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 struct bkey_alloc_buf {
  42         struct bkey_i   k;
  43         struct bch_alloc_v3 v;
  44
  45 #define x(_name,  _bits)                + _bits / 8
  46         u8              _pad[0 + BCH_ALLOC_FIELDS_V2()];
  47 #undef  x
  48 } __attribute__((packed, aligned(8)));
  49
  50 /* Persistent alloc info: */
  51
  52 static inline u64 alloc_field_v1_get(const struct bch_alloc *a,
  53                                      const void **p, unsigned field)
  54 {
  55         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  56         u64 v;
  57
  58         if (!(a->fields & (1 << field)))
  59                 return 0;
  60
  61         switch (bytes) {
  62         case 1:
  63                 v = *((const u8 *) *p);
  64                 break;
  65         case 2:
  66                 v = le16_to_cpup(*p);
  67                 break;
  68         case 4:
  69                 v = le32_to_cpup(*p);
  70                 break;
  71         case 8:
  72                 v = le64_to_cpup(*p);
  73                 break;
  74         default:
  75                 BUG();
  76         }
  77
  78         *p += bytes;
  79         return v;
  80 }
  81
  82 static inline void alloc_field_v1_put(struct bkey_i_alloc *a, void **p,
  83                                       unsigned field, u64 v)
  84 {
  85         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  86
  87         if (!v)
  88                 return;
  89
  90         a->v.fields |= 1 << field;
  91
  92         switch (bytes) {
  93         case 1:
  94                 *((u8 *) *p) = v;
  95                 break;
  96         case 2:
  97                 *((__le16 *) *p) = cpu_to_le16(v);
  98                 break;
  99         case 4:
 100                 *((__le32 *) *p) = cpu_to_le32(v);
 101                 break;
 102         case 8:
 103                 *((__le64 *) *p) = cpu_to_le64(v);
 104                 break;
 105         default:
 106                 BUG();
 107         }
 108
 109         *p += bytes;
 110 }
 111
 112 static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out,
 113                                  struct bkey_s_c k)
 114 {
 115         const struct bch_alloc *in = bkey_s_c_to_alloc(k).v;
 116         const void *d = in->data;
 117         unsigned idx = 0;
 118
 119         out->gen = in->gen;
 120
 121 #define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++);
 122         BCH_ALLOC_FIELDS_V1()
 123 #undef  x
 124 }
 125
 126 static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out,
 127                                 struct bkey_s_c k)
 128 {
 129         struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k);
 130         const u8 *in = a.v->data;
 131         const u8 *end = bkey_val_end(a);
 132         unsigned fieldnr = 0;
 133         int ret;
 134         u64 v;
 135
 136         out->gen        = a.v->gen;
 137         out->oldest_gen = a.v->oldest_gen;
 138         out->data_type  = a.v->data_type;
 139
 140 #define x(_name, _bits)                                                 \
 141         if (fieldnr < a.v->nr_fields) {                                 \
 142                 ret = bch2_varint_decode_fast(in, end, &v);             \
 143                 if (ret < 0)                                            \
 144                         return ret;                                     \
 145                 in += ret;                                              \
 146         } else {                                                        \
 147                 v = 0;                                                  \
 148         }                                                               \
 149         out->_name = v;                                                 \
 150         if (v != out->_name)                                            \
 151                 return -1;                                              \
 152         fieldnr++;
 153
 154         BCH_ALLOC_FIELDS_V2()
 155 #undef  x
 156         return 0;
 157 }
 158
 159 static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out,
 160                                 struct bkey_s_c k)
 161 {
 162         struct bkey_s_c_alloc_v3 a = bkey_s_c_to_alloc_v3(k);
 163         const u8 *in = a.v->data;
 164         const u8 *end = bkey_val_end(a);
 165         unsigned fieldnr = 0;
 166         int ret;
 167         u64 v;
 168
 169         out->gen        = a.v->gen;
 170         out->oldest_gen = a.v->oldest_gen;
 171         out->data_type  = a.v->data_type;
 172         out->journal_seq = le64_to_cpu(a.v->journal_seq);
 173
 174 #define x(_name, _bits)                                                 \
 175         if (fieldnr < a.v->nr_fields) {                                 \
 176                 ret = bch2_varint_decode_fast(in, end, &v);             \
 177                 if (ret < 0)                                            \
 178                         return ret;                                     \
 179                 in += ret;                                              \
 180         } else {                                                        \
 181                 v = 0;                                                  \
 182         }                                                               \
 183         out->_name = v;                                                 \
 184         if (v != out->_name)                                            \
 185                 return -1;                                              \
