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 "buckets_waiting_for_journal.h"
  13 #include "clock.h"
  14 #include "debug.h"
  15 #include "ec.h"
  16 #include "error.h"
  17 #include "lru.h"
  18 #include "recovery.h"
  19 #include "varint.h"
  20
  21 #include <linux/kthread.h>
  22 #include <linux/math64.h>
  23 #include <linux/random.h>
  24 #include <linux/rculist.h>
  25 #include <linux/rcupdate.h>
  26 #include <linux/sched/task.h>
  27 #include <linux/sort.h>
  28 #include <trace/events/bcachefs.h>
  29
  30 /* Persistent alloc info: */
  31
  32 static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
  33 #define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8,
  34         BCH_ALLOC_FIELDS_V1()
  35 #undef x
  36 };
  37
  38 const char * const bch2_bucket_states[] = {
  39         "free",
  40         "need gc gens",
  41         "need discard",
  42         "cached",
  43         "dirty",
  44         NULL
  45 };
  46
  47 struct bkey_alloc_unpacked {
  48         u64             journal_seq;
  49         u64             bucket;
  50         u8              dev;
  51         u8              gen;
  52         u8              oldest_gen;
  53         u8              data_type;
  54         bool            need_discard:1;
  55         bool            need_inc_gen:1;
  56 #define x(_name, _bits) u##_bits _name;
  57         BCH_ALLOC_FIELDS_V2()
  58 #undef  x
  59 };
  60
  61 static inline u64 alloc_field_v1_get(const struct bch_alloc *a,
  62                                      const void **p, unsigned field)
  63 {
  64         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  65         u64 v;
  66
  67         if (!(a->fields & (1 << field)))
  68                 return 0;
  69
  70         switch (bytes) {
  71         case 1:
  72                 v = *((const u8 *) *p);
  73                 break;
  74         case 2:
  75                 v = le16_to_cpup(*p);
  76                 break;
  77         case 4:
  78                 v = le32_to_cpup(*p);
  79                 break;
  80         case 8:
  81                 v = le64_to_cpup(*p);
  82                 break;
  83         default:
  84                 BUG();
  85         }
  86
  87         *p += bytes;
  88         return v;
  89 }
  90
  91 static inline void alloc_field_v1_put(struct bkey_i_alloc *a, void **p,
  92                                       unsigned field, u64 v)
  93 {
  94         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  95
  96         if (!v)
  97                 return;
  98
  99         a->v.fields |= 1 << field;
 100
 101         switch (bytes) {
 102         case 1:
 103                 *((u8 *) *p) = v;
 104                 break;
 105         case 2:
 106                 *((__le16 *) *p) = cpu_to_le16(v);
 107                 break;
 108         case 4:
 109                 *((__le32 *) *p) = cpu_to_le32(v);
 110                 break;
 111         case 8:
 112                 *((__le64 *) *p) = cpu_to_le64(v);
 113                 break;
 114         default:
 115                 BUG();
 116         }
 117
 118         *p += bytes;
 119 }
 120
 121 static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out,
 122                                  struct bkey_s_c k)
 123 {
 124         const struct bch_alloc *in = bkey_s_c_to_alloc(k).v;
 125         const void *d = in->data;
 126         unsigned idx = 0;
 127
 128         out->gen = in->gen;
 129
 130 #define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++);
 131         BCH_ALLOC_FIELDS_V1()
 132 #undef  x
 133 }
 134
 135 static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out,
 136                                 struct bkey_s_c k)
 137 {
 138         struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k);
 139         const u8 *in = a.v->data;
 140         const u8 *end = bkey_val_end(a);
 141         unsigned fieldnr = 0;
 142         int ret;
 143         u64 v;
 144
 145         out->gen        = a.v->gen;
 146         out->oldest_gen = a.v->oldest_gen;
 147         out->data_type  = a.v->data_type;
 148
 149 #define x(_name, _bits)                                                 \
 150         if (fieldnr < a.v->nr_fields) {                                 \
 151                 ret = bch2_varint_decode_fast(in, end, &v);             \
 152                 if (ret < 0)                                            \
 153                         return ret;                                     \
 154                 in += ret;                                              \
 155         } else {                                                        \
 156                 v = 0;                                                  \
 157         }                                                               \
 158         out->_name = v;                                                 \
 159         if (v != out->_name)                                            \
 160                 return -1;                                              \
 161         fieldnr++;
 162
 163         BCH_ALLOC_FIELDS_V2()
 164 #undef  x
 165         return 0;
 166 }
 167
 168 static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out,
 169                                 struct bkey_s_c k)
 170 {
 171         struct bkey_s_c_alloc_v3 a = bkey_s_c_to_alloc_v3(k);
 172         const u8 *in = a.v->data;
 173         const u8 *end = bkey_val_end(a);
 174         unsigned fieldnr = 0;
 175         int ret;
 176         u64 v;
 177
 178         out->gen        = a.v->gen;
 179         out->oldest_gen = a.v->oldest_gen;
 180         out->data_type  = a.v->data_type;
 181         out->need_discard = BCH_ALLOC_V3_NEED_DISCARD(a.v);
 182         out->need_inc_gen = BCH_ALLOC_V3_NEED_INC_GEN(a.v);
 183         out->journal_seq = le64_to_cpu(a.v->journal_seq);
 184
 185 #define x(_name, _bits)                                                 \
 186         if (fieldnr < a.v->nr_fields) {                                 \
 187                 ret = bch2_varint_decode_fast(in, end, &v);             \
 188                 if (ret < 0)                                            \
 189                         return ret;                                     \
 190                 in += ret;                                              \
 191         } else {                                                        \
 192                 v = 0;                                                  \
 193         }                                                               \
 194         out->_name = v;                                                 \
