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(a.data_type != BCH_DATA_free, c,
 304                         "non free bucket in freespace btree (state %s)\n"
 305                         "  %s\n"
 306                         "  at %llu (genbits %u)",
 307                         bch2_data_types[a.data_type],
 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 (a.data_type != BCH_DATA_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         struct bch_dev_usage usage;
 493         u64 avail;
 494         u64 cur_bucket = 0;
 495         u64 buckets_seen = 0;
 496         u64 skipped_open = 0;
 497         u64 skipped_need_journal_commit = 0;
 498         u64 skipped_nouse = 0;
 499         bool waiting = false;
 500         int ret;
 501 again:
 502         usage = bch2_dev_usage_read(ca);
 503         avail = __dev_buckets_available(ca, usage,reserve);
 504
 505         if (usage.d[BCH_DATA_need_discard].buckets > avail)
 506                 bch2_do_discards(c);
 507
 508         if (usage.d[BCH_DATA_need_gc_gens].buckets > avail)
 509                 bch2_do_gc_gens(c);
 510
 511         if (should_invalidate_buckets(ca, usage))
 512                 bch2_do_invalidates(c);
 513
 514         if (!avail) {
 515                 if (cl && !waiting) {
 516                         closure_wait(&c->freelist_wait, cl);
 517                         waiting = true;
 518                         goto again;
 519                 }
 520
 521                 if (!c->blocked_allocate)
 522                         c->blocked_allocate = local_clock();
 523
 524                 ob = ERR_PTR(-FREELIST_EMPTY);
 525                 goto err;
 526         }
 527
 528         if (waiting)
 529                 closure_wake_up(&c->freelist_wait);
 530
 531         if (may_alloc_partial) {
 532                 ob = try_alloc_partial_bucket(c, ca, reserve);
 533                 if (ob)
 534                         return ob;
 535         }
 536
 537         ret = bch2_trans_do(c, NULL, NULL, 0,
 538                         PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(&trans, ca, reserve,
 539                                                         &cur_bucket,
 540                                                         &buckets_seen,
 541                                                         &skipped_open,
 542                                                         &skipped_need_journal_commit,
 543                                                         &skipped_nouse,
 544                                                         cl)));
 545
 546         if (skipped_need_journal_commit * 2 > avail)
 547                 bch2_journal_flush_async(&c->journal, NULL);
 548 err:
 549         if (!ob)
 550                 ob = ERR_PTR(ret ?: -FREELIST_EMPTY);
 551
 552         if (IS_ERR(ob)) {
 553                 trace_bucket_alloc_fail(ca, bch2_alloc_reserves[reserve], avail,
 554                                         buckets_seen,
 555                                         skipped_open,
 556                                         skipped_need_journal_commit,
 557                                         skipped_nouse,
 558                                         cl == NULL, PTR_ERR(ob));
 559                 atomic_long_inc(&c->bucket_alloc_fail);
 560         }
 561
 562         return ob;
 563 }
 564
 565 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 566                             unsigned l, unsigned r)
 567 {
 568         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 569                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 570 }
 571
 572 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 573
 574 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 575                                           struct dev_stripe_state *stripe,
 576                                           struct bch_devs_mask *devs)
 577 {
 578         struct dev_alloc_list ret = { .nr = 0 };
 579         unsigned i;
 580
 581         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 582                 ret.devs[ret.nr++] = i;
 583
 584         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 585         return ret;
 586 }
 587
 588 void bch2_dev_stripe_increment(struct bch_dev *ca,
 589                                struct dev_stripe_state *stripe)
 590 {
