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

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "bcachefs.h"
   3 #include "alloc_background.h"
   4 #include "alloc_foreground.h"
   5 #include "btree_cache.h"
   6 #include "btree_io.h"
   7 #include "btree_key_cache.h"
   8 #include "btree_update.h"
   9 #include "btree_update_interior.h"
  10 #include "btree_gc.h"
  11 #include "buckets.h"
  12 #include "clock.h"
  13 #include "debug.h"
  14 #include "ec.h"
  15 #include "error.h"
  16 #include "recovery.h"
  17 #include "varint.h"
  18
  19 #include <linux/kthread.h>
  20 #include <linux/math64.h>
  21 #include <linux/random.h>
  22 #include <linux/rculist.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/sched/task.h>
  25 #include <linux/sort.h>
  26 #include <trace/events/bcachefs.h>
  27
  28 const char * const bch2_allocator_states[] = {
  29 #define x(n)    #n,
  30         ALLOC_THREAD_STATES()
  31 #undef x
  32         NULL
  33 };
  34
  35 static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
  36 #define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8,
  37         BCH_ALLOC_FIELDS_V1()
  38 #undef x
  39 };
  40
  41 /* Persistent alloc info: */
  42
  43 static inline u64 alloc_field_v1_get(const struct bch_alloc *a,
  44                                      const void **p, unsigned field)
  45 {
  46         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  47         u64 v;
  48
  49         if (!(a->fields & (1 << field)))
  50                 return 0;
  51
  52         switch (bytes) {
  53         case 1:
  54                 v = *((const u8 *) *p);
  55                 break;
  56         case 2:
  57                 v = le16_to_cpup(*p);
  58                 break;
  59         case 4:
  60                 v = le32_to_cpup(*p);
  61                 break;
  62         case 8:
  63                 v = le64_to_cpup(*p);
  64                 break;
  65         default:
  66                 BUG();
  67         }
  68
  69         *p += bytes;
  70         return v;
  71 }
  72
  73 static inline void alloc_field_v1_put(struct bkey_i_alloc *a, void **p,
  74                                       unsigned field, u64 v)
  75 {
  76         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  77
  78         if (!v)
  79                 return;
  80
  81         a->v.fields |= 1 << field;
  82
  83         switch (bytes) {
  84         case 1:
  85                 *((u8 *) *p) = v;
  86                 break;
  87         case 2:
  88                 *((__le16 *) *p) = cpu_to_le16(v);
  89                 break;
  90         case 4:
  91                 *((__le32 *) *p) = cpu_to_le32(v);
  92                 break;
  93         case 8:
  94                 *((__le64 *) *p) = cpu_to_le64(v);
  95                 break;
  96         default:
  97                 BUG();
  98         }
  99
 100         *p += bytes;
 101 }
 102
 103 static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out,
 104                                  struct bkey_s_c k)
 105 {
 106         const struct bch_alloc *in = bkey_s_c_to_alloc(k).v;
 107         const void *d = in->data;
 108         unsigned idx = 0;
 109
 110         out->gen = in->gen;
 111
 112 #define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++);
 113         BCH_ALLOC_FIELDS_V1()
 114 #undef  x
 115 }
 116
 117 static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out,
 118                                 struct bkey_s_c k)
 119 {
 120         struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k);
 121         const u8 *in = a.v->data;
 122         const u8 *end = bkey_val_end(a);
 123         unsigned fieldnr = 0;
 124         int ret;
 125         u64 v;
 126
 127         out->gen        = a.v->gen;
 128         out->oldest_gen = a.v->oldest_gen;
 129         out->data_type  = a.v->data_type;
 130
 131 #define x(_name, _bits)                                                 \
 132         if (fieldnr < a.v->nr_fields) {                                 \
 133                 ret = bch2_varint_decode(in, end, &v);                  \
 134                 if (ret < 0)                                            \
 135                         return ret;                                     \
 136                 in += ret;                                              \
 137         } else {                                                        \
 138                 v = 0;                                                  \
 139         }                                                               \
 140         out->_name = v;                                                 \
 141         if (v != out->_name)                                            \
 142                 return -1;                                              \
 143         fieldnr++;
 144
 145         BCH_ALLOC_FIELDS_V2()
 146 #undef  x
 147         return 0;
 148 }
 149
 150 static void bch2_alloc_pack_v2(struct bkey_alloc_buf *dst,
 151                                const struct bkey_alloc_unpacked src)
 152 {
 153         struct bkey_i_alloc_v2 *a = bkey_alloc_v2_init(&dst->k);
 154         unsigned nr_fields = 0, last_nonzero_fieldnr = 0;
 155         u8 *out = a->v.data;
 156         u8 *end = (void *) &dst[1];
 157         u8 *last_nonzero_field = out;
 158         unsigned bytes;
 159
 160         a->k.p          = POS(src.dev, src.bucket);
 161         a->v.gen        = src.gen;