 195         if (v != out->_name)                                            \
 196                 return -1;                                              \
 197         fieldnr++;
 198
 199         BCH_ALLOC_FIELDS_V2()
 200 #undef  x
 201         return 0;
 202 }
 203
 204 static struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 205 {
 206         struct bkey_alloc_unpacked ret = {
 207                 .dev    = k.k->p.inode,
 208                 .bucket = k.k->p.offset,
 209                 .gen    = 0,
 210         };
 211
 212         switch (k.k->type) {
 213         case KEY_TYPE_alloc:
 214                 bch2_alloc_unpack_v1(&ret, k);
 215                 break;
 216         case KEY_TYPE_alloc_v2:
 217                 bch2_alloc_unpack_v2(&ret, k);
 218                 break;
 219         case KEY_TYPE_alloc_v3:
 220                 bch2_alloc_unpack_v3(&ret, k);
 221                 break;
 222         }
 223
 224         return ret;
 225 }
 226
 227 void bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out)
 228 {
 229         if (k.k->type == KEY_TYPE_alloc_v4) {
 230                 *out = *bkey_s_c_to_alloc_v4(k).v;
 231         } else {
 232                 struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 233
 234                 *out = (struct bch_alloc_v4) {
 235                         .journal_seq            = u.journal_seq,
 236                         .flags                  = u.need_discard,
 237                         .gen                    = u.gen,
 238                         .oldest_gen             = u.oldest_gen,
 239                         .data_type              = u.data_type,
 240                         .stripe_redundancy      = u.stripe_redundancy,
 241                         .dirty_sectors          = u.dirty_sectors,
 242                         .cached_sectors         = u.cached_sectors,
 243                         .io_time[READ]          = u.read_time,
 244                         .io_time[WRITE]         = u.write_time,
 245                         .stripe                 = u.stripe,
 246                 };
 247         }
 248 }
 249
 250 struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 251 {
 252         struct bkey_i_alloc_v4 *ret;
 253
 254         if (k.k->type == KEY_TYPE_alloc_v4) {
 255                 ret = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 256                 if (!IS_ERR(ret))
 257                         bkey_reassemble(&ret->k_i, k);
 258         } else {
 259                 ret = bch2_trans_kmalloc(trans, sizeof(*ret));
 260                 if (!IS_ERR(ret)) {
 261                         bkey_alloc_v4_init(&ret->k_i);
 262                         ret->k.p = k.k->p;
 263                         bch2_alloc_to_v4(k, &ret->v);
 264                 }
 265         }
 266         return ret;
 267 }
 268
 269 struct bkey_i_alloc_v4 *
 270 bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
 271                               struct bpos pos)
 272 {
 273         struct bkey_s_c k;
 274         struct bkey_i_alloc_v4 *a;
 275         int ret;
 276
 277         bch2_trans_iter_init(trans, iter, BTREE_ID_alloc, pos,
 278                              BTREE_ITER_WITH_UPDATES|
 279                              BTREE_ITER_CACHED|
 280                              BTREE_ITER_INTENT);
 281         k = bch2_btree_iter_peek_slot(iter);
 282         ret = bkey_err(k);
 283         if (ret) {
 284                 bch2_trans_iter_exit(trans, iter);
 285                 return ERR_PTR(ret);
 286         }
 287
 288         a = bch2_alloc_to_v4_mut(trans, k);
 289         if (IS_ERR(a))
 290                 bch2_trans_iter_exit(trans, iter);
 291         return a;
 292 }
 293
 294 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
 295 {
 296         unsigned i, bytes = offsetof(struct bch_alloc, data);
 297
 298         for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++)
 299                 if (a->fields & (1 << i))
 300                         bytes += BCH_ALLOC_V1_FIELD_BYTES[i];
 301
 302         return DIV_ROUND_UP(bytes, sizeof(u64));
 303 }
 304
 305 int bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k,
 306                           int rw, struct printbuf *err)
 307 {
 308         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 309
 310         /* allow for unknown fields */
 311         if (bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v)) {
 312                 pr_buf(err, "incorrect value size (%zu < %u)",
 313                        bkey_val_u64s(a.k), bch_alloc_v1_val_u64s(a.v));
 314                 return -EINVAL;
 315         }
 316
 317         return 0;
 318 }
 319
 320 int bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k,
 321                           int rw, struct printbuf *err)
 322 {
 323         struct bkey_alloc_unpacked u;
 324
 325         if (bch2_alloc_unpack_v2(&u, k)) {
 326                 pr_buf(err, "unpack error");
 327                 return -EINVAL;
 328         }
 329
 330         return 0;
 331 }
 332
 333 int bch2_alloc_v3_invalid(const struct bch_fs *c, struct bkey_s_c k,
 334                           int rw, struct printbuf *err)
 335 {
 336         struct bkey_alloc_unpacked u;
 337
 338         if (bch2_alloc_unpack_v3(&u, k)) {
 339                 pr_buf(err, "unpack error");
 340                 return -EINVAL;
 341         }
 342
 343         return 0;
 344 }
 345
 346 int bch2_alloc_v4_invalid(const struct bch_fs *c, struct bkey_s_c k,
 347                           int rw, struct printbuf *err)
 348 {
 349         if (bkey_val_bytes(k.k) != sizeof(struct bch_alloc_v4)) {
 350                 pr_buf(err, "bad val size (%zu != %zu)",
 351                        bkey_val_bytes(k.k), sizeof(struct bch_alloc_v4));
 352                 return -EINVAL;
 353         }
 354
 355         return 0;
 356 }
 357
 358 void bch2_alloc_v4_swab(struct bkey_s k)
 359 {
 360         struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v;
 361
 362         a->journal_seq          = swab64(a->journal_seq);
 363         a->flags                = swab32(a->flags);
 364         a->dirty_sectors        = swab32(a->dirty_sectors);
 365         a->cached_sectors       = swab32(a->cached_sectors);
 366         a->io_time[0]           = swab64(a->io_time[0]);
 367         a->io_time[1]           = swab64(a->io_time[1]);
 368         a->stripe               = swab32(a->stripe);
 369         a->nr_external_backpointers = swab32(a->nr_external_backpointers);
 370 }
 371
 372 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 373 {
