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
 230                 trace_open_bucket_alloc_fail(ca, bch2_alloc_reserves[reserve]);
 231                 return ERR_PTR(-OPEN_BUCKETS_EMPTY);
 232         }
 233
 234         /* Recheck under lock: */
 235         if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
 236                 spin_unlock(&c->freelist_lock);
 237                 (*skipped_open)++;
 238                 return NULL;
 239         }
 240
 241         ob = bch2_open_bucket_alloc(c);
 242
 243         spin_lock(&ob->lock);
 244
 245         ob->valid       = true;
 246         ob->sectors_free = ca->mi.bucket_size;
 247         ob->alloc_reserve = reserve;
 248         ob->dev         = ca->dev_idx;
 249         ob->gen         = a->gen;
 250         ob->bucket      = bucket;
 251         spin_unlock(&ob->lock);
 252
 253         ca->nr_open_buckets++;
 254         bch2_open_bucket_hash_add(c, ob);
 255
 256         if (c->blocked_allocate_open_bucket) {
 257                 bch2_time_stats_update(
 258                         &c->times[BCH_TIME_blocked_allocate_open_bucket],
 259                         c->blocked_allocate_open_bucket);
 260                 c->blocked_allocate_open_bucket = 0;
 261         }
 262
 263         if (c->blocked_allocate) {
 264                 bch2_time_stats_update(
 265                         &c->times[BCH_TIME_blocked_allocate],
 266                         c->blocked_allocate);
 267                 c->blocked_allocate = 0;
 268         }
 269
 270         spin_unlock(&c->freelist_lock);
 271
 272         trace_bucket_alloc(ca, bch2_alloc_reserves[reserve]);
 273         return ob;
 274 }
 275
 276 static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bch_dev *ca,
 277                                             enum alloc_reserve reserve, u64 free_entry,
 278                                             u64 *skipped_open,
 279                                             u64 *skipped_need_journal_commit,
 280                                             u64 *skipped_nouse,
 281                                             struct closure *cl)
 282 {
 283         struct bch_fs *c = trans->c;
 284         struct btree_iter iter;
 285         struct bkey_s_c k;
 286         struct open_bucket *ob;
 287         struct bch_alloc_v4 a;
 288         u64 b = free_entry & ~(~0ULL << 56);
 289         unsigned genbits = free_entry >> 56;
 290         struct printbuf buf = PRINTBUF;
 291         int ret;
 292
 293         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS(ca->dev_idx, b), BTREE_ITER_CACHED);
 294         k = bch2_btree_iter_peek_slot(&iter);
 295         ret = bkey_err(k);
 296         if (ret) {
 297                 ob = ERR_PTR(ret);
 298                 goto err;
 299         }
 300
 301         bch2_alloc_to_v4(k, &a);
 302
 303         if (bch2_fs_inconsistent_on(bucket_state(a) != BUCKET_free, c,
 304                         "non free bucket in freespace btree (state %s)\n"
 305                         "  %s\n"
 306                         "  at %llu (genbits %u)",
 307                         bch2_bucket_states[bucket_state(a)],
 308                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
 309                         free_entry, genbits)) {
 310                 ob = ERR_PTR(-EIO);
 311                 goto err;
 312         }
 313
 314         if (bch2_fs_inconsistent_on(genbits != (alloc_freespace_genbits(a) >> 56), c,
 315                         "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
 316                         "  %s",
 317                         genbits, alloc_freespace_genbits(a) >> 56,
 318                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 319                 ob = ERR_PTR(-EIO);
 320                 goto err;
 321         }
 322
 323         if (bch2_fs_inconsistent_on(b < ca->mi.first_bucket || b >= ca->mi.nbuckets, c,
 324                         "freespace btree has bucket outside allowed range (got %llu, valid %u-%llu)",
 325                         b, ca->mi.first_bucket, ca->mi.nbuckets)) {
 326                 ob = ERR_PTR(-EIO);
 327                 goto err;
 328         }
 329
 330         ob = __try_alloc_bucket(c, ca, b, reserve, &a,
 331                                 skipped_open,
 332                                 skipped_need_journal_commit,
 333                                 skipped_nouse,
 334                                 cl);
 335 err:
 336         bch2_trans_iter_exit(trans, &iter);
 337         printbuf_exit(&buf);
 338         return ob;
 339 }
 340
 341 static struct open_bucket *try_alloc_partial_bucket(struct bch_fs *c, struct bch_dev *ca,
 342                                                     enum alloc_reserve reserve)
 343 {
 344         struct open_bucket *ob;
 345         int i;
 346
 347         spin_lock(&c->freelist_lock);
 348
 349         for (i = ca->open_buckets_partial_nr - 1; i >= 0; --i) {
 350                 ob = c->open_buckets + ca->open_buckets_partial[i];
 351
 352                 if (reserve <= ob->alloc_reserve) {
 353                         array_remove_item(ca->open_buckets_partial,