 162         a->v.oldest_gen = src.oldest_gen;
 163         a->v.data_type  = src.data_type;
 164
 165 #define x(_name, _bits)                                                 \
 166         nr_fields++;                                                    \
 167                                                                         \
 168         if (src._name) {                                                \
 169                 out += bch2_varint_encode(out, src._name);              \
 170                                                                         \
 171                 last_nonzero_field = out;                               \
 172                 last_nonzero_fieldnr = nr_fields;                       \
 173         } else {                                                        \
 174                 *out++ = 0;                                             \
 175         }
 176
 177         BCH_ALLOC_FIELDS_V2()
 178 #undef  x
 179         BUG_ON(out > end);
 180
 181         out = last_nonzero_field;
 182         a->v.nr_fields = last_nonzero_fieldnr;
 183
 184         bytes = (u8 *) out - (u8 *) &a->v;
 185         set_bkey_val_bytes(&a->k, bytes);
 186         memset_u64s_tail(&a->v, 0, bytes);
 187 }
 188
 189 struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 190 {
 191         struct bkey_alloc_unpacked ret = {
 192                 .dev    = k.k->p.inode,
 193                 .bucket = k.k->p.offset,
 194                 .gen    = 0,
 195         };
 196
 197         if (k.k->type == KEY_TYPE_alloc_v2)
 198                 bch2_alloc_unpack_v2(&ret, k);
 199         else if (k.k->type == KEY_TYPE_alloc)
 200                 bch2_alloc_unpack_v1(&ret, k);
 201
 202         return ret;
 203 }
 204
 205 void bch2_alloc_pack(struct bch_fs *c,
 206                      struct bkey_alloc_buf *dst,
 207                      const struct bkey_alloc_unpacked src)
 208 {
 209         bch2_alloc_pack_v2(dst, src);
 210 }
 211
 212 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
 213 {
 214         unsigned i, bytes = offsetof(struct bch_alloc, data);
 215
 216         for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++)
 217                 if (a->fields & (1 << i))
 218                         bytes += BCH_ALLOC_V1_FIELD_BYTES[i];
 219
 220         return DIV_ROUND_UP(bytes, sizeof(u64));
 221 }
 222
 223 const char *bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k)
 224 {
 225         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 226
 227         if (k.k->p.inode >= c->sb.nr_devices ||
 228             !c->devs[k.k->p.inode])
 229                 return "invalid device";
 230
 231         /* allow for unknown fields */
 232         if (bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v))
 233                 return "incorrect value size";
 234
 235         return NULL;
 236 }
 237
 238 const char *bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k)
 239 {
 240         struct bkey_alloc_unpacked u;
 241
 242         if (k.k->p.inode >= c->sb.nr_devices ||
 243             !c->devs[k.k->p.inode])
 244                 return "invalid device";
 245
 246         if (bch2_alloc_unpack_v2(&u, k))
 247                 return "unpack error";
 248
 249         return NULL;
 250 }
 251
 252 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c,
 253                            struct bkey_s_c k)
 254 {
 255         struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 256
 257         pr_buf(out, "gen %u oldest_gen %u data_type %s",
 258                u.gen, u.oldest_gen, bch2_data_types[u.data_type]);
 259 #define x(_name, ...)   pr_buf(out, " " #_name " %llu", (u64) u._name);
 260         BCH_ALLOC_FIELDS_V2()
 261 #undef  x
 262 }
 263
 264 static int bch2_alloc_read_fn(struct bch_fs *c, struct bkey_s_c k)
 265 {
 266         struct bch_dev *ca;
 267         struct bucket *g;
 268         struct bkey_alloc_unpacked u;
 269
 270         if (k.k->type != KEY_TYPE_alloc &&
 271             k.k->type != KEY_TYPE_alloc_v2)
 272                 return 0;
 273
 274         ca = bch_dev_bkey_exists(c, k.k->p.inode);
 275         g = bucket(ca, k.k->p.offset);
 276         u = bch2_alloc_unpack(k);
 277
 278         g->_mark.gen            = u.gen;
 279         g->_mark.data_type      = u.data_type;
 280         g->_mark.dirty_sectors  = u.dirty_sectors;
 281         g->_mark.cached_sectors = u.cached_sectors;
 282         g->io_time[READ]        = u.read_time;
 283         g->io_time[WRITE]       = u.write_time;
 284         g->oldest_gen           = u.oldest_gen;
 285         g->gen_valid            = 1;
 286
 287         return 0;
 288 }
 289
 290 int bch2_alloc_read(struct bch_fs *c)
 291 {
 292         int ret;
 293
 294         down_read(&c->gc_lock);
 295         ret = bch2_btree_and_journal_walk(c, BTREE_ID_alloc, bch2_alloc_read_fn);
 296         up_read(&c->gc_lock);
 297         if (ret) {
 298                 bch_err(c, "error reading alloc info: %i", ret);
 299                 return ret;
 300         }
 301
 302         return 0;
 303 }
 304
 305 static int bch2_alloc_write_key(struct btree_trans *trans,
 306                                 struct btree_iter *iter,
 307                                 unsigned flags)
 308 {
 309         struct bch_fs *c = trans->c;
 310         struct bkey_s_c k;
 311         struct bch_dev *ca;
 312         struct bucket *g;
 313         struct bucket_mark m;