 374         struct bch_alloc_v4 a;
 375
 376         bch2_alloc_to_v4(k, &a);
 377
 378         pr_buf(out, "gen %u oldest_gen %u data_type %s journal_seq %llu need_discard %llu",
 379                a.gen, a.oldest_gen, bch2_data_types[a.data_type],
 380                a.journal_seq, BCH_ALLOC_V4_NEED_DISCARD(&a));
 381         pr_buf(out, " dirty_sectors %u",        a.dirty_sectors);
 382         pr_buf(out, " cached_sectors %u",       a.cached_sectors);
 383         pr_buf(out, " stripe %u",               a.stripe);
 384         pr_buf(out, " stripe_redundancy %u",    a.stripe_redundancy);
 385         pr_buf(out, " read_time %llu",          a.io_time[READ]);
 386         pr_buf(out, " write_time %llu",         a.io_time[WRITE]);
 387 }
 388
 389 int bch2_alloc_read(struct bch_fs *c)
 390 {
 391         struct btree_trans trans;
 392         struct btree_iter iter;
 393         struct bkey_s_c k;
 394         struct bch_alloc_v4 a;
 395         struct bch_dev *ca;
 396         int ret;
 397
 398         bch2_trans_init(&trans, c, 0, 0);
 399
 400         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 401                            BTREE_ITER_PREFETCH, k, ret) {
 402                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 403                 bch2_alloc_to_v4(k, &a);
 404
 405                 *bucket_gen(ca, k.k->p.offset) = a.gen;
 406         }
 407         bch2_trans_iter_exit(&trans, &iter);
 408
 409         bch2_trans_exit(&trans);
 410
 411         if (ret)
 412                 bch_err(c, "error reading alloc info: %i", ret);
 413
 414         return ret;
 415 }
 416
 417 /* Free space/discard btree: */
 418
 419 static int bch2_bucket_do_index(struct btree_trans *trans,
 420                                 struct bkey_s_c alloc_k,
 421                                 struct bch_alloc_v4 a,
 422                                 bool set)
 423 {
 424         struct bch_fs *c = trans->c;
 425         struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 426         struct btree_iter iter;
 427         struct bkey_s_c old;
 428         struct bkey_i *k;
 429         enum bucket_state state = bucket_state(a);
 430         enum btree_id btree;
 431         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 432         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 433         struct printbuf buf = PRINTBUF;
 434         int ret;
 435
 436         if (state != BUCKET_free &&
 437             state != BUCKET_need_discard)
 438                 return 0;
 439
 440         k = bch2_trans_kmalloc(trans, sizeof(*k));
 441         if (IS_ERR(k))
 442                 return PTR_ERR(k);
 443
 444         bkey_init(&k->k);
 445         k->k.type = new_type;
 446
 447         switch (state) {
 448         case BUCKET_free:
 449                 btree = BTREE_ID_freespace;
 450                 k->k.p = alloc_freespace_pos(alloc_k.k->p, a);
 451                 bch2_key_resize(&k->k, 1);
 452                 break;
 453         case BUCKET_need_discard:
 454                 btree = BTREE_ID_need_discard;
 455                 k->k.p = alloc_k.k->p;
 456                 break;
 457         default:
 458                 return 0;
 459         }
 460
 461         bch2_trans_iter_init(trans, &iter, btree,
 462                              bkey_start_pos(&k->k),
 463                              BTREE_ITER_INTENT);
 464         old = bch2_btree_iter_peek_slot(&iter);
 465         ret = bkey_err(old);
 466         if (ret)
 467                 goto err;
 468
 469         if (ca->mi.freespace_initialized &&
 470             bch2_fs_inconsistent_on(old.k->type != old_type, c,
 471                         "incorrect key when %s %s btree (got %s should be %s)\n"
 472                         "  for %s",
 473                         set ? "setting" : "clearing",
 474                         bch2_btree_ids[btree],
 475                         bch2_bkey_types[old.k->type],
 476                         bch2_bkey_types[old_type],
 477                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 478                 ret = -EIO;
 479                 goto err;
 480         }
 481
 482         ret = bch2_trans_update(trans, &iter, k, 0);
 483 err:
 484         bch2_trans_iter_exit(trans, &iter);
 485         printbuf_exit(&buf);
 486         return ret;
 487 }
 488
 489 int bch2_trans_mark_alloc(struct btree_trans *trans,
 490                           struct bkey_s_c old, struct bkey_i *new,
 491                           unsigned flags)
 492 {
 493         struct bch_fs *c = trans->c;
 494         struct bch_alloc_v4 old_a, *new_a;
 495         u64 old_lru, new_lru;
 496         int ret = 0;
 497
 498         /*
 499          * Deletion only happens in the device removal path, with
 500          * BTREE_TRIGGER_NORUN:
 501          */
 502         BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
 503
 504         bch2_alloc_to_v4(old, &old_a);
 505         new_a = &bkey_i_to_alloc_v4(new)->v;
 506
 507         if (new_a->dirty_sectors > old_a.dirty_sectors ||
 508             new_a->cached_sectors > old_a.cached_sectors) {
 509                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 510                 new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 511                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
 512                 SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
 513         }
 514
 515         if (old_a.data_type && !new_a->data_type &&
 516             old_a.gen == new_a->gen &&
 517             !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
 518                 new_a->gen++;
 519                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
 520         }
 521
 522         if (bucket_state(old_a) != bucket_state(*new_a) ||
 523             (bucket_state(*new_a) == BUCKET_free &&
 524              alloc_freespace_genbits(old_a) != alloc_freespace_genbits(*new_a))) {
 525                 ret =   bch2_bucket_do_index(trans, old, old_a, false) ?:
 526                         bch2_bucket_do_index(trans, bkey_i_to_s_c(new), *new_a, true);
 527                 if (ret)
 528                         return ret;
 529         }
 530
 531         old_lru = alloc_lru_idx(old_a);
 532         new_lru = alloc_lru_idx(*new_a);
 533
 534         if (old_lru != new_lru) {
 535                 ret = bch2_lru_change(trans, new->k.p.inode, new->k.p.offset,
 536                                       old_lru, &new_lru);
 537                 if (ret)
 538                         return ret;
 539
 540                 if (new_lru && new_a->io_time[READ] != new_lru)