 354                                           ca->open_buckets_partial_nr,
 355                                           i);
 356                         ob->on_partial_list = false;
 357                         ob->alloc_reserve = reserve;
 358                         spin_unlock(&c->freelist_lock);
 359                         return ob;
 360                 }
 361         }
 362
 363         spin_unlock(&c->freelist_lock);
 364         return NULL;
 365 }
 366
 367 /*
 368  * This path is for before the freespace btree is initialized:
 369  *
 370  * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
 371  * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
 372  */
 373 static noinline struct open_bucket *
 374 bch2_bucket_alloc_trans_early(struct btree_trans *trans,
 375                               struct bch_dev *ca,
 376                               enum alloc_reserve reserve,
 377                               u64 *cur_bucket,
 378                               u64 *buckets_seen,
 379                               u64 *skipped_open,
 380                               u64 *skipped_need_journal_commit,
 381                               u64 *skipped_nouse,
 382                               struct closure *cl)
 383 {
 384         struct btree_iter iter;
 385         struct bkey_s_c k;
 386         struct open_bucket *ob = NULL;
 387         int ret;
 388
 389         *cur_bucket = max_t(u64, *cur_bucket, ca->mi.first_bucket);
 390         *cur_bucket = max_t(u64, *cur_bucket, ca->new_fs_bucket_idx);
 391
 392         for_each_btree_key(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, *cur_bucket),
 393                            BTREE_ITER_SLOTS, k, ret) {
 394                 struct bch_alloc_v4 a;
 395
 396                 if (bkey_cmp(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)) >= 0)
 397                         break;
 398
 399                 if (ca->new_fs_bucket_idx &&
 400                     is_superblock_bucket(ca, k.k->p.offset))
 401                         continue;
 402
 403                 bch2_alloc_to_v4(k, &a);
 404
 405                 if (bucket_state(a) != BUCKET_free)
 406                         continue;
 407
 408                 (*buckets_seen)++;
 409
 410                 ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, reserve, &a,
 411                                         skipped_open,
 412                                         skipped_need_journal_commit,
 413                                         skipped_nouse,
 414                                         cl);
 415                 if (ob)
 416                         break;
 417         }
 418         bch2_trans_iter_exit(trans, &iter);
 419
 420         *cur_bucket = iter.pos.offset;
 421
 422         return ob ?: ERR_PTR(ret ?: -FREELIST_EMPTY);
 423 }
 424
 425 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
 426                                                    struct bch_dev *ca,
 427                                                    enum alloc_reserve reserve,
 428                                                    u64 *cur_bucket,
 429                                                    u64 *buckets_seen,
 430                                                    u64 *skipped_open,
 431                                                    u64 *skipped_need_journal_commit,
 432                                                    u64 *skipped_nouse,
 433                                                    struct closure *cl)
 434 {
 435         struct btree_iter iter;
 436         struct bkey_s_c k;
 437         struct open_bucket *ob = NULL;
 438         int ret;
 439
 440         if (unlikely(!ca->mi.freespace_initialized))
 441                 return bch2_bucket_alloc_trans_early(trans, ca, reserve,
 442                                                      cur_bucket,
 443                                                      buckets_seen,
 444                                                      skipped_open,
 445                                                      skipped_need_journal_commit,
 446                                                      skipped_nouse,
 447                                                      cl);
 448
 449         BUG_ON(ca->new_fs_bucket_idx);
 450
 451         for_each_btree_key(trans, iter, BTREE_ID_freespace,
 452                            POS(ca->dev_idx, *cur_bucket), 0, k, ret) {
 453                 if (k.k->p.inode != ca->dev_idx)
 454                         break;
 455
 456                 for (*cur_bucket = max(*cur_bucket, bkey_start_offset(k.k));
 457                      *cur_bucket != k.k->p.offset && !ob;
 458                      (*cur_bucket)++) {
 459                         if (btree_trans_too_many_iters(trans)) {
 460                                 ob = ERR_PTR(-EINTR);
 461                                 break;
 462                         }
 463
 464                         (*buckets_seen)++;
 465
 466                         ob = try_alloc_bucket(trans, ca, reserve,
 467                                               *cur_bucket,