 314         struct bkey_alloc_unpacked old_u, new_u;
 315         struct bkey_alloc_buf a;
 316         int ret;
 317 retry:
 318         bch2_trans_begin(trans);
 319
 320         ret = bch2_btree_key_cache_flush(trans,
 321                         BTREE_ID_alloc, iter->pos);
 322         if (ret)
 323                 goto err;
 324
 325         k = bch2_btree_iter_peek_slot(iter);
 326         ret = bkey_err(k);
 327         if (ret)
 328                 goto err;
 329
 330         old_u = bch2_alloc_unpack(k);
 331
 332         percpu_down_read(&c->mark_lock);
 333         ca      = bch_dev_bkey_exists(c, iter->pos.inode);
 334         g       = bucket(ca, iter->pos.offset);
 335         m       = READ_ONCE(g->mark);
 336         new_u   = alloc_mem_to_key(iter, g, m);
 337         percpu_up_read(&c->mark_lock);
 338
 339         if (!bkey_alloc_unpacked_cmp(old_u, new_u))
 340                 return 0;
 341
 342         bch2_alloc_pack(c, &a, new_u);
 343         bch2_trans_update(trans, iter, &a.k,
 344                           BTREE_TRIGGER_NORUN);
 345         ret = bch2_trans_commit(trans, NULL, NULL,
 346                                 BTREE_INSERT_NOFAIL|flags);
 347 err:
 348         if (ret == -EINTR)
 349                 goto retry;
 350         return ret;
 351 }
 352
 353 int bch2_alloc_write(struct bch_fs *c, unsigned flags)
 354 {
 355         struct btree_trans trans;
 356         struct btree_iter *iter;
 357         struct bch_dev *ca;
 358         unsigned i;
 359         int ret = 0;
 360
 361         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
 362         iter = bch2_trans_get_iter(&trans, BTREE_ID_alloc, POS_MIN,
 363                                    BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 364
 365         for_each_member_device(ca, c, i) {
 366                 bch2_btree_iter_set_pos(iter,
 367                         POS(ca->dev_idx, ca->mi.first_bucket));
 368
 369                 while (iter->pos.offset < ca->mi.nbuckets) {
 370                         bch2_trans_cond_resched(&trans);
 371
 372                         ret = bch2_alloc_write_key(&trans, iter, flags);
 373                         if (ret) {
 374                                 percpu_ref_put(&ca->ref);
 375                                 goto err;
 376                         }
 377                         bch2_btree_iter_next_slot(iter);
 378                 }
 379         }
 380 err:
 381         bch2_trans_iter_put(&trans, iter);
 382         bch2_trans_exit(&trans);
 383         return ret;
 384 }
 385
 386 /* Bucket IO clocks: */
 387
 388 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
 389                               size_t bucket_nr, int rw)
 390 {
 391         struct bch_fs *c = trans->c;
 392         struct bch_dev *ca = bch_dev_bkey_exists(c, dev);
 393         struct btree_iter *iter;
 394         struct bucket *g;
 395         struct bkey_alloc_buf *a;
 396         struct bkey_alloc_unpacked u;
 397         u64 *time, now;
 398         int ret = 0;
 399
 400         iter = bch2_trans_get_iter(trans, BTREE_ID_alloc, POS(dev, bucket_nr),
 401                                    BTREE_ITER_CACHED|
 402                                    BTREE_ITER_CACHED_NOFILL|
 403                                    BTREE_ITER_INTENT);
 404         ret = bch2_btree_iter_traverse(iter);
 405         if (ret)
 406                 goto out;
 407
 408         a = bch2_trans_kmalloc(trans, sizeof(struct bkey_alloc_buf));
 409         ret = PTR_ERR_OR_ZERO(a);
 410         if (ret)
 411                 goto out;
 412
 413         percpu_down_read(&c->mark_lock);
 414         g = bucket(ca, bucket_nr);
 415         u = alloc_mem_to_key(iter, g, READ_ONCE(g->mark));
 416         percpu_up_read(&c->mark_lock);
 417
 418         time = rw == READ ? &u.read_time : &u.write_time;
 419         now = atomic64_read(&c->io_clock[rw].now);
 420         if (*time == now)
 421                 goto out;
 422
 423         *time = now;
 424
 425         bch2_alloc_pack(c, a, u);
 426         ret   = bch2_trans_update(trans, iter, &a->k, 0) ?:
 427                 bch2_trans_commit(trans, NULL, NULL, 0);
 428 out:
 429         bch2_trans_iter_put(trans, iter);
 430         return ret;
 431 }
 432
 433 /* Background allocator thread: */
 434
 435 /*
 436  * Scans for buckets to be invalidated, invalidates them, rewrites prios/gens
 437  * (marking them as invalidated on disk), then optionally issues discard
 438  * commands to the newly free buckets, then puts them on the various freelists.
 439  */
 440
 441 static bool bch2_can_invalidate_bucket(struct bch_dev *ca, size_t b,
 442                                        struct bucket_mark m)
 443 {
 444         u8 gc_gen;
 445
 446         if (!is_available_bucket(m))
 447                 return false;
 448
 449         if (m.owned_by_allocator)
 450                 return false;
 451
 452         if (ca->buckets_nouse &&
 453             test_bit(b, ca->buckets_nouse))
 454                 return false;
 455
 456         gc_gen = bucket_gc_gen(bucket(ca, b));
 457
 458         ca->inc_gen_needs_gc            += gc_gen >= BUCKET_GC_GEN_MAX / 2;
 459         ca->inc_gen_really_needs_gc     += gc_gen >= BUCKET_GC_GEN_MAX;
 460
 461         return gc_gen < BUCKET_GC_GEN_MAX;
 462 }
 463
 464 /*
 465  * Determines what order we're going to reuse buckets, smallest bucket_key()
 466  * first.