 541                         new_a->io_time[READ] = new_lru;
 542         }
 543
 544         return 0;
 545 }
 546
 547 static int bch2_check_alloc_key(struct btree_trans *trans,
 548                                 struct btree_iter *alloc_iter)
 549 {
 550         struct bch_fs *c = trans->c;
 551         struct bch_dev *ca;
 552         struct btree_iter discard_iter, freespace_iter;
 553         struct bch_alloc_v4 a;
 554         unsigned discard_key_type, freespace_key_type;
 555         struct bkey_s_c alloc_k, k;
 556         struct printbuf buf = PRINTBUF;
 557         struct printbuf buf2 = PRINTBUF;
 558         int ret;
 559
 560         alloc_k = bch2_btree_iter_peek(alloc_iter);
 561         if (!alloc_k.k)
 562                 return 0;
 563
 564         ret = bkey_err(alloc_k);
 565         if (ret)
 566                 return ret;
 567
 568         if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
 569                         "alloc key for invalid device or bucket"))
 570                 return bch2_btree_delete_at(trans, alloc_iter, 0);
 571
 572         ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 573         if (!ca->mi.freespace_initialized)
 574                 return 0;
 575
 576         bch2_alloc_to_v4(alloc_k, &a);
 577
 578         discard_key_type = bucket_state(a) == BUCKET_need_discard
 579                 ? KEY_TYPE_set : 0;
 580         freespace_key_type = bucket_state(a) == BUCKET_free
 581                 ? KEY_TYPE_set : 0;
 582
 583         bch2_trans_iter_init(trans, &discard_iter, BTREE_ID_need_discard,
 584                              alloc_k.k->p, 0);
 585         bch2_trans_iter_init(trans, &freespace_iter, BTREE_ID_freespace,
 586                              alloc_freespace_pos(alloc_k.k->p, a), 0);
 587
 588         k = bch2_btree_iter_peek_slot(&discard_iter);
 589         ret = bkey_err(k);
 590         if (ret)
 591                 goto err;
 592
 593         if (fsck_err_on(k.k->type != discard_key_type, c,
 594                         "incorrect key in need_discard btree (got %s should be %s)\n"
 595                         "  %s",
 596                         bch2_bkey_types[k.k->type],
 597                         bch2_bkey_types[discard_key_type],
 598                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 599                 struct bkey_i *update =
 600                         bch2_trans_kmalloc(trans, sizeof(*update));
 601
 602                 ret = PTR_ERR_OR_ZERO(update);
 603                 if (ret)
 604                         goto err;
 605
 606                 bkey_init(&update->k);
 607                 update->k.type  = discard_key_type;
 608                 update->k.p     = discard_iter.pos;
 609
 610                 ret = bch2_trans_update(trans, &discard_iter, update, 0);
 611                 if (ret)
 612                         goto err;
 613         }
 614
 615         k = bch2_btree_iter_peek_slot(&freespace_iter);
 616         ret = bkey_err(k);
 617         if (ret)
 618                 goto err;
 619
 620         if (fsck_err_on(k.k->type != freespace_key_type, c,
 621                         "incorrect key in freespace btree (got %s should be %s)\n"
 622                         "  %s",
 623                         bch2_bkey_types[k.k->type],
 624                         bch2_bkey_types[freespace_key_type],
 625                         (printbuf_reset(&buf),
 626                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 627                 struct bkey_i *update =
 628                         bch2_trans_kmalloc(trans, sizeof(*update));
 629
 630                 ret = PTR_ERR_OR_ZERO(update);
 631                 if (ret)
 632                         goto err;
 633
 634                 bkey_init(&update->k);
 635                 update->k.type  = freespace_key_type;
 636                 update->k.p     = freespace_iter.pos;
 637                 bch2_key_resize(&update->k, 1);
 638
 639                 ret = bch2_trans_update(trans, &freespace_iter, update, 0);
 640                 if (ret)
 641                         goto err;
 642         }
 643 err:
 644 fsck_err:
 645         bch2_trans_iter_exit(trans, &freespace_iter);
 646         bch2_trans_iter_exit(trans, &discard_iter);
 647         printbuf_exit(&buf2);
 648         printbuf_exit(&buf);
 649         return ret;
 650 }
 651
 652 static int bch2_check_discard_freespace_key(struct btree_trans *trans,
 653                                             struct btree_iter *iter)
 654 {
 655         struct bch_fs *c = trans->c;
 656         struct btree_iter alloc_iter;
 657         struct bkey_s_c k, freespace_k;
 658         struct bch_alloc_v4 a;
 659         u64 genbits;
 660         struct bpos pos;
 661         struct bkey_i *update;
 662         enum bucket_state state = iter->btree_id == BTREE_ID_need_discard
 663                 ? BUCKET_need_discard
 664                 : BUCKET_free;
 665         struct printbuf buf = PRINTBUF;
 666         int ret;
 667
 668         freespace_k = bch2_btree_iter_peek(iter);
 669         if (!freespace_k.k)
 670                 return 1;
 671
 672         ret = bkey_err(freespace_k);
 673         if (ret)
 674                 return ret;
 675
 676         pos = iter->pos;
 677         pos.offset &= ~(~0ULL << 56);
 678         genbits = iter->pos.offset & (~0ULL << 56);
 679
 680         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
 681
 682         if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
 683                         "%llu:%llu set in %s btree but device or bucket does not exist",
 684                         pos.inode, pos.offset,
 685                         bch2_btree_ids[iter->btree_id]))
 686                 goto delete;
 687
 688         k = bch2_btree_iter_peek_slot(&alloc_iter);
 689         ret = bkey_err(k);
 690         if (ret)
 691                 goto err;
 692
 693         bch2_alloc_to_v4(k, &a);
 694
 695         if (fsck_err_on(bucket_state(a) != state ||
 696                         (state == BUCKET_free &&
 697                          genbits != alloc_freespace_genbits(a)), c,
 698                         "%s\n  incorrectly set in %s index (free %u, genbits %llu should be %llu)",
 699                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
 700                         bch2_btree_ids[iter->btree_id],
 701                         bucket_state(a) == state,
 702                         genbits >> 56, alloc_freespace_genbits(a) >> 56))
 703                 goto delete;
 704 out:
 705 err:
 706 fsck_err:
 707         bch2_trans_iter_exit(trans, &alloc_iter);
