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

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright 2012 Google, Inc.
   4  *
   5  * Foreground allocator code: allocate buckets from freelist, and allocate in
   6  * sector granularity from writepoints.
   7  *
   8  * bch2_bucket_alloc() allocates a single bucket from a specific device.
   9  *
  10  * bch2_bucket_alloc_set() allocates one or more buckets from different devices
  11  * in a given filesystem.
  12  */
  13
  14 #include "bcachefs.h"
  15 #include "alloc_background.h"
  16 #include "alloc_foreground.h"
  17 #include "backpointers.h"
  18 #include "btree_iter.h"
  19 #include "btree_update.h"
  20 #include "btree_gc.h"
  21 #include "buckets.h"
  22 #include "buckets_waiting_for_journal.h"
  23 #include "clock.h"
  24 #include "debug.h"
  25 #include "disk_groups.h"
  26 #include "ec.h"
  27 #include "error.h"
  28 #include "io_write.h"
  29 #include "journal.h"
  30 #include "movinggc.h"
  31 #include "nocow_locking.h"
  32 #include "trace.h"
  33
  34 #include <linux/math64.h>
  35 #include <linux/rculist.h>
  36 #include <linux/rcupdate.h>
  37
  38 static void bch2_trans_mutex_lock_norelock(struct btree_trans *trans,
  39                                            struct mutex *lock)
  40 {
  41         if (!mutex_trylock(lock)) {
  42                 bch2_trans_unlock(trans);
  43                 mutex_lock(lock);
  44         }
  45 }
  46
  47 const char * const bch2_watermarks[] = {
  48 #define x(t) #t,
  49         BCH_WATERMARKS()
  50 #undef x
  51         NULL
  52 };
  53
  54 /*
  55  * Open buckets represent a bucket that's currently being allocated from.  They
  56  * serve two purposes:
  57  *
  58  *  - They track buckets that have been partially allocated, allowing for
  59  *    sub-bucket sized allocations - they're used by the sector allocator below
  60  *
  61  *  - They provide a reference to the buckets they own that mark and sweep GC
  62  *    can find, until the new allocation has a pointer to it inserted into the
  63  *    btree
  64  *
  65  * When allocating some space with the sector allocator, the allocation comes
  66  * with a reference to an open bucket - the caller is required to put that
  67  * reference _after_ doing the index update that makes its allocation reachable.
  68  */
  69
  70 void bch2_reset_alloc_cursors(struct bch_fs *c)
  71 {
  72         struct bch_dev *ca;
  73         unsigned i;
  74
  75         rcu_read_lock();
  76         for_each_member_device_rcu(ca, c, i, NULL)
  77                 ca->alloc_cursor = 0;
  78         rcu_read_unlock();
  79 }
  80
  81 static void bch2_open_bucket_hash_add(struct bch_fs *c, struct open_bucket *ob)
  82 {
  83         open_bucket_idx_t idx = ob - c->open_buckets;
  84         open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
  85
  86         ob->hash = *slot;
  87         *slot = idx;
  88 }
  89
  90 static void bch2_open_bucket_hash_remove(struct bch_fs *c, struct open_bucket *ob)
  91 {
  92         open_bucket_idx_t idx = ob - c->open_buckets;
  93         open_bucket_idx_t *slot = open_bucket_hashslot(c, ob->dev, ob->bucket);
  94
  95         while (*slot != idx) {
  96                 BUG_ON(!*slot);
  97                 slot = &c->open_buckets[*slot].hash;
  98         }
  99
 100         *slot = ob->hash;
 101         ob->hash = 0;
 102 }
 103
 104 void __bch2_open_bucket_put(struct bch_fs *c, struct open_bucket *ob)
 105 {
 106         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 107
 108         if (ob->ec) {
 109                 ec_stripe_new_put(c, ob->ec, STRIPE_REF_io);
 110                 return;
 111         }
 112
 113         percpu_down_read(&c->mark_lock);
 114         spin_lock(&ob->lock);
 115
 116         ob->valid = false;
 117         ob->data_type = 0;
 118
 119         spin_unlock(&ob->lock);
 120         percpu_up_read(&c->mark_lock);
 121
 122         spin_lock(&c->freelist_lock);
 123         bch2_open_bucket_hash_remove(c, ob);
 124
 125         ob->freelist = c->open_buckets_freelist;
 126         c->open_buckets_freelist = ob - c->open_buckets;
 127
 128         c->open_buckets_nr_free++;
 129         ca->nr_open_buckets--;
 130         spin_unlock(&c->freelist_lock);
 131
 132         closure_wake_up(&c->open_buckets_wait);
 133 }
 134
 135 void bch2_open_bucket_write_error(struct bch_fs *c,
 136                                   struct open_buckets *obs,
 137                                   unsigned dev)
 138 {
 139         struct open_bucket *ob;
 140         unsigned i;
 141
 142         open_bucket_for_each(c, obs, ob, i)
 143                 if (ob->dev == dev && ob->ec)
 144                         bch2_ec_bucket_cancel(c, ob);
 145 }
 146
 147 static struct open_bucket *bch2_open_bucket_alloc(struct bch_fs *c)
 148 {
 149         struct open_bucket *ob;
 150
 151         BUG_ON(!c->open_buckets_freelist || !c->open_buckets_nr_free);
 152
 153         ob = c->open_buckets + c->open_buckets_freelist;
 154         c->open_buckets_freelist = ob->freelist;
 155         atomic_set(&ob->pin, 1);
 156         ob->data_type = 0;
 157
 158         c->open_buckets_nr_free--;
 159         return ob;
 160 }
 161
 162 static void open_bucket_free_unused(struct bch_fs *c, struct open_bucket *ob)
 163 {
 164         BUG_ON(c->open_buckets_partial_nr >=
 165                ARRAY_SIZE(c->open_buckets_partial));
 166
 167         spin_lock(&c->freelist_lock);
 168         ob->on_partial_list = true;
 169         c->open_buckets_partial[c->open_buckets_partial_nr++] =
 170                 ob - c->open_buckets;
 171         spin_unlock(&c->freelist_lock);
 172
 173         closure_wake_up(&c->open_buckets_wait);
 174         closure_wake_up(&c->freelist_wait);
 175 }
 176
 177 /* _only_ for allocating the journal on a new device: */
 178 long bch2_bucket_alloc_new_fs(struct bch_dev *ca)
 179 {
 180         while (ca->new_fs_bucket_idx < ca->mi.nbuckets) {
 181                 u64 b = ca->new_fs_bucket_idx++;
 182
 183                 if (!is_superblock_bucket(ca, b) &&
 184                     (!ca->buckets_nouse || !test_bit(b, ca->buckets_nouse)))
 185                         return b;
 186         }
 187
 188         return -1;
 189 }
 190
 191 static inline unsigned open_buckets_reserved(enum bch_watermark watermark)
 192 {
 193         switch (watermark) {
 194         case BCH_WATERMARK_reclaim:
 195                 return 0;
 196         case BCH_WATERMARK_btree:
 197         case BCH_WATERMARK_btree_copygc:
 198                 return OPEN_BUCKETS_COUNT / 4;
 199         case BCH_WATERMARK_copygc:
 200                 return OPEN_BUCKETS_COUNT / 3;
 201         default:
 202                 return OPEN_BUCKETS_COUNT / 2;
 203         }
 204 }
 205
 206 static struct open_bucket *__try_alloc_bucket(struct bch_fs *c, struct bch_dev *ca,
 207                                               u64 bucket,
 208                                               enum bch_watermark watermark,
 209                                               const struct bch_alloc_v4 *a,
 210                                               struct bucket_alloc_state *s,
 211                                               struct closure *cl)
 212 {
 213         struct open_bucket *ob;
 214
 215         if (unlikely(ca->buckets_nouse && test_bit(bucket, ca->buckets_nouse))) {
 216                 s->skipped_nouse++;
 217                 return NULL;
 218         }
 219
 220         if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
 221                 s->skipped_open++;
 222                 return NULL;
 223         }
 224
 225         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 226                         c->journal.flushed_seq_ondisk, ca->dev_idx, bucket)) {
 227                 s->skipped_need_journal_commit++;
 228                 return NULL;
 229         }
 230
 231         if (bch2_bucket_nocow_is_locked(&c->nocow_locks, POS(ca->dev_idx, bucket))) {
