                         bch2_ec_bucket_cancel(c, ob);
 122 }
 123
 124 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
 125 {
 126         struct open_bucket *ob;
 127
 128         BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
 129
 130         ob = c->open_buckets + c->open_buckets_freelist;
 131         c->open_buckets_freelist = ob->freelist;
 132         atomic_set(&ob->pin, 1);
 133         ob->data_type = 0;
 134
 135         c->open_buckets_nr_free--;
 136         return ob;
 137 }
 138
 139 static void open_bucket_free_unused(struct bch_fs *c,
 140                                     struct write_point *wp,
 141                                     struct open_bucket *ob)
 142 {
 143         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 144         bool may_realloc = wp->data_type == BCH_DATA_user;
 145
 146         BUG_ON(ca->open_buckets_partial_nr >
 147                ARRAY_SIZE(ca->open_buckets_partial));
 148
 149         if (ca->open_buckets_partial_nr <
 150             ARRAY_SIZE(ca->open_buckets_partial) &&
 151             may_realloc) {
 152                 spin_lock(&c->freelist_lock);
 153                 ob->on_partial_list = true;
 154                 ca->open_buckets_partial[ca->open_buckets_partial_nr++] =
 155                         ob - c->open_buckets;
 156                 spin_unlock(&c->freelist_lock);
 157
 158                 closure_wake_up(&c->open_buckets_wait);
 159                 closure_wake_up(&c->freelist_wait);
 160         } else {
 161                 bch2_open_bucket_put(c, ob);
 162         }
 163 }
 164
 165 /* _only_ for allocating the journal on a new device: */
 166 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
 167 {
 168         while (ca->new_fs_bucket_idx < ca->mi.nbuckets) {
 169                 u64 b = ca->new_fs_bucket_idx++;
 170
 171                 if (!is_superblock_bucket(ca, b) &&
 172                     (!ca->buckets_nouse || !test_bit(b, ca->buckets_nouse)))
 173                         return b;
 174         }
 175
 176         return -1;
 177 }
 178
 179 static inline unsigned open_buckets_reserved(enum alloc_reserve reserve)
 180 {
 181         switch (reserve) {
 182         case RESERVE_btree:
 183         case RESERVE_btree_movinggc:
 184                 return 0;
 185         case RESERVE_movinggc:
 186                 return OPEN_BUCKETS_COUNT / 4;
 187         default:
 188                 return OPEN_BUCKETS_COUNT / 2;
 189         }
 190 }
 191
 192 static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 193                                               u64 bucket,
 194                                               enum alloc_reserve reserve,
 195                                               struct bch_alloc_v4 *a,
 196                                               u64 *skipped_open,
 197                                               u64 *skipped_need_journal_commit,
 198                                               u64 *skipped_nouse,
 199                                               struct closure *cl)
 200 {
 201         struct open_bucket *ob;
 202
 203         if (unlikely(ca->buckets_nouse && test_bit(bucket, ca->buckets_nouse))) {
 204                 (*skipped_nouse)++;
 205                 return NULL;
 206         }
 207
 208         if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
 209                 (*skipped_open)++;
 210                 return NULL;
 211         }
 212
 213         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 214                         c->journal.flushed_seq_ondisk, ca->dev_idx, bucket)) {
 215                 (*skipped_need_journal_commit)++;
 216                 return NULL;
 217         }
 218
 219         spin_lock(&c->freelist_lock);
 220
 221         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(reserve))) {
 222                 if (cl)
 223                         closure_wait(&c->open_buckets_wait, cl);
 224
 225                 if (!c->blocked_allocate_open_bucket)
 226                         c->blocked_allocate_open_bucket = local_clock();
 227
 228                 spin_unlock(&c->freelist_lock);
 229                 return ERR_PTR(-OPEN_BUCKETS_EMPTY);
 230         }
 231
 232         /* Recheck under lock: */
 233         if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
 234                 spin_unlock(&c->freelist_lock);
 235                 (*skipped_open)++;
 236                 return NULL;
 237         }
 238
 239         ob = bch2_open_bucket_alloc(c);
 240
 241         spin_lock(&ob->lock);
 242
 243         ob->valid       = true;
 244         ob->sectors_free = ca->mi.bucket_size;
 245         ob->alloc_reserve = reserve;
 246         ob->dev         = ca->dev_idx;
 247         ob->gen         = a->gen;
 248         ob->bucket      = bucket;
 249         spin_unlock(&ob->lock);
 250
 251         ca->nr_open_buckets++;
 252         bch2_open_bucket_hash_add(c, ob);
 253
 254         if (c->blocked_allocate_open_bucket) {
 255                 bch2_time_stats_update(
 256                         &c->times[BCH_TIME_blocked_allocate_open_bucket],
 257                         c->blocked_allocate_open_bucket);
 258                 c->blocked_allocate_open_bucket = 0;
 259         }
 260
 261         if (c->blocked_allocate) {
 262                 bch2_time_stats_update(
 263                         &c->times[BCH_TIME_blocked_allocate],
 264                         c->blocked_allocate);
 265                 c->blocked_allocate = 0;
 266         }
 267
 268         spin_unlock(&c->freelist_lock);
 269
 270         trace_bucket_alloc(ca, bch2_alloc_reserves[reserve]);
 271         return ob;
 272 }
 273
 274 static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bch_dev *ca,