 468                                               skipped_open,
 469                                               skipped_need_journal_commit,
 470                                               skipped_nouse,
 471                                               cl);
 472                 }
 473                 if (ob)
 474                         break;
 475         }
 476         bch2_trans_iter_exit(trans, &iter);
 477
 478         return ob ?: ERR_PTR(ret);
 479 }
 480
 481 /**
 482  * bch_bucket_alloc - allocate a single bucket from a specific device
 483  *
 484  * Returns index of bucket on success, 0 on failure
 485  * */
 486 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 487                                       enum alloc_reserve reserve,
 488                                       bool may_alloc_partial,
 489                                       struct closure *cl)
 490 {
 491         struct open_bucket *ob = NULL;
 492         u64 avail = dev_buckets_available(ca, reserve);
 493         u64 cur_bucket = 0;
 494         u64 buckets_seen = 0;
 495         u64 skipped_open = 0;
 496         u64 skipped_need_journal_commit = 0;
 497         u64 skipped_nouse = 0;
 498         int ret;
 499
 500         if (may_alloc_partial) {
 501                 ob = try_alloc_partial_bucket(c, ca, reserve);
 502                 if (ob)
 503                         return ob;
 504         }
 505 again:
 506         if (!avail) {
 507                 if (cl) {
 508                         closure_wait(&c->freelist_wait, cl);
 509                         /* recheck after putting ourself on waitlist */
 510                         avail = dev_buckets_available(ca, reserve);
 511                         if (avail) {
 512                                 closure_wake_up(&c->freelist_wait);
 513                                 goto again;
 514                         }
 515                 }
 516
 517                 if (!c->blocked_allocate)
 518                         c->blocked_allocate = local_clock();
 519
 520                 ob = ERR_PTR(-FREELIST_EMPTY);
 521                 goto err;
 522         }
 523
 524         ret = bch2_trans_do(c, NULL, NULL, 0,
 525                         PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, reserve,
 526                                                         &cur_bucket,
 527                                                         &buckets_seen,
 528                                                         &skipped_open,
 529                                                         &skipped_need_journal_commit,
 530                                                         &skipped_nouse,
 531                                                         cl)));
 532
 533         if (skipped_need_journal_commit * 2 > avail)
 534                 bch2_journal_flush_async(&c->journal, NULL);
 535 err:
 536         if (!ob)
 537                 ob = ERR_PTR(ret ?: -FREELIST_EMPTY);
 538
 539         if (IS_ERR(ob)) {
 540                 trace_bucket_alloc_fail(ca, bch2_alloc_reserves[reserve], avail,
 541                                         buckets_seen,
 542                                         skipped_open,
 543                                         skipped_need_journal_commit,
 544                                         skipped_nouse,
 545                                         cl == NULL, PTR_ERR(ob));
 546                 atomic_long_inc(&c->bucket_alloc_fail);
 547         }
 548
 549         return ob;
 550 }
 551
 552 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 553                             unsigned l, unsigned r)
 554 {
 555         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 556                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 557 }
 558
 559 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 560
 561 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 562                                           struct dev_stripe_state *stripe,
 563                                           struct bch_devs_mask *devs)
 564 {
 565         struct dev_alloc_list ret = { .nr = 0 };
 566         unsigned i;
 567
 568         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 569                 ret.devs[ret.nr++] = i;
 570
 571         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 572         return ret;
 573 }
 574
 575 void bch2_dev_stripe_increment(struct bch_dev *ca,
 576                                struct dev_stripe_state *stripe)
 577 {
 578         u64 *v = stripe->next_alloc + ca->dev_idx;
 579         u64 free_space = dev_buckets_available(ca, RESERVE_none);
 580         u64 free_space_inv = free_space
 581                 ? div64_u64(1ULL << 48, free_space)
 582                 : 1ULL << 48;
 583         u64 scale = *v / 4;
 584
 585         if (*v + free_space_inv >= *v)
 586                 *v += free_space_inv;
 587         else
 588                 *v = U64_MAX;
 589
 590         for (v = stripe->next_alloc;
 591              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 592                 *v = *v < scale ? 0 : *v - scale;
 593 }
 594