 232                 s->skipped_nocow++;
 233                 return NULL;
 234         }
 235
 236         spin_lock(&c->freelist_lock);
 237
 238         if (unlikely(c->open_buckets_nr_free <= open_buckets_reserved(watermark))) {
 239                 if (cl)
 240                         closure_wait(&c->open_buckets_wait, cl);
 241
 242                 if (!c->blocked_allocate_open_bucket)
 243                         c->blocked_allocate_open_bucket = local_clock();
 244
 245                 spin_unlock(&c->freelist_lock);
 246                 return ERR_PTR(-BCH_ERR_open_buckets_empty);
 247         }
 248
 249         /* Recheck under lock: */
 250         if (bch2_bucket_is_open(c, ca->dev_idx, bucket)) {
 251                 spin_unlock(&c->freelist_lock);
 252                 s->skipped_open++;
 253                 return NULL;
 254         }
 255
 256         ob = bch2_open_bucket_alloc(c);
 257
 258         spin_lock(&ob->lock);
 259
 260         ob->valid       = true;
 261         ob->sectors_free = ca->mi.bucket_size;
 262         ob->dev         = ca->dev_idx;
 263         ob->gen         = a->gen;
 264         ob->bucket      = bucket;
 265         spin_unlock(&ob->lock);
 266
 267         ca->nr_open_buckets++;
 268         bch2_open_bucket_hash_add(c, ob);
 269
 270         if (c->blocked_allocate_open_bucket) {
 271                 bch2_time_stats_update(
 272                         &c->times[BCH_TIME_blocked_allocate_open_bucket],
 273                         c->blocked_allocate_open_bucket);
 274                 c->blocked_allocate_open_bucket = 0;
 275         }
 276
 277         if (c->blocked_allocate) {
 278                 bch2_time_stats_update(
 279                         &c->times[BCH_TIME_blocked_allocate],
 280                         c->blocked_allocate);
 281                 c->blocked_allocate = 0;
 282         }
 283
 284         spin_unlock(&c->freelist_lock);
 285         return ob;
 286 }
 287
 288 static struct open_bucket *try_alloc_bucket(struct btree_trans *trans, struct bch_dev *ca,
 289                                             enum bch_watermark watermark, u64 free_entry,
 290                                             struct bucket_alloc_state *s,
 291                                             struct bkey_s_c freespace_k,
 292                                             struct closure *cl)
 293 {
 294         struct bch_fs *c = trans->c;
 295         struct btree_iter iter = { NULL };
 296         struct bkey_s_c k;
 297         struct open_bucket *ob;
 298         struct bch_alloc_v4 a_convert;
 299         const struct bch_alloc_v4 *a;
 300         u64 b = free_entry & ~(~0ULL << 56);
 301         unsigned genbits = free_entry >> 56;
 302         struct printbuf buf = PRINTBUF;
 303         int ret;
 304
 305         if (b < ca->mi.first_bucket || b >= ca->mi.nbuckets) {
 306                 prt_printf(&buf, "freespace btree has bucket outside allowed range %u-%llu\n"
 307                        "  freespace key ",
 308                         ca->mi.first_bucket, ca->mi.nbuckets);
 309                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 310                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 311                 ob = ERR_PTR(-EIO);
 312                 goto err;
 313         }
 314
 315         k = bch2_bkey_get_iter(trans, &iter,
 316                                BTREE_ID_alloc, POS(ca->dev_idx, b),
 317                                BTREE_ITER_CACHED);
 318         ret = bkey_err(k);
 319         if (ret) {
 320                 ob = ERR_PTR(ret);
 321                 goto err;
 322         }
 323
 324         a = bch2_alloc_to_v4(k, &a_convert);
 325
 326         if (a->data_type != BCH_DATA_free) {
 327                 if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_alloc_info) {
 328                         ob = NULL;
 329                         goto err;
 330                 }
 331
 332                 prt_printf(&buf, "non free bucket in freespace btree\n"
 333                        "  freespace key ");
 334                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 335                 prt_printf(&buf, "\n  ");
 336                 bch2_bkey_val_to_text(&buf, c, k);
 337                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 338                 ob = ERR_PTR(-EIO);
 339                 goto err;
 340         }
 341
 342         if (genbits != (alloc_freespace_genbits(*a) >> 56) &&
 343             c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
 344                 prt_printf(&buf, "bucket in freespace btree with wrong genbits (got %u should be %llu)\n"
 345                        "  freespace key ",
 346                        genbits, alloc_freespace_genbits(*a) >> 56);
 347                 bch2_bkey_val_to_text(&buf, c, freespace_k);
 348                 prt_printf(&buf, "\n  ");
 349                 bch2_bkey_val_to_text(&buf, c, k);
 350                 bch2_trans_inconsistent(trans, "%s", buf.buf);
 351                 ob = ERR_PTR(-EIO);
 352                 goto err;
 353         }
 354
 355         if (c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_extents_to_backpointers) {
 356                 struct bch_backpointer bp;
 357                 struct bpos bp_pos = POS_MIN;
 358
 359                 ret = bch2_get_next_backpointer(trans, POS(ca->dev_idx, b), -1,
 360                                                 &bp_pos, &bp,
 361                                                 BTREE_ITER_NOPRESERVE);
 362                 if (ret) {
 363                         ob = ERR_PTR(ret);
 364                         goto err;
 365                 }
 366
 367                 if (!bkey_eq(bp_pos, POS_MAX)) {
 368                         /*
 369                          * Bucket may have data in it - we don't call
 370                          * bc2h_trans_inconnsistent() because fsck hasn't
 371                          * finished yet
 372                          */
 373                         ob = NULL;
 374                         goto err;
 375                 }
 376         }
 377
 378         ob = __try_alloc_bucket(c, ca, b, watermark, a, s, cl);
 379         if (!ob)
 380                 iter.path->preserve = false;
 381 err:
 382         if (iter.trans && iter.path)
 383                 set_btree_iter_dontneed(&iter);
 384         bch2_trans_iter_exit(trans, &iter);
 385         printbuf_exit(&buf);
 386         return ob;
 387 }
 388
 389 /*
 390  * This path is for before the freespace btree is initialized:
 391  *
 392  * If ca->new_fs_bucket_idx is nonzero, we haven't yet marked superblock &
 393  * journal buckets - journal buckets will be < ca->new_fs_bucket_idx
 394  */
 395 static noinline struct open_bucket *
 396 bch2_bucket_alloc_early(struct btree_trans *trans,
 397                         struct bch_dev *ca,
 398                         enum bch_watermark watermark,
 399                         struct bucket_alloc_state *s,
 400                         struct closure *cl)
 401 {
 402         struct btree_iter iter, citer;
 403         struct bkey_s_c k, ck;
 404         struct open_bucket *ob = NULL;
 405         u64 first_bucket = max_t(u64, ca->mi.first_bucket, ca->new_fs_bucket_idx);
 406         u64 alloc_start = max(first_bucket, READ_ONCE(ca->alloc_cursor));
 407         u64 alloc_cursor = alloc_start;
 408         int ret;
 409
 410         /*
 411          * Scan with an uncached iterator to avoid polluting the key cache. An
 412          * uncached iter will return a cached key if one exists, but if not
 413          * there is no other underlying protection for the associated key cache
 414          * slot. To avoid racing bucket allocations, look up the cached key slot
 415          * of any likely allocation candidate before attempting to proceed with
 416          * the allocation. This provides proper exclusion on the associated
 417          * bucket.
