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 const char *bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k)
 306 {
 307         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 308
 309         if (k.k->p.inode >= c->sb.nr_devices ||
 310             !c->devs[k.k->p.inode])
 311                 return "invalid device";
 312
 313         /* allow for unknown fields */
 314         if (bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v))
 315                 return "incorrect value size";
 316
 317         return NULL;
 318 }
 319
 320 const char *bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k)
 321 {
 322         struct bkey_alloc_unpacked u;
 323
 324         if (k.k->p.inode >= c->sb.nr_devices ||
 325             !c->devs[k.k->p.inode])
 326                 return "invalid device";
 327
 328         if (bch2_alloc_unpack_v2(&u, k))
 329                 return "unpack error";
 330
 331         return NULL;
 332 }
 333
 334 const char *bch2_alloc_v3_invalid(const struct bch_fs *c, struct bkey_s_c k)
 335 {
 336         struct bkey_alloc_unpacked u;
 337         struct bch_dev *ca;
 338
 339         if (k.k->p.inode >= c->sb.nr_devices ||
 340             !c->devs[k.k->p.inode])
 341                 return "invalid device";
 342
 343         ca = bch_dev_bkey_exists(c, k.k->p.inode);
 344
 345         if (k.k->p.offset < ca->mi.first_bucket ||
 346             k.k->p.offset >= ca->mi.nbuckets)
 347                 return "invalid bucket";
 348
 349         if (bch2_alloc_unpack_v3(&u, k))
 350                 return "unpack error";
 351
 352         return NULL;
 353 }
 354
 355 const char *bch2_alloc_v4_invalid(const struct bch_fs *c, struct bkey_s_c k)
 356 {
 357         struct bch_dev *ca;
 358
 359         if (k.k->p.inode >= c->sb.nr_devices ||
 360             !c->devs[k.k->p.inode])
 361                 return "invalid device";
 362
 363         ca = bch_dev_bkey_exists(c, k.k->p.inode);
 364
 365         if (k.k->p.offset < ca->mi.first_bucket ||
 366             k.k->p.offset >= ca->mi.nbuckets)
 367                 return "invalid bucket";
 368
 369         return NULL;
 370 }
 371
 372 void bch2_alloc_v4_swab(struct bkey_s k)
 373 {
 374         struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v;
 375
 376         a->journal_seq          = swab64(a->journal_seq);
 377         a->flags                = swab32(a->flags);
 378         a->dirty_sectors        = swab32(a->dirty_sectors);
 379         a->cached_sectors       = swab32(a->cached_sectors);
 380         a->io_time[0]           = swab64(a->io_time[0]);
 381         a->io_time[1]           = swab64(a->io_time[1]);
 382         a->stripe               = swab32(a->stripe);
 383         a->nr_external_backpointers = swab32(a->nr_external_backpointers);
 384 }
 385
 386 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 387 {
 388         struct bch_alloc_v4 a;
 389
 390         bch2_alloc_to_v4(k, &a);
 391
 392         pr_buf(out, "gen %u oldest_gen %u data_type %s journal_seq %llu need_discard %llu",
 393                a.gen, a.oldest_gen, bch2_data_types[a.data_type],
 394                a.journal_seq, BCH_ALLOC_V4_NEED_DISCARD(&a));
 395         pr_buf(out, " dirty_sectors %u",        a.dirty_sectors);
 396         pr_buf(out, " cached_sectors %u",       a.cached_sectors);
 397         pr_buf(out, " stripe %u",               a.stripe);
 398         pr_buf(out, " stripe_redundancy %u",    a.stripe_redundancy);
 399         pr_buf(out, " read_time %llu",          a.io_time[READ]);
 400         pr_buf(out, " write_time %llu",         a.io_time[WRITE]);
 401 }
 402
 403 int bch2_alloc_read(struct bch_fs *c)
 404 {
 405         struct btree_trans trans;
 406         struct btree_iter iter;
 407         struct bkey_s_c k;
 408         struct bch_alloc_v4 a;
 409         struct bch_dev *ca;
 410         int ret;
 411
 412         bch2_trans_init(&trans, c, 0, 0);
 413
 414         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 415                            BTREE_ITER_PREFETCH, k, ret) {
 416                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 417                 bch2_alloc_to_v4(k, &a);