 591         u64 *v = stripe->next_alloc + ca->dev_idx;
 592         u64 free_space = dev_buckets_available(ca, RESERVE_none);
 593         u64 free_space_inv = free_space
 594                 ? div64_u64(1ULL << 48, free_space)
 595                 : 1ULL << 48;
 596         u64 scale = *v / 4;
 597
 598         if (*v + free_space_inv >= *v)
 599                 *v += free_space_inv;
 600         else
 601                 *v = U64_MAX;
 602
 603         for (v = stripe->next_alloc;
 604              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 605                 *v = *v < scale ? 0 : *v - scale;
 606 }
 607
 608 #define BUCKET_MAY_ALLOC_PARTIAL        (1 << 0)
 609 #define BUCKET_ALLOC_USE_DURABILITY     (1 << 1)
 610
 611 static void add_new_bucket(struct bch_fs *c,
 612                            struct open_buckets *ptrs,
 613                            struct bch_devs_mask *devs_may_alloc,
 614                            unsigned *nr_effective,
 615                            bool *have_cache,
 616                            unsigned flags,
 617                            struct open_bucket *ob)
 618 {
 619         unsigned durability =
 620                 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
 621
 622         __clear_bit(ob->dev, devs_may_alloc->d);
 623         *nr_effective   += (flags & BUCKET_ALLOC_USE_DURABILITY)
 624                 ? durability : 1;
 625         *have_cache     |= !durability;
 626
 627         ob_push(c, ptrs, ob);
 628 }
 629
 630 int bch2_bucket_alloc_set(struct bch_fs *c,
 631                       struct open_buckets *ptrs,
 632                       struct dev_stripe_state *stripe,
 633                       struct bch_devs_mask *devs_may_alloc,
 634                       unsigned nr_replicas,
 635                       unsigned *nr_effective,
 636                       bool *have_cache,
 637                       enum alloc_reserve reserve,
 638                       unsigned flags,
 639                       struct closure *cl)
 640 {
 641         struct dev_alloc_list devs_sorted =
 642                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 643         unsigned dev;
 644         struct bch_dev *ca;
 645         int ret = -INSUFFICIENT_DEVICES;
 646         unsigned i;
 647
 648         BUG_ON(*nr_effective >= nr_replicas);
 649
 650         for (i = 0; i < devs_sorted.nr; i++) {
 651                 struct open_bucket *ob;
 652
 653                 dev = devs_sorted.devs[i];
 654
 655                 rcu_read_lock();
 656                 ca = rcu_dereference(c->devs[dev]);
 657                 if (ca)
 658                         percpu_ref_get(&ca->ref);
 659                 rcu_read_unlock();
 660
 661                 if (!ca)
 662                         continue;
 663
 664                 if (!ca->mi.durability && *have_cache) {
 665                         percpu_ref_put(&ca->ref);
 666                         continue;
 667                 }
 668
 669                 ob = bch2_bucket_alloc(c, ca, reserve,
 670                                 flags & BUCKET_MAY_ALLOC_PARTIAL, cl);
 671                 if (!IS_ERR(ob))
 672                         bch2_dev_stripe_increment(ca, stripe);
 673                 percpu_ref_put(&ca->ref);
 674
 675                 if (IS_ERR(ob)) {
 676                         ret = PTR_ERR(ob);
 677
 678                         if (cl)
 679                                 break;
 680                         continue;
 681                 }
 682
 683                 add_new_bucket(c, ptrs, devs_may_alloc,
 684                                nr_effective, have_cache, flags, ob);
 685
 686                 if (*nr_effective >= nr_replicas) {
 687                         ret = 0;
 688                         break;
 689                 }
 690         }
 691
 692         return ret;
 693 }
 694
 695 /* Allocate from stripes: */
 696
 697 /*
 698  * if we can't allocate a new stripe because there are already too many
 699  * partially filled stripes, force allocating from an existing stripe even when
 700  * it's to a device we don't want:
 701  */
 702
 703 static int bucket_alloc_from_stripe(struct bch_fs *c,
 704                          struct open_buckets *ptrs,
 705                          struct write_point *wp,
 706                          struct bch_devs_mask *devs_may_alloc,