 418          */
 419 again:
 420         for_each_btree_key_norestart(trans, iter, BTREE_ID_alloc, POS(ca->dev_idx, alloc_cursor),
 421                            BTREE_ITER_SLOTS, k, ret) {
 422                 struct bch_alloc_v4 a_convert;
 423                 const struct bch_alloc_v4 *a;
 424
 425                 if (bkey_ge(k.k->p, POS(ca->dev_idx, ca->mi.nbuckets)))
 426                         break;
 427
 428                 if (ca->new_fs_bucket_idx &&
 429                     is_superblock_bucket(ca, k.k->p.offset))
 430                         continue;
 431
 432                 a = bch2_alloc_to_v4(k, &a_convert);
 433                 if (a->data_type != BCH_DATA_free)
 434                         continue;
 435
 436                 /* now check the cached key to serialize concurrent allocs of the bucket */
 437                 ck = bch2_bkey_get_iter(trans, &citer, BTREE_ID_alloc, k.k->p, BTREE_ITER_CACHED);
 438                 ret = bkey_err(ck);
 439                 if (ret)
 440                         break;
 441
 442                 a = bch2_alloc_to_v4(ck, &a_convert);
 443                 if (a->data_type != BCH_DATA_free)
 444                         goto next;
 445
 446                 s->buckets_seen++;
 447
 448                 ob = __try_alloc_bucket(trans->c, ca, k.k->p.offset, watermark, a, s, cl);
 449 next:
 450                 citer.path->preserve = false;
 451                 bch2_trans_iter_exit(trans, &citer);
 452                 if (ob)
 453                         break;
 454         }
 455         bch2_trans_iter_exit(trans, &iter);
 456
 457         alloc_cursor = iter.pos.offset;
 458         ca->alloc_cursor = alloc_cursor;
 459
 460         if (!ob && ret)
 461                 ob = ERR_PTR(ret);
 462
 463         if (!ob && alloc_start > first_bucket) {
 464                 alloc_cursor = alloc_start = first_bucket;
 465                 goto again;
 466         }
 467
 468         return ob;
 469 }
 470
 471 static struct open_bucket *bch2_bucket_alloc_freelist(struct btree_trans *trans,
 472                                                    struct bch_dev *ca,
 473                                                    enum bch_watermark watermark,
 474                                                    struct bucket_alloc_state *s,
 475                                                    struct closure *cl)
 476 {
 477         struct btree_iter iter;
 478         struct bkey_s_c k;
 479         struct open_bucket *ob = NULL;
 480         u64 alloc_start = max_t(u64, ca->mi.first_bucket, READ_ONCE(ca->alloc_cursor));
 481         u64 alloc_cursor = alloc_start;
 482         int ret;
 483
 484         BUG_ON(ca->new_fs_bucket_idx);
 485 again:
 486         for_each_btree_key_norestart(trans, iter, BTREE_ID_freespace,
 487                                      POS(ca->dev_idx, alloc_cursor), 0, k, ret) {
 488                 if (k.k->p.inode != ca->dev_idx)
 489                         break;
 490
 491                 for (alloc_cursor = max(alloc_cursor, bkey_start_offset(k.k));
 492                      alloc_cursor < k.k->p.offset;
 493                      alloc_cursor++) {
 494                         ret = btree_trans_too_many_iters(trans);
 495                         if (ret) {
 496                                 ob = ERR_PTR(ret);
 497                                 break;
 498                         }
 499
 500                         s->buckets_seen++;
 501
 502                         ob = try_alloc_bucket(trans, ca, watermark,
 503                                               alloc_cursor, s, k, cl);
 504                         if (ob) {
 505                                 iter.path->preserve = false;
 506                                 break;
 507                         }
 508                 }
 509
 510                 if (ob || ret)
 511                         break;
 512         }
 513         bch2_trans_iter_exit(trans, &iter);
 514
 515         ca->alloc_cursor = alloc_cursor;
 516
 517         if (!ob && ret)
 518                 ob = ERR_PTR(ret);
 519
 520         if (!ob && alloc_start > ca->mi.first_bucket) {
 521                 alloc_cursor = alloc_start = ca->mi.first_bucket;
 522                 goto again;
 523         }
 524
 525         return ob;
 526 }
 527
 528 /**
 529  * bch2_bucket_alloc_trans - allocate a single bucket from a specific device
 530  * @trans:      transaction object
 531  * @ca:         device to allocate from
 532  * @watermark:  how important is this allocation?
 533  * @cl:         if not NULL, closure to be used to wait if buckets not available
 534  * @usage:      for secondarily also returning the current device usage
 535  *
 536  * Returns:     an open_bucket on success, or an ERR_PTR() on failure.
 537  */
 538 static struct open_bucket *bch2_bucket_alloc_trans(struct btree_trans *trans,
 539                                       struct bch_dev *ca,
 540                                       enum bch_watermark watermark,
 541                                       struct closure *cl,
 542                                       struct bch_dev_usage *usage)
 543 {
 544         struct bch_fs *c = trans->c;
 545         struct open_bucket *ob = NULL;
 546         bool freespace = READ_ONCE(ca->mi.freespace_initialized);
 547         u64 avail;
 548         struct bucket_alloc_state s = { 0 };
 549         bool waiting = false;
 550 again:
 551         bch2_dev_usage_read_fast(ca, usage);
 552         avail = dev_buckets_free(ca, *usage, watermark);
 553
 554         if (usage->d[BCH_DATA_need_discard].buckets > avail)
 555                 bch2_do_discards(c);
 556
 557         if (usage->d[BCH_DATA_need_gc_gens].buckets > avail)
 558                 bch2_do_gc_gens(c);
 559
 560         if (should_invalidate_buckets(ca, *usage))
 561                 bch2_do_invalidates(c);
 562
 563         if (!avail) {
 564                 if (cl && !waiting) {
 565                         closure_wait(&c->freelist_wait, cl);
 566                         waiting = true;
 567                         goto again;
 568                 }
 569
 570                 if (!c->blocked_allocate)
 571                         c->blocked_allocate = local_clock();
 572
 573                 ob = ERR_PTR(-BCH_ERR_freelist_empty);
 574                 goto err;
 575         }
 576
 577         if (waiting)
 578                 closure_wake_up(&c->freelist_wait);
 579 alloc:
 580         ob = likely(freespace)
 581                 ? bch2_bucket_alloc_freelist(trans, ca, watermark, &s, cl)
 582                 : bch2_bucket_alloc_early(trans, ca, watermark, &s, cl);
 583
 584         if (s.skipped_need_journal_commit * 2 > avail)
 585                 bch2_journal_flush_async(&c->journal, NULL);
 586
 587         if (!ob && freespace && c->curr_recovery_pass <= BCH_RECOVERY_PASS_check_alloc_info) {
 588                 freespace = false;
 589                 goto alloc;
 590         }
 591 err:
 592         if (!ob)
 593                 ob = ERR_PTR(-BCH_ERR_no_buckets_found);
 594
 595         if (!IS_ERR(ob))
 596                 trace_and_count(c, bucket_alloc, ca,
 597                                 bch2_watermarks[watermark],
 598                                 ob->bucket,
 599                                 usage->d[BCH_DATA_free].buckets,
 600                                 avail,
 601                                 bch2_copygc_wait_amount(c),
 602                                 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
 603                                 &s,
 604                                 cl == NULL,
 605                                 "");
 606         else if (!bch2_err_matches(PTR_ERR(ob), BCH_ERR_transaction_restart))
 607                 trace_and_count(c, bucket_alloc_fail, ca,
 608                                 bch2_watermarks[watermark],
 609                                 0,
 610                                 usage->d[BCH_DATA_free].buckets,
 611                                 avail,
 612                                 bch2_copygc_wait_amount(c),