 595 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 596 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 597
 598 static void add_new_bucket(struct bch_fs *c,
 599                            struct open_buckets *ptrs,
 600                            struct bch_devs_mask *devs_may_alloc,
 601                            unsigned *nr_effective,
 602                            bool *have_cache,
 603                            unsigned flags,
 604                            struct open_bucket *ob)
 605 {
 606         unsigned durability =
 607                 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
 608
 609         __clear_bit(ob->dev, devs_may_alloc->d);
 610         *nr_effective   += (flags & BUCKET_ALLOC_USE_DURABILITY)
 611                 ? durability : 1;
 612         *have_cache     |= !durability;
 613
 614         ob_push(c, ptrs, ob);
 615 }
 616
 617 int bch2_bucket_alloc_set(struct bch_fs *c,
 618                       struct open_buckets *ptrs,
 619                       struct dev_stripe_state *stripe,
 620                       struct bch_devs_mask *devs_may_alloc,
 621                       unsigned nr_replicas,
 622                       unsigned *nr_effective,
 623                       bool *have_cache,
 624                       enum alloc_reserve reserve,
 625                       unsigned flags,
 626                       struct closure *cl)
 627 {
 628         struct dev_alloc_list devs_sorted =
 629                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 630         unsigned dev;
 631         struct bch_dev *ca;
 632         int ret = -INSUFFICIENT_DEVICES;
 633         unsigned i;
 634
 635         BUG_ON(*nr_effective >= nr_replicas);
 636
 637         for (i = 0; i < devs_sorted.nr; i++) {
 638                 struct open_bucket *ob;
 639
 640                 dev = devs_sorted.devs[i];
 641
 642                 rcu_read_lock();
 643                 ca = rcu_dereference(c->devs[dev]);
 644                 if (ca)
 645                         percpu_ref_get(&ca->ref);
 646                 rcu_read_unlock();
 647
 648                 if (!ca)
 649                         continue;
 650
 651                 if (!ca->mi.durability && *have_cache) {
 652                         percpu_ref_put(&ca->ref);
 653                         continue;
 654                 }
 655
 656                 ob = bch2_bucket_alloc(c, ca, reserve,
 657                                 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
 658                 if (!IS_ERR(ob))
 659                         bch2_dev_stripe_increment(ca, stripe);
 660                 percpu_ref_put(&ca->ref);
 661
 662                 if (IS_ERR(ob)) {
 663                         ret = PTR_ERR(ob);
 664
 665                         if (cl)
 666                                 break;
 667                         continue;
 668                 }
 669
 670                 add_new_bucket(c, ptrs, devs_may_alloc,
 671                                nr_effective, have_cache, flags, ob);
 672
 673                 if (*nr_effective >= nr_replicas) {
 674                         ret = 0;
 675                         break;
 676                 }
 677         }
 678
 679         return ret;
 680 }
 681
 682 /* Allocate from stripes: */
 683
 684 /*
 685  * if we can't allocate a new stripe because there are already too many
 686  * partially filled stripes, force allocating from an existing stripe even when
 687  * it's to a device we don't want:
 688  */
 689
 690 static int bucket_alloc_from_stripe(struct bch_fs *c,
 691                          struct open_buckets *ptrs,
 692                          struct write_point *wp,
 693                          struct bch_devs_mask *devs_may_alloc,
 694                          u16 target,
 695                          unsigned erasure_code,
 696                          unsigned nr_replicas,
 697                          unsigned *nr_effective,
 698                          bool *have_cache,
 699                          unsigned flags,
 700                          struct closure *cl)
 701 {
 702         struct dev_alloc_list devs_sorted;
 703         struct ec_stripe_head *h;
 704         struct open_bucket *ob;
 705         struct bch_dev *ca;
 706         unsigned i, ec_idx;
 707
 708         if (!erasure_code)
 709                 return 0;
 710
 711         if (nr_replicas < 2)
 712                 return 0;
 713
 714         if (ec_open_bucket(c, ptrs))
 715                 return 0;
 716
 717         h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1,
 718                                     wp == &c->copygc_write_point,
 719                                     cl);
 720         if (IS_ERR(h))
 721                 return -PTR_ERR(h);
 722         if (!h)
 723                 return 0;
 724
 725         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 726
 727         for (i = 0; i < devs_sorted.nr; i++)
 728                 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 729                         if (!h->s->blocks[ec_idx])
 730                                 continue;
 731