 418
 419                 *bucket_gen(ca, k.k->p.offset) = a.gen;
 420         }
 421         bch2_trans_iter_exit(&trans, &iter);
 422
 423         bch2_trans_exit(&trans);
 424
 425         if (ret)
 426                 bch_err(c, "error reading alloc info: %i", ret);
 427
 428         return ret;
 429 }
 430
 431 /* Free space/discard btree: */
 432
 433 static int bch2_bucket_do_index(struct btree_trans *trans,
 434                                 struct bkey_s_c alloc_k,
 435                                 struct bch_alloc_v4 a,
 436                                 bool set)
 437 {
 438         struct bch_fs *c = trans->c;
 439         struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 440         struct btree_iter iter;
 441         struct bkey_s_c old;
 442         struct bkey_i *k;
 443         enum bucket_state state = bucket_state(a);
 444         enum btree_id btree;
 445         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 446         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 447         struct printbuf buf = PRINTBUF;
 448         int ret;
 449
 450         if (state != BUCKET_free &&
 451             state != BUCKET_need_discard)
 452                 return 0;
 453
 454         k = bch2_trans_kmalloc(trans, sizeof(*k));
 455         if (IS_ERR(k))
 456                 return PTR_ERR(k);
 457
 458         bkey_init(&k->k);
 459         k->k.type = new_type;
 460
 461         switch (state) {
 462         case BUCKET_free:
 463                 btree = BTREE_ID_freespace;
 464                 k->k.p = alloc_freespace_pos(alloc_k.k->p, a);
 465                 bch2_key_resize(&k->k, 1);
 466                 break;
 467         case BUCKET_need_discard:
 468                 btree = BTREE_ID_need_discard;
 469                 k->k.p = alloc_k.k->p;
 470                 break;
 471         default:
 472                 return 0;
 473         }
 474
 475         bch2_trans_iter_init(trans, &iter, btree,
 476                              bkey_start_pos(&k->k),
 477                              BTREE_ITER_INTENT);
 478         old = bch2_btree_iter_peek_slot(&iter);
 479         ret = bkey_err(old);
 480         if (ret)
 481                 goto err;
 482
 483         if (ca->mi.freespace_initialized &&
 484             bch2_fs_inconsistent_on(old.k->type != old_type, c,
 485                         "incorrect key when %s %s btree (got %s should be %s)\n"
 486                         "  for %s",
 487                         set ? "setting" : "clearing",
 488                         bch2_btree_ids[btree],
 489                         bch2_bkey_types[old.k->type],
 490                         bch2_bkey_types[old_type],
 491                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 492                 ret = -EIO;
 493                 goto err;
 494         }
 495
 496         ret = bch2_trans_update(trans, &iter, k, 0);
 497 err:
 498         bch2_trans_iter_exit(trans, &iter);
 499         printbuf_exit(&buf);
 500         return ret;
 501 }
 502
 503 int bch2_trans_mark_alloc(struct btree_trans *trans,
 504                           struct bkey_s_c old, struct bkey_i *new,
 505                           unsigned flags)
 506 {
 507         struct bch_fs *c = trans->c;
 508         struct bch_alloc_v4 old_a, *new_a;
 509         u64 old_lru, new_lru;
 510         int ret = 0;
 511
 512         /*
 513          * Deletion only happens in the device removal path, with
 514          * BTREE_TRIGGER_NORUN:
 515          */
 516         BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
 517
 518         bch2_alloc_to_v4(old, &old_a);
 519         new_a = &bkey_i_to_alloc_v4(new)->v;
 520
 521         if (new_a->dirty_sectors > old_a.dirty_sectors ||
 522             new_a->cached_sectors > old_a.cached_sectors) {
 523                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 524                 new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 525                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