 613                                 c->copygc_wait - atomic64_read(&c->io_clock[WRITE].now),
 614                                 &s,
 615                                 cl == NULL,
 616                                 bch2_err_str(PTR_ERR(ob)));
 617
 618         return ob;
 619 }
 620
 621 struct open_bucket *bch2_bucket_alloc(struct bch_fs *c, struct bch_dev *ca,
 622                                       enum bch_watermark watermark,
 623                                       struct closure *cl)
 624 {
 625         struct bch_dev_usage usage;
 626         struct open_bucket *ob;
 627
 628         bch2_trans_do(c, NULL, NULL, 0,
 629                       PTR_ERR_OR_ZERO(ob = bch2_bucket_alloc_trans(trans, ca, watermark,
 630                                                         cl, &usage)));
 631         return ob;
 632 }
 633
 634 static int __dev_stripe_cmp(struct dev_stripe_state *stripe,
 635                             unsigned l, unsigned r)
 636 {
 637         return ((stripe->next_alloc[l] > stripe->next_alloc[r]) -
 638                 (stripe->next_alloc[l] < stripe->next_alloc[r]));
 639 }
 640
 641 #define dev_stripe_cmp(l, r) __dev_stripe_cmp(stripe, l, r)
 642
 643 struct dev_alloc_list bch2_dev_alloc_list(struct bch_fs *c,
 644                                           struct dev_stripe_state *stripe,
 645                                           struct bch_devs_mask *devs)
 646 {
 647         struct dev_alloc_list ret = { .nr = 0 };
 648         unsigned i;
 649
 650         for_each_set_bit(i, devs->d, BCH_SB_MEMBERS_MAX)
 651                 ret.devs[ret.nr++] = i;
 652
 653         bubble_sort(ret.devs, ret.nr, dev_stripe_cmp);
 654         return ret;
 655 }
 656
 657 static inline void bch2_dev_stripe_increment_inlined(struct bch_dev *ca,
 658                                struct dev_stripe_state *stripe,
 659                                struct bch_dev_usage *usage)
 660 {
 661         u64 *v = stripe->next_alloc + ca->dev_idx;
 662         u64 free_space = dev_buckets_available(ca, BCH_WATERMARK_normal);
 663         u64 free_space_inv = free_space
 664                 ? div64_u64(1ULL << 48, free_space)
 665                 : 1ULL << 48;
 666         u64 scale = *v / 4;
 667
 668         if (*v + free_space_inv >= *v)
 669                 *v += free_space_inv;
 670         else
 671                 *v = U64_MAX;
 672
 673         for (v = stripe->next_alloc;
 674              v < stripe->next_alloc + ARRAY_SIZE(stripe->next_alloc); v++)
 675                 *v = *v < scale ? 0 : *v - scale;
 676 }
 677
 678 void bch2_dev_stripe_increment(struct bch_dev *ca,
 679                                struct dev_stripe_state *stripe)
 680 {
 681         struct bch_dev_usage usage;
 682
 683         bch2_dev_usage_read_fast(ca, &usage);
 684         bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
 685 }
 686
 687 static int add_new_bucket(struct bch_fs *c,
 688                            struct open_buckets *ptrs,
 689                            struct bch_devs_mask *devs_may_alloc,
 690                            unsigned nr_replicas,
 691                            unsigned *nr_effective,
 692                            bool *have_cache,
 693                            unsigned flags,
 694                            struct open_bucket *ob)
 695 {
 696         unsigned durability =
 697                 bch_dev_bkey_exists(c, ob->dev)->mi.durability;
 698
 699         BUG_ON(*nr_effective >= nr_replicas);
 700
 701         __clear_bit(ob->dev, devs_may_alloc->d);
 702         *nr_effective   += durability;
 703         *have_cache     |= !durability;
 704
 705         ob_push(c, ptrs, ob);
 706
 707         if (*nr_effective >= nr_replicas)
 708                 return 1;
 709         if (ob->ec)
 710                 return 1;
 711         return 0;
 712 }
 713
 714 int bch2_bucket_alloc_set_trans(struct btree_trans *trans,
 715                       struct open_buckets *ptrs,
 716                       struct dev_stripe_state *stripe,
 717                       struct bch_devs_mask *devs_may_alloc,
 718                       unsigned nr_replicas,
 719                       unsigned *nr_effective,
 720                       bool *have_cache,
 721                       unsigned flags,
 722                       enum bch_data_type data_type,
 723                       enum bch_watermark watermark,
 724                       struct closure *cl)
 725 {
 726         struct bch_fs *c = trans->c;
 727         struct dev_alloc_list devs_sorted =
 728                 bch2_dev_alloc_list(c, stripe, devs_may_alloc);
 729         unsigned dev;
 730         struct bch_dev *ca;
 731         int ret = -BCH_ERR_insufficient_devices;
 732         unsigned i;
 733
 734         BUG_ON(*nr_effective >= nr_replicas);
 735
 736         for (i = 0; i < devs_sorted.nr; i++) {
 737                 struct bch_dev_usage usage;
 738                 struct open_bucket *ob;
 739
 740                 dev = devs_sorted.devs[i];
 741
 742                 rcu_read_lock();
 743                 ca = rcu_dereference(c->devs[dev]);
 744                 if (ca)
 745                         percpu_ref_get(&ca->ref);
 746                 rcu_read_unlock();
 747
 748                 if (!ca)
 749                         continue;
 750
 751                 if (!ca->mi.durability && *have_cache) {
 752                         percpu_ref_put(&ca->ref);
 753                         continue;
 754                 }
 755
 756                 ob = bch2_bucket_alloc_trans(trans, ca, watermark, cl, &usage);
 757                 if (!IS_ERR(ob))
 758                         bch2_dev_stripe_increment_inlined(ca, stripe, &usage);
 759                 percpu_ref_put(&ca->ref);
 760
 761                 if (IS_ERR(ob)) {
 762                         ret = PTR_ERR(ob);
 763                         if (bch2_err_matches(ret, BCH_ERR_transaction_restart) || cl)
 764                                 break;
 765                         continue;
 766                 }
 767
 768                 ob->data_type = data_type;
 769
 770                 if (add_new_bucket(c, ptrs, devs_may_alloc,
 771                                    nr_replicas, nr_effective,
 772                                    have_cache, flags, ob)) {
 773                         ret = 0;
 774                         break;
 775                 }
 776         }
 777
 778         return ret;
 779 }
 780
 781 /* Allocate from stripes: */
 782
 783 /*
 784  * if we can't allocate a new stripe because there are already too many
 785  * partially filled stripes, force allocating from an existing stripe even when
 786  * it's to a device we don't want:
 787  */
 788
 789 static int bucket_alloc_from_stripe(struct btree_trans *trans,
 790                          struct open_buckets *ptrs,
 791                          struct write_point *wp,
 792                          struct bch_devs_mask *devs_may_alloc,
 793                          u16 target,
 794                          unsigned nr_replicas,
 795                          unsigned *nr_effective,
 796                          bool *have_cache,
 797                          enum bch_watermark watermark,
 798                          unsigned flags,
 799                          struct closure *cl)
 800 {
 801         struct bch_fs *c = trans->c;
 802         struct dev_alloc_list devs_sorted;
 803         struct ec_stripe_head *h;
 804         struct open_bucket *ob;
 805         unsigned i, ec_idx;
 806         int ret = 0;
 807
 808         if (nr_replicas < 2)
 809                 return 0;
 810
 811         if (ec_open_bucket(c, ptrs))
 812                 return 0;
 813
 814         h = bch2_ec_stripe_head_get(trans, target, 0, nr_replicas - 1, watermark, cl);
 815         if (IS_ERR(h))
 816                 return PTR_ERR(h);
 817         if (!h)
 818                 return 0;
 819
 820         devs_sorted = bch2_dev_alloc_list(c, &wp->stripe, devs_may_alloc);
 821
 822         for (i = 0; i < devs_sorted.nr; i++)
 823                 for (ec_idx = 0; ec_idx < h->s->nr_data; ec_idx++) {