 732                         ob = c->open_buckets + h->s->blocks[ec_idx];
 733                         if (ob->dev == devs_sorted.devs[i] &&
 734                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 735                                 goto got_bucket;
 736                 }
 737         goto out_put_head;
 738 got_bucket:
 739         ca = bch_dev_bkey_exists(c, ob->dev);
 740
 741         ob->ec_idx      = ec_idx;
 742         ob->ec          = h->s;
 743
 744         add_new_bucket(c, ptrs, devs_may_alloc,
 745                        nr_effective, have_cache, flags, ob);
 746         atomic_inc(&h->s->pin);
 747 out_put_head:
 748         bch2_ec_stripe_head_put(c, h);
 749         return 0;
 750 }
 751
 752 /* Sector allocator */
 753
 754 static void get_buckets_from_writepoint(struct bch_fs *c,
 755                                         struct open_buckets *ptrs,
 756                                         struct write_point *wp,
 757                                         struct bch_devs_mask *devs_may_alloc,
 758                                         unsigned nr_replicas,
 759                                         unsigned *nr_effective,
 760                                         bool *have_cache,
 761                                         unsigned flags,
 762                                         bool need_ec)
 763 {
 764         struct open_buckets ptrs_skip = { .nr = 0 };
 765         struct open_bucket *ob;
 766         unsigned i;
 767
 768         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 769                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 770
 771                 if (*nr_effective < nr_replicas &&
 772                     test_bit(ob->dev, devs_may_alloc->d) &&
 773                     (ca->mi.durability ||
 774                      (wp->data_type == BCH_DATA_user && !*have_cache)) &&
 775                     (ob->ec || !need_ec)) {
 776                         add_new_bucket(c, ptrs, devs_may_alloc,
 777                                        nr_effective, have_cache,
 778                                        flags, ob);
 779                 } else {
 780                         ob_push(c, &ptrs_skip, ob);
 781                 }
 782         }
 783         wp->ptrs = ptrs_skip;
 784 }
 785
 786 static int open_bucket_add_buckets(struct bch_fs *c,
 787                         struct open_buckets *ptrs,
 788                         struct write_point *wp,
 789                         struct bch_devs_list *devs_have,
 790                         u16 target,
 791                         unsigned erasure_code,
 792                         unsigned nr_replicas,
 793                         unsigned *nr_effective,
 794                         bool *have_cache,
 795                         enum alloc_reserve reserve,
 796                         unsigned flags,
 797                         struct closure *_cl)
 798 {
 799         struct bch_devs_mask devs;
 800         struct open_bucket *ob;
 801         struct closure *cl = NULL;
 802         int ret;
 803         unsigned i;
 804
 805         rcu_read_lock();
 806         devs = target_rw_devs(c, wp->data_type, target);
 807         rcu_read_unlock();
 808
 809         /* Don't allocate from devices we already have pointers to: */
 810         for (i = 0; i < devs_have->nr; i++)
 811                 __clear_bit(devs_have->devs[i], devs.d);
 812
 813         open_bucket_for_each(c, ptrs, ob, i)
 814                 __clear_bit(ob->dev, devs.d);
 815
 816         if (erasure_code) {
 817                 if (!ec_open_bucket(c, ptrs)) {
 818                         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 819                                                     nr_replicas, nr_effective,
 820                                                     have_cache, flags, true);
 821                         if (*nr_effective >= nr_replicas)
 822                                 return 0;
 823                 }
 824
 825                 if (!ec_open_bucket(c, ptrs)) {
 826                         ret = bucket_alloc_from_stripe(c, ptrs, wp, &devs,
 827                                                  target, erasure_code,
 828                                                  nr_replicas, nr_effective,
 829                                                  have_cache, flags, _cl);
 830                         if (ret == -FREELIST_EMPTY ||
 831                             ret == -OPEN_BUCKETS_EMPTY)
 832                                 return ret;
 833                         if (*nr_effective >= nr_replicas)
 834                                 return 0;
 835                 }
 836         }
 837
 838         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 839                                     nr_replicas, nr_effective,
 840                                     have_cache, flags, false);
 841         if (*nr_effective >= nr_replicas)
 842                 return 0;
 843
 844 retry_blocking:
 845         /*
 846          * Try nonblocking first, so that if one device is full we'll try from
 847          * other devices:
 848          */
 849         ret = bch2_bucket_alloc_set(c, ptrs, &wp->stripe, &devs,