 275                                             enum alloc_reserve reserve, u64 free_entry,
 276                                             u64 *skipped_open,
 277                                             u64 *skipped_need_journal_commit,
 278                                             u64 *skipped_nouse,
 279                                             struct bkey_s_c freespace_k,
 280                                             struct closure *cl)
 281 {
 282         struct bch_fs *c = trans->c;
 283         struct btree_iter iter = { NULL };
 284         struct bkey_s_c k;
 285         struct open_bucket *ob;
 286         struct bch_alloc_v4 a;
 287         u64 b = free_entry & ~(~0ULL << 56);
 288         unsigned genbits = free_entry >> 56;
 289         struct printbuf buf = PRINTBUF;
 290         int ret;
 291
 292         if (b < ca->mi.first_bucket || b >= ca->mi.nbuckets) {
 293                 prt_printf(&buf, "freespace btree has bucket outside allowed range %u-%llu\n"
 294                        "  freespace key ",
 295                         ca->mi.first_bucket, ca->mi.nbuckets);
 296                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 297                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 298                 ob = ERR_PTR(-EIO);
 299                 goto err;
 300         }
 301
 302         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS(ca->dev_idx, b), BTREE_ITER_CACHED);
 303         k = bch2_btree_iter_peek_slot(&iter);
 304         ret = bkey_err(k);
 305         if (ret) {
 306                 ob = ERR_PTR(ret);
 307                 goto err;
 308         }
 309
 310         bch2_alloc_to_v4(k, &a);
 311
 312         if (genbits != (alloc_freespace_genbits(a) >> 56)) {
 313                 prt_printf(&buf, "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
 314                        "  freespace key ",
 315                        genbits, alloc_freespace_genbits(a) >> 56);
 316                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 317                 prt_printf(&buf, "\n  ");
 318                 bch2_bkey_val_to_text(&buf, c, k);
 319                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 320                 ob = ERR_PTR(-EIO);
 321                 goto err;
 322
 323         }
 324
 325         if (a.data_type != BCH_DATA_free) {
 326                 prt_printf(&buf, "non free bucket in freespace btree\n"
 327                        "  freespace key ");
 328                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 329                 prt_printf(&buf, "\n  ");
 330                 bch2_bkey_val_to_text(&buf, c, k);
 331                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 332                 ob = ERR_PTR(-EIO);
 333                 goto err;
 334         }
 335
 336         ob = __try_alloc_bucket(c, ca, b, reserve, &a,
 337                                 skipped_open,
 338                                 skipped_need_journal_commit,
 339                                 skipped_nouse,
 340                                 cl);
 341 err:
 342         bch2_trans_iter_exit(trans, &iter);
 343         printbuf_exit(&buf);
 344         return ob;
 345 }
 346
 347 static struct open_bucket *try_alloc_partial_bucket(struct bch_fs *c, struct bch_dev *ca,
 348                                                     enum alloc_reserve reserve)
 349 {
 350         struct open_bucket *ob;
 351         int i;
 352
 353         spin_lock(&c->freelist_lock);
 354
 355         for (i = ca->open_buckets_partial_nr - 1; i >= 0; --i) {
 356                 ob = c->open_buckets + ca->open_buckets_partial[i];
 357
 358                 if (reserve <= ob->alloc_reserve) {
 359                         array_remove_item(ca->open_buckets_partial,
 360                                           ca->open_buckets_partial_nr,
 361                                           i);
 362                         ob->on_partial_list = false;
 363                         ob->alloc_reserve = reserve;
 364                         spin_unlock(&c->freelist_lock);
 365                         return ob;
 366                 }
 367         }
 368
 369         spin_unlock(&c->freelist_lock);
 370         return NULL;
 371 }
 372
 373 /*
 374  * This path is for before the freespace btree is initialized:
 375  *
 376  * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
 377  * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
 378  */
 379 static noinline struct open_bucket *
 380 bch2_bucket_alloc_trans_early(struct btree_trans *trans,
 381                               struct bch_dev *ca,
 382                               enum alloc_reserve reserve,
 383                               u64 *cur_bucket,
 384                               u64 *buckets_seen,
 385                               u64 *skipped_open,
 386                               u64 *skipped_need_journal_commit,
 387                               u64 *skipped_nouse,
 388                               struct closure *cl)
 389 {
 390         struct btree_iter iter;
 391         struct bkey_s_c k;
 392         struct open_bucket *ob = NULL;
 393         int ret;
 394
 395         *cur_bucket = max_t(u64, *cur_bucket, ca->mi.first_bucket);
 396         *cur_bucket = max_t(u64, *cur_bucket, ca->new_fs_bucket_idx);
 397
 398         for_each_btree_key(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, *cur_bucket),
 399                            BTREE_ITER_SLOTS, k, ret) {
 400                 struct bch_alloc_v4 a;
 401
 402                 if (bkey_cmp(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)) >= 0)
 403                         break;
 404
 405                 if (ca->new_fs_bucket_idx &&