 708         printbuf_exit(&buf);
 709         return ret;
 710 delete:
 711         if (iter->btree_id == BTREE_ID_freespace) {
 712                 /* should probably add a helper for deleting extents */
 713                 update = bch2_trans_kmalloc(trans, sizeof(*update));
 714                 ret = PTR_ERR_OR_ZERO(update);
 715                 if (ret)
 716                         goto err;
 717
 718                 bkey_init(&update->k);
 719                 update->k.p = iter->pos;
 720                 bch2_key_resize(&update->k, 1);
 721
 722                 ret = bch2_trans_update(trans, iter, update, 0);
 723         } else {
 724                 ret = bch2_btree_delete_at(trans, iter, 0);
 725         }
 726         goto out;
 727 }
 728
 729 int bch2_check_alloc_info(struct bch_fs *c)
 730 {
 731         struct btree_trans trans;
 732         struct btree_iter iter;
 733         struct bkey_s_c k;
 734         int ret = 0;
 735
 736         bch2_trans_init(&trans, c, 0, 0);
 737
 738         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 739                            BTREE_ITER_PREFETCH, k, ret) {
 740                 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
 741                         bch2_check_alloc_key(&trans, &iter));
 742                 if (ret)
 743                         break;
 744         }
 745         bch2_trans_iter_exit(&trans, &iter);
 746
 747         if (ret)
 748                 goto err;
 749
 750         bch2_trans_iter_init(&trans, &iter, BTREE_ID_need_discard, POS_MIN,
 751                              BTREE_ITER_PREFETCH);
 752         while (1) {
 753                 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
 754                         bch2_check_discard_freespace_key(&trans, &iter));
 755                 if (ret)
 756                         break;
 757
 758                 bch2_btree_iter_set_pos(&iter, bpos_nosnap_successor(iter.pos));
 759         }
 760         bch2_trans_iter_exit(&trans, &iter);
 761
 762         if (ret)
 763                 goto err;
 764
 765         bch2_trans_iter_init(&trans, &iter, BTREE_ID_freespace, POS_MIN,
 766                              BTREE_ITER_PREFETCH);
 767         while (1) {
 768                 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
 769                         bch2_check_discard_freespace_key(&trans, &iter));
 770                 if (ret)
 771                         break;
 772
 773                 bch2_btree_iter_set_pos(&iter, bpos_nosnap_successor(iter.pos));
 774         }
 775         bch2_trans_iter_exit(&trans, &iter);
 776 err:
 777         bch2_trans_exit(&trans);
 778         return ret < 0 ? ret : 0;
 779 }
 780
 781 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
 782                                        struct btree_iter *alloc_iter)
 783 {
 784         struct bch_fs *c = trans->c;
 785         struct btree_iter lru_iter;
 786         struct bch_alloc_v4 a;
 787         struct bkey_s_c alloc_k, k;
 788         struct printbuf buf = PRINTBUF;
 789         struct printbuf buf2 = PRINTBUF;
 790         int ret;
 791
 792         alloc_k = bch2_btree_iter_peek(alloc_iter);
 793         if (!alloc_k.k)
 794                 return 0;
 795
 796         ret = bkey_err(alloc_k);
 797         if (ret)
 798                 return ret;
 799
 800         bch2_alloc_to_v4(alloc_k, &a);
 801
 802         if (bucket_state(a) != BUCKET_cached)
 803                 return 0;
 804
 805         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
 806                              POS(alloc_k.k->p.inode, a.io_time[READ]), 0);
 807
 808         k = bch2_btree_iter_peek_slot(&lru_iter);
 809         ret = bkey_err(k);
 810         if (ret)
 811                 goto err;
 812
 813         if (fsck_err_on(!a.io_time[READ], c,
 814                         "cached bucket with read_time 0\n"
 815                         "  %s",
 816                 (printbuf_reset(&buf),
 817                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) ||
 818             fsck_err_on(k.k->type != KEY_TYPE_lru ||
 819                         le64_to_cpu(bkey_s_c_to_lru(k).v->idx) != alloc_k.k->p.offset, c,
 820                         "incorrect/missing lru entry\n"
 821                         "  %s\n"
 822                         "  %s",
 823                         (printbuf_reset(&buf),
 824                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
 825                         (bch2_bkey_val_to_text(&buf2, c, k), buf2.buf))) {
 826                 u64 read_time = a.io_time[READ];
 827
 828                 if (!a.io_time[READ])
 829                         a.io_time[READ] = atomic64_read(&c->io_clock[READ].now);
 830
 831                 ret = bch2_lru_change(trans,
 832                                       alloc_k.k->p.inode,
 833                                       alloc_k.k->p.offset,
 834                                       0, &a.io_time[READ]);
 835                 if (ret)
 836                         goto err;
 837
 838                 if (a.io_time[READ] != read_time) {
 839                         struct bkey_i_alloc_v4 *a_mut =
 840                                 bch2_alloc_to_v4_mut(trans, alloc_k);
 841                         ret = PTR_ERR_OR_ZERO(a_mut);
 842                         if (ret)
 843                                 goto err;
 844
 845                         a_mut->v.io_time[READ] = a.io_time[READ];
 846                         ret = bch2_trans_update(trans, alloc_iter,
 847                                                 &a_mut->k_i, BTREE_TRIGGER_NORUN);
 848                         if (ret)
 849                                 goto err;
 850                 }
 851         }
 852 err:
 853 fsck_err:
 854         bch2_trans_iter_exit(trans, &lru_iter);
 855         printbuf_exit(&buf2);
 856         printbuf_exit(&buf);
 857         return ret;
 858 }
 859
 860 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
 861 {
 862         struct btree_trans trans;
 863         struct btree_iter iter;
 864         struct bkey_s_c k;
 865         int ret = 0;
 866
 867         bch2_trans_init(&trans, c, 0, 0);
 868
 869         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 870                            BTREE_ITER_PREFETCH, k, ret) {
 871                 ret = __bch2_trans_do(&trans, NULL, NULL,
 872                                       BTREE_INSERT_NOFAIL|
 873                                       BTREE_INSERT_LAZY_RW,
 874                         bch2_check_alloc_to_lru_ref(&trans, &iter));
 875                 if (ret)