 850                                 nr_replicas, nr_effective, have_cache,
 851                                 reserve, flags, cl);
 852         if (ret && ret != -INSUFFICIENT_DEVICES && !cl && _cl) {
 853                 cl = _cl;
 854                 goto retry_blocking;
 855         }
 856
 857         return ret;
 858 }
 859
 860 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 861                                 struct open_buckets *obs)
 862 {
 863         struct open_buckets ptrs = { .nr = 0 };
 864         struct open_bucket *ob, *ob2;
 865         unsigned i, j;
 866
 867         open_bucket_for_each(c, obs, ob, i) {
 868                 bool drop = !ca || ob->dev == ca->dev_idx;
 869
 870                 if (!drop && ob->ec) {
 871                         mutex_lock(&ob->ec->lock);
 872                         for (j = 0; j < ob->ec->new_stripe.key.v.nr_blocks; j++) {
 873                                 if (!ob->ec->blocks[j])
 874                                         continue;
 875
 876                                 ob2 = c->open_buckets + ob->ec->blocks[j];
 877                                 drop |= ob2->dev == ca->dev_idx;
 878                         }
 879                         mutex_unlock(&ob->ec->lock);
 880                 }
 881
 882                 if (drop)
 883                         bch2_open_bucket_put(c, ob);
 884                 else
 885                         ob_push(c, &ptrs, ob);
 886         }
 887
 888         *obs = ptrs;
 889 }
 890
 891 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 892                           struct write_point *wp)
 893 {
 894         mutex_lock(&wp->lock);
 895         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
 896         mutex_unlock(&wp->lock);
 897 }
 898
 899 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 900                                                  unsigned long write_point)
 901 {
 902         unsigned hash =
 903                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
 904
 905         return &c->write_points_hash[hash];
 906 }
 907
 908 static struct write_point *__writepoint_find(struct hlist_head *head,
 909                                              unsigned long write_point)
 910 {
 911         struct write_point *wp;
 912
 913         rcu_read_lock();
 914         hlist_for_each_entry_rcu(wp, head, node)
 915                 if (wp->write_point == write_point)
 916                         goto out;
 917         wp = NULL;
 918 out:
 919         rcu_read_unlock();
 920         return wp;
 921 }
 922
 923 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
 924 {
 925         u64 stranded    = c->write_points_nr * c->bucket_size_max;
 926         u64 free        = bch2_fs_usage_read_short(c).free;
 927
 928         return stranded * factor > free;
 929 }
 930
 931 static bool try_increase_writepoints(struct bch_fs *c)
 932 {
 933         struct write_point *wp;
 934
 935         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
 936             too_many_writepoints(c, 32))
 937                 return false;
 938
 939         wp = c->write_points + c->write_points_nr++;
 940         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
 941         return true;
 942 }
 943
 944 static bool try_decrease_writepoints(struct bch_fs *c,
 945                                      unsigned old_nr)
 946 {
 947         struct write_point *wp;
 948
 949         mutex_lock(&c->write_points_hash_lock);
 950         if (c->write_points_nr < old_nr) {
 951                 mutex_unlock(&c->write_points_hash_lock);
 952                 return true;
 953         }
 954
 955         if (c->write_points_nr == 1 ||
 956             !too_many_writepoints(c, 8)) {
 957                 mutex_unlock(&c->write_points_hash_lock);
 958                 return false;
 959         }
 960
 961         wp = c->write_points + --c->write_points_nr;
 962
 963         hlist_del_rcu(&wp->node);
 964         mutex_unlock(&c->write_points_hash_lock);
 965
 966         bch2_writepoint_stop(c, NULL, wp);
 967         return true;
 968 }
 969
 970 static struct write_point *writepoint_find(struct bch_fs *c,
 971                                            unsigned long write_point)
 972 {
 973         struct write_point *wp, *oldest;
 974         struct hlist_head *head;
 975
 976         if (!(write_point & 1UL)) {
 977                 wp = (struct write_point *) write_point;
 978                 mutex_lock(&wp->lock);
 979                 return wp;
 980         }
 981
 982         head = writepoint_hash(c, write_point);
 983 restart_find:
 984         wp = __writepoint_find(head, write_point);
 985         if (wp) {
 986 lock_wp:
 987                 mutex_lock(&wp->lock);
 988                 if (wp->write_point == write_point)
 989                         goto out;
 990                 mutex_unlock(&wp->lock);
 991                 goto restart_find;
 992         }
 993 restart_find_oldest:
 994         oldest = NULL;
 995         for (wp = c->write_points;
 996              wp < c->write_points + c->write_points_nr; wp++)