 707                          u16 target,
 708                          unsigned erasure_code,
 709                          unsigned nr_replicas,
 710                          unsigned *nr_effective,
 711                          bool *have_cache,
 712                          unsigned flags,
 713                          struct closure *cl)
 714 {
 715         struct dev_alloc_list devs_sorted;
 716         struct ec_stripe_head *h;
 717         struct open_bucket *ob;
 718         struct bch_dev *ca;
 719         unsigned i, ec_idx;
 720
 721         if (!erasure_code)
 722                 return 0;
 723
 724         if (nr_replicas < 2)
 725                 return 0;
 726
 727         if (ec_open_bucket(c, ptrs))
 728                 return 0;
 729
 730         h = bch2_ec_stripe_head_get(c, target, 0, nr_replicas - 1,
 731                                     wp == &c->copygc_write_point,
 732                                     cl);
 733         if (IS_ERR(h))
 734                 return -PTR_ERR(h);
 735         if (!h)
 736                 return 0;
 737
 738         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 739
 740         for (i = 0; i < devs_sorted.nr; i++)
 741                 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 742                         if (!h->s->blocks[ec_idx])
 743                                 continue;
 744
 745                         ob = c->open_buckets + h->s->blocks[ec_idx];
 746                         if (ob->dev == devs_sorted.devs[i] &&
 747                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 748                                 goto got_bucket;
 749                 }
 750         goto out_put_head;
 751 got_bucket:
 752         ca = bch_dev_bkey_exists(c, ob->dev);
 753
 754         ob->ec_idx      = ec_idx;
 755         ob->ec          = h->s;
 756
 757         add_new_bucket(c, ptrs, devs_may_alloc,
 758                        nr_effective, have_cache, flags, ob);
 759         atomic_inc(&h->s->pin);
 760 out_put_head:
 761         bch2_ec_stripe_head_put(c, h);
 762         return 0;
 763 }
 764
 765 /* Sector allocator */
 766
 767 static void get_buckets_from_writepoint(struct bch_fs *c,
 768                                         struct open_buckets *ptrs,
 769                                         struct write_point *wp,
 770                                         struct bch_devs_mask *devs_may_alloc,
 771                                         unsigned nr_replicas,
 772                                         unsigned *nr_effective,
 773                                         bool *have_cache,
 774                                         unsigned flags,
 775                                         bool need_ec)
 776 {
 777         struct open_buckets ptrs_skip = { .nr = 0 };
 778         struct open_bucket *ob;
 779         unsigned i;
 780
 781         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 782                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 783
 784                 if (*nr_effective < nr_replicas &&
 785                     test_bit(ob->dev, devs_may_alloc->d) &&
 786                     (ca->mi.durability ||
 787                      (wp->data_type == BCH_DATA_user && !*have_cache)) &&
 788                     (ob->ec || !need_ec)) {
 789                         add_new_bucket(c, ptrs, devs_may_alloc,
 790                                        nr_effective, have_cache,
 791                                        flags, ob);
 792                 } else {
 793                         ob_push(c, &ptrs_skip, ob);
 794                 }
 795         }
 796         wp->ptrs = ptrs_skip;
 797 }
 798
 799 static int open_bucket_add_buckets(struct bch_fs *c,
 800                         struct open_buckets *ptrs,
 801                         struct write_point *wp,
 802                         struct bch_devs_list *devs_have,
 803                         u16 target,
 804                         unsigned erasure_code,
 805                         unsigned nr_replicas,
 806                         unsigned *nr_effective,
 807                         bool *have_cache,
 808                         enum alloc_reserve reserve,
 809                         unsigned flags,
 810                         struct closure *_cl)
 811 {
 812         struct bch_devs_mask devs;