 526                 SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
 527         }
 528
 529         if (old_a.data_type && !new_a->data_type &&
 530             old_a.gen == new_a->gen &&
 531             !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
 532                 new_a->gen++;
 533                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
 534         }
 535
 536         if (bucket_state(old_a) != bucket_state(*new_a) ||
 537             (bucket_state(*new_a) == BUCKET_free &&
 538              alloc_freespace_genbits(old_a) != alloc_freespace_genbits(*new_a))) {
 539                 ret =   bch2_bucket_do_index(trans, old, old_a, false) ?:
 540                         bch2_bucket_do_index(trans, bkey_i_to_s_c(new), *new_a, true);
 541                 if (ret)
 542                         return ret;
 543         }
 544
 545         old_lru = alloc_lru_idx(old_a);
 546         new_lru = alloc_lru_idx(*new_a);
 547
 548         if (old_lru != new_lru) {
 549                 ret = bch2_lru_change(trans, new->k.p.inode, new->k.p.offset,
 550                                       old_lru, &new_lru);
 551                 if (ret)
 552                         return ret;
 553
 554                 if (new_lru && new_a->io_time[READ] != new_lru)
 555                         new_a->io_time[READ] = new_lru;
 556         }
 557
 558         return 0;
 559 }
 560
 561 static int bch2_check_alloc_key(struct btree_trans *trans,
 562                                 struct btree_iter *alloc_iter)
 563 {
 564         struct bch_fs *c = trans->c;
 565         struct btree_iter discard_iter, freespace_iter, lru_iter;
 566         struct bch_alloc_v4 a;
 567         unsigned discard_key_type, freespace_key_type;
 568         struct bkey_s_c alloc_k, k;
 569         struct printbuf buf = PRINTBUF;
 570         struct printbuf buf2 = PRINTBUF;
 571         int ret;
 572
 573         alloc_k = bch2_btree_iter_peek(alloc_iter);
 574         if (!alloc_k.k)
 575                 return 0;
 576
 577         ret = bkey_err(alloc_k);
 578         if (ret)
 579                 return ret;
 580
 581         bch2_alloc_to_v4(alloc_k, &a);
 582         discard_key_type = bucket_state(a) == BUCKET_need_discard
 583                 ? KEY_TYPE_set : 0;
 584         freespace_key_type = bucket_state(a) == BUCKET_free
 585                 ? KEY_TYPE_set : 0;
 586
 587         bch2_trans_iter_init(trans, &discard_iter, BTREE_ID_need_discard,
 588                              alloc_k.k->p, 0);
 589         bch2_trans_iter_init(trans, &freespace_iter, BTREE_ID_freespace,
 590                              alloc_freespace_pos(alloc_k.k->p, a), 0);
 591         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
 592                              POS(alloc_k.k->p.inode, a.io_time[READ]), 0);
 593
 594         k = bch2_btree_iter_peek_slot(&discard_iter);
 595         ret = bkey_err(k);
 596         if (ret)
 597                 goto err;
 598
 599         if (fsck_err_on(k.k->type != discard_key_type, c,
 600                         "incorrect key in need_discard btree (got %s should be %s)\n"
 601                         "  %s",
 602                         bch2_bkey_types[k.k->type],
 603                         bch2_bkey_types[discard_key_type],
 604                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 605                 struct bkey_i *update =
 606                         bch2_trans_kmalloc(trans, sizeof(*update));
 607
 608                 ret = PTR_ERR_OR_ZERO(update);
 609                 if (ret)
 610                         goto err;
 611
 612                 bkey_init(&update->k);
 613                 update->k.type  = discard_key_type;
 614                 update->k.p     = discard_iter.pos;
 615
 616                 ret =   bch2_trans_update(trans, &discard_iter, update, 0) ?:
 617                         bch2_trans_commit(trans, NULL, NULL, 0);
 618                 if (ret)
 619                         goto err;
 620         }
 621
 622         k = bch2_btree_iter_peek_slot(&freespace_iter);
 623         ret = bkey_err(k);
 624         if (ret)
 625                 goto err;