 467  */
 468
 469 static unsigned bucket_sort_key(struct bucket *g, struct bucket_mark m,
 470                                 u64 now, u64 last_seq_ondisk)
 471 {
 472         unsigned used = bucket_sectors_used(m);
 473
 474         if (used) {
 475                 /*
 476                  * Prefer to keep buckets that have been read more recently, and
 477                  * buckets that have more data in them:
 478                  */
 479                 u64 last_read = max_t(s64, 0, now - g->io_time[READ]);
 480                 u32 last_read_scaled = max_t(u64, U32_MAX, div_u64(last_read, used));
 481
 482                 return -last_read_scaled;
 483         } else {
 484                 /*
 485                  * Prefer to use buckets with smaller gc_gen so that we don't
 486                  * have to walk the btree and recalculate oldest_gen - but shift
 487                  * off the low bits so that buckets will still have equal sort
 488                  * keys when there's only a small difference, so that we can
 489                  * keep sequential buckets together:
 490                  */
 491                 return  (bucket_needs_journal_commit(m, last_seq_ondisk) << 4)|
 492                         (bucket_gc_gen(g) >> 4);
 493         }
 494 }
 495
 496 static inline int bucket_alloc_cmp(alloc_heap *h,
 497                                    struct alloc_heap_entry l,
 498                                    struct alloc_heap_entry r)
 499 {
 500         return  cmp_int(l.key, r.key) ?:
 501                 cmp_int(r.nr, l.nr) ?:
 502                 cmp_int(l.bucket, r.bucket);
 503 }
 504
 505 static inline int bucket_idx_cmp(const void *_l, const void *_r)
 506 {
 507         const struct alloc_heap_entry *l = _l, *r = _r;
 508
 509         return cmp_int(l->bucket, r->bucket);
 510 }
 511
 512 static void find_reclaimable_buckets_lru(struct bch_fs *c, struct bch_dev *ca)
 513 {
 514         struct bucket_array *buckets;
 515         struct alloc_heap_entry e = { 0 };
 516         u64 now, last_seq_ondisk;
 517         size_t b, i, nr = 0;
 518
 519         down_read(&ca->bucket_lock);
 520
 521         buckets = bucket_array(ca);
 522         ca->alloc_heap.used = 0;
 523         now = atomic64_read(&c->io_clock[READ].now);
 524         last_seq_ondisk = c->journal.last_seq_ondisk;
 525
 526         /*
 527          * Find buckets with lowest read priority, by building a maxheap sorted
 528          * by read priority and repeatedly replacing the maximum element until
 529          * all buckets have been visited.
 530          */
 531         for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++) {
 532                 struct bucket *g = &buckets->b[b];
 533                 struct bucket_mark m = READ_ONCE(g->mark);
 534                 unsigned key = bucket_sort_key(g, m, now, last_seq_ondisk);
 535
 536                 cond_resched();
 537
 538                 if (!bch2_can_invalidate_bucket(ca, b, m))
 539                         continue;
 540
 541                 if (e.nr && e.bucket + e.nr == b && e.key == key) {
 542                         e.nr++;
 543                 } else {
 544                         if (e.nr)
 545                                 heap_add_or_replace(&ca->alloc_heap, e,
 546                                         -bucket_alloc_cmp, NULL);
 547
 548                         e = (struct alloc_heap_entry) {
 549                                 .bucket = b,
 550                                 .nr     = 1,
 551                                 .key    = key,
 552                         };
 553                 }
 554         }
 555
 556         if (e.nr)
 557                 heap_add_or_replace(&ca->alloc_heap, e,
 558                                 -bucket_alloc_cmp, NULL);
 559
 560         for (i = 0; i < ca->alloc_heap.used; i++)
 561                 nr += ca->alloc_heap.data[i].nr;
 562
 563         while (nr - ca->alloc_heap.data[0].nr >= ALLOC_SCAN_BATCH(ca)) {
 564                 nr -= ca->alloc_heap.data[0].nr;
 565                 heap_pop(&ca->alloc_heap, e, -bucket_alloc_cmp, NULL);
 566         }
 567
 568         up_read(&ca->bucket_lock);
 569 }
 570
 571 static void find_reclaimable_buckets_fifo(struct bch_fs *c, struct bch_dev *ca)
 572 {
 573         struct bucket_array *buckets = bucket_array(ca);
 574         struct bucket_mark m;
 575         size_t b, start;
 576
 577         if (ca->fifo_last_bucket <  ca->mi.first_bucket ||
 578             ca->fifo_last_bucket >= ca->mi.nbuckets)
 579                 ca->fifo_last_bucket = ca->mi.first_bucket;
 580
 581         start = ca->fifo_last_bucket;
 582
 583         do {
 584                 ca->fifo_last_bucket++;
 585                 if (ca->fifo_last_bucket == ca->mi.nbuckets)
 586                         ca->fifo_last_bucket = ca->mi.first_bucket;
 587
 588                 b = ca->fifo_last_bucket;
 589                 m = READ_ONCE(buckets->b[b].mark);
 590
 591                 if (bch2_can_invalidate_bucket(ca, b, m)) {
 592                         struct alloc_heap_entry e = { .bucket = b, .nr = 1, };
 593
 594                         heap_add(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
 595                         if (heap_full(&ca->alloc_heap))
 596                                 break;
 597                 }
 598
 599                 cond_resched();
 600         } while (ca->fifo_last_bucket != start);
 601 }
 602
 603 static void find_reclaimable_buckets_random(struct bch_fs *c, struct bch_dev *ca)
 604 {
 605         struct bucket_array *buckets = bucket_array(ca);
 606         struct bucket_mark m;
 607         size_t checked, i;
 608
 609         for (checked = 0;
 610              checked < ca->mi.nbuckets / 2;
 611              checked++) {
 612                 size_t b = bch2_rand_range(ca->mi.nbuckets -
 613                                            ca->mi.first_bucket) +
 614                         ca->mi.first_bucket;
 615
 616                 m = READ_ONCE(buckets->b[b].mark);