 406                     is_superblock_bucket(ca, k.k->p.offset))
 407                         continue;
 408
 409                 bch2_alloc_to_v4(k, &a);
 410
 411                 if (a.data_type != BCH_DATA_free)
 412                         continue;
 413
 414                 (*buckets_seen)++;
 415
 416                 ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, reserve, &a,
 417                                         skipped_open,
 418                                         skipped_need_journal_commit,
 419                                         skipped_nouse,
 420                                         cl);
 421                 if (ob)
 422                         break;
 423         }
 424         bch2_trans_iter_exit(trans, &iter);
 425
 426         *cur_bucket = iter.pos.offset;
 427
 428         return ob ?: ERR_PTR(ret ?: -FREELIST_EMPTY);
 429 }
 430
 431 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
 432                                                    struct bch_dev *ca,
 433                                                    enum alloc_reserve reserve,
 434                                                    u64 *cur_bucket,
 435                                                    u64 *buckets_seen,
 436                                                    u64 *skipped_open,
 437                                                    u64 *skipped_need_journal_commit,
 438                                                    u64 *skipped_nouse,
 439                                                    struct closure *cl)
 440 {
 441         struct btree_iter iter;
 442         struct bkey_s_c k;
 443         struct open_bucket *ob = NULL;
 444         int ret;
 445
 446         if (unlikely(!ca->mi.freespace_initialized))
 447                 return bch2_bucket_alloc_trans_early(trans, ca, reserve,
 448                                                      cur_bucket,
 449                                                      buckets_seen,
 450                                                      skipped_open,
 451                                                      skipped_need_journal_commit,
 452                                                      skipped_nouse,
 453                                                      cl);
 454
 455         BUG_ON(ca->new_fs_bucket_idx);
 456
 457         for_each_btree_key_norestart(trans, iter, BTREE_ID_freespace,
 458                                      POS(ca->dev_idx, *cur_bucket), 0, k, ret) {
 459                 if (k.k->p.inode != ca->dev_idx)
 460                         break;
 461
 462                 for (*cur_bucket = max(*cur_bucket, bkey_start_offset(k.k));
 463                      *cur_bucket < k.k->p.offset && !ob;
 464                      (*cur_bucket)++) {
 465                         if (btree_trans_too_many_iters(trans)) {
 466                                 ob = ERR_PTR(-EINTR);
 467                                 break;
 468                         }
 469
 470                         (*buckets_seen)++;
 471
 472                         ob = try_alloc_bucket(trans, ca, reserve,
 473                                               *cur_bucket,
 474                                               skipped_open,
 475                                               skipped_need_journal_commit,
 476                                               skipped_nouse,
 477                                               k, cl);
 478                 }
 479                 if (ob)
 480                         break;
 481         }
 482         bch2_trans_iter_exit(trans, &iter);
 483
 484         return ob ?: ERR_PTR(ret);
 485 }
 486
 487 /**
 488  * bch_bucket_alloc - allocate a single bucket from a specific device
 489  *
 490  * Returns index of bucket on success, 0 on failure
 491  * */
 492 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 493                                       enum alloc_reserve reserve,
 494                                       bool may_alloc_partial,
 495                                       struct closure *cl)
 496 {
 497         struct open_bucket *ob = NULL;
 498         struct bch_dev_usage usage;
 499         u64 avail;
 500         u64 cur_bucket = 0;
 501         u64 buckets_seen = 0;
 502         u64 skipped_open = 0;
 503         u64 skipped_need_journal_commit = 0;
 504         u64 skipped_nouse = 0;
 505         bool waiting = false;
 506         int ret;
 507 again:
 508         usage = bch2_dev_usage_read(ca);
 509         avail = __dev_buckets_available(ca, usage,reserve);
 510
 511         if (usage.d[BCH_DATA_need_discard].buckets > avail)
 512                 bch2_do_discards(c);
 513
 514         if (usage.d[BCH_DATA_need_gc_gens].buckets > avail)
 515                 bch2_do_gc_gens(c);
 516
 517         if (should_invalidate_buckets(ca, usage))
 518                 bch2_do_invalidates(c);
 519
 520         if (!avail) {
 521                 if (cl && !waiting) {
 522                         closure_wait(&c->freelist_wait, cl);
 523                         waiting = true;
 524                         goto again;
 525                 }
 526
 527                 if (!c->blocked_allocate)
 528                         c->blocked_allocate = local_clock();
 529
 530                 ob = ERR_PTR(-FREELIST_EMPTY);
 531                 goto err;
 532         }
 533
 534         if (waiting)
 535                 closure_wake_up(&c->freelist_wait);
 536
 537         if (may_alloc_partial) {
 538                 ob = try_alloc_partial_bucket(c, ca, reserve);
 539                 if (ob)
 540                         return ob;
 541         }
 542
 543         ret = bch2_trans_do(c, NULL, NULL, 0,