 186         fieldnr++;
 187
 188         BCH_ALLOC_FIELDS_V2()
 189 #undef  x
 190         return 0;
 191 }
 192
 193 static void bch2_alloc_pack_v3(struct bkey_alloc_buf *dst,
 194                                const struct bkey_alloc_unpacked src)
 195 {
 196         struct bkey_i_alloc_v3 *a = bkey_alloc_v3_init(&dst->k);
 197         unsigned nr_fields = 0, last_nonzero_fieldnr = 0;
 198         u8 *out = a->v.data;
 199         u8 *end = (void *) &dst[1];
 200         u8 *last_nonzero_field = out;
 201         unsigned bytes;
 202
 203         a->k.p          = POS(src.dev, src.bucket);
 204         a->v.gen        = src.gen;
 205         a->v.oldest_gen = src.oldest_gen;
 206         a->v.data_type  = src.data_type;
 207         a->v.journal_seq = cpu_to_le64(src.journal_seq);
 208
 209 #define x(_name, _bits)                                                 \
 210         nr_fields++;                                                    \
 211                                                                         \
 212         if (src._name) {                                                \
 213                 out += bch2_varint_encode_fast(out, src._name);         \
 214                                                                         \
 215                 last_nonzero_field = out;                               \
 216                 last_nonzero_fieldnr = nr_fields;                       \
 217         } else {                                                        \
 218                 *out++ = 0;                                             \
 219         }
 220
 221         BCH_ALLOC_FIELDS_V2()
 222 #undef  x
 223         BUG_ON(out > end);
 224
 225         out = last_nonzero_field;
 226         a->v.nr_fields = last_nonzero_fieldnr;
 227
 228         bytes = (u8 *) out - (u8 *) &a->v;
 229         set_bkey_val_bytes(&a->k, bytes);
 230         memset_u64s_tail(&a->v, 0, bytes);
 231 }
 232
 233 struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 234 {
 235         struct bkey_alloc_unpacked ret = {
 236                 .dev    = k.k->p.inode,
 237                 .bucket = k.k->p.offset,
 238                 .gen    = 0,
 239         };
 240
 241         switch (k.k->type) {
 242         case KEY_TYPE_alloc:
 243                 bch2_alloc_unpack_v1(&ret, k);
 244                 break;
 245         case KEY_TYPE_alloc_v2:
 246                 bch2_alloc_unpack_v2(&ret, k);
 247                 break;
 248         case KEY_TYPE_alloc_v3:
 249                 bch2_alloc_unpack_v3(&ret, k);
 250                 break;
 251         }
 252
 253         return ret;
 254 }
 255
 256 static void bch2_alloc_pack(struct bch_fs *c,
 257                             struct bkey_alloc_buf *dst,
 258                             const struct bkey_alloc_unpacked src)
 259 {
 260         bch2_alloc_pack_v3(dst, src);
 261 }
 262
 263 int bch2_alloc_write(struct btree_trans *trans, struct btree_iter *iter,
 264                      struct bkey_alloc_unpacked *u, unsigned trigger_flags)
 265 {
 266         struct bkey_alloc_buf *a;
 267
 268         a = bch2_trans_kmalloc(trans, sizeof(struct bkey_alloc_buf));
 269         if (IS_ERR(a))
 270                 return PTR_ERR(a);
 271
 272         bch2_alloc_pack(trans->c, a, *u);
 273         return bch2_trans_update(trans, iter, &a->k, trigger_flags);
 274 }
 275
 276 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
 277 {
 278         unsigned i, bytes = offsetof(struct bch_alloc, data);
 279
 280         for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++)
 281                 if (a->fields & (1 << i))
 282                         bytes += BCH_ALLOC_V1_FIELD_BYTES[i];
 283
 284         return DIV_ROUND_UP(bytes, sizeof(u64));
 285 }
 286
 287 const char *bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k)
 288 {
 289         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 290
 291         if (k.k->p.inode >= c->sb.nr_devices ||
 292             !c->devs[k.k->p.inode])
 293                 return "invalid device";
 294
 295         /* allow for unknown fields */
 296         if (bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v))
 297                 return "incorrect value size";
 298
 299         return NULL;
 300 }
 301
 302 const char *bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k)
 303 {
 304         struct bkey_alloc_unpacked u;
 305
 306         if (k.k->p.inode >= c->sb.nr_devices ||
 307             !c->devs[k.k->p.inode])
 308                 return "invalid device";
 309
 310         if (bch2_alloc_unpack_v2(&u, k))
 311                 return "unpack error";
 312
 313         return NULL;
 314 }
 315
 316 const char *bch2_alloc_v3_invalid(const struct bch_fs *c, struct bkey_s_c k)
 317 {
 318         struct bkey_alloc_unpacked u;
 319
 320         if (k.k->p.inode >= c->sb.nr_devices ||
 321             !c->devs[k.k->p.inode])
 322                 return "invalid device";
 323
 324         if (bch2_alloc_unpack_v3(&u, k))
 325                 return "unpack error";
 326
 327         return NULL;
 328 }
 329
 330 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c,
 331                            struct bkey_s_c k)
 332 {
 333         struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 334
 335         pr_buf(out, "gen %u oldest_gen %u data_type %s journal_seq %llu",
 336                u.gen, u.oldest_gen, bch2_data_types[u.data_type],
 337                u.journal_seq);
 338 #define x(_name, ...)   pr_buf(out, " " #_name " %llu", (u64) u._name);
 339         BCH_ALLOC_FIELDS_V2()
 340 #undef  x