 617
 618                 if (bch2_can_invalidate_bucket(ca, b, m)) {
 619                         struct alloc_heap_entry e = { .bucket = b, .nr = 1, };
 620
 621                         heap_add(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
 622                         if (heap_full(&ca->alloc_heap))
 623                                 break;
 624                 }
 625
 626                 cond_resched();
 627         }
 628
 629         sort(ca->alloc_heap.data,
 630              ca->alloc_heap.used,
 631              sizeof(ca->alloc_heap.data[0]),
 632              bucket_idx_cmp, NULL);
 633
 634         /* remove duplicates: */
 635         for (i = 0; i + 1 < ca->alloc_heap.used; i++)
 636                 if (ca->alloc_heap.data[i].bucket ==
 637                     ca->alloc_heap.data[i + 1].bucket)
 638                         ca->alloc_heap.data[i].nr = 0;
 639 }
 640
 641 static size_t find_reclaimable_buckets(struct bch_fs *c, struct bch_dev *ca)
 642 {
 643         size_t i, nr = 0;
 644
 645         ca->inc_gen_needs_gc                    = 0;
 646         ca->inc_gen_really_needs_gc             = 0;
 647
 648         switch (ca->mi.replacement) {
 649         case BCH_CACHE_REPLACEMENT_lru:
 650                 find_reclaimable_buckets_lru(c, ca);
 651                 break;
 652         case BCH_CACHE_REPLACEMENT_fifo:
 653                 find_reclaimable_buckets_fifo(c, ca);
 654                 break;
 655         case BCH_CACHE_REPLACEMENT_random:
 656                 find_reclaimable_buckets_random(c, ca);
 657                 break;
 658         }
 659
 660         heap_resort(&ca->alloc_heap, bucket_alloc_cmp, NULL);
 661
 662         for (i = 0; i < ca->alloc_heap.used; i++)
 663                 nr += ca->alloc_heap.data[i].nr;
 664
 665         return nr;
 666 }
 667
 668 /*
 669  * returns sequence number of most recent journal entry that updated this
 670  * bucket:
 671  */
 672 static u64 bucket_journal_seq(struct bch_fs *c, struct bucket_mark m)
 673 {
 674         if (m.journal_seq_valid) {
 675                 u64 journal_seq = atomic64_read(&c->journal.seq);
 676                 u64 bucket_seq  = journal_seq;
 677
 678                 bucket_seq &= ~((u64) U16_MAX);
 679                 bucket_seq |= m.journal_seq;
 680
 681                 if (bucket_seq > journal_seq)
 682                         bucket_seq -= 1 << 16;
 683
 684                 return bucket_seq;
 685         } else {
 686                 return 0;
 687         }
 688 }
 689
 690 static int bucket_invalidate_btree(struct btree_trans *trans,
 691                                    struct bch_dev *ca, u64 b)
 692 {
 693         struct bch_fs *c = trans->c;
 694         struct bkey_alloc_buf *a;
 695         struct bkey_alloc_unpacked u;
 696         struct bucket *g;
 697         struct bucket_mark m;
 698         struct btree_iter *iter =
 699                 bch2_trans_get_iter(trans, BTREE_ID_alloc,
 700                                     POS(ca->dev_idx, b),
 701                                     BTREE_ITER_CACHED|
 702                                     BTREE_ITER_CACHED_NOFILL|
 703                                     BTREE_ITER_INTENT);
 704         int ret;
 705
 706         a = bch2_trans_kmalloc(trans, sizeof(*a));
 707         ret = PTR_ERR_OR_ZERO(a);
 708         if (ret)
 709                 goto err;
 710
 711         ret = bch2_btree_iter_traverse(iter);
 712         if (ret)
 713                 goto err;
 714
 715         percpu_down_read(&c->mark_lock);
 716         g = bucket(ca, b);
 717         m = READ_ONCE(g->mark);
 718         u = alloc_mem_to_key(iter, g, m);
 719         percpu_up_read(&c->mark_lock);
 720
 721         u.gen++;
 722         u.data_type     = 0;
 723         u.dirty_sectors = 0;
 724         u.cached_sectors = 0;
 725         u.read_time     = atomic64_read(&c->io_clock[READ].now);
 726         u.write_time    = atomic64_read(&c->io_clock[WRITE].now);
 727
 728         bch2_alloc_pack(c, a, u);
 729         bch2_trans_update(trans, iter, &a->k, BTREE_TRIGGER_BUCKET_INVALIDATE);
 730 err:
 731         bch2_trans_iter_put(trans, iter);
 732         return ret;
 733 }
 734
 735 static int bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca,
 736                                       u64 *journal_seq, unsigned flags)
 737 {
 738         struct bucket *g;
 739         struct bucket_mark m;
 740         size_t b;
 741         int ret = 0;
 742
 743         BUG_ON(!ca->alloc_heap.used ||
 744                !ca->alloc_heap.data[0].nr);
 745         b = ca->alloc_heap.data[0].bucket;
 746
 747         /* first, put on free_inc and mark as owned by allocator: */
 748         percpu_down_read(&c->mark_lock);
 749         g = bucket(ca, b);
 750         m = READ_ONCE(g->mark);
 751
 752         BUG_ON(m.dirty_sectors);
 753
 754         bch2_mark_alloc_bucket(c, ca, b, true);
 755
 756         spin_lock(&c->freelist_lock);
 757         verify_not_on_freelist(c, ca, b);
 758         BUG_ON(!fifo_push(&ca->free_inc, b));
 759         spin_unlock(&c->freelist_lock);
 760
 761         /*
 762          * If we're not invalidating cached data, we only increment the bucket
 763          * gen in memory here, the incremented gen will be updated in the btree
 764          * by bch2_trans_mark_pointer():
 765          */
 766         if (!m.cached_sectors &&
 767             !bucket_needs_journal_commit(m, c->journal.last_seq_ondisk)) {
 768                 BUG_ON(m.data_type);
 769                 bucket_cmpxchg(g, m, m.gen++);
 770                 percpu_up_read(&c->mark_lock);
 771                 goto out;
 772         }
 773
 774         percpu_up_read(&c->mark_lock);
 775
 776         /*
 777          * If the read-only path is trying to shut down, we can't be generating
 778          * new btree updates:
 779          */
 780         if (test_bit(BCH_FS_ALLOCATOR_STOPPING, &c->flags)) {
 781                 ret = 1;
 782                 goto out;
 783         }