 997                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
 998                         oldest = wp;
 999
1000         mutex_lock(&oldest->lock);
1001         mutex_lock(&c->write_points_hash_lock);
1002         if (oldest >= c->write_points + c->write_points_nr ||
1003             try_increase_writepoints(c)) {
1004                 mutex_unlock(&c->write_points_hash_lock);
1005                 mutex_unlock(&oldest->lock);
1006                 goto restart_find_oldest;
1007         }
1008
1009         wp = __writepoint_find(head, write_point);
1010         if (wp && wp != oldest) {
1011                 mutex_unlock(&c->write_points_hash_lock);
1012                 mutex_unlock(&oldest->lock);
1013                 goto lock_wp;
1014         }
1015
1016         wp = oldest;
1017         hlist_del_rcu(&wp->node);
1018         wp->write_point = write_point;
1019         hlist_add_head_rcu(&wp->node, head);
1020         mutex_unlock(&c->write_points_hash_lock);
1021 out:
1022         wp->last_used = sched_clock();
1023         return wp;
1024 }
1025
1026 /*
1027  * Get us an open_bucket we can allocate from, return with it locked:
1028  */
1029 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
1030                                 unsigned target,
1031                                 unsigned erasure_code,
1032                                 struct write_point_specifier write_point,
1033                                 struct bch_devs_list *devs_have,
1034                                 unsigned nr_replicas,
1035                                 unsigned nr_replicas_required,
1036                                 enum alloc_reserve reserve,
1037                                 unsigned flags,
1038                                 struct closure *cl)
1039 {
1040         struct write_point *wp;
1041         struct open_bucket *ob;
1042         struct open_buckets ptrs;
1043         unsigned nr_effective, write_points_nr;
1044         unsigned ob_flags = 0;
1045         bool have_cache;
1046         int ret;
1047         int i;
1048
1049         if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
1050                 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
1051
1052         BUG_ON(!nr_replicas || !nr_replicas_required);
1053 retry:
1054         ptrs.nr         = 0;
1055         nr_effective    = 0;
1056         write_points_nr = c->write_points_nr;
1057         have_cache      = false;
1058
1059         wp = writepoint_find(c, write_point.v);
1060
1061         if (wp->data_type == BCH_DATA_user)
1062                 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
1063
1064         /* metadata may not allocate on cache devices: */
1065         if (wp->data_type != BCH_DATA_user)
1066                 have_cache = true;
1067
1068         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1069                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
1070                                               target, erasure_code,
1071                                               nr_replicas, &nr_effective,
1072                                               &have_cache, reserve,
1073                                               ob_flags, cl);
1074         } else {
1075                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
1076                                               target, erasure_code,
1077                                               nr_replicas, &nr_effective,
1078                                               &have_cache, reserve,
1079                                               ob_flags, NULL);
1080                 if (!ret)
1081                         goto alloc_done;
1082
1083                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
1084                                               0, erasure_code,
1085                                               nr_replicas, &nr_effective,
1086                                               &have_cache, reserve,
1087                                               ob_flags, cl);
1088         }
1089 alloc_done:
1090         BUG_ON(!ret && nr_effective < nr_replicas);
1091
1092         if (erasure_code && !ec_open_bucket(c, &ptrs))
1093                 pr_debug("failed to get ec bucket: ret %u", ret);
1094
1095         if (ret == -INSUFFICIENT_DEVICES &&
1096             nr_effective >= nr_replicas_required)
1097                 ret = 0;
1098
1099         if (ret)
1100                 goto err;
1101
1102         /* Free buckets we didn't use: */
1103         open_bucket_for_each(c, &wp->ptrs, ob, i)
1104                 open_bucket_free_unused(c, wp, ob);
1105
1106         wp->ptrs = ptrs;
1107
1108         wp->sectors_free = UINT_MAX;
1109
1110         open_bucket_for_each(c, &wp->ptrs, ob, i)
1111                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1112
1113         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1114
1115         return wp;
1116 err:
1117         open_bucket_for_each(c, &wp->ptrs, ob, i)
1118                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1119                         ob_push(c, &ptrs, ob);