 813         struct open_bucket *ob;
 814         struct closure *cl = NULL;
 815         int ret;
 816         unsigned i;
 817
 818         rcu_read_lock();
 819         devs = target_rw_devs(c, wp->data_type, target);
 820         rcu_read_unlock();
 821
 822         /* Don't allocate from devices we already have pointers to: */
 823         for (i = 0; i < devs_have->nr; i++)
 824                 __clear_bit(devs_have->devs[i], devs.d);
 825
 826         open_bucket_for_each(c, ptrs, ob, i)
 827                 __clear_bit(ob->dev, devs.d);
 828
 829         if (erasure_code) {
 830                 if (!ec_open_bucket(c, ptrs)) {
 831                         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 832                                                     nr_replicas, nr_effective,
 833                                                     have_cache, flags, true);
 834                         if (*nr_effective >= nr_replicas)
 835                                 return 0;
 836                 }
 837
 838                 if (!ec_open_bucket(c, ptrs)) {
 839                         ret = bucket_alloc_from_stripe(c, ptrs, wp, &devs,
 840                                                  target, erasure_code,
 841                                                  nr_replicas, nr_effective,
 842                                                  have_cache, flags, _cl);
 843                         if (ret == -FREELIST_EMPTY ||
 844                             ret == -OPEN_BUCKETS_EMPTY)
 845                                 return ret;
 846                         if (*nr_effective >= nr_replicas)
 847                                 return 0;
 848                 }
 849         }
 850
 851         get_buckets_from_writepoint(c, ptrs, wp, &devs,
 852                                     nr_replicas, nr_effective,
 853                                     have_cache, flags, false);
 854         if (*nr_effective >= nr_replicas)
 855                 return 0;
 856
 857 retry_blocking:
 858         /*
 859          * Try nonblocking first, so that if one device is full we'll try from
 860          * other devices:
 861          */
 862         ret = bch2_bucket_alloc_set(c, ptrs, &wp->stripe, &devs,
 863                                 nr_replicas, nr_effective, have_cache,
 864                                 reserve, flags, cl);
 865         if (ret && ret != -INSUFFICIENT_DEVICES && !cl && _cl) {
 866                 cl = _cl;
 867                 goto retry_blocking;
 868         }
 869
 870         return ret;
 871 }
 872
 873 void bch2_open_buckets_stop_dev(struct bch_fs *c, struct bch_dev *ca,
 874                                 struct open_buckets *obs)
 875 {
 876         struct open_buckets ptrs = { .nr = 0 };
 877         struct open_bucket *ob, *ob2;
 878         unsigned i, j;
 879
 880         open_bucket_for_each(c, obs, ob, i) {
 881                 bool drop = !ca || ob->dev == ca->dev_idx;
 882
 883                 if (!drop && ob->ec) {
 884                         mutex_lock(&ob->ec->lock);
 885                         for (j = 0; j < ob->ec->new_stripe.key.v.nr_blocks; j++) {
 886                                 if (!ob->ec->blocks[j])
 887                                         continue;
 888
 889                                 ob2 = c->open_buckets + ob->ec->blocks[j];
 890                                 drop |= ob2->dev == ca->dev_idx;
 891                         }
 892                         mutex_unlock(&ob->ec->lock);
 893                 }
 894
 895                 if (drop)
 896                         bch2_open_bucket_put(c, ob);
 897                 else
 898                         ob_push(c, &ptrs, ob);
 899         }
 900
 901         *obs = ptrs;
 902 }
 903
 904 void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
 905                           struct write_point *wp)
 906 {
 907         mutex_lock(&wp->lock);
 908         bch2_open_buckets_stop_dev(c, ca, &wp->ptrs);
 909         mutex_unlock(&wp->lock);
 910 }
 911
 912 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
 913                                                  unsigned long write_point)
 914 {
 915         unsigned hash =
 916                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
 917
 918         return &c->write_points_hash[hash];
 919 }
 920