 784
 785         ret = bch2_trans_do(c, NULL, journal_seq,
 786                             BTREE_INSERT_NOCHECK_RW|
 787                             BTREE_INSERT_NOFAIL|
 788                             BTREE_INSERT_JOURNAL_RESERVED|
 789                             flags,
 790                             bucket_invalidate_btree(&trans, ca, b));
 791 out:
 792         if (!ret) {
 793                 /* remove from alloc_heap: */
 794                 struct alloc_heap_entry e, *top = ca->alloc_heap.data;
 795
 796                 top->bucket++;
 797                 top->nr--;
 798
 799                 if (!top->nr)
 800                         heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
 801
 802                 /*
 803                  * Make sure we flush the last journal entry that updated this
 804                  * bucket (i.e. deleting the last reference) before writing to
 805                  * this bucket again:
 806                  */
 807                 *journal_seq = max(*journal_seq, bucket_journal_seq(c, m));
 808         } else {
 809                 size_t b2;
 810
 811                 /* remove from free_inc: */
 812                 percpu_down_read(&c->mark_lock);
 813                 spin_lock(&c->freelist_lock);
 814
 815                 bch2_mark_alloc_bucket(c, ca, b, false);
 816
 817                 BUG_ON(!fifo_pop_back(&ca->free_inc, b2));
 818                 BUG_ON(b != b2);
 819
 820                 spin_unlock(&c->freelist_lock);
 821                 percpu_up_read(&c->mark_lock);
 822         }
 823
 824         return ret < 0 ? ret : 0;
 825 }
 826
 827 /*
 828  * Pull buckets off ca->alloc_heap, invalidate them, move them to ca->free_inc:
 829  */
 830 static int bch2_invalidate_buckets(struct bch_fs *c, struct bch_dev *ca)
 831 {
 832         u64 journal_seq = 0;
 833         int ret = 0;
 834
 835         /* Only use nowait if we've already invalidated at least one bucket: */
 836         while (!ret &&
 837                !fifo_full(&ca->free_inc) &&
 838                ca->alloc_heap.used) {
 839                 ret = bch2_invalidate_one_bucket(c, ca, &journal_seq,
 840                                 (!fifo_empty(&ca->free_inc)
 841                                  ? BTREE_INSERT_NOWAIT : 0));
 842                 /*
 843                  * We only want to batch up invalidates when they're going to
 844                  * require flushing the journal:
 845                  */
 846                 if (!journal_seq)
 847                         break;
 848         }
 849
 850         /* If we used NOWAIT, don't return the error: */
 851         if (!fifo_empty(&ca->free_inc))
 852                 ret = 0;
 853         if (ret) {
 854                 bch_err(ca, "error invalidating buckets: %i", ret);
 855                 return ret;
 856         }
 857
 858         if (journal_seq)
 859                 ret = bch2_journal_flush_seq(&c->journal, journal_seq);
 860         if (ret) {
 861                 bch_err(ca, "journal error: %i", ret);
 862                 return ret;
 863         }
 864
 865         return 0;
 866 }
 867
 868 static void alloc_thread_set_state(struct bch_dev *ca, unsigned new_state)
 869 {
 870         if (ca->allocator_state != new_state) {
 871                 ca->allocator_state = new_state;
 872                 closure_wake_up(&ca->fs->freelist_wait);
 873         }
 874 }
 875
 876 static int push_invalidated_bucket(struct bch_fs *c, struct bch_dev *ca, u64 b)
 877 {
 878         unsigned i;
 879         int ret = 0;
 880
 881         spin_lock(&c->freelist_lock);
 882         for (i = 0; i < RESERVE_NR; i++) {
 883                 /*
 884                  * Don't strand buckets on the copygc freelist until
 885                  * after recovery is finished:
 886                  */
 887                 if (i == RESERVE_MOVINGGC &&
 888                     !test_bit(BCH_FS_STARTED, &c->flags))
 889                         continue;
 890
 891                 if (fifo_push(&ca->free[i], b)) {
 892                         fifo_pop(&ca->free_inc, b);
 893                         ret = 1;
 894                         break;
 895                 }
 896         }
 897         spin_unlock(&c->freelist_lock);
 898
 899         ca->allocator_state = ret
 900                 ? ALLOCATOR_running
 901                 : ALLOCATOR_blocked_full;
 902         closure_wake_up(&c->freelist_wait);
 903         return ret;
 904 }
 905
 906 static void discard_one_bucket(struct bch_fs *c, struct bch_dev *ca, u64 b)
 907 {
 908         if (ca->mi.discard &&
 909             blk_queue_discard(bdev_get_queue(ca->disk_sb.bdev)))
 910                 blkdev_issue_discard(ca->disk_sb.bdev, bucket_to_sector(ca, b),
 911                                      ca->mi.bucket_size, GFP_NOFS, 0);
 912 }
 913
 914 static bool allocator_thread_running(struct bch_dev *ca)
 915 {
 916         unsigned state = ca->mi.state == BCH_MEMBER_STATE_rw &&
 917                 test_bit(BCH_FS_ALLOCATOR_RUNNING, &ca->fs->flags)
 918                 ? ALLOCATOR_running
 919                 : ALLOCATOR_stopped;
 920         alloc_thread_set_state(ca, state);
 921         return state == ALLOCATOR_running;
 922 }
 923
 924 static int buckets_available(struct bch_dev *ca, unsigned long gc_count)
 925 {
 926         s64 available = dev_buckets_reclaimable(ca) -
 927                 (gc_count == ca->fs->gc_count ? ca->inc_gen_really_needs_gc : 0);
 928         bool ret = available > 0;
 929
 930         alloc_thread_set_state(ca, ret
 931                                ? ALLOCATOR_running
 932                                : ALLOCATOR_blocked);
 933         return ret;
 934 }
 935
 936 /**
 937  * bch_allocator_thread - move buckets from free_inc to reserves
 938  *
 939  * The free_inc FIFO is populated by find_reclaimable_buckets(), and
 940  * the reserves are depleted by bucket allocation. When we run out
 941  * of free_inc, try to invalidate some buckets and write out
 942  * prios and gens.