 341 }
 342
 343 static int bch2_alloc_read_fn(struct btree_trans *trans, struct bkey_s_c k)
 344 {
 345         struct bch_fs *c = trans->c;
 346         struct bch_dev *ca;
 347         struct bucket *g;
 348         struct bkey_alloc_unpacked u;
 349
 350         if (!bkey_is_alloc(k.k))
 351                 return 0;
 352
 353         ca = bch_dev_bkey_exists(c, k.k->p.inode);
 354         g = bucket(ca, k.k->p.offset);
 355         u = bch2_alloc_unpack(k);
 356
 357         *bucket_gen(ca, k.k->p.offset) = u.gen;
 358         g->_mark.gen            = u.gen;
 359         g->_mark.data_type      = u.data_type;
 360         g->_mark.dirty_sectors  = u.dirty_sectors;
 361         g->_mark.cached_sectors = u.cached_sectors;
 362         g->_mark.stripe         = u.stripe != 0;
 363         g->stripe               = u.stripe;
 364         g->stripe_redundancy    = u.stripe_redundancy;
 365         g->io_time[READ]        = u.read_time;
 366         g->io_time[WRITE]       = u.write_time;
 367         g->oldest_gen           = u.oldest_gen;
 368         g->gen_valid            = 1;
 369
 370         return 0;
 371 }
 372
 373 int bch2_alloc_read(struct bch_fs *c)
 374 {
 375         struct btree_trans trans;
 376         int ret;
 377
 378         bch2_trans_init(&trans, c, 0, 0);
 379         down_read(&c->gc_lock);
 380         ret = bch2_btree_and_journal_walk(&trans, BTREE_ID_alloc, bch2_alloc_read_fn);
 381         up_read(&c->gc_lock);
 382         bch2_trans_exit(&trans);
 383         if (ret) {
 384                 bch_err(c, "error reading alloc info: %i", ret);
 385                 return ret;
 386         }
 387
 388         return 0;
 389 }
 390
 391 static int bch2_alloc_write_key(struct btree_trans *trans,
 392                                 struct btree_iter *iter,
 393                                 unsigned flags)
 394 {
 395         struct bch_fs *c = trans->c;
 396         struct bkey_s_c k;
 397         struct bkey_alloc_unpacked old_u, new_u;
 398         int ret;
 399 retry:
 400         bch2_trans_begin(trans);
 401
 402         ret = bch2_btree_key_cache_flush(trans,
 403                         BTREE_ID_alloc, iter->pos);
 404         if (ret)
 405                 goto err;
 406
 407         k = bch2_btree_iter_peek_slot(iter);
 408         ret = bkey_err(k);
 409         if (ret)
 410                 goto err;
 411
 412         old_u   = bch2_alloc_unpack(k);
 413         new_u   = alloc_mem_to_key(c, iter);
 414
 415         if (!bkey_alloc_unpacked_cmp(old_u, new_u))
 416                 return 0;
 417
 418         ret   = bch2_alloc_write(trans, iter, &new_u,
 419                                   BTREE_TRIGGER_NORUN) ?:
 420                 bch2_trans_commit(trans, NULL, NULL,
 421                                 BTREE_INSERT_NOFAIL|flags);
 422 err:
 423         if (ret == -EINTR)
 424                 goto retry;
 425         return ret;
 426 }
 427
 428 int bch2_alloc_write_all(struct bch_fs *c, unsigned flags)
 429 {
 430         struct btree_trans trans;
 431         struct btree_iter iter;
 432         struct bch_dev *ca;
 433         unsigned i;
 434         int ret = 0;
 435
 436         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 0);
 437         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc, POS_MIN,
 438                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 439
 440         for_each_member_device(ca, c, i) {
 441                 bch2_btree_iter_set_pos(&iter,
 442                         POS(ca->dev_idx, ca->mi.first_bucket));
 443
 444                 while (iter.pos.offset < ca->mi.nbuckets) {
 445                         ret = bch2_alloc_write_key(&trans, &iter, flags);
 446                         if (ret) {
 447                                 percpu_ref_put(&ca->ref);
 448                                 goto err;
 449                         }
 450                         bch2_btree_iter_advance(&iter);
 451                 }
 452         }
 453 err:
 454         bch2_trans_iter_exit(&trans, &iter);
 455         bch2_trans_exit(&trans);
 456         return ret;
 457 }
 458
 459 /* Bucket IO clocks: */
 460
 461 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
 462                               size_t bucket_nr, int rw)
 463 {
 464         struct bch_fs *c = trans->c;
 465         struct btree_iter iter;
 466         struct bkey_alloc_unpacked u;
 467         u64 *time, now;
 468         int ret = 0;
 469
 470         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS(dev, bucket_nr),
 471                              BTREE_ITER_CACHED|
 472                              BTREE_ITER_CACHED_NOFILL|
 473                              BTREE_ITER_INTENT);
 474         ret = bch2_btree_iter_traverse(&iter);
 475         if (ret)
 476                 goto out;
 477
 478         u = alloc_mem_to_key(c, &iter);
 479
 480         time = rw == READ ? &u.read_time : &u.write_time;
 481         now = atomic64_read(&c->io_clock[rw].now);
 482         if (*time == now)
 483                 goto out;
 484
 485         *time = now;
 486
 487         ret   = bch2_alloc_write(trans, &iter, &u, 0) ?:
 488                 bch2_trans_commit(trans, NULL, NULL, 0);
 489 out:
 490         bch2_trans_iter_exit(trans, &iter);
 491         return ret;
 492 }
 493
 494 /* Background allocator thread: */
 495
 496 /*
 497  * Scans for buckets to be invalidated, invalidates them, rewrites prios/gens
 498  * (marking them as invalidated on disk), then optionally issues discard
 499  * commands to the newly free buckets, then puts them on the various freelists.
 500  */
 501
 502 static bool bch2_can_invalidate_bucket(struct bch_dev *ca, size_t b,
 503                                        struct bucket_mark m)
 504 {
 505         u8 gc_gen;
 506
 507         if (!is_available_bucket(m))
 508                 return false;
 509
 510         if (m.owned_by_allocator)
 511                 return false;
 512
 513         if (ca->buckets_nouse &&
 514             test_bit(b, ca->buckets_nouse))
 515                 return false;
 516
 517         if (ca->new_fs_bucket_idx) {
 518                 /*
 519                  * Device or filesystem is still being initialized, and we
 520                  * haven't fully marked superblocks & journal:
 521                  */
 522                 if (is_superblock_bucket(ca, b))
 523                         return false;
 524
 525                 if (b < ca->new_fs_bucket_idx)
 526                         return false;
 527         }
 528
 529         gc_gen = bucket_gc_gen(bucket(ca, b));
 530
 531         ca->inc_gen_needs_gc            += gc_gen >= BUCKET_GC_GEN_MAX / 2;
 532         ca->inc_gen_really_needs_gc     += gc_gen >= BUCKET_GC_GEN_MAX;
 533
 534         return gc_gen < BUCKET_GC_GEN_MAX;
 535 }
 536
 537 /*
 538  * Determines what order we're going to reuse buckets, smallest bucket_key()
 539  * first.
 540  */
 541
 542 static unsigned bucket_sort_key(struct bucket *g, struct bucket_mark m,
 543                                 u64 now, u64 last_seq_ondisk)
 544 {
 545         unsigned used = bucket_sectors_used(m);
 546
 547         if (used) {
 548                 /*
 549                  * Prefer to keep buckets that have been read more recently, and
 550                  * buckets that have more data in them:
 551                  */
 552                 u64 last_read = max_t(s64, 0, now - g->io_time[READ]);
 553                 u32 last_read_scaled = max_t(u64, U32_MAX, div_u64(last_read, used));
 554
 555                 return -last_read_scaled;
 556         } else {
 557                 /*
 558                  * Prefer to use buckets with smaller gc_gen so that we don't
 559                  * have to walk the btree and recalculate oldest_gen - but shift
 560                  * off the low bits so that buckets will still have equal sort
 561                  * keys when there's only a small difference, so that we can
 562                  * keep sequential buckets together:
 563                  */
 564                 return  (bucket_needs_journal_commit(m, last_seq_ondisk) << 4)|
 565                         (bucket_gc_gen(g) >> 4);
 566         }
 567 }
 568
 569 static inline int bucket_alloc_cmp(alloc_heap *h,
 570                                    struct alloc_heap_entry l,
 571                                    struct alloc_heap_entry r)
 572 {
 573         return  cmp_int(l.key, r.key) ?:
 574                 cmp_int(r.nr, l.nr) ?:
 575                 cmp_int(l.bucket, r.bucket);
 576 }
 577
 578 static inline int bucket_idx_cmp(const void *_l, const void *_r)
 579 {
 580         const struct alloc_heap_entry *l = _l, *r = _r;
 581
 582         return cmp_int(l->bucket, r->bucket);
 583 }
 584
 585 static void find_reclaimable_buckets_lru(struct bch_fs *c, struct bch_dev *ca)
 586 {
 587         struct bucket_array *buckets;
 588         struct alloc_heap_entry e = { 0 };
 589         u64 now, last_seq_ondisk;
 590         size_t b, i, nr = 0;
 591
 592         down_read(&ca->bucket_lock);
 593
 594         buckets = bucket_array(ca);
 595         ca->alloc_heap.used = 0;
 596         now = atomic64_read(&c->io_clock[READ].now);
 597         last_seq_ondisk = c->journal.flushed_seq_ondisk;
 598
 599         /*
 600          * Find buckets with lowest read priority, by building a maxheap sorted
 601          * by read priority and repeatedly replacing the maximum element until
 602          * all buckets have been visited.
 603          */
 604         for (b = ca->mi.first_bucket; b < ca->mi.nbuckets; b++) {
 605                 struct bucket *g = &buckets->b[b];
 606                 struct bucket_mark m = READ_ONCE(g->mark);
 607                 unsigned key = bucket_sort_key(g, m, now, last_seq_ondisk);
 608
 609                 cond_resched();
 610
 611                 if (!bch2_can_invalidate_bucket(ca, b, m))
 612                         continue;
 613
 614                 if (e.nr && e.bucket + e.nr == b && e.key == key) {
 615                         e.nr++;
 616                 } else {
 617                         if (e.nr)
 618                                 heap_add_or_replace(&ca->alloc_heap, e,
 619                                         -bucket_alloc_cmp, NULL);
 620
 621                         e = (struct alloc_heap_entry) {
 622                                 .bucket = b,
 623                                 .nr     = 1,
 624                                 .key    = key,
 625                         };
 626                 }
 627         }
 628
 629         if (e.nr)
 630                 heap_add_or_replace(&ca->alloc_heap, e,
 631                                 -bucket_alloc_cmp, NULL);
 632
 633         for (i = 0; i < ca->alloc_heap.used; i++)
 634                 nr += ca->alloc_heap.data[i].nr;
 635
 636         while (nr - ca->alloc_heap.data[0].nr >= ALLOC_SCAN_BATCH(ca)) {
 637                 nr -= ca->alloc_heap.data[0].nr;
 638                 heap_pop(&ca->alloc_heap, e, -bucket_alloc_cmp, NULL);
 639         }
 640
 641         up_read(&ca->bucket_lock);
 642 }
 643
 644 static size_t find_reclaimable_buckets(struct bch_fs *c, struct bch_dev *ca)
 645 {
 646         size_t i, nr = 0;
 647
 648         ca->inc_gen_needs_gc                    = 0;
 649         ca->inc_gen_really_needs_gc             = 0;
 650
 651         find_reclaimable_buckets_lru(c, ca);
 652
 653         heap_resort(&ca->alloc_heap, bucket_alloc_cmp, NULL);
 654
 655         for (i = 0; i < ca->alloc_heap.used; i++)
 656                 nr += ca->alloc_heap.data[i].nr;
 657
 658         return nr;
 659 }
 660
 661 static int bucket_invalidate_btree(struct btree_trans *trans,
 662                                    struct bch_dev *ca, u64 b,
 663                                    struct bkey_alloc_unpacked *u)
 664 {
 665         struct bch_fs *c = trans->c;
 666         struct btree_iter iter;
 667         int ret;
 668
 669         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
 670                              POS(ca->dev_idx, b),
 671                              BTREE_ITER_CACHED|
 672                              BTREE_ITER_CACHED_NOFILL|
 673                              BTREE_ITER_INTENT);
 674
 675         ret = bch2_btree_iter_traverse(&iter);
 676         if (ret)
 677                 goto err;
 678
 679         *u = alloc_mem_to_key(c, &iter);
 680
 681         u->gen++;
 682         u->data_type            = 0;
 683         u->dirty_sectors        = 0;
 684         u->cached_sectors       = 0;
 685         u->read_time            = atomic64_read(&c->io_clock[READ].now);
 686         u->write_time           = atomic64_read(&c->io_clock[WRITE].now);
 687
 688         ret = bch2_alloc_write(trans, &iter, u,
 689                                BTREE_TRIGGER_BUCKET_INVALIDATE);
 690 err:
 691         bch2_trans_iter_exit(trans, &iter);
 692         return ret;
 693 }
 694
 695 static int bch2_invalidate_one_bucket(struct bch_fs *c, struct bch_dev *ca,
 696                                       u64 *journal_seq, unsigned flags)
 697 {
 698         struct bkey_alloc_unpacked u;
 699         size_t b;
 700         int ret = 0;
 701
 702         /*
 703          * If the read-only path is trying to shut down, we can't be generating
 704          * new btree updates:
 705          */
 706         if (test_bit(BCH_FS_ALLOCATOR_STOPPING, &c->flags))
 707                 return 1;
 708
 709         BUG_ON(!ca->alloc_heap.used ||
 710                !ca->alloc_heap.data[0].nr);
 711         b = ca->alloc_heap.data[0].bucket;
 712
 713         /* first, put on free_inc and mark as owned by allocator: */
 714         percpu_down_read(&c->mark_lock);
 715
 716         bch2_mark_alloc_bucket(c, ca, b, true);
 717
 718         spin_lock(&c->freelist_lock);
 719         verify_not_on_freelist(c, ca, b);
 720         BUG_ON(!fifo_push(&ca->free_inc, b));
 721         spin_unlock(&c->freelist_lock);
 722
 723         percpu_up_read(&c->mark_lock);
 724
 725         ret = bch2_trans_do(c, NULL, journal_seq,
 726                             BTREE_INSERT_NOCHECK_RW|
 727                             BTREE_INSERT_NOFAIL|
 728                             BTREE_INSERT_JOURNAL_RESERVED|
 729                             flags,
 730                             bucket_invalidate_btree(&trans, ca, b, &u));
 731
 732         if (!ret) {
 733                 /* remove from alloc_heap: */
 734                 struct alloc_heap_entry e, *top = ca->alloc_heap.data;
 735
 736                 top->bucket++;
 737                 top->nr--;
 738
 739                 if (!top->nr)
 740                         heap_pop(&ca->alloc_heap, e, bucket_alloc_cmp, NULL);
 741
 742                 /*
 743                  * Make sure we flush the last journal entry that updated this
 744                  * bucket (i.e. deleting the last reference) before writing to
 745                  * this bucket again:
 746                  */
 747                 *journal_seq = max(*journal_seq, u.journal_seq);
 748         } else {
 749                 size_t b2;
 750
 751                 /* remove from free_inc: */
 752                 percpu_down_read(&c->mark_lock);
 753                 spin_lock(&c->freelist_lock);
 754
 755                 bch2_mark_alloc_bucket(c, ca, b, false);
 756
 757                 BUG_ON(!fifo_pop_back(&ca->free_inc, b2));
 758                 BUG_ON(b != b2);
 759
 760                 spin_unlock(&c->freelist_lock);
 761                 percpu_up_read(&c->mark_lock);
 762         }
 763
 764         return ret < 0 ? ret : 0;
 765 }
 766
 767 /*
 768  * Pull buckets off ca->alloc_heap, invalidate them, move them to ca->free_inc:
 769  */
 770 static int bch2_invalidate_buckets(struct bch_fs *c, struct bch_dev *ca)
 771 {
 772         u64 journal_seq = 0;
 773         int ret = 0;
 774
 775         /* Only use nowait if we've already invalidated at least one bucket: */
 776         while (!ret &&
 777                !fifo_full(&ca->free_inc) &&
 778                ca->alloc_heap.used) {
 779                 if (kthread_should_stop()) {
 780                         ret = 1;
 781                         break;
 782                 }
 783
 784                 ret = bch2_invalidate_one_bucket(c, ca, &journal_seq,
 785                                 (!fifo_empty(&ca->free_inc)
 786                                  ? BTREE_INSERT_NOWAIT : 0));
 787                 /*
 788                  * We only want to batch up invalidates when they're going to
 789                  * require flushing the journal:
 790                  */
 791                 if (!journal_seq)
 792                         break;
 793         }
 794
 795         /* If we used NOWAIT, don't return the error: */
 796         if (!fifo_empty(&ca->free_inc))
 797                 ret = 0;
 798         if (ret < 0)
 799                 bch_err(ca, "error invalidating buckets: %i", ret);
 800         if (ret)
 801                 return ret;
 802
 803         if (journal_seq)
 804                 ret = bch2_journal_flush_seq(&c->journal, journal_seq);
 805         if (ret) {
 806                 bch_err(ca, "journal error: %i", ret);
 807                 return ret;
 808         }
 809
 810         return 0;
 811 }
 812
 813 static void alloc_thread_set_state(struct bch_dev *ca, unsigned new_state)
 814 {
 815         if (ca->allocator_state != new_state) {
 816                 ca->allocator_state = new_state;
 817                 closure_wake_up(&ca->fs->freelist_wait);
 818         }
 819 }
 820
 821 static int push_invalidated_bucket(struct bch_fs *c, struct bch_dev *ca, u64 b)
 822 {
 823         unsigned i;
 824         int ret = 0;
 825
 826         spin_lock(&c->freelist_lock);
 827         for (i = 0; i < RESERVE_NR; i++) {
 828                 /*
 829                  * Don't strand buckets on the copygc freelist until
 830                  * after recovery is finished:
 831                  */
 832                 if (i == RESERVE_MOVINGGC &&
 833                     !test_bit(BCH_FS_STARTED, &c->flags))
 834                         continue;
 835
 836                 if (fifo_push(&ca->free[i], b)) {
 837                         fifo_pop(&ca->free_inc, b);
 838                         ret = 1;
 839                         break;
 840                 }
 841         }
 842         spin_unlock(&c->freelist_lock);
 843
 844         ca->allocator_state = ret
 845                 ? ALLOCATOR_running
 846                 : ALLOCATOR_blocked_full;
 847         closure_wake_up(&c->freelist_wait);
 848         return ret;
 849 }
 850
 851 static void discard_one_bucket(struct bch_fs *c, struct bch_dev *ca, u64 b)
 852 {
 853         if (ca->mi.discard &&
 854             blk_queue_discard(bdev_get_queue(ca->disk_sb.bdev)))
 855                 blkdev_issue_discard(ca->disk_sb.bdev, bucket_to_sector(ca, b),
 856                                      ca->mi.bucket_size, GFP_NOFS, 0);
 857 }
 858
 859 static bool allocator_thread_running(struct bch_dev *ca)
 860 {
 861         unsigned state = ca->mi.state == BCH_MEMBER_STATE_rw &&
 862                 test_bit(BCH_FS_ALLOCATOR_RUNNING, &ca->fs->flags)
 863                 ? ALLOCATOR_running
 864                 : ALLOCATOR_stopped;
 865         alloc_thread_set_state(ca, state);
 866         return state == ALLOCATOR_running;
 867 }
 868
 869 static int buckets_available(struct bch_dev *ca, unsigned long gc_count)
 870 {
 871         s64 available = dev_buckets_reclaimable(ca) -
 872                 (gc_count == ca->fs->gc_count ? ca->inc_gen_really_needs_gc : 0);
 873         bool ret = available > 0;
 874
 875         alloc_thread_set_state(ca, ret
 876                                ? ALLOCATOR_running
 877                                : ALLOCATOR_blocked);
 878         return ret;
 879 }
 880
 881 /**
 882  * bch_allocator_thread - move buckets from free_inc to reserves
 883  *
 884  * The free_inc FIFO is populated by find_reclaimable_buckets(), and
 885  * the reserves are depleted by bucket allocation. When we run out
 886  * of free_inc, try to invalidate some buckets and write out
 887  * prios and gens.
 888  */
 889 static int bch2_allocator_thread(void *arg)
 890 {
 891         struct bch_dev *ca = arg;
 892         struct bch_fs *c = ca->fs;
 893         unsigned long gc_count = c->gc_count;
 894         size_t nr;
 895         int ret;
 896
 897         set_freezable();
 898
 899         while (1) {
 900                 ret = kthread_wait_freezable(allocator_thread_running(ca));
 901                 if (ret)
 902                         goto stop;
 903
 904                 while (!ca->alloc_heap.used) {
 905                         cond_resched();
 906
 907                         ret = kthread_wait_freezable(buckets_available(ca, gc_count));
 908                         if (ret)
 909                                 goto stop;
 910
 911                         gc_count = c->gc_count;
 912                         nr = find_reclaimable_buckets(c, ca);
 913
 914                         trace_alloc_scan(ca, nr, ca->inc_gen_needs_gc,
 915                                          ca->inc_gen_really_needs_gc);
 916
 917                         if ((ca->inc_gen_needs_gc >= ALLOC_SCAN_BATCH(ca) ||
 918                              ca->inc_gen_really_needs_gc) &&
 919                             c->gc_thread) {
 920                                 atomic_inc(&c->kick_gc);
 921                                 wake_up_process(c->gc_thread);
 922                         }
 923                 }
 924
 925                 ret = bch2_invalidate_buckets(c, ca);
 926                 if (ret)
 927                         goto stop;
 928
 929                 while (!fifo_empty(&ca->free_inc)) {
 930                         u64 b = fifo_peek(&ca->free_inc);
 931
 932                         discard_one_bucket(c, ca, b);
 933
 934                         ret = kthread_wait_freezable(push_invalidated_bucket(c, ca, b));
 935                         if (ret)
 936                                 goto stop;
 937                 }
 938         }
 939 stop:
 940         alloc_thread_set_state(ca, ALLOCATOR_stopped);
 941         return 0;
 942 }
 943
 944 /* Startup/shutdown (ro/rw): */
 945
 946 void bch2_recalc_capacity(struct bch_fs *c)
 947 {
 948         struct bch_dev *ca;
 949         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
 950         unsigned bucket_size_max = 0;
 951         unsigned long ra_pages = 0;
 952         unsigned i, j;
 953
 954         lockdep_assert_held(&c->state_lock);
 955
 956         for_each_online_member(ca, c, i) {
 957                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
 958
 959                 ra_pages += bdi->ra_pages;
 960         }
 961
 962         bch2_set_ra_pages(c, ra_pages);
 963
 964         for_each_rw_member(ca, c, i) {
 965                 u64 dev_reserve = 0;
 966
 967                 /*
 968                  * We need to reserve buckets (from the number
 969                  * of currently available buckets) against
 970                  * foreground writes so that mainly copygc can
 971                  * make forward progress.
 972                  *
 973                  * We need enough to refill the various reserves
 974                  * from scratch - copygc will use its entire
 975                  * reserve all at once, then run against when
 976                  * its reserve is refilled (from the formerly
 977                  * available buckets).
 978                  *
 979                  * This reserve is just used when considering if
 980                  * allocations for foreground writes must wait -
 981                  * not -ENOSPC calculations.
 982                  */
 983                 for (j = 0; j < RESERVE_NONE; j++)
 984                         dev_reserve += ca->free[j].size;
 985
 986                 dev_reserve += 1;       /* btree write point */
 987                 dev_reserve += 1;       /* copygc write point */
 988                 dev_reserve += 1;       /* rebalance write point */
 989
 990                 dev_reserve *= ca->mi.bucket_size;
 991
 992                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
 993                                              ca->mi.first_bucket);
 994
 995                 reserved_sectors += dev_reserve * 2;
 996
 997                 bucket_size_max = max_t(unsigned, bucket_size_max,
 998                                         ca->mi.bucket_size);
 999         }
1000
1001         gc_reserve = c->opts.gc_reserve_bytes
1002                 ? c->opts.gc_reserve_bytes >> 9
1003                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
1004
1005         reserved_sectors = max(gc_reserve, reserved_sectors);
1006
1007         reserved_sectors = min(reserved_sectors, capacity);
1008
1009         c->capacity = capacity - reserved_sectors;
1010
1011         c->bucket_size_max = bucket_size_max;
1012
1013         /* Wake up case someone was waiting for buckets */
1014         closure_wake_up(&c->freelist_wait);
1015 }
1016
1017 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
1018 {
1019         struct open_bucket *ob;
1020         bool ret = false;
1021
1022         for (ob = c->open_buckets;
1023              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1024              ob++) {
1025                 spin_lock(&ob->lock);
1026                 if (ob->valid && !ob->on_partial_list &&
1027                     ob->dev == ca->dev_idx)
1028                         ret = true;
1029                 spin_unlock(&ob->lock);
1030         }
1031
1032         return ret;
1033 }
1034
1035 /* device goes ro: */
1036 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
1037 {
1038         unsigned i;
1039
1040         BUG_ON(ca->alloc_thread);
1041
1042         /* First, remove device from allocation groups: */
1043
1044         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1045                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
1046
1047         /*
1048          * Capacity is calculated based off of devices in allocation groups:
1049          */
1050         bch2_recalc_capacity(c);
1051
1052         /* Next, close write points that point to this device... */
1053         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1054                 bch2_writepoint_stop(c, ca, &c->write_points[i]);
1055
1056         bch2_writepoint_stop(c, ca, &c->copygc_write_point);
1057         bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
1058         bch2_writepoint_stop(c, ca, &c->btree_write_point);
1059
1060         mutex_lock(&c->btree_reserve_cache_lock);
1061         while (c->btree_reserve_cache_nr) {
1062                 struct btree_alloc *a =
1063                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1064
1065                 bch2_open_buckets_put(c, &a->ob);
1066         }
1067         mutex_unlock(&c->btree_reserve_cache_lock);
1068
1069         while (1) {
1070                 struct open_bucket *ob;
1071
1072                 spin_lock(&c->freelist_lock);
1073                 if (!ca->open_buckets_partial_nr) {
1074                         spin_unlock(&c->freelist_lock);
1075                         break;
1076                 }
1077                 ob = c->open_buckets +
1078                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
1079                 ob->on_partial_list = false;
1080                 spin_unlock(&c->freelist_lock);
1081
1082                 bch2_open_bucket_put(c, ob);
1083         }
1084
1085         bch2_ec_stop_dev(c, ca);
1086
1087         /*
1088          * Wake up threads that were blocked on allocation, so they can notice
1089          * the device can no longer be removed and the capacity has changed:
1090          */
1091         closure_wake_up(&c->freelist_wait);
1092
1093         /*
1094          * journal_res_get() can block waiting for free space in the journal -
1095          * it needs to notice there may not be devices to allocate from anymore:
1096          */
1097         wake_up(&c->journal.wait);
1098
1099         /* Now wait for any in flight writes: */
1100
1101         closure_wait_event(&c->open_buckets_wait,
1102                            !bch2_dev_has_open_write_point(c, ca));
1103 }
1104
1105 /* device goes rw: */
1106 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
1107 {
1108         unsigned i;
1109
1110         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1111                 if (ca->mi.data_allowed & (1 << i))
1112                         set_bit(ca->dev_idx, c->rw_devs[i].d);
1113 }
1114
1115 void bch2_dev_allocator_quiesce(struct bch_fs *c, struct bch_dev *ca)
1116 {
1117         if (ca->alloc_thread)
1118                 closure_wait_event(&c->freelist_wait,
1119                                    ca->allocator_state != ALLOCATOR_running);
1120 }
1121
1122 /* stop allocator thread: */
1123 void bch2_dev_allocator_stop(struct bch_dev *ca)
1124 {
1125         struct task_struct *p;
1126
1127         p = rcu_dereference_protected(ca->alloc_thread, 1);
1128         ca->alloc_thread = NULL;
1129
1130         /*
1131          * We need an rcu barrier between setting ca->alloc_thread = NULL and
1132          * the thread shutting down to avoid bch2_wake_allocator() racing:
1133          *
1134          * XXX: it would be better to have the rcu barrier be asynchronous
1135          * instead of blocking us here
1136          */
1137         synchronize_rcu();
1138
1139         if (p) {
1140                 kthread_stop(p);
1141                 put_task_struct(p);
1142         }
1143 }
1144
1145 /* start allocator thread: */
1146 int bch2_dev_allocator_start(struct bch_dev *ca)
1147 {
1148         struct task_struct *p;
1149
1150         /*
1151          * allocator thread already started?
1152          */
1153         if (ca->alloc_thread)
1154                 return 0;
1155
1156         p = kthread_create(bch2_allocator_thread, ca,
1157                            "bch-alloc/%s", ca->name);
1158         if (IS_ERR(p)) {
1159                 bch_err(ca->fs, "error creating allocator thread: %li",
1160                         PTR_ERR(p));
1161                 return PTR_ERR(p);
1162         }
1163
1164         get_task_struct(p);
1165         rcu_assign_pointer(ca->alloc_thread, p);
1166         wake_up_process(p);
1167         return 0;
1168 }
1169
1170 void bch2_fs_allocator_background_init(struct bch_fs *c)
1171 {
1172         spin_lock_init(&c->freelist_lock);
1173 }