 544                         PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, reserve,
 545                                                         &cur_bucket,
 546                                                         &buckets_seen,
 547                                                         &skipped_open,
 548                                                         &skipped_need_journal_commit,
 549                                                         &skipped_nouse,
 550                                                         cl)));
 551
 552         if (skipped_need_journal_commit * 2 > avail)
 553                 bch2_journal_flush_async(&c->journal, NULL);
 554 err:
 555         if (!ob)
 556                 ob = ERR_PTR(ret ?: -FREELIST_EMPTY);
 557
 558         if (IS_ERR(ob)) {
 559                 trace_bucket_alloc_fail(ca, bch2_alloc_reserves[reserve], avail,
 560                                         buckets_seen,
 561                                         skipped_open,
 562                                         skipped_need_journal_commit,
 563                                         skipped_nouse,
 564                                         cl == NULL, PTR_ERR(ob));
 565                 atomic_long_inc(&c->bucket_alloc_fail);
 566         }
 567
 568         return ob;
 569 }
 570
 571 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 572                             unsigned l, unsigned r)
 573 {
 574         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 575                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 576 }
 577
 578 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 579
 580 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 581                                           struct dev_stripe_state *stripe,
 582                                           struct bch_devs_mask *devs)
 583 {
 584         struct dev_alloc_list ret = { .nr = 0 };
 585         unsigned i;
 586
 587         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 588                 ret.devs[ret.nr++] = i;
 589
 590         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 591         return ret;
 592 }
 593
 594 void bch2_dev_stripe_increment(struct bch_dev *ca,
 595                                struct dev_stripe_state *stripe)
 596 {
 597         u64 *v = stripe->next_alloc + ca->dev_idx;
 598         u64 free_space = dev_buckets_available(ca, RESERVE_none);
 599         u64 free_space_inv = free_space
 600                 ? div64_u64(1ULL << 48, free_space)
 601                 : 1ULL << 48;
 602         u64 scale = *v / 4;
 603
 604         if (*v + free_space_inv >= *v)
 605                 *v += free_space_inv;
 606         else
 607                 *v = U64_MAX;
 608
 609         for (v = stripe->next_alloc;
 610              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 611                 *v = *v < scale ? 0 : *v - scale;
 612 }
 613
 614 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 615 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 616
 617 static void add_new_bucket(struct bch_fs *c,
 618                            struct open_buckets *ptrs,
 619                            struct bch_devs_mask *devs_may_alloc,
 620                            unsigned *nr_effective,
 621                            bool *have_cache,
 622                            unsigned flags,
 623                            struct open_bucket *ob)
 624 {
 625         unsigned durability =
 626                 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
 627
 628         __clear_bit(ob->dev, devs_may_alloc->d);
 629         *nr_effective   += (flags & BUCKET_ALLOC_USE_DURABILITY)
 630                 ? durability : 1;
 631         *have_cache     |= !durability;
 632
 633         ob_push(c, ptrs, ob);
 634 }
 635
 636 int bch2_bucket_alloc_set(struct bch_fs *c,
 637                       struct open_buckets *ptrs,
 638                       struct dev_stripe_state *stripe,
 639                       struct bch_devs_mask *devs_may_alloc,
 640                       unsigned nr_replicas,
 641                       unsigned *nr_effective,
 642                       bool *have_cache,
 643                       enum alloc_reserve reserve,
 644                       unsigned flags,
 645                       struct closure *cl)
 646 {
 647         struct dev_alloc_list devs_sorted =
 648                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 649         unsigned dev;
 650         struct bch_dev *ca;
 651         int ret = -INSUFFICIENT_DEVICES;
 652         unsigned i;
 653
 654         BUG_ON(*nr_effective >= nr_replicas);
 655
 656         for (i = 0; i < devs_sorted.nr; i++) {
 657                 struct open_bucket *ob;
 658
 659                 dev = devs_sorted.devs[i];
 660
 661                 rcu_read_lock();
 662                 ca = rcu_dereference(c->devs[dev]);
 663                 if (ca)
 664                         percpu_ref_get(&ca->ref);
 665                 rcu_read_unlock();
 666
 667                 if (!ca)
 668                         continue;
 669
 670                 if (!ca->mi.durability && *have_cache) {
 671                         percpu_ref_put(&ca->ref);
 672                         continue;
 673                 }
 674
 675                 ob = bch2_bucket_alloc(c, ca, reserve,
 676                                 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
 677                 if (!IS_ERR(ob))
 678                         bch2_dev_stripe_increment(ca, stripe);
 679                 percpu_ref_put(&ca->ref);
 680
 681                 if (IS_ERR(ob)) {
 682                         ret = PTR_ERR(ob);
 683
 684                         if (cl)
 685                                 break;