 943  */
 944 static int bch2_allocator_thread(void *arg)
 945 {
 946         struct bch_dev *ca = arg;
 947         struct bch_fs *c = ca->fs;
 948         unsigned long gc_count = c->gc_count;
 949         size_t nr;
 950         int ret;
 951
 952         set_freezable();
 953
 954         while (1) {
 955                 ret = kthread_wait_freezable(allocator_thread_running(ca));
 956                 if (ret)
 957                         goto stop;
 958
 959                 while (!ca->alloc_heap.used) {
 960                         cond_resched();
 961
 962                         ret = kthread_wait_freezable(buckets_available(ca, gc_count));
 963                         if (ret)
 964                                 goto stop;
 965
 966                         gc_count = c->gc_count;
 967                         nr = find_reclaimable_buckets(c, ca);
 968
 969                         trace_alloc_scan(ca, nr, ca->inc_gen_needs_gc,
 970                                          ca->inc_gen_really_needs_gc);
 971
 972                         if ((ca->inc_gen_needs_gc >= ALLOC_SCAN_BATCH(ca) ||
 973                              ca->inc_gen_really_needs_gc) &&
 974                             c->gc_thread) {
 975                                 atomic_inc(&c->kick_gc);
 976                                 wake_up_process(c->gc_thread);
 977                         }
 978                 }
 979
 980                 ret = bch2_invalidate_buckets(c, ca);
 981                 if (ret)
 982                         goto stop;
 983
 984                 while (!fifo_empty(&ca->free_inc)) {
 985                         u64 b = fifo_peek(&ca->free_inc);
 986
 987                         discard_one_bucket(c, ca, b);
 988
 989                         ret = kthread_wait_freezable(push_invalidated_bucket(c, ca, b));
 990                         if (ret)
 991                                 goto stop;
 992                 }
 993         }
 994 stop:
 995         alloc_thread_set_state(ca, ALLOCATOR_stopped);
 996         return 0;
 997 }
 998
 999 /* Startup/shutdown (ro/rw): */
1000
1001 void bch2_recalc_capacity(struct bch_fs *c)
1002 {
1003         struct bch_dev *ca;
1004         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
1005         unsigned bucket_size_max = 0;
1006         unsigned long ra_pages = 0;
1007         unsigned i, j;
1008
1009         lockdep_assert_held(&c->state_lock);
1010
1011         for_each_online_member(ca, c, i) {
1012                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_bdi;
1013
1014                 ra_pages += bdi->ra_pages;
1015         }
1016
1017         bch2_set_ra_pages(c, ra_pages);
1018
1019         for_each_rw_member(ca, c, i) {
1020                 u64 dev_reserve = 0;
1021
1022                 /*
1023                  * We need to reserve buckets (from the number
1024                  * of currently available buckets) against
1025                  * foreground writes so that mainly copygc can
1026                  * make forward progress.
1027                  *
1028                  * We need enough to refill the various reserves
1029                  * from scratch - copygc will use its entire
1030                  * reserve all at once, then run against when
1031                  * its reserve is refilled (from the formerly
1032                  * available buckets).
1033                  *
1034                  * This reserve is just used when considering if
1035                  * allocations for foreground writes must wait -
1036                  * not -ENOSPC calculations.