 921 static struct write_point *__writepoint_find(struct hlist_head *head,
 922                                              unsigned long write_point)
 923 {
 924         struct write_point *wp;
 925
 926         rcu_read_lock();
 927         hlist_for_each_entry_rcu(wp, head, node)
 928                 if (wp->write_point == write_point)
 929                         goto out;
 930         wp = NULL;
 931 out:
 932         rcu_read_unlock();
 933         return wp;
 934 }
 935
 936 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
 937 {
 938         u64 stranded    = c->write_points_nr * c->bucket_size_max;
 939         u64 free        = bch2_fs_usage_read_short(c).free;
 940
 941         return stranded * factor > free;
 942 }
 943
 944 static bool try_increase_writepoints(struct bch_fs *c)
 945 {
 946         struct write_point *wp;
 947
 948         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
 949             too_many_writepoints(c, 32))
 950                 return false;
 951
 952         wp = c->write_points + c->write_points_nr++;
 953         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
 954         return true;
 955 }
 956
 957 static bool try_decrease_writepoints(struct bch_fs *c,
 958                                      unsigned old_nr)
 959 {
 960         struct write_point *wp;
 961
 962         mutex_lock(&c->write_points_hash_lock);
 963         if (c->write_points_nr < old_nr) {
 964                 mutex_unlock(&c->write_points_hash_lock);
 965                 return true;
 966         }
 967
 968         if (c->write_points_nr == 1 ||
 969             !too_many_writepoints(c, 8)) {
 970                 mutex_unlock(&c->write_points_hash_lock);
 971                 return false;
 972         }
 973
 974         wp = c->write_points + --c->write_points_nr;
 975
 976         hlist_del_rcu(&wp->node);
 977         mutex_unlock(&c->write_points_hash_lock);
 978
 979         bch2_writepoint_stop(c, NULL, wp);
 980         return true;
 981 }
 982
 983 static struct write_point *writepoint_find(struct bch_fs *c,
 984                                            unsigned long write_point)
 985 {
 986         struct write_point *wp, *oldest;
 987         struct hlist_head *head;
 988
 989         if (!(write_point & 1UL)) {
 990                 wp = (struct write_point *) write_point;
 991                 mutex_lock(&wp->lock);
 992                 return wp;
 993         }
 994
 995         head = writepoint_hash(c, write_point);
 996 restart_find:
 997         wp = __writepoint_find(head, write_point);
 998         if (wp) {
 999 lock_wp:
1000                 mutex_lock(&wp->lock);
1001                 if (wp->write_point == write_point)
1002                         goto out;
1003                 mutex_unlock(&wp->lock);
1004                 goto restart_find;
1005         }
1006 restart_find_oldest:
1007         oldest = NULL;
1008         for (wp = c->write_points;
1009              wp < c->write_points + c->write_points_nr; wp++)
1010                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1011                         oldest = wp;
1012
1013         mutex_lock(&oldest->lock);
1014         mutex_lock(&c->write_points_hash_lock);
1015         if (oldest >= c->write_points + c->write_points_nr ||
1016             try_increase_writepoints(c)) {
1017                 mutex_unlock(&c->write_points_hash_lock);
1018                 mutex_unlock(&oldest->lock);
1019                 goto restart_find_oldest;
1020         }
1021
1022         wp = __writepoint_find(head, write_point);
1023         if (wp && wp != oldest) {
1024                 mutex_unlock(&c->write_points_hash_lock);
1025                 mutex_unlock(&oldest->lock);
1026                 goto lock_wp;
1027         }
1028
1029         wp = oldest;
1030         hlist_del_rcu(&wp->node);
1031         wp->write_point = write_point;
1032         hlist_add_head_rcu(&wp->node, head);
1033         mutex_unlock(&c->write_points_hash_lock);
1034 out:
1035         wp->last_used = sched_clock();
1036         return wp;
1037 }
1038
1039 /*
1040  * Get us an open_bucket we can allocate from, return with it locked:
1041  */
1042 struct write_point *bch2_alloc_sectors_start(struct bch_fs *c,
1043                                 unsigned target,
1044                                 unsigned erasure_code,