 824                         if (!h->s->blocks[ec_idx])
 825                                 continue;
 826
 827                         ob = c->open_buckets + h->s->blocks[ec_idx];
 828                         if (ob->dev == devs_sorted.devs[i] &&
 829                             !test_and_set_bit(ec_idx, h->s->blocks_allocated))
 830                                 goto got_bucket;
 831                 }
 832         goto out_put_head;
 833 got_bucket:
 834         ob->ec_idx      = ec_idx;
 835         ob->ec          = h->s;
 836         ec_stripe_new_get(h->s, STRIPE_REF_io);
 837
 838         ret = add_new_bucket(c, ptrs, devs_may_alloc,
 839                              nr_replicas, nr_effective,
 840                              have_cache, flags, ob);
 841 out_put_head:
 842         bch2_ec_stripe_head_put(c, h);
 843         return ret;
 844 }
 845
 846 /* Sector allocator */
 847
 848 static bool want_bucket(struct bch_fs *c,
 849                         struct write_point *wp,
 850                         struct bch_devs_mask *devs_may_alloc,
 851                         bool *have_cache, bool ec,
 852                         struct open_bucket *ob)
 853 {
 854         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 855
 856         if (!test_bit(ob->dev, devs_may_alloc->d))
 857                 return false;
 858
 859         if (ob->data_type != wp->data_type)
 860                 return false;
 861
 862         if (!ca->mi.durability &&
 863             (wp->data_type == BCH_DATA_btree || ec || *have_cache))
 864                 return false;
 865
 866         if (ec != (ob->ec != NULL))
 867                 return false;
 868
 869         return true;
 870 }
 871
 872 static int bucket_alloc_set_writepoint(struct bch_fs *c,
 873                                        struct open_buckets *ptrs,
 874                                        struct write_point *wp,
 875                                        struct bch_devs_mask *devs_may_alloc,
 876                                        unsigned nr_replicas,
 877                                        unsigned *nr_effective,
 878                                        bool *have_cache,
 879                                        bool ec, unsigned flags)
 880 {
 881         struct open_buckets ptrs_skip = { .nr = 0 };
 882         struct open_bucket *ob;
 883         unsigned i;
 884         int ret = 0;
 885
 886         open_bucket_for_each(c, &wp->ptrs, ob, i) {
 887                 if (!ret && want_bucket(c, wp, devs_may_alloc,
 888                                         have_cache, ec, ob))
 889                         ret = add_new_bucket(c, ptrs, devs_may_alloc,
 890                                        nr_replicas, nr_effective,
 891                                        have_cache, flags, ob);
 892                 else
 893                         ob_push(c, &ptrs_skip, ob);
 894         }
 895         wp->ptrs = ptrs_skip;
 896
 897         return ret;
 898 }
 899
 900 static int bucket_alloc_set_partial(struct bch_fs *c,
 901                                     struct open_buckets *ptrs,
 902                                     struct write_point *wp,
 903                                     struct bch_devs_mask *devs_may_alloc,
 904                                     unsigned nr_replicas,
 905                                     unsigned *nr_effective,
 906                                     bool *have_cache, bool ec,
 907                                     enum bch_watermark watermark,
 908                                     unsigned flags)
 909 {
 910         int i, ret = 0;
 911
 912         if (!c->open_buckets_partial_nr)
 913                 return 0;
 914
 915         spin_lock(&c->freelist_lock);
 916
 917         if (!c->open_buckets_partial_nr)
 918                 goto unlock;
 919
 920         for (i = c->open_buckets_partial_nr - 1; i >= 0; --i) {
 921                 struct open_bucket *ob = c->open_buckets + c->open_buckets_partial[i];
 922
 923                 if (want_bucket(c, wp, devs_may_alloc, have_cache, ec, ob)) {
 924                         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
 925                         struct bch_dev_usage usage;
 926                         u64 avail;
 927
 928                         bch2_dev_usage_read_fast(ca, &usage);
 929                         avail = dev_buckets_free(ca, usage, watermark);
 930                         if (!avail)
 931                                 continue;
 932
 933                         array_remove_item(c->open_buckets_partial,
 934                                           c->open_buckets_partial_nr,
 935                                           i);
 936                         ob->on_partial_list = false;
 937
 938                         ret = add_new_bucket(c, ptrs, devs_may_alloc,
 939                                              nr_replicas, nr_effective,
 940                                              have_cache, flags, ob);
 941                         if (ret)
 942                                 break;
 943                 }
 944         }
 945 unlock:
 946         spin_unlock(&c->freelist_lock);
 947         return ret;
 948 }
 949
 950 static int __open_bucket_add_buckets(struct btree_trans *trans,
 951                         struct open_buckets *ptrs,
 952                         struct write_point *wp,
 953                         struct bch_devs_list *devs_have,
 954                         u16 target,
 955                         bool erasure_code,
 956                         unsigned nr_replicas,
 957                         unsigned *nr_effective,
 958                         bool *have_cache,
 959                         enum bch_watermark watermark,
 960                         unsigned flags,
 961                         struct closure *_cl)
 962 {
 963         struct bch_fs *c = trans->c;
 964         struct bch_devs_mask devs;
 965         struct open_bucket *ob;
 966         struct closure *cl = NULL;
 967         unsigned i;
 968         int ret;
 969
 970         devs = target_rw_devs(c, wp->data_type, target);
 971
 972         /* Don't allocate from devices we already have pointers to: */
 973         for (i = 0; i < devs_have->nr; i++)
 974                 __clear_bit(devs_have->devs[i], devs.d);
 975
 976         open_bucket_for_each(c, ptrs, ob, i)
 977                 __clear_bit(ob->dev, devs.d);
 978
 979         if (erasure_code && ec_open_bucket(c, ptrs))
 980                 return 0;
 981
 982         ret = bucket_alloc_set_writepoint(c, ptrs, wp, &devs,
 983                                  nr_replicas, nr_effective,
 984                                  have_cache, erasure_code, flags);
 985         if (ret)
 986                 return ret;
 987
 988         ret = bucket_alloc_set_partial(c, ptrs, wp, &devs,
 989                                  nr_replicas, nr_effective,
 990                                  have_cache, erasure_code, watermark, flags);
 991         if (ret)
 992                 return ret;
 993
 994         if (erasure_code) {
 995                 ret = bucket_alloc_from_stripe(trans, ptrs, wp, &devs,
 996                                          target,
 997                                          nr_replicas, nr_effective,
 998                                          have_cache,
 999                                          watermark, flags, _cl);
1000         } else {
1001 retry_blocking:
1002                 /*
1003                  * Try nonblocking first, so that if one device is full we'll try from
1004                  * other devices:
1005                  */
1006                 ret = bch2_bucket_alloc_set_trans(trans, ptrs, &wp->stripe, &devs,
1007                                         nr_replicas, nr_effective, have_cache,
1008                                         flags, wp->data_type, watermark, cl);
1009                 if (ret &&
1010                     !bch2_err_matches(ret, BCH_ERR_transaction_restart) &&
1011                     !bch2_err_matches(ret, BCH_ERR_insufficient_devices) &&