1120                 else
1121                         open_bucket_free_unused(c, wp, ob);
1122         wp->ptrs = ptrs;
1123
1124         mutex_unlock(&wp->lock);
1125
1126         if (ret == -FREELIST_EMPTY &&
1127             try_decrease_writepoints(c, write_points_nr))
1128                 goto retry;
1129
1130         switch (ret) {
1131         case -OPEN_BUCKETS_EMPTY:
1132         case -FREELIST_EMPTY:
1133                 return cl ? ERR_PTR(-EAGAIN) : ERR_PTR(-ENOSPC);
1134         case -INSUFFICIENT_DEVICES:
1135                 return ERR_PTR(-EROFS);
1136         default:
1137                 return ERR_PTR(ret);
1138         }
1139 }
1140
1141 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1142 {
1143         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1144
1145         return (struct bch_extent_ptr) {
1146                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
1147                 .gen    = ob->gen,
1148                 .dev    = ob->dev,
1149                 .offset = bucket_to_sector(ca, ob->bucket) +
1150                         ca->mi.bucket_size -
1151                         ob->sectors_free,
1152         };
1153 }
1154
1155 /*
1156  * Append pointers to the space we just allocated to @k, and mark @sectors space
1157  * as allocated out of @ob
1158  */
1159 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1160                                     struct bkey_i *k, unsigned sectors,
1161                                     bool cached)
1162
1163 {
1164         struct open_bucket *ob;
1165         unsigned i;
1166
1167         BUG_ON(sectors > wp->sectors_free);
1168         wp->sectors_free -= sectors;
1169
1170         open_bucket_for_each(c, &wp->ptrs, ob, i) {
1171                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1172                 struct bch_extent_ptr ptr = bch2_ob_ptr(c, ob);
1173
1174                 ptr.cached = cached ||
1175                         (!ca->mi.durability &&
1176                          wp->data_type == BCH_DATA_user);
1177
1178                 bch2_bkey_append_ptr(k, ptr);
1179
1180                 BUG_ON(sectors > ob->sectors_free);
1181                 ob->sectors_free -= sectors;
1182         }
1183 }
1184
1185 /*
1186  * Append pointers to the space we just allocated to @k, and mark @sectors space
1187  * as allocated out of @ob
1188  */
1189 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1190 {
1191         struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
1192         struct open_bucket *ob;
1193         unsigned i;
1194
1195         open_bucket_for_each(c, &wp->ptrs, ob, i)
1196                 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
1197         wp->ptrs = keep;
1198
1199         mutex_unlock(&wp->lock);
1200
1201         bch2_open_buckets_put(c, &ptrs);
1202 }
1203
1204 static inline void writepoint_init(struct write_point *wp,
1205                                    enum bch_data_type type)
1206 {
1207         mutex_init(&wp->lock);
1208         wp->data_type = type;
1209 }
1210
1211 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1212 {
1213         struct open_bucket *ob;
1214         struct write_point *wp;
1215
1216         mutex_init(&c->write_points_hash_lock);
1217         c->write_points_nr = ARRAY_SIZE(c->write_points);
1218
1219         /* open bucket 0 is a sentinal NULL: */
1220         spin_lock_init(&c->open_buckets[0].lock);
1221
1222         for (ob = c->open_buckets + 1;
1223              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1224                 spin_lock_init(&ob->lock);
1225                 c->open_buckets_nr_free++;
1226
1227                 ob->freelist = c->open_buckets_freelist;
1228                 c->open_buckets_freelist = ob - c->open_buckets;
1229         }
1230
1231         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
1232         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
1233         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
1234
1235         for (wp = c->write_points;
1236              wp < c->write_points + c->write_points_nr; wp++) {
1237                 writepoint_init(wp, BCH_DATA_user);
1238
1239                 wp->last_used   = sched_clock();
1240                 wp->write_point = (unsigned long) wp;
1241                 hlist_add_head_rcu(&wp->node,
1242                                    writepoint_hash(c, wp->write_point));
1243         }
1244 }
1245
1246 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1247 {
1248         struct open_bucket *ob;
1249
1250         for (ob = c->open_buckets;
1251              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1252              ob++) {
1253                 spin_lock(&ob->lock);
1254                 if (ob->valid && !ob->on_partial_list) {
1255                         pr_buf(out, "%zu ref %u type %s\n",
1256                                ob - c->open_buckets,
1257                                atomic_read(&ob->pin),
1258                                bch2_data_types[ob->data_type]);
1259                 }
1260                 spin_unlock(&ob->lock);
1261         }
1262
1263 }