1037                  */
1038                 for (j = 0; j < RESERVE_NONE; j++)
1039                         dev_reserve += ca->free[j].size;
1040
1041                 dev_reserve += 1;       /* btree write point */
1042                 dev_reserve += 1;       /* copygc write point */
1043                 dev_reserve += 1;       /* rebalance write point */
1044
1045                 dev_reserve *= ca->mi.bucket_size;
1046
1047                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
1048                                              ca->mi.first_bucket);
1049
1050                 reserved_sectors += dev_reserve * 2;
1051
1052                 bucket_size_max = max_t(unsigned, bucket_size_max,
1053                                         ca->mi.bucket_size);
1054         }
1055
1056         gc_reserve = c->opts.gc_reserve_bytes
1057                 ? c->opts.gc_reserve_bytes >> 9
1058                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
1059
1060         reserved_sectors = max(gc_reserve, reserved_sectors);
1061
1062         reserved_sectors = min(reserved_sectors, capacity);
1063
1064         c->capacity = capacity - reserved_sectors;
1065
1066         c->bucket_size_max = bucket_size_max;
1067
1068         /* Wake up case someone was waiting for buckets */
1069         closure_wake_up(&c->freelist_wait);
1070 }
1071
1072 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
1073 {
1074         struct open_bucket *ob;
1075         bool ret = false;
1076
1077         for (ob = c->open_buckets;
1078              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1079              ob++) {
1080                 spin_lock(&ob->lock);
1081                 if (ob->valid && !ob->on_partial_list &&
1082                     ob->ptr.dev == ca->dev_idx)
1083                         ret = true;
1084                 spin_unlock(&ob->lock);
1085         }
1086
1087         return ret;
1088 }
1089
1090 /* device goes ro: */
1091 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
1092 {
1093         unsigned i;
1094
1095         BUG_ON(ca->alloc_thread);
1096
1097         /* First, remove device from allocation groups: */
1098
1099         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1100                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
1101
1102         /*
1103          * Capacity is calculated based off of devices in allocation groups:
1104          */
1105         bch2_recalc_capacity(c);
1106
1107         /* Next, close write points that point to this device... */
1108         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1109                 bch2_writepoint_stop(c, ca, &c->write_points[i]);
1110
1111         bch2_writepoint_stop(c, ca, &c->copygc_write_point);
1112         bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
1113         bch2_writepoint_stop(c, ca, &c->btree_write_point);
1114
1115         mutex_lock(&c->btree_reserve_cache_lock);
1116         while (c->btree_reserve_cache_nr) {
1117                 struct btree_alloc *a =
1118                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1119
1120                 bch2_open_buckets_put(c, &a->ob);
1121         }
1122         mutex_unlock(&c->btree_reserve_cache_lock);
1123
1124         while (1) {
1125                 struct open_bucket *ob;
1126
1127                 spin_lock(&c->freelist_lock);
1128                 if (!ca->open_buckets_partial_nr) {
1129                         spin_unlock(&c->freelist_lock);
1130                         break;
1131                 }
1132                 ob = c->open_buckets +
1133                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
1134                 ob->on_partial_list = false;
1135                 spin_unlock(&c->freelist_lock);
1136
1137                 bch2_open_bucket_put(c, ob);
1138         }
1139
1140         bch2_ec_stop_dev(c, ca);
1141
1142         /*
1143          * Wake up threads that were blocked on allocation, so they can notice
1144          * the device can no longer be removed and the capacity has changed:
1145          */
1146         closure_wake_up(&c->freelist_wait);
1147
1148         /*
1149          * journal_res_get() can block waiting for free space in the journal -
1150          * it needs to notice there may not be devices to allocate from anymore:
1151          */
1152         wake_up(&c->journal.wait);
1153
1154         /* Now wait for any in flight writes: */
1155
1156         closure_wait_event(&c->open_buckets_wait,
1157                            !bch2_dev_has_open_write_point(c, ca));
1158 }
1159
1160 /* device goes rw: */
1161 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
1162 {
1163         unsigned i;
1164
1165         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1166                 if (ca->mi.data_allowed & (1 << i))
1167                         set_bit(ca->dev_idx, c->rw_devs[i].d);
1168 }
1169
1170 void bch2_dev_allocator_quiesce(struct bch_fs *c, struct bch_dev *ca)
1171 {
1172         if (ca->alloc_thread)
1173                 closure_wait_event(&c->freelist_wait,
1174                                    ca->allocator_state != ALLOCATOR_running);
1175 }
1176
1177 /* stop allocator thread: */
1178 void bch2_dev_allocator_stop(struct bch_dev *ca)
1179 {
1180         struct task_struct *p;
1181
1182         p = rcu_dereference_protected(ca->alloc_thread, 1);
1183         ca->alloc_thread = NULL;
1184
1185         /*
1186          * We need an rcu barrier between setting ca->alloc_thread = NULL and
1187          * the thread shutting down to avoid bch2_wake_allocator() racing:
1188          *
1189          * XXX: it would be better to have the rcu barrier be asynchronous
1190          * instead of blocking us here
1191          */
1192         synchronize_rcu();
1193
1194         if (p) {
1195                 kthread_stop(p);
1196                 put_task_struct(p);
1197         }
1198 }
1199
1200 /* start allocator thread: */
1201 int bch2_dev_allocator_start(struct bch_dev *ca)
1202 {
1203         struct task_struct *p;
1204
1205         /*
1206          * allocator thread already started?
1207          */
1208         if (ca->alloc_thread)
1209                 return 0;
1210
1211         p = kthread_create(bch2_allocator_thread, ca,
1212                            "bch-alloc/%s", ca->name);
1213         if (IS_ERR(p)) {
1214                 bch_err(ca->fs, "error creating allocator thread: %li",
1215                         PTR_ERR(p));
1216                 return PTR_ERR(p);
1217         }
1218
1219         get_task_struct(p);
1220         rcu_assign_pointer(ca->alloc_thread, p);
1221         wake_up_process(p);
1222         return 0;
1223 }
1224
1225 void bch2_fs_allocator_background_init(struct bch_fs *c)
1226 {
1227         spin_lock_init(&c->freelist_lock);
1228 }