1012                     !cl && _cl) {
1013                         cl = _cl;
1014                         goto retry_blocking;
1015                 }
1016         }
1017
1018         return ret;
1019 }
1020
1021 static int open_bucket_add_buckets(struct btree_trans *trans,
1022                         struct open_buckets *ptrs,
1023                         struct write_point *wp,
1024                         struct bch_devs_list *devs_have,
1025                         u16 target,
1026                         unsigned erasure_code,
1027                         unsigned nr_replicas,
1028                         unsigned *nr_effective,
1029                         bool *have_cache,
1030                         enum bch_watermark watermark,
1031                         unsigned flags,
1032                         struct closure *cl)
1033 {
1034         int ret;
1035
1036         if (erasure_code) {
1037                 ret = __open_bucket_add_buckets(trans, ptrs, wp,
1038                                 devs_have, target, erasure_code,
1039                                 nr_replicas, nr_effective, have_cache,
1040                                 watermark, flags, cl);
1041                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart) ||
1042                     bch2_err_matches(ret, BCH_ERR_operation_blocked) ||
1043                     bch2_err_matches(ret, BCH_ERR_freelist_empty) ||
1044                     bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
1045                         return ret;
1046                 if (*nr_effective >= nr_replicas)
1047                         return 0;
1048         }
1049
1050         ret = __open_bucket_add_buckets(trans, ptrs, wp,
1051                         devs_have, target, false,
1052                         nr_replicas, nr_effective, have_cache,
1053                         watermark, flags, cl);
1054         return ret < 0 ? ret : 0;
1055 }
1056
1057 /**
1058  * should_drop_bucket - check if this is open_bucket should go away
1059  * @ob:         open_bucket to predicate on
1060  * @c:          filesystem handle
1061  * @ca:         if set, we're killing buckets for a particular device
1062  * @ec:         if true, we're shutting down erasure coding and killing all ec
1063  *              open_buckets
1064  *              otherwise, return true
1065  * Returns: true if we should kill this open_bucket
1066  *
1067  * We're killing open_buckets because we're shutting down a device, erasure
1068  * coding, or the entire filesystem - check if this open_bucket matches:
1069  */
1070 static bool should_drop_bucket(struct open_bucket *ob, struct bch_fs *c,
1071                                struct bch_dev *ca, bool ec)
1072 {
1073         if (ec) {
1074                 return ob->ec != NULL;
1075         } else if (ca) {
1076                 bool drop = ob->dev == ca->dev_idx;
1077                 struct open_bucket *ob2;
1078                 unsigned i;
1079
1080                 if (!drop && ob->ec) {
1081                         unsigned nr_blocks;
1082
1083                         mutex_lock(&ob->ec->lock);
1084                         nr_blocks = bkey_i_to_stripe(&ob->ec->new_stripe.key)->v.nr_blocks;
1085
1086                         for (i = 0; i < nr_blocks; i++) {
1087                                 if (!ob->ec->blocks[i])
1088                                         continue;
1089
1090                                 ob2 = c->open_buckets + ob->ec->blocks[i];
1091                                 drop |= ob2->dev == ca->dev_idx;
1092                         }
1093                         mutex_unlock(&ob->ec->lock);
1094                 }
1095
1096                 return drop;
1097         } else {
1098                 return true;
1099         }
1100 }
1101
1102 static void bch2_writepoint_stop(struct bch_fs *c, struct bch_dev *ca,
1103                                  bool ec, struct write_point *wp)
1104 {
1105         struct open_buckets ptrs = { .nr = 0 };
1106         struct open_bucket *ob;
1107         unsigned i;
1108
1109         mutex_lock(&wp->lock);
1110         open_bucket_for_each(c, &wp->ptrs, ob, i)
1111                 if (should_drop_bucket(ob, c, ca, ec))
1112                         bch2_open_bucket_put(c, ob);
1113                 else
1114                         ob_push(c, &ptrs, ob);
1115         wp->ptrs = ptrs;
1116         mutex_unlock(&wp->lock);
1117 }
1118
1119 void bch2_open_buckets_stop(struct bch_fs *c, struct bch_dev *ca,
1120                             bool ec)
1121 {
1122         unsigned i;
1123
1124         /* Next, close write points that point to this device... */
1125         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1126                 bch2_writepoint_stop(c, ca, ec, &c->write_points[i]);
1127
1128         bch2_writepoint_stop(c, ca, ec, &c->copygc_write_point);
1129         bch2_writepoint_stop(c, ca, ec, &c->rebalance_write_point);
1130         bch2_writepoint_stop(c, ca, ec, &c->btree_write_point);
1131
1132         mutex_lock(&c->btree_reserve_cache_lock);
1133         while (c->btree_reserve_cache_nr) {
1134                 struct btree_alloc *a =
1135                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1136
1137                 bch2_open_buckets_put(c, &a->ob);
1138         }
1139         mutex_unlock(&c->btree_reserve_cache_lock);
1140
1141         spin_lock(&c->freelist_lock);
1142         i = 0;
1143         while (i < c->open_buckets_partial_nr) {
1144                 struct open_bucket *ob =
1145                         c->open_buckets + c->open_buckets_partial[i];
1146
1147                 if (should_drop_bucket(ob, c, ca, ec)) {
1148                         --c->open_buckets_partial_nr;
1149                         swap(c->open_buckets_partial[i],
1150                              c->open_buckets_partial[c->open_buckets_partial_nr]);
1151                         ob->on_partial_list = false;
1152                         spin_unlock(&c->freelist_lock);
1153                         bch2_open_bucket_put(c, ob);
1154                         spin_lock(&c->freelist_lock);
1155                 } else {
1156                         i++;
1157                 }
1158         }
1159         spin_unlock(&c->freelist_lock);
1160
1161         bch2_ec_stop_dev(c, ca);
1162 }
1163
1164 static inline struct hlist_head *writepoint_hash(struct bch_fs *c,
1165                                                  unsigned long write_point)
1166 {
1167         unsigned hash =
1168                 hash_long(write_point, ilog2(ARRAY_SIZE(c->write_points_hash)));
1169
1170         return &c->write_points_hash[hash];
1171 }
1172
1173 static struct write_point *__writepoint_find(struct hlist_head *head,
1174                                              unsigned long write_point)
1175 {
1176         struct write_point *wp;
1177
1178         rcu_read_lock();
1179         hlist_for_each_entry_rcu(wp, head, node)
1180                 if (wp->write_point == write_point)
1181                         goto out;
1182         wp = NULL;
1183 out:
1184         rcu_read_unlock();
1185         return wp;
1186 }
1187
1188 static inline bool too_many_writepoints(struct bch_fs *c, unsigned factor)
1189 {
1190         u64 stranded    = c->write_points_nr * c->bucket_size_max;
1191         u64 free        = bch2_fs_usage_read_short(c).free;
1192
1193         return stranded * factor > free;
1194 }
1195
1196 static bool try_increase_writepoints(struct bch_fs *c)
1197 {
1198         struct write_point *wp;
1199
1200         if (c->write_points_nr == ARRAY_SIZE(c->write_points) ||
1201             too_many_writepoints(c, 32))
1202                 return false;
1203
1204         wp = c->write_points + c->write_points_nr++;
1205         hlist_add_head_rcu(&wp->node, writepoint_hash(c, wp->write_point));
1206         return true;
1207 }
1208
1209 static bool try_decrease_writepoints(struct btree_trans *trans, unsigned old_nr)
1210 {
1211         struct bch_fs *c = trans->c;
1212         struct write_point *wp;
1213         struct open_bucket *ob;