 626
 627         if (fsck_err_on(k.k->type != freespace_key_type, c,
 628                         "incorrect key in freespace btree (got %s should be %s)\n"
 629                         "  %s",
 630                         bch2_bkey_types[k.k->type],
 631                         bch2_bkey_types[freespace_key_type],
 632                         (printbuf_reset(&buf),
 633                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 634                 struct bkey_i *update =
 635                         bch2_trans_kmalloc(trans, sizeof(*update));
 636
 637                 ret = PTR_ERR_OR_ZERO(update);
 638                 if (ret)
 639                         goto err;
 640
 641                 bkey_init(&update->k);
 642                 update->k.type  = freespace_key_type;
 643                 update->k.p     = freespace_iter.pos;
 644                 bch2_key_resize(&update->k, 1);
 645
 646                 ret   = bch2_trans_update(trans, &freespace_iter, update, 0) ?:
 647                         bch2_trans_commit(trans, NULL, NULL, 0);
 648                 if (ret)
 649                         goto err;
 650         }
 651
 652         if (bucket_state(a) == BUCKET_cached) {
 653                 k = bch2_btree_iter_peek_slot(&lru_iter);
 654                 ret = bkey_err(k);
 655                 if (ret)
 656                         goto err;
 657
 658                 if (fsck_err_on(!a.io_time[READ], c,
 659                                 "cached bucket with read_time 0\n"
 660                                 "  %s",
 661                         (printbuf_reset(&buf),
 662                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) ||
 663                     fsck_err_on(k.k->type != KEY_TYPE_lru ||
 664                                 le64_to_cpu(bkey_s_c_to_lru(k).v->idx) != alloc_k.k->p.offset, c,
 665                                 "incorrect/missing lru entry\n"
 666                                 "  %s\n"
 667                                 "  %s",
 668                                 (printbuf_reset(&buf),
 669                                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
 670                                 (bch2_bkey_val_to_text(&buf2, c, k), buf2.buf))) {
 671                         u64 read_time = a.io_time[READ];
 672
 673                         if (!a.io_time[READ])
 674                                 a.io_time[READ] = atomic64_read(&c->io_clock[READ].now);
 675
 676                         ret   = bch2_lru_change(trans,
 677                                                 alloc_k.k->p.inode,
 678                                                 alloc_k.k->p.offset,
 679                                                 0, &a.io_time[READ]);
 680                         if (ret)
 681                                 goto err;
 682
 683                         if (a.io_time[READ] != read_time) {
 684                                 struct bkey_i_alloc_v4 *a_mut =
 685                                         bch2_alloc_to_v4_mut(trans, alloc_k);
 686                                 ret = PTR_ERR_OR_ZERO(a_mut);
 687                                 if (ret)
 688                                         goto err;
 689
 690                                 a_mut->v.io_time[READ] = a.io_time[READ];
 691                                 ret = bch2_trans_update(trans, alloc_iter,
 692                                                         &a_mut->k_i, BTREE_TRIGGER_NORUN);
 693                                 if (ret)
 694                                         goto err;
 695                         }
 696
 697                         ret = bch2_trans_commit(trans, NULL, NULL, 0);
 698                         if (ret)
 699                                 goto err;
 700                 }
 701         }
 702 err:
 703 fsck_err:
 704         bch2_trans_iter_exit(trans, &lru_iter);
 705         bch2_trans_iter_exit(trans, &freespace_iter);
 706         bch2_trans_iter_exit(trans, &discard_iter);
 707         printbuf_exit(&buf2);
 708         printbuf_exit(&buf);
 709         return ret;
 710 }
 711