 876                         break;
 877         }
 878         bch2_trans_iter_exit(&trans, &iter);
 879
 880         bch2_trans_exit(&trans);
 881         return ret < 0 ? ret : 0;
 882 }
 883
 884 static int bch2_clear_need_discard(struct btree_trans *trans, struct bpos pos,
 885                                    struct bch_dev *ca, bool *discard_done)
 886 {
 887         struct bch_fs *c = trans->c;
 888         struct btree_iter iter;
 889         struct bkey_s_c k;
 890         struct bkey_i_alloc_v4 *a;
 891         struct printbuf buf = PRINTBUF;
 892         int ret;
 893
 894         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, pos,
 895                              BTREE_ITER_CACHED);
 896         k = bch2_btree_iter_peek_slot(&iter);
 897         ret = bkey_err(k);
 898         if (ret)
 899                 goto out;
 900
 901         a = bch2_alloc_to_v4_mut(trans, k);
 902         ret = PTR_ERR_OR_ZERO(a);
 903         if (ret)
 904                 goto out;
 905
 906         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
 907                 a->v.gen++;
 908                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
 909                 goto write;
 910         }
 911
 912         BUG_ON(a->v.journal_seq > c->journal.flushed_seq_ondisk);
 913
 914         if (bch2_fs_inconsistent_on(!BCH_ALLOC_V4_NEED_DISCARD(&a->v), c,
 915                         "%s\n  incorrectly set in need_discard btree",
 916                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 917                 ret = -EIO;
 918                 goto out;
 919         }
 920
 921         if (!*discard_done && ca->mi.discard && !c->opts.nochanges) {
 922                 /*
 923                  * This works without any other locks because this is the only
 924                  * thread that removes items from the need_discard tree
 925                  */
 926                 bch2_trans_unlock(trans);
 927                 blkdev_issue_discard(ca->disk_sb.bdev,
 928                                      k.k->p.offset * ca->mi.bucket_size,
 929                                      ca->mi.bucket_size,
 930                                      GFP_KERNEL, 0);
 931                 *discard_done = true;
 932
 933                 ret = bch2_trans_relock(trans) ? 0 : -EINTR;
 934                 if (ret)
 935                         goto out;
 936         }
 937
 938         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
 939 write:
 940         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
 941 out:
 942         bch2_trans_iter_exit(trans, &iter);
 943         printbuf_exit(&buf);
 944         return ret;
 945 }
 946
 947 static void bch2_do_discards_work(struct work_struct *work)
 948 {
 949         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
 950         struct bch_dev *ca = NULL;
 951         struct btree_trans trans;
 952         struct btree_iter iter;
 953         struct bkey_s_c k;
 954         u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
 955         int ret;
 956
 957         bch2_trans_init(&trans, c, 0, 0);
 958
 959         for_each_btree_key(&trans, iter, BTREE_ID_need_discard,
 960                            POS_MIN, 0, k, ret) {
 961                 bool discard_done = false;
 962
 963                 if (ca && k.k->p.inode != ca->dev_idx) {
 964                         percpu_ref_put(&ca->io_ref);
 965                         ca = NULL;
 966                 }
 967
 968                 if (!ca) {
 969                         ca = bch_dev_bkey_exists(c, k.k->p.inode);
 970                         if (!percpu_ref_tryget(&ca->io_ref)) {
 971                                 ca = NULL;
 972                                 bch2_btree_iter_set_pos(&iter, POS(k.k->p.inode + 1, 0));
 973                                 continue;
 974                         }
 975                 }
 976
 977                 seen++;
 978
 979                 if (bch2_bucket_is_open_safe(c, k.k->p.inode, k.k->p.offset)) {
 980                         open++;
 981                         continue;
 982                 }
 983
 984                 if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 985                                 c->journal.flushed_seq_ondisk,
 986                                 k.k->p.inode, k.k->p.offset)) {
 987                         need_journal_commit++;
 988                         continue;
 989                 }
 990
 991                 ret = __bch2_trans_do(&trans, NULL, NULL,
 992                                       BTREE_INSERT_USE_RESERVE|
 993                                       BTREE_INSERT_NOFAIL,
 994                                 bch2_clear_need_discard(&trans, k.k->p, ca, &discard_done));
 995                 if (ret)
 996                         break;
 997
 998                 discarded++;
 999         }
1000         bch2_trans_iter_exit(&trans, &iter);
1001
1002         if (ca)
1003                 percpu_ref_put(&ca->io_ref);
1004
1005         bch2_trans_exit(&trans);
1006
1007         if (need_journal_commit * 2 > seen)
1008                 bch2_journal_flush_async(&c->journal, NULL);
1009
1010         percpu_ref_put(&c->writes);
1011
1012         trace_do_discards(c, seen, open, need_journal_commit, discarded, ret);
1013 }
1014
1015 void bch2_do_discards(struct bch_fs *c)
1016 {
1017         if (percpu_ref_tryget(&c->writes) &&
1018             !queue_work(system_long_wq, &c->discard_work))
1019                 percpu_ref_put(&c->writes);
1020 }
1021
1022 static int invalidate_one_bucket(struct btree_trans *trans, struct bch_dev *ca)
1023 {
1024         struct bch_fs *c = trans->c;
1025         struct btree_iter lru_iter, alloc_iter = { NULL };
1026         struct bkey_s_c k;
1027         struct bkey_i_alloc_v4 *a;
1028         u64 bucket, idx;
1029         int ret;
1030
1031         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
1032                              POS(ca->dev_idx, 0), 0);
1033         k = bch2_btree_iter_peek(&lru_iter);
1034         ret = bkey_err(k);
1035         if (ret)
1036                 goto out;
1037
1038         if (!k.k || k.k->p.inode != ca->dev_idx)
1039                 goto out;
1040
1041         if (bch2_fs_inconsistent_on(k.k->type != KEY_TYPE_lru, c,
1042                                     "non lru key in lru btree"))
1043                 goto out;
1044
1045         idx     = k.k->p.offset;
1046         bucket  = le64_to_cpu(bkey_s_c_to_lru(k).v->idx);
1047
1048         a = bch2_trans_start_alloc_update(trans, &alloc_iter,
1049                                           POS(ca->dev_idx, bucket));