1214         unsigned i;
1215
1216         mutex_lock(&c->write_points_hash_lock);
1217         if (c->write_points_nr < old_nr) {
1218                 mutex_unlock(&c->write_points_hash_lock);
1219                 return true;
1220         }
1221
1222         if (c->write_points_nr == 1 ||
1223             !too_many_writepoints(c, 8)) {
1224                 mutex_unlock(&c->write_points_hash_lock);
1225                 return false;
1226         }
1227
1228         wp = c->write_points + --c->write_points_nr;
1229
1230         hlist_del_rcu(&wp->node);
1231         mutex_unlock(&c->write_points_hash_lock);
1232
1233         bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1234         open_bucket_for_each(c, &wp->ptrs, ob, i)
1235                 open_bucket_free_unused(c, ob);
1236         wp->ptrs.nr = 0;
1237         mutex_unlock(&wp->lock);
1238         return true;
1239 }
1240
1241 static struct write_point *writepoint_find(struct btree_trans *trans,
1242                                            unsigned long write_point)
1243 {
1244         struct bch_fs *c = trans->c;
1245         struct write_point *wp, *oldest;
1246         struct hlist_head *head;
1247
1248         if (!(write_point & 1UL)) {
1249                 wp = (struct write_point *) write_point;
1250                 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1251                 return wp;
1252         }
1253
1254         head = writepoint_hash(c, write_point);
1255 restart_find:
1256         wp = __writepoint_find(head, write_point);
1257         if (wp) {
1258 lock_wp:
1259                 bch2_trans_mutex_lock_norelock(trans, &wp->lock);
1260                 if (wp->write_point == write_point)
1261                         goto out;
1262                 mutex_unlock(&wp->lock);
1263                 goto restart_find;
1264         }
1265 restart_find_oldest:
1266         oldest = NULL;
1267         for (wp = c->write_points;
1268              wp < c->write_points + c->write_points_nr; wp++)
1269                 if (!oldest || time_before64(wp->last_used, oldest->last_used))
1270                         oldest = wp;
1271
1272         bch2_trans_mutex_lock_norelock(trans, &oldest->lock);
1273         bch2_trans_mutex_lock_norelock(trans, &c->write_points_hash_lock);
1274         if (oldest >= c->write_points + c->write_points_nr ||
1275             try_increase_writepoints(c)) {
1276                 mutex_unlock(&c->write_points_hash_lock);
1277                 mutex_unlock(&oldest->lock);
1278                 goto restart_find_oldest;
1279         }
1280
1281         wp = __writepoint_find(head, write_point);
1282         if (wp && wp != oldest) {
1283                 mutex_unlock(&c->write_points_hash_lock);
1284                 mutex_unlock(&oldest->lock);
1285                 goto lock_wp;
1286         }
1287
1288         wp = oldest;
1289         hlist_del_rcu(&wp->node);
1290         wp->write_point = write_point;
1291         hlist_add_head_rcu(&wp->node, head);
1292         mutex_unlock(&c->write_points_hash_lock);
1293 out:
1294         wp->last_used = local_clock();
1295         return wp;
1296 }
1297
1298 static noinline void
1299 deallocate_extra_replicas(struct bch_fs *c,
1300                           struct open_buckets *ptrs,
1301                           struct open_buckets *ptrs_no_use,
1302                           unsigned extra_replicas)
1303 {
1304         struct open_buckets ptrs2 = { 0 };
1305         struct open_bucket *ob;
1306         unsigned i;
1307
1308         open_bucket_for_each(c, ptrs, ob, i) {
1309                 unsigned d = bch_dev_bkey_exists(c, ob->dev)->mi.durability;
1310
1311                 if (d && d <= extra_replicas) {
1312                         extra_replicas -= d;
1313                         ob_push(c, ptrs_no_use, ob);
1314                 } else {
1315                         ob_push(c, &ptrs2, ob);
1316                 }
1317         }
1318
1319         *ptrs = ptrs2;
1320 }
1321
1322 /*
1323  * Get us an open_bucket we can allocate from, return with it locked:
1324  */
1325 int bch2_alloc_sectors_start_trans(struct btree_trans *trans,
1326                              unsigned target,
1327                              unsigned erasure_code,
1328                              struct write_point_specifier write_point,
1329                              struct bch_devs_list *devs_have,
1330                              unsigned nr_replicas,
1331                              unsigned nr_replicas_required,
1332                              enum bch_watermark watermark,
1333                              unsigned flags,
1334                              struct closure *cl,
1335                              struct write_point **wp_ret)
1336 {
1337         struct bch_fs *c = trans->c;
1338         struct write_point *wp;
1339         struct open_bucket *ob;
1340         struct open_buckets ptrs;
1341         unsigned nr_effective, write_points_nr;
1342         bool have_cache;
1343         int ret;
1344         int i;
1345
1346         if (!IS_ENABLED(CONFIG_BCACHEFS_ERASURE_CODING))
1347                 erasure_code = false;
1348
1349         BUG_ON(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS);
1350
1351         BUG_ON(!nr_replicas || !nr_replicas_required);
1352 retry:
1353         ptrs.nr         = 0;
1354         nr_effective    = 0;
1355         write_points_nr = c->write_points_nr;
1356         have_cache      = false;
1357
1358         *wp_ret = wp = writepoint_find(trans, write_point.v);
1359
1360         /* metadata may not allocate on cache devices: */
1361         if (wp->data_type != BCH_DATA_user)
1362                 have_cache = true;
1363
1364         if (target && !(flags & BCH_WRITE_ONLY_SPECIFIED_DEVS)) {
1365                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1366                                               target, erasure_code,
1367                                               nr_replicas, &nr_effective,
1368                                               &have_cache, watermark,
1369                                               flags, NULL);
1370                 if (!ret ||
1371                     bch2_err_matches(ret, BCH_ERR_transaction_restart))
1372                         goto alloc_done;
1373
1374                 /* Don't retry from all devices if we're out of open buckets: */
1375                 if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty))
1376                         goto allocate_blocking;
1377
1378                 /*
1379                  * Only try to allocate cache (durability = 0 devices) from the
1380                  * specified target:
1381                  */
1382                 have_cache = true;
1383
1384                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1385                                               0, erasure_code,
1386                                               nr_replicas, &nr_effective,
1387                                               &have_cache, watermark,
1388                                               flags, cl);
1389         } else {
1390 allocate_blocking:
1391                 ret = open_bucket_add_buckets(trans, &ptrs, wp, devs_have,
1392                                               target, erasure_code,
1393                                               nr_replicas, &nr_effective,
1394                                               &have_cache, watermark,
1395                                               flags, cl);
1396         }
1397 alloc_done:
1398         BUG_ON(!ret && nr_effective < nr_replicas);
1399
1400         if (erasure_code && !ec_open_bucket(c, &ptrs))
1401                 pr_debug("failed to get ec bucket: ret %u", ret);
1402
1403         if (ret == -BCH_ERR_insufficient_devices &&
1404             nr_effective >= nr_replicas_required)
1405                 ret = 0;
1406
1407         if (ret)
1408                 goto err;
1409
1410         if (nr_effective > nr_replicas)
1411                 deallocate_extra_replicas(c, &ptrs, &wp->ptrs, nr_effective - nr_replicas);
1412