 712 static inline bool bch2_dev_bucket_exists(struct bch_fs *c, struct bpos pos)
 713 {
 714         struct bch_dev *ca;
 715
 716         if (pos.inode >= c->sb.nr_devices || !c->devs[pos.inode])
 717                 return false;
 718
 719         ca = bch_dev_bkey_exists(c, pos.inode);
 720         return pos.offset >= ca->mi.first_bucket &&
 721                 pos.offset < ca->mi.nbuckets;
 722 }
 723
 724 static int bch2_check_freespace_key(struct btree_trans *trans,
 725                                     struct btree_iter *freespace_iter,
 726                                     bool initial)
 727 {
 728         struct bch_fs *c = trans->c;
 729         struct btree_iter alloc_iter;
 730         struct bkey_s_c k, freespace_k;
 731         struct bch_alloc_v4 a;
 732         u64 genbits;
 733         struct bpos pos;
 734         struct bkey_i *update;
 735         struct printbuf buf = PRINTBUF;
 736         int ret;
 737
 738         freespace_k = bch2_btree_iter_peek(freespace_iter);
 739         if (!freespace_k.k)
 740                 return 1;
 741
 742         ret = bkey_err(freespace_k);
 743         if (ret)
 744                 return ret;
 745
 746         pos = freespace_iter->pos;
 747         pos.offset &= ~(~0ULL << 56);
 748         genbits = freespace_iter->pos.offset & (~0ULL << 56);
 749
 750         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
 751
 752         if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
 753                         "%llu:%llu set in freespace btree but device or bucket does not exist",
 754                         pos.inode, pos.offset))
 755                 goto delete;
 756
 757         k = bch2_btree_iter_peek_slot(&alloc_iter);
 758         ret = bkey_err(k);
 759         if (ret)
 760                 goto err;
 761
 762         bch2_alloc_to_v4(k, &a);
 763
 764         if (fsck_err_on(bucket_state(a) != BUCKET_free ||
 765                         genbits != alloc_freespace_genbits(a), c,
 766                         "%s\n  incorrectly set in freespace index (free %u, genbits %llu should be %llu)",
 767                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
 768                         bucket_state(a) == BUCKET_free,
 769                         genbits >> 56, alloc_freespace_genbits(a) >> 56))
 770                 goto delete;
 771 out:
 772 err:
 773 fsck_err:
 774         bch2_trans_iter_exit(trans, &alloc_iter);
 775         printbuf_exit(&buf);
 776         return ret;
 777 delete:
 778         update = bch2_trans_kmalloc(trans, sizeof(*update));
 779         ret = PTR_ERR_OR_ZERO(update);
 780         if (ret)
 781                 goto err;
 782
 783         bkey_init(&update->k);
 784         update->k.p = freespace_iter->pos;
 785         bch2_key_resize(&update->k, 1);
 786
 787         ret   = bch2_trans_update(trans, freespace_iter, update, 0) ?:
 788                 bch2_trans_commit(trans, NULL, NULL, 0);
 789         goto out;
 790 }
 791
 792 int bch2_check_alloc_info(struct bch_fs *c, bool initial)
 793 {
 794         struct btree_trans trans;
 795         struct btree_iter iter;
 796         struct bkey_s_c k;
 797         int ret = 0, last_dev = -1;
 798
 799         bch2_trans_init(&trans, c, 0, 0);
 800
 801         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 802                            BTREE_ITER_PREFETCH, k, ret) {
 803                 if (k.k->p.inode != last_dev) {
 804                         struct bch_dev *ca = bch_dev_bkey_exists(c, k.k->p.inode);
 805
 806                         if (!ca->mi.freespace_initialized) {
 807                                 bch2_btree_iter_set_pos(&iter, POS(k.k->p.inode + 1, 0));
 808                                 continue;
 809                         }
 810
 811                         last_dev = k.k->p.inode;
 812                 }
 813
 814                 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
 815                         bch2_check_alloc_key(&trans, &iter));
 816                 if (ret)
 817                         break;
 818         }
 819         bch2_trans_iter_exit(&trans, &iter);