 686                         continue;
 687                 }
 688
 689                 add_new_bucket(c, ptrs, devs_may_alloc,
 690                                nr_effective, have_cache, flags, ob);
 691
 692                 if (*nr_effective >= nr_replicas) {
 693                         ret = 0;
 694                         break;
 695                 }
 696         }
 697
 698         return ret;
 699 }
 700
 701 /* Allocate from stripes: */
 702
 703 /*
 704  * if we can't allocate a new stripe because there are already too many
 705  * partially filled stripes, force allocating from an existing stripe even when
 706  * it's to a device we don't want:
 707  */
 708
 709 static int bucket_alloc_from_stripe(struct bch_fs *c,
 710                          struct open_buckets *ptrs,
 711                          struct write_point *wp,
 712                          struct bch_devs_mask *devs_may_alloc,
 713                          u16 target,
 714                          unsigned erasure_code,
 715                          unsigned nr_replicas,
 716                          unsigned *nr_effective,
 717                          bool *have_cache,
 718                          unsigned flags,
 719                          struct closure *cl)
 720 {
 721         struct dev_alloc_list devs_sorted;
 722         struct ec_stripe_head *h;
 723         struct open_bucket *ob;
 724         struct bch_dev *ca;
 725         unsigned i, ec_idx;
 726
 727         if (!erasure_code)
 728                 return 0;
 729
 730         if (nr_replicas < 2)
 731                 return 0;
 732
 733         if (ec_open_bucket(c, ptrs))
 734                 return 0;
 735
 736         h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1,
 737                                     wp == &c->copygc_write_point,
 738                                     cl);
 739         if (IS_ERR(h))
 740                 return -PTR_ERR(h);
 741         if (!h)
 742                 return 0;
 743
 744         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 745
 746         for (i = 0; i < devs_sorted.nr; i++)
 747                 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 748                         if (!h->s->blocks[ec_idx])
 749                                 continue;
 750
 751                         ob = c->open_buckets + h->s->blocks[ec_idx];
 752                         if (ob->dev == devs_sorted.devs[i] &&
 753                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 754                                 goto got_bucket;
 755                 }
 756         goto out_put_head;
 757 got_bucket:
 758         ca = bch_dev_bkey_exists(c, ob->dev);
 759
 760         ob->ec_idx      = ec_idx;
 761         ob->ec          = h->s;
 762
 763         add_new_bucket(c, ptrs, devs_may_alloc,
 764                        nr_effective, have_cache, flags, ob);
 765         atomic_inc(&h->s->pin);
 766 out_put_head:
 767         bch2_ec_stripe_head_put(c, h);
 768         return 0;
 769 }
 770
 771 /* Sector allocator */
 772
 773 static void get_buckets_from_writepoint(struct bch_fs *c,
 774                                         struct open_buckets *ptrs,
 775                                         struct write_point *wp,
 776                                         struct bch_devs_mask *devs_may_alloc,
 777                                         unsigned nr_replicas,
 778                                         unsigned *nr_effective,
 779                                         bool *have_cache,
 780                                         unsigned flags,
 781                                         bool need_ec)
 782 {
 783         struct open_buckets ptrs_skip = { .nr = 0 };
 784         struct open_bucket *ob;
 785         unsigned i;
 786
 787         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 788                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 789
 790                 if (*nr_effective < nr_replicas &&
 791                     test_bit(ob->dev, devs_may_alloc->d) &&
 792                     (ca->mi.durability ||
 793                      (wp->data_type == BCH_DATA_user && !*have_cache)) &&
 794                     (ob->ec || !need_ec)) {
 795                         add_new_bucket(c, ptrs, devs_may_alloc,
 796                                        nr_effective, have_cache,
 797                                        flags, ob);
 798                 } else {
 799                         ob_push(c, &ptrs_skip, ob);
 800                 }
 801         }
 802         wp->ptrs = ptrs_skip;
 803 }
 804
 805 static int open_bucket_add_buckets(struct bch_fs *c,
 806                         struct open_buckets *ptrs,
 807                         struct write_point *wp,
 808                         struct bch_devs_list *devs_have,
 809                         u16 target,
 810                         unsigned erasure_code,
 811                         unsigned nr_replicas,
 812                         unsigned *nr_effective,
 813                         bool *have_cache,
 814                         enum alloc_reserve reserve,
 815                         unsigned flags,
 816                         struct closure *_cl)
 817 {
 818         struct bch_devs_mask devs;
 819         struct open_bucket *ob;
 820         struct closure *cl = NULL;
 821         int ret;
 822         unsigned i;
 823
 824         rcu_read_lock();
 825         devs = target_rw_devs(c, wp->data_type, target);
 826         rcu_read_unlock();
 827
 828         /* Don't allocate from devices we already have pointers to: */
 829         for (i = 0; i < devs_have->nr; i++)
 830                 __clear_bit(devs_have->devs[i], devs.d);
 831