1413         /* Free buckets we didn't use: */
1414         open_bucket_for_each(c, &wp->ptrs, ob, i)
1415                 open_bucket_free_unused(c, ob);
1416
1417         wp->ptrs = ptrs;
1418
1419         wp->sectors_free = UINT_MAX;
1420
1421         open_bucket_for_each(c, &wp->ptrs, ob, i)
1422                 wp->sectors_free = min(wp->sectors_free, ob->sectors_free);
1423
1424         BUG_ON(!wp->sectors_free || wp->sectors_free == UINT_MAX);
1425
1426         return 0;
1427 err:
1428         open_bucket_for_each(c, &wp->ptrs, ob, i)
1429                 if (ptrs.nr < ARRAY_SIZE(ptrs.v))
1430                         ob_push(c, &ptrs, ob);
1431                 else
1432                         open_bucket_free_unused(c, ob);
1433         wp->ptrs = ptrs;
1434
1435         mutex_unlock(&wp->lock);
1436
1437         if (bch2_err_matches(ret, BCH_ERR_freelist_empty) &&
1438             try_decrease_writepoints(trans, write_points_nr))
1439                 goto retry;
1440
1441         if (bch2_err_matches(ret, BCH_ERR_open_buckets_empty) ||
1442             bch2_err_matches(ret, BCH_ERR_freelist_empty))
1443                 return cl
1444                         ? -BCH_ERR_bucket_alloc_blocked
1445                         : -BCH_ERR_ENOSPC_bucket_alloc;
1446
1447         return ret;
1448 }
1449
1450 struct bch_extent_ptr bch2_ob_ptr(struct bch_fs *c, struct open_bucket *ob)
1451 {
1452         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1453
1454         return (struct bch_extent_ptr) {
1455                 .type   = 1 << BCH_EXTENT_ENTRY_ptr,
1456                 .gen    = ob->gen,
1457                 .dev    = ob->dev,
1458                 .offset = bucket_to_sector(ca, ob->bucket) +
1459                         ca->mi.bucket_size -
1460                         ob->sectors_free,
1461         };
1462 }
1463
1464 void bch2_alloc_sectors_append_ptrs(struct bch_fs *c, struct write_point *wp,
1465                                     struct bkey_i *k, unsigned sectors,
1466                                     bool cached)
1467 {
1468         bch2_alloc_sectors_append_ptrs_inlined(c, wp, k, sectors, cached);
1469 }
1470
1471 /*
1472  * Append pointers to the space we just allocated to @k, and mark @sectors space
1473  * as allocated out of @ob
1474  */
1475 void bch2_alloc_sectors_done(struct bch_fs *c, struct write_point *wp)
1476 {
1477         bch2_alloc_sectors_done_inlined(c, wp);
1478 }
1479
1480 static inline void writepoint_init(struct write_point *wp,
1481                                    enum bch_data_type type)
1482 {
1483         mutex_init(&wp->lock);
1484         wp->data_type = type;
1485
1486         INIT_WORK(&wp->index_update_work, bch2_write_point_do_index_updates);
1487         INIT_LIST_HEAD(&wp->writes);
1488         spin_lock_init(&wp->writes_lock);
1489 }
1490
1491 void bch2_fs_allocator_foreground_init(struct bch_fs *c)
1492 {
1493         struct open_bucket *ob;
1494         struct write_point *wp;
1495
1496         mutex_init(&c->write_points_hash_lock);
1497         c->write_points_nr = ARRAY_SIZE(c->write_points);
1498
1499         /* open bucket 0 is a sentinal NULL: */
1500         spin_lock_init(&c->open_buckets[0].lock);
1501
1502         for (ob = c->open_buckets + 1;
1503              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets); ob++) {
1504                 spin_lock_init(&ob->lock);
1505                 c->open_buckets_nr_free++;
1506
1507                 ob->freelist = c->open_buckets_freelist;
1508                 c->open_buckets_freelist = ob - c->open_buckets;
1509         }
1510
1511         writepoint_init(&c->btree_write_point,          BCH_DATA_btree);
1512         writepoint_init(&c->rebalance_write_point,      BCH_DATA_user);
1513         writepoint_init(&c->copygc_write_point,         BCH_DATA_user);
1514
1515         for (wp = c->write_points;
1516              wp < c->write_points + c->write_points_nr; wp++) {
1517                 writepoint_init(wp, BCH_DATA_user);
1518
1519                 wp->last_used   = local_clock();
1520                 wp->write_point = (unsigned long) wp;
1521                 hlist_add_head_rcu(&wp->node,
1522                                    writepoint_hash(c, wp->write_point));
1523         }
1524 }
1525
1526 static void bch2_open_bucket_to_text(struct printbuf *out, struct bch_fs *c, struct open_bucket *ob)
1527 {
1528         struct bch_dev *ca = bch_dev_bkey_exists(c, ob->dev);
1529         unsigned data_type = ob->data_type;
1530         barrier(); /* READ_ONCE() doesn't work on bitfields */
1531
1532         prt_printf(out, "%zu ref %u %s %u:%llu gen %u allocated %u/%u",
1533                    ob - c->open_buckets,
1534                    atomic_read(&ob->pin),
1535                    data_type < BCH_DATA_NR ? bch2_data_types[data_type] : "invalid data type",
1536                    ob->dev, ob->bucket, ob->gen,
1537                    ca->mi.bucket_size - ob->sectors_free, ca->mi.bucket_size);
1538         if (ob->ec)
1539                 prt_printf(out, " ec idx %llu", ob->ec->idx);
1540         if (ob->on_partial_list)
1541                 prt_str(out, " partial");
1542         prt_newline(out);
1543 }
1544
1545 void bch2_open_buckets_to_text(struct printbuf *out, struct bch_fs *c)
1546 {
1547         struct open_bucket *ob;
1548
1549         out->atomic++;
1550
1551         for (ob = c->open_buckets;
1552              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1553              ob++) {
1554                 spin_lock(&ob->lock);
1555                 if (ob->valid && !ob->on_partial_list)
1556                         bch2_open_bucket_to_text(out, c, ob);
1557                 spin_unlock(&ob->lock);
1558         }
1559
1560         --out->atomic;
1561 }
1562
1563 void bch2_open_buckets_partial_to_text(struct printbuf *out, struct bch_fs *c)
1564 {
1565         unsigned i;
1566
1567         out->atomic++;
1568         spin_lock(&c->freelist_lock);
1569
1570         for (i = 0; i < c->open_buckets_partial_nr; i++)
1571                 bch2_open_bucket_to_text(out, c,
1572                                 c->open_buckets + c->open_buckets_partial[i]);
1573
1574         spin_unlock(&c->freelist_lock);
1575         --out->atomic;
1576 }
1577
1578 static const char * const bch2_write_point_states[] = {
1579 #define x(n)    #n,
1580         WRITE_POINT_STATES()
1581 #undef x
1582         NULL
1583 };
1584
1585 static void bch2_write_point_to_text(struct printbuf *out, struct bch_fs *c,
1586                                      struct write_point *wp)
1587 {
1588         struct open_bucket *ob;
1589         unsigned i;
1590
1591         prt_printf(out, "%lu: ", wp->write_point);
1592         prt_human_readable_u64(out, wp->sectors_allocated);
1593
1594         prt_printf(out, " last wrote: ");
1595         bch2_pr_time_units(out, sched_clock() - wp->last_used);
1596
1597         for (i = 0; i < WRITE_POINT_STATE_NR; i++) {
1598                 prt_printf(out, " %s: ", bch2_write_point_states[i]);
1599                 bch2_pr_time_units(out, wp->time[i]);
1600         }
1601
1602         prt_newline(out);
1603
1604         printbuf_indent_add(out, 2);
1605         open_bucket_for_each(c, &wp->ptrs, ob, i)
1606                 bch2_open_bucket_to_text(out, c, ob);
1607         printbuf_indent_sub(out, 2);
1608 }
1609
1610 void bch2_write_points_to_text(struct printbuf *out, struct bch_fs *c)
1611 {
1612         struct write_point *wp;
1613
1614         prt_str(out, "Foreground write points\n");
1615         for (wp = c->write_points;
1616              wp < c->write_points + ARRAY_SIZE(c->write_points);
1617              wp++)
1618                 bch2_write_point_to_text(out, c, wp);
1619
1620         prt_str(out, "Copygc write point\n");
1621         bch2_write_point_to_text(out, c, &c->copygc_write_point);
1622
1623         prt_str(out, "Rebalance write point\n");
1624         bch2_write_point_to_text(out, c, &c->rebalance_write_point);
1625
1626         prt_str(out, "Btree write point\n");
1627         bch2_write_point_to_text(out, c, &c->btree_write_point);
1628 }