 820
 821         if (ret)
 822                 goto err;
 823
 824         bch2_trans_iter_init(&trans, &iter, BTREE_ID_freespace, POS_MIN,
 825                              BTREE_ITER_PREFETCH);
 826         while (1) {
 827                 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
 828                         bch2_check_freespace_key(&trans, &iter, initial));
 829                 if (ret)
 830                         break;
 831
 832                 bch2_btree_iter_set_pos(&iter, bpos_nosnap_successor(iter.pos));
 833         }
 834         bch2_trans_iter_exit(&trans, &iter);
 835 err:
 836         bch2_trans_exit(&trans);
 837         return ret < 0 ? ret : 0;
 838 }
 839
 840 static int bch2_clear_need_discard(struct btree_trans *trans, struct bpos pos,
 841                                    struct bch_dev *ca, bool *discard_done)
 842 {
 843         struct bch_fs *c = trans->c;
 844         struct btree_iter iter;
 845         struct bkey_s_c k;
 846         struct bkey_i_alloc_v4 *a;
 847         struct printbuf buf = PRINTBUF;
 848         int ret;
 849
 850         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, pos,
 851                              BTREE_ITER_CACHED);
 852         k = bch2_btree_iter_peek_slot(&iter);
 853         ret = bkey_err(k);
 854         if (ret)
 855                 goto out;
 856
 857         a = bch2_alloc_to_v4_mut(trans, k);
 858         ret = PTR_ERR_OR_ZERO(a);
 859         if (ret)
 860                 goto out;
 861
 862         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
 863                 a->v.gen++;
 864                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
 865                 goto write;
 866         }
 867
 868         BUG_ON(a->v.journal_seq > c->journal.flushed_seq_ondisk);
 869
 870         if (bch2_fs_inconsistent_on(!BCH_ALLOC_V4_NEED_DISCARD(&a->v), c,
 871                         "%s\n  incorrectly set in need_discard btree",
 872                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 873                 ret = -EIO;
 874                 goto out;
 875         }
 876
 877         if (!*discard_done && ca->mi.discard && !c->opts.nochanges) {
 878                 /*
 879                  * This works without any other locks because this is the only
 880                  * thread that removes items from the need_discard tree
 881                  */
 882                 bch2_trans_unlock(trans);
 883                 blkdev_issue_discard(ca->disk_sb.bdev,
 884                                      k.k->p.offset * ca->mi.bucket_size,
 885                                      ca->mi.bucket_size,
 886                                      GFP_KERNEL, 0);
 887                 *discard_done = true;
 888
 889                 ret = bch2_trans_relock(trans) ? 0 : -EINTR;
 890                 if (ret)
 891                         goto out;
 892         }
 893
 894         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
 895 write:
 896         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
 897 out:
 898         bch2_trans_iter_exit(trans, &iter);
 899         printbuf_exit(&buf);
 900         return ret;
 901 }
 902
 903 static void bch2_do_discards_work(struct work_struct *work)
 904 {
 905         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
 906         struct bch_dev *ca = NULL;
 907         struct btree_trans trans;
 908         struct btree_iter iter;
 909         struct bkey_s_c k;
 910         u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
 911         int ret;
 912
 913         bch2_trans_init(&trans, c, 0, 0);
 914
 915         for_each_btree_key(&trans, iter, BTREE_ID_need_discard,
 916                            POS_MIN, 0, k, ret) {
 917                 bool discard_done = false;
 918
 919                 if (ca && k.k->p.inode != ca->dev_idx) {
 920                         percpu_ref_put(&ca->io_ref);
 921                         ca = NULL;
 922                 }
 923
 924                 if (!ca) {
 925                         ca = bch_dev_bkey_exists(c, k.k->p.inode);
 926                         if (!percpu_ref_tryget(&ca->io_ref)) {