 832         open_bucket_for_each(c, ptrs, ob, i)
 833                 __clear_bit(ob->dev, devs.d);
 834
 835         if (erasure_code) {
 836                 if (!ec_open_bucket(c, ptrs)) {
 837                         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 838                                                     nr_replicas, nr_effective,
 839                                                     have_cache, flags, true);
 840                         if (*nr_effective >= nr_replicas)
 841                                 return 0;
 842                 }
 843
 844                 if (!ec_open_bucket(c, ptrs)) {
 845                         ret = bucket_alloc_from_stripe(c, ptrs, wp, &devs,
 846                                                  target, erasure_code,
 847                                                  nr_replicas, nr_effective,
 848                                                  have_cache, flags, _cl);
 849                         if (ret == -FREELIST_EMPTY ||
 850                             ret == -OPEN_BUCKETS_EMPTY)
 851                                 return ret;
 852                         if (*nr_effective >= nr_replicas)
 853                                 return 0;
 854                 }
 855         }
 856
 857         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 858                                     nr_replicas, nr_effective,
 859                                     have_cache, flags, false);
 860         if (*nr_effective >= nr_replicas)
 861                 return 0;
 862
 863 retry_blocking:
 864         /*
 865          * Try nonblocking first, so that if one device is full we'll try from
 866          * other devices:
 867          */
 868         ret = bch2_bucket_alloc_set(c, ptrs, &wp->stripe, &devs,
 869                                 nr_replicas, nr_effective, have_cache,
 870                                 reserve, flags, cl);
 871         if (ret && ret != -INSUFFICIENT_DEVICES && !cl && _cl) {
 872                 cl = _cl;
 873                 goto retry_blocking;
 874         }
 875
 876         return ret;
 877 }
 878
 879 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 880                                 struct open_buckets *obs)
 881 {
 882         struct open_buckets ptrs = { .nr = 0 };
 883         struct open_bucket *ob, *ob2;
 884         unsigned i, j;
 885
 886         open_bucket_for_each(c, obs, ob, i) {
 887                 bool drop = !ca || ob->dev == ca->dev_idx;
 888
 889                 if (!drop && ob->ec) {
 890                         mutex_lock(&ob->ec->lock);
 891                         for (j = 0; j < ob->ec->new_stripe.key.v.nr_blocks; j++) {
 892                                 if (!ob->ec->blocks[j])
 893                                         continue;
 894
 895                                 ob2 = c->open_buckets + ob->ec->blocks[j];
 896                                 drop |= ob2->dev == ca->dev_idx;
 897                         }
 898                         mutex_unlock(&ob->ec->lock);
 899                 }
 900
 901                 if (drop)
 902                         bch2_open_bucket_put(c, ob);
 903                 else
 904                         ob_push(c, &ptrs, ob);
 905         }
 906
 907         *obs = ptrs;
 908 }
 909
 910 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 911                           struct write_point *wp)
 912 {
 913         mutex_lock(&wp->lock);
 914         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
 915         mutex_unlock(&wp->lock);
 916 }
 917
 918 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 919                                                  unsigned long write_point)
 920 {
 921         unsigned hash =
 922                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
 923
 924         return &c->write_points_hash[hash];
 925 }
 926
 927 static struct write_point *__writepoint_find(struct hlist_head *head,
 928                                              unsigned long write_point)
 929 {
 930         struct write_point *wp;
 931
 932         rcu_read_lock();
 933         hlist_for_each_entry_rcu(wp, head, node)
 934                 if (wp->write_point == write_point)
 935                         goto out;
 936         wp = NULL;
 937 out:
 938         rcu_read_unlock();
 939         return wp;
 940 }
 941
 942 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
 943 {
 944         u64 stranded    = c->write_points_nr * c->bucket_size_max;
 945         u64 free        = bch2_fs_usage_read_short(c).free;
 946
 947         return stranded * factor > free;
 948 }
 949
 950 static bool try_increase_writepoints(struct bch_fs *c)
 951 {
 952         struct write_point *wp;
 953
 954         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
 955             too_many_writepoints(c, 32))
 956                 return false;
 957
 958         wp = c->write_points + c->write_points_nr++;
 959         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
 960         return true;
 961 }
 962
 963 static bool try_decrease_writepoints(struct bch_fs *c,
 964                                      unsigned old_nr)
 965 {
 966         struct write_point *wp;
 967
 968         mutex_lock(&c->write_points_hash_lock);
 969         if (c->write_points_nr < old_nr) {
 970                 mutex_unlock(&c->write_points_hash_lock);
 971                 return true;
 972         }
 973
 974         if (c->write_points_nr == 1 ||
 975             !too_many_writepoints(c, 8)) {
 976                 mutex_unlock(&c->write_points_hash_lock);
 977                 return false;
 978         }
 979
 980         wp = c->write_points + --c->write_points_nr;
 981
 982         hlist_del_rcu(&wp->node);