1045                                 struct write_point_specifier write_point,
1046                                 struct bch_devs_list *devs_have,
1047                                 unsigned nr_replicas,
1048                                 unsigned nr_replicas_required,
1049                                 enum alloc_reserve reserve,
1050                                 unsigned flags,
1051                                 struct closure *cl)
1052 {
1053         struct write_point *wp;
1054         struct open_bucket *ob;
1055         struct open_buckets ptrs;
1056         unsigned nr_effective, write_points_nr;
1057         unsigned ob_flags = 0;
1058         bool have_cache;
1059         int ret;
1060         int i;
1061
1062         if (!(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS))
1063                 ob_flags |= BUCKET_ALLOC_USE_DURABILITY;
1064
1065         BUG_ON(!nr_replicas || !nr_replicas_required);
1066 retry:
1067         ptrs.nr         = 0;
1068         nr_effective    = 0;
1069         write_points_nr = c->write_points_nr;
1070         have_cache      = false;
1071
1072         wp = writepoint_find(c, write_point.v);
1073
1074         if (wp->data_type == BCH_DATA_user)
1075                 ob_flags |= BUCKET_MAY_ALLOC_PARTIAL;
1076
1077         /* metadata may not allocate on cache devices: */
1078         if (wp->data_type != BCH_DATA_user)
1079                 have_cache = true;
1080
1081         if (!target || (flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1082                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
1083                                               target, erasure_code,
1084                                               nr_replicas, &nr_effective,
1085                                               &have_cache, reserve,
1086                                               ob_flags, cl);
1087         } else {
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, NULL);
1093                 if (!ret)
1094                         goto alloc_done;
1095
1096                 ret = open_bucket_add_buckets(c, &ptrs, wp, devs_have,
1097                                               0, erasure_code,
1098                                               nr_replicas, &nr_effective,
1099                                               &have_cache, reserve,
1100                                               ob_flags, cl);
1101         }
1102 alloc_done:
1103         BUG_ON(!ret && nr_effective < nr_replicas);
1104
1105         if (erasure_code && !ec_open_bucket(c, &ptrs))
1106                 pr_debug("failed to get ec bucket: ret %u", ret);
1107
1108         if (ret == -INSUFFICIENT_DEVICES &&
1109             nr_effective >= nr_replicas_required)
1110                 ret = 0;
1111
1112         if (ret)
1113                 goto err;
1114
1115         /* Free buckets we didn't use: */
1116         open_bucket_for_each(c, &wp->ptrs, ob, i)
1117                 open_bucket_free_unused(c, wp, ob);
1118
1119         wp->ptrs = ptrs;
1120
1121         wp->sectors_free = UINT_MAX;
1122
1123         open_bucket_for_each(c, &wp->ptrs, ob, i)
1124                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1125
1126         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1127
1128         return wp;
1129 err:
1130         open_bucket_for_each(c, &wp->ptrs, ob, i)
1131                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1132                         ob_push(c, &ptrs, ob);
1133                 else
1134                         open_bucket_free_unused(c, wp, ob);
1135         wp->ptrs = ptrs;
1136
1137         mutex_unlock(&wp->lock);
1138
1139         if (ret == -FREELIST_EMPTY &&
1140             try_decrease_writepoints(c, write_points_nr))
1141                 goto retry;
1142
1143         switch (ret) {
1144         case -OPEN_BUCKETS_EMPTY:
1145         case -FREELIST_EMPTY:
1146                 return cl ? ERR_PTR(-EAGAIN) : ERR_PTR(-ENOSPC);
1147         case -INSUFFICIENT_DEVICES:
1148                 return ERR_PTR(-EROFS);
1149         default:
1150                 return ERR_PTR(ret);
1151         }
1152 }
1153
1154 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1155 {
1156         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1157
1158         return (struct bch_extent_ptr) {