 927                                 ca = NULL;
 928                                 bch2_btree_iter_set_pos(&iter, POS(k.k->p.inode + 1, 0));
 929                                 continue;
 930                         }
 931                 }
 932
 933                 seen++;
 934
 935                 if (bch2_bucket_is_open_safe(c, k.k->p.inode, k.k->p.offset)) {
 936                         open++;
 937                         continue;
 938                 }
 939
 940                 if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 941                                 c->journal.flushed_seq_ondisk,
 942                                 k.k->p.inode, k.k->p.offset)) {
 943                         need_journal_commit++;
 944                         continue;
 945                 }
 946
 947                 ret = __bch2_trans_do(&trans, NULL, NULL,
 948                                       BTREE_INSERT_USE_RESERVE|
 949                                       BTREE_INSERT_NOFAIL,
 950                                 bch2_clear_need_discard(&trans, k.k->p, ca, &discard_done));
 951                 if (ret)
 952                         break;
 953
 954                 discarded++;
 955         }
 956         bch2_trans_iter_exit(&trans, &iter);
 957
 958         if (ca)
 959                 percpu_ref_put(&ca->io_ref);
 960
 961         bch2_trans_exit(&trans);
 962
 963         if (need_journal_commit * 2 > seen)
 964                 bch2_journal_flush_async(&c->journal, NULL);
 965
 966         percpu_ref_put(&c->writes);
 967
 968         trace_do_discards(c, seen, open, need_journal_commit, discarded, ret);
 969 }
 970
 971 void bch2_do_discards(struct bch_fs *c)
 972 {
 973         if (percpu_ref_tryget(&c->writes) &&
 974             !queue_work(system_long_wq, &c->discard_work))
 975                 percpu_ref_put(&c->writes);
 976 }
 977
 978 static int invalidate_one_bucket(struct btree_trans *trans, struct bch_dev *ca)
 979 {
 980         struct bch_fs *c = trans->c;
 981         struct btree_iter lru_iter, alloc_iter = { NULL };
 982         struct bkey_s_c k;
 983         struct bkey_i_alloc_v4 *a;
 984         u64 bucket, idx;
 985         int ret;
 986
 987         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
 988                              POS(ca->dev_idx, 0), 0);
 989         k = bch2_btree_iter_peek(&lru_iter);
 990         ret = bkey_err(k);
 991         if (ret)
 992                 goto out;
 993
 994         if (!k.k || k.k->p.inode != ca->dev_idx)
 995                 goto out;
 996
 997         if (bch2_fs_inconsistent_on(k.k->type != KEY_TYPE_lru, c,
 998                                     "non lru key in lru btree"))
 999                 goto out;
1000
1001         idx     = k.k->p.offset;
1002         bucket  = le64_to_cpu(bkey_s_c_to_lru(k).v->idx);
1003
1004         a = bch2_trans_start_alloc_update(trans, &alloc_iter,
1005                                           POS(ca->dev_idx, bucket));
1006         ret = PTR_ERR_OR_ZERO(a);
1007         if (ret)
1008                 goto out;
1009
1010         if (bch2_fs_inconsistent_on(idx != alloc_lru_idx(a->v), c,
1011                         "invalidating bucket with wrong lru idx (got %llu should be %llu",
1012                         idx, alloc_lru_idx(a->v)))
1013                 goto out;
1014
1015         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1016         a->v.gen++;
1017         a->v.data_type          = 0;
1018         a->v.dirty_sectors      = 0;
1019         a->v.cached_sectors     = 0;
1020         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1021         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1022
1023         ret = bch2_trans_update(trans, &alloc_iter, &a->k_i,
1024                                 BTREE_TRIGGER_BUCKET_INVALIDATE);
1025 out:
1026         bch2_trans_iter_exit(trans, &alloc_iter);
1027         bch2_trans_iter_exit(trans, &lru_iter);
1028         return ret;
1029 }
1030
1031 static void bch2_do_invalidates_work(struct work_struct *work)
1032 {
1033         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1034         struct bch_dev *ca;