 983         mutex_unlock(&c->write_points_hash_lock);
 984
 985         bch2_writepoint_stop(c, NULL, wp);
 986         return true;
 987 }
 988
 989 static struct write_point *writepoint_find(struct bch_fs *c,
 990                                            unsigned long write_point)
 991 {
 992         struct write_point *wp, *oldest;
 993         struct hlist_head *head;
 994
 995         if (!(write_point & 1UL)) {
 996                 wp = (struct write_point *) write_point;
 997                 mutex_lock(&wp->lock);
 998                 return wp;
 999         }
1000
1001         head = writepoint_hash(c, write_point);
1002 restart_find:
1003         wp = __writepoint_find(head, write_point);
1004         if (wp) {
1005 lock_wp:
1006                 mutex_lock(&wp->lock);
1007                 if (wp->write_point == write_point)
1008                         goto out;
1009                 mutex_unlock(&wp->lock);
1010                 goto restart_find;
1011         }
1012 restart_find_oldest:
1013         oldest = NULL;
1014         for (wp = c->write_points;
1015              wp < c->write_points + c->write_points_nr; wp++)
1016                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1017                         oldest = wp;
1018
1019         mutex_lock(&oldest->lock);
1020         mutex_lock(&c->write_points_hash_lock);
1021         if (oldest >= c->write_points + c->write_points_nr ||
1022             try_increase_writepoints(c)) {
1023                 mutex_unlock(&c->write_points_hash_lock);
1024                 mutex_unlock(&oldest->lock);
1025                 goto restart_find_oldest;
1026         }
1027
1028         wp = __writepoint_find(head, write_point);
1029         if (wp && wp != oldest) {
1030                 mutex_unlock(&c->write_points_hash_lock);
1031                 mutex_unlock(&oldest->lock);
1032                 goto lock_wp;
1033         }
1034
1035         wp = oldest;
1036         hlist_del_rcu(&wp->node);
1037         wp->write_point = write_point;
1038         hlist_add_head_rcu(&wp->node, head);
1039         mutex_unlock(&c->write_points_hash_lock);
1040 out:
1041         wp->last_used = sched_clock();
1042         return wp;
1043 }
1044
1045 /*
1046  * Get us an open_bucket we can allocate from, return with it locked:
1047  */
1048 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
1049                                 unsigned target,
1050                                 unsigned erasure_code,
1051                                 struct write_point_specifier write_point,
1052                                 struct bch_devs_list *devs_have,
1053                                 unsigned nr_replicas,
1054                                 unsigned nr_replicas_required,
1055                                 enum alloc_reserve reserve,
1056                                 unsigned flags,
1057                                 struct closure *cl)
1058 {
1059         struct write_point *wp;
1060         struct open_bucket *ob;
1061         struct open_buckets ptrs;
1062         unsigned nr_effective, write_points_nr;
1063         unsigned ob_flags = 0;
1064         bool have_cache;
1065         int ret;
1066         int i;
1067
1068         if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
1069                 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
1070
1071         BUG_ON(!nr_replicas || !nr_replicas_required);
1072 retry:
1073         ptrs.nr         = 0;
1074         nr_effective    = 0;
1075         write_points_nr = c->write_points_nr;
1076         have_cache      = false;
1077
1078         wp = writepoint_find(c, write_point.v);
1079
1080         if (wp->data_type == BCH_DATA_user)
1081                 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
1082
1083         /* metadata may not allocate on cache devices: */
1084         if (wp->data_type != BCH_DATA_user)
1085                 have_cache = true;
1086
1087         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1088                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
1089                                               target, erasure_code,
1090                                               nr_replicas, &nr_effective,
1091                                               &have_cache, reserve,
1092                                               ob_flags, cl);
1093         } else {
1094                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
1095                                               target, erasure_code,
1096                                               nr_replicas, &nr_effective,
1097                                               &have_cache, reserve,
1098                                               ob_flags, NULL);
1099                 if (!ret)
1100                         goto alloc_done;
1101
1102                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
1103                                               0, erasure_code,
1104                                               nr_replicas, &nr_effective,
1105                                               &have_cache, reserve,
1106                                               ob_flags, cl);
1107         }
1108 alloc_done:
1109         BUG_ON(!ret && nr_effective < nr_replicas);
1110
1111         if (erasure_code && !ec_open_bucket(c, &ptrs))
1112                 pr_debug("failed to get ec bucket: ret %u", ret);
1113
1114         if (ret == -INSUFFICIENT_DEVICES &&
1115             nr_effective >= nr_replicas_required)
1116                 ret = 0;
1117
1118         if (ret)
1119                 goto err;
1120
1121         /* Free buckets we didn't use: */
1122         open_bucket_for_each(c, &wp->ptrs, ob, i)
1123                 open_bucket_free_unused(c, wp, ob);