1050         ret = PTR_ERR_OR_ZERO(a);
1051         if (ret)
1052                 goto out;
1053
1054         if (bch2_fs_inconsistent_on(idx != alloc_lru_idx(a->v), c,
1055                         "invalidating bucket with wrong lru idx (got %llu should be %llu",
1056                         idx, alloc_lru_idx(a->v)))
1057                 goto out;
1058
1059         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1060         a->v.gen++;
1061         a->v.data_type          = 0;
1062         a->v.dirty_sectors      = 0;
1063         a->v.cached_sectors     = 0;
1064         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1065         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1066
1067         ret = bch2_trans_update(trans, &alloc_iter, &a->k_i,
1068                                 BTREE_TRIGGER_BUCKET_INVALIDATE);
1069 out:
1070         bch2_trans_iter_exit(trans, &alloc_iter);
1071         bch2_trans_iter_exit(trans, &lru_iter);
1072         return ret;
1073 }
1074
1075 static void bch2_do_invalidates_work(struct work_struct *work)
1076 {
1077         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1078         struct bch_dev *ca;
1079         struct btree_trans trans;
1080         unsigned i;
1081         int ret = 0;
1082
1083         bch2_trans_init(&trans, c, 0, 0);
1084
1085         for_each_member_device(ca, c, i)
1086                 while (!ret && should_invalidate_buckets(ca))
1087                         ret = __bch2_trans_do(&trans, NULL, NULL,
1088                                               BTREE_INSERT_USE_RESERVE|
1089                                               BTREE_INSERT_NOFAIL,
1090                                         invalidate_one_bucket(&trans, ca));
1091
1092         bch2_trans_exit(&trans);
1093         percpu_ref_put(&c->writes);
1094 }
1095
1096 void bch2_do_invalidates(struct bch_fs *c)
1097 {
1098         if (percpu_ref_tryget(&c->writes))
1099                 queue_work(system_long_wq, &c->invalidate_work);
1100 }
1101
1102 static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca)
1103 {
1104         struct btree_trans trans;
1105         struct btree_iter iter;
1106         struct bkey_s_c k;
1107         struct bch_alloc_v4 a;
1108         struct bch_member *m;
1109         int ret;
1110
1111         bch2_trans_init(&trans, c, 0, 0);
1112
1113         for_each_btree_key(&trans, iter, BTREE_ID_alloc,
1114                            POS(ca->dev_idx, ca->mi.first_bucket),
1115                            BTREE_ITER_SLOTS|
1116                            BTREE_ITER_PREFETCH, k, ret) {
1117                 if (iter.pos.offset >= ca->mi.nbuckets)
1118                         break;
1119
1120                 bch2_alloc_to_v4(k, &a);
1121                 ret = __bch2_trans_do(&trans, NULL, NULL,
1122                                       BTREE_INSERT_LAZY_RW,
1123                                  bch2_bucket_do_index(&trans, k, a, true));
1124                 if (ret)
1125                         break;
1126         }
1127         bch2_trans_iter_exit(&trans, &iter);
1128
1129         bch2_trans_exit(&trans);
1130
1131         if (ret) {
1132                 bch_err(ca, "error initializing free space: %i", ret);
1133                 return ret;
1134         }
1135
1136         mutex_lock(&c->sb_lock);
1137         m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx;
1138         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1139         mutex_unlock(&c->sb_lock);
1140
1141         return ret;
1142 }
1143
1144 int bch2_fs_freespace_init(struct bch_fs *c)
1145 {
1146         struct bch_dev *ca;
1147         unsigned i;
1148         int ret = 0;
1149         bool doing_init = false;
1150
1151         /*
1152          * We can crash during the device add path, so we need to check this on
1153          * every mount:
1154          */
1155
1156         for_each_member_device(ca, c, i) {
1157                 if (ca->mi.freespace_initialized)
1158                         continue;
1159
1160                 if (!doing_init) {
1161                         bch_info(c, "initializing freespace");
1162                         doing_init = true;
1163                 }
1164
1165                 ret = bch2_dev_freespace_init(c, ca);
1166                 if (ret) {
1167                         percpu_ref_put(&ca->ref);
1168                         return ret;
1169                 }
1170         }
1171
1172         if (doing_init) {
1173                 mutex_lock(&c->sb_lock);
1174                 bch2_write_super(c);
1175                 mutex_unlock(&c->sb_lock);
1176
1177                 bch_verbose(c, "done initializing freespace");
1178         }
1179
1180         return ret;
1181 }
1182
1183 /* Bucket IO clocks: */
1184
1185 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
1186                               size_t bucket_nr, int rw)
1187 {
1188         struct bch_fs *c = trans->c;
1189         struct btree_iter iter;
1190         struct bkey_i_alloc_v4 *a;
1191         u64 now;
1192         int ret = 0;
1193
1194         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
1195         ret = PTR_ERR_OR_ZERO(a);
1196         if (ret)
1197                 return ret;
1198
1199         now = atomic64_read(&c->io_clock[rw].now);
1200         if (a->v.io_time[rw] == now)
1201                 goto out;
1202
1203         a->v.io_time[rw] = now;
1204
1205         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1206                 bch2_trans_commit(trans, NULL, NULL, 0);
1207 out:
1208         bch2_trans_iter_exit(trans, &iter);
1209         return ret;
1210 }
1211
1212 /* Startup/shutdown (ro/rw): */
1213
1214 void bch2_recalc_capacity(struct bch_fs *c)
1215 {
1216         struct bch_dev *ca;
1217         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
1218         unsigned bucket_size_max = 0;
1219         unsigned long ra_pages = 0;
1220         unsigned i;
1221
1222         lockdep_assert_held(&c->state_lock);
1223
1224         for_each_online_member(ca, c, i) {
1225                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
1226
1227                 ra_pages += bdi->ra_pages;
1228         }
1229
1230         bch2_set_ra_pages(c, ra_pages);
1231
1232         for_each_rw_member(ca, c, i) {
1233                 u64 dev_reserve = 0;
1234
1235                 /*
1236                  * We need to reserve buckets (from the number
1237                  * of currently available buckets) against
1238                  * foreground writes so that mainly copygc can
1239                  * make forward progress.