1035         struct btree_trans trans;
1036         unsigned i;
1037         int ret = 0;
1038
1039         bch2_trans_init(&trans, c, 0, 0);
1040
1041         for_each_member_device(ca, c, i)
1042                 while (!ret && should_invalidate_buckets(ca))
1043                         ret = __bch2_trans_do(&trans, NULL, NULL,
1044                                               BTREE_INSERT_USE_RESERVE|
1045                                               BTREE_INSERT_NOFAIL,
1046                                         invalidate_one_bucket(&trans, ca));
1047
1048         bch2_trans_exit(&trans);
1049         percpu_ref_put(&c->writes);
1050 }
1051
1052 void bch2_do_invalidates(struct bch_fs *c)
1053 {
1054         if (percpu_ref_tryget(&c->writes))
1055                 queue_work(system_long_wq, &c->invalidate_work);
1056 }
1057
1058 static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca)
1059 {
1060         struct btree_trans trans;
1061         struct btree_iter iter;
1062         struct bkey_s_c k;
1063         struct bch_alloc_v4 a;
1064         struct bch_member *m;
1065         int ret;
1066
1067         bch2_trans_init(&trans, c, 0, 0);
1068
1069         for_each_btree_key(&trans, iter, BTREE_ID_alloc,
1070                            POS(ca->dev_idx, ca->mi.first_bucket),
1071                            BTREE_ITER_SLOTS|
1072                            BTREE_ITER_PREFETCH, k, ret) {
1073                 if (iter.pos.offset >= ca->mi.nbuckets)
1074                         break;
1075
1076                 bch2_alloc_to_v4(k, &a);
1077                 ret = __bch2_trans_do(&trans, NULL, NULL,
1078                                       BTREE_INSERT_LAZY_RW,
1079                                  bch2_bucket_do_index(&trans, k, a, true));
1080                 if (ret)
1081                         break;
1082         }
1083         bch2_trans_iter_exit(&trans, &iter);
1084
1085         bch2_trans_exit(&trans);
1086
1087         if (ret) {
1088                 bch_err(ca, "error initializing free space: %i", ret);
1089                 return ret;
1090         }
1091
1092         mutex_lock(&c->sb_lock);
1093         m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx;
1094         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1095         mutex_unlock(&c->sb_lock);
1096
1097         return ret;
1098 }
1099
1100 int bch2_fs_freespace_init(struct bch_fs *c)
1101 {
1102         struct bch_dev *ca;
1103         unsigned i;
1104         int ret = 0;
1105         bool doing_init = false;
1106
1107         /*
1108          * We can crash during the device add path, so we need to check this on
1109          * every mount:
1110          */
1111
1112         for_each_member_device(ca, c, i) {
1113                 if (ca->mi.freespace_initialized)
1114                         continue;
1115
1116                 if (!doing_init) {
1117                         bch_info(c, "initializing freespace");
1118                         doing_init = true;
1119                 }
1120
1121                 ret = bch2_dev_freespace_init(c, ca);
1122                 if (ret) {
1123                         percpu_ref_put(&ca->ref);
1124                         return ret;
1125                 }
1126         }
1127
1128         if (doing_init) {
1129                 mutex_lock(&c->sb_lock);
1130                 bch2_write_super(c);
1131                 mutex_unlock(&c->sb_lock);
1132
1133                 bch_verbose(c, "done initializing freespace");
1134         }
1135
1136         return ret;
1137 }
1138
1139 /* Bucket IO clocks: */
1140
1141 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
1142                               size_t bucket_nr, int rw)
1143 {
1144         struct bch_fs *c = trans->c;
1145         struct btree_iter iter;
1146         struct bkey_i_alloc_v4 *a;
1147         u64 now;
1148         int ret = 0;
1149
1150         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
1151         ret = PTR_ERR_OR_ZERO(a);
1152         if (ret)
1153                 return ret;
1154
1155         now = atomic64_read(