1124
1125         wp->ptrs = ptrs;
1126
1127         wp->sectors_free = UINT_MAX;
1128
1129         open_bucket_for_each(c, &wp->ptrs, ob, i)
1130                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1131
1132         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1133
1134         return wp;
1135 err:
1136         open_bucket_for_each(c, &wp->ptrs, ob, i)
1137                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1138                         ob_push(c, &ptrs, ob);
1139                 else
1140                         open_bucket_free_unused(c, wp, ob);
1141         wp->ptrs = ptrs;
1142
1143         mutex_unlock(&wp->lock);
1144
1145         if (ret == -FREELIST_EMPTY &&
1146             try_decrease_writepoints(c, write_points_nr))
1147                 goto retry;
1148
1149         switch (ret) {
1150         case -OPEN_BUCKETS_EMPTY:
1151         case -FREELIST_EMPTY:
1152                 return cl ? ERR_PTR(-EAGAIN) : ERR_PTR(-ENOSPC);
1153         case -INSUFFICIENT_DEVICES:
1154                 return ERR_PTR(-EROFS);
1155         default:
1156                 return ERR_PTR(ret);
1157         }
1158 }
1159
1160 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1161 {
1162         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1163
1164         return (struct bch_extent_ptr) {
1165                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
1166                 .gen    = ob->gen,
1167                 .dev    = ob->dev,
1168                 .offset = bucket_to_sector(ca, ob->bucket) +
1169                         ca->mi.bucket_size -
1170                         ob->sectors_free,
1171         };
1172 }
1173
1174 /*
1175  * Append pointers to the space we just allocated to @k, and mark @sectors space
1176  * as allocated out of @ob
1177  */
1178 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1179                                     struct bkey_i *k, unsigned sectors,
1180                                     bool cached)
1181
1182 {
1183         struct open_bucket *ob;
1184         unsigned i;
1185
1186         BUG_ON(sectors > wp->sectors_free);
1187         wp->sectors_free -= sectors;
1188
1189         open_bucket_for_each(c, &wp->ptrs, ob, i) {
1190                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1191                 struct bch_extent_ptr ptr = bch2_ob_ptr(c, ob);
1192
1193                 ptr.cached = cached ||
1194                         (!ca->mi.durability &&
1195                          wp->data_type == BCH_DATA_user);
1196
1197                 bch2_bkey_append_ptr(k, ptr);
1198
1199                 BUG_ON(sectors > ob->sectors_free);
1200                 ob->sectors_free -= sectors;
1201         }
1202 }
1203
1204 /*
1205  * Append pointers to the space we just allocated to @k, and mark @sectors space
1206  * as allocated out of @ob
1207  */
1208 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1209 {
1210         struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
1211         struct open_bucket *ob;
1212         unsigned i;
1213
1214         open_bucket_for_each(c, &wp->ptrs, ob, i)
1215                 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
1216         wp->ptrs = keep;
1217
1218         mutex_unlock(&wp->lock);
1219
1220         bch2_open_buckets_put(c, &ptrs);
1221 }
1222
1223 static inline void writepoint_init(struct write_point *wp,
1224                                    enum bch_data_type type)
1225 {
1226         mutex_init(&wp->lock);
1227         wp->data_type = type;
1228 }
1229
1230 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1231 {
1232         struct open_bucket *ob;
1233         struct write_point *wp;
1234
1235         mutex_init(&c->write_points_hash_lock);
1236         c->write_points_nr = ARRAY_SIZE(c->write_points);
1237
1238         /* open bucket 0 is a sentinal NULL: */
1239         spin_lock_init(&c->open_buckets[0].lock);
1240
1241         for (ob = c->open_buckets + 1;
1242              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1243                 spin_lock_init(&ob->lock);
1244                 c->open_buckets_nr_free++;
1245
1246                 ob->freelist = c->open_buckets_freelist;
1247                 c->open_buckets_freelist = ob - c->open_buckets;
1248         }
1249
1250         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
1251         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
1252         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
1253
1254         for (wp = c->write_points;
1255              wp < c->write_points + c->write_points_nr; wp++) {
1256                 writepoint_init(wp, BCH_DATA_user);
1257
1258                 wp->last_used   = sched_clock();
1259                 wp->write_point = (unsigned long) wp;
1260                 hlist_add_head_rcu(&wp->node,
1261                                    writepoint_hash(c, wp->write_point));
1262         }
1263 }
1264
1265 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1266 {
1267         struct open_bucket *ob;
1268
1269         for (ob = c->open_buckets;
1270              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1271              ob++) {
1272                 spin_lock(&ob->lock);
1273                 if (ob->valid && !ob->on_partial_list) {
1274                         prt_printf(out, "%zu ref %u type %s %u:%llu:%u\n",
1275                                ob - c->open_buckets,
1276                                atomic_read(&ob->pin),
1277                                bch2_data_types[ob->data_type],
1278                                ob->dev, ob->bucket, ob->gen);
1279                 }
1280                 spin_unlock(&ob->lock);
1281         }
1282 }