1159                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
1160                 .gen    = ob->gen,
1161                 .dev    = ob->dev,
1162                 .offset = bucket_to_sector(ca, ob->bucket) +
1163                         ca->mi.bucket_size -
1164                         ob->sectors_free,
1165         };
1166 }
1167
1168 /*
1169  * Append pointers to the space we just allocated to @k, and mark @sectors space
1170  * as allocated out of @ob
1171  */
1172 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1173                                     struct bkey_i *k, unsigned sectors,
1174                                     bool cached)
1175
1176 {
1177         struct open_bucket *ob;
1178         unsigned i;
1179
1180         BUG_ON(sectors > wp->sectors_free);
1181         wp->sectors_free -= sectors;
1182
1183         open_bucket_for_each(c, &wp->ptrs, ob, i) {
1184                 struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1185                 struct bch_extent_ptr ptr = bch2_ob_ptr(c, ob);
1186
1187                 ptr.cached = cached ||
1188                         (!ca->mi.durability &&
1189                          wp->data_type == BCH_DATA_user);
1190
1191                 bch2_bkey_append_ptr(k, ptr);
1192
1193                 BUG_ON(sectors > ob->sectors_free);
1194                 ob->sectors_free -= sectors;
1195         }
1196 }
1197
1198 /*
1199  * Append pointers to the space we just allocated to @k, and mark @sectors space
1200  * as allocated out of @ob
1201  */
1202 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1203 {
1204         struct open_buckets ptrs = { .nr = 0 }, keep = { .nr = 0 };
1205         struct open_bucket *ob;
1206         unsigned i;
1207
1208         open_bucket_for_each(c, &wp->ptrs, ob, i)
1209                 ob_push(c, !ob->sectors_free ? &ptrs : &keep, ob);
1210         wp->ptrs = keep;
1211
1212         mutex_unlock(&wp->lock);
1213
1214         bch2_open_buckets_put(c, &ptrs);
1215 }
1216
1217 static inline void writepoint_init(struct write_point *wp,
1218                                    enum bch_data_type type)
1219 {
1220         mutex_init(&wp->lock);
1221         wp->data_type = type;
1222 }
1223
1224 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1225 {
1226         struct open_bucket *ob;
1227         struct write_point *wp;
1228
1229         mutex_init(&c->write_points_hash_lock);
1230         c->write_points_nr = ARRAY_SIZE(c->write_points);
1231
1232         /* open bucket 0 is a sentinal NULL: */
1233         spin_lock_init(&c->open_buckets[0].lock);
1234
1235         for (ob = c->open_buckets + 1;
1236              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1237                 spin_lock_init(&ob->lock);
1238                 c->open_buckets_nr_free++;
1239
1240                 ob->freelist = c->open_buckets_freelist;
1241                 c->open_buckets_freelist = ob - c->open_buckets;
1242         }
1243
1244         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
1245         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
1246         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
1247
1248         for (wp = c->write_points;
1249              wp < c->write_points + c->write_points_nr; wp++) {
1250                 writepoint_init(wp, BCH_DATA_user);
1251
1252                 wp->last_used   = sched_clock();
1253                 wp->write_point = (unsigned long) wp;
1254                 hlist_add_head_rcu(&wp->node,
1255                                    writepoint_hash(c, wp->write_point));
1256         }
1257 }
1258
1259 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1260 {
1261         struct open_bucket *ob;
1262
1263         for (ob = c->open_buckets;
1264              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1265              ob++) {
1266                 spin_lock(&ob->lock);
1267                 if (ob->valid && !ob->on_partial_list) {
1268                         pr_buf(out, "%zu ref %u type %s %u:%llu:%u\n",
1269                                ob - c->open_buckets,
1270                                atomic_read(&ob->pin),
1271                                bch2_data_types[ob->data_type],
1272                                ob->dev, ob->bucket, ob->gen);
1273                 }
1274                 spin_unlock(&ob->lock);
1275         }
1276 }