1240                  *
1241                  * We need enough to refill the various reserves
1242                  * from scratch - copygc will use its entire
1243                  * reserve all at once, then run against when
1244                  * its reserve is refilled (from the formerly
1245                  * available buckets).
1246                  *
1247                  * This reserve is just used when considering if
1248                  * allocations for foreground writes must wait -
1249                  * not -ENOSPC calculations.
1250                  */
1251
1252                 dev_reserve += ca->nr_btree_reserve * 2;
1253                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
1254
1255                 dev_reserve += 1;       /* btree write point */
1256                 dev_reserve += 1;       /* copygc write point */
1257                 dev_reserve += 1;       /* rebalance write point */
1258
1259                 dev_reserve *= ca->mi.bucket_size;
1260
1261                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
1262                                              ca->mi.first_bucket);
1263
1264                 reserved_sectors += dev_reserve * 2;
1265
1266                 bucket_size_max = max_t(unsigned, bucket_size_max,
1267                                         ca->mi.bucket_size);
1268         }
1269
1270         gc_reserve = c->opts.gc_reserve_bytes
1271                 ? c->opts.gc_reserve_bytes >> 9
1272                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
1273
1274         reserved_sectors = max(gc_reserve, reserved_sectors);
1275
1276         reserved_sectors = min(reserved_sectors, capacity);
1277
1278         c->capacity = capacity - reserved_sectors;
1279
1280         c->bucket_size_max = bucket_size_max;
1281
1282         /* Wake up case someone was waiting for buckets */
1283         closure_wake_up(&c->freelist_wait);
1284 }
1285
1286 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
1287 {
1288         struct open_bucket *ob;
1289         bool ret = false;
1290
1291         for (ob = c->open_buckets;
1292              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1293              ob++) {
1294                 spin_lock(&ob->lock);
1295                 if (ob->valid && !ob->on_partial_list &&
1296                     ob->dev == ca->dev_idx)
1297                         ret = true;
1298                 spin_unlock(&ob->lock);
1299         }
1300
1301         return ret;
1302 }
1303
1304 /* device goes ro: */
1305 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
1306 {
1307         unsigned i;
1308
1309         /* First, remove device from allocation groups: */
1310
1311         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1312                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
1313
1314         /*
1315          * Capacity is calculated based off of devices in allocation groups:
1316          */
1317         bch2_recalc_capacity(c);
1318
1319         /* Next, close write points that point to this device... */
1320         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1321                 bch2_writepoint_stop(c, ca, &c->write_points[i]);
1322
1323         bch2_writepoint_stop(c, ca, &c->copygc_write_point);
1324         bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
1325         bch2_writepoint_stop(c, ca, &c->btree_write_point);
1326
1327         mutex_lock(&c->btree_reserve_cache_lock);
1328         while (c->btree_reserve_cache_nr) {
1329                 struct btree_alloc *a =
1330                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1331
1332                 bch2_open_buckets_put(c, &a->ob);
1333         }
1334         mutex_unlock(&c->btree_reserve_cache_lock);
1335
1336         while (1) {
1337                 struct open_bucket *ob;
1338
1339                 spin_lock(&c->freelist_lock);
1340                 if (!ca->open_buckets_partial_nr) {
1341                         spin_unlock(&c->freelist_lock);
1342                         break;
1343                 }
1344                 ob = c->open_buckets +
1345                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
1346                 ob->on_partial_list = false;
1347                 spin_unlock(&c->freelist_lock);
1348
1349                 bch2_open_bucket_put(c, ob);
1350         }
1351
1352         bch2_ec_stop_dev(c, ca);
1353
1354         /*
1355          * Wake up threads that were blocked on allocation, so they can notice
1356          * the device can no longer be removed and the capacity has changed:
1357          */
1358         closure_wake_up(&c->freelist_wait);
1359
1360         /*
1361          * journal_res_get() can block waiting for free space in the journal -
1362          * it needs to notice there may not be devices to allocate from anymore:
1363          */
1364         wake_up(&c->journal.wait);
1365
1366         /* Now wait for any in flight writes: */
1367
1368         closure_wait_event(&c->open_buckets_wait,
1369                            !bch2_dev_has_open_write_point(c, ca));
1370 }
1371
1372 /* device goes rw: */
1373 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
1374 {
1375         unsigned i;
1376
1377         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1378                 if (ca->mi.data_allowed & (1 << i))
1379                         set_bit(ca->dev_idx, c->rw_devs[i].d);
1380 }
1381
1382 void bch2_fs_allocator_background_init(struct bch_fs *c)
1383 {
1384         spin_lock_init(&c->freelist_lock);
1385         INIT_WORK(&c->discard_work, bch2_do_discards_work);
1386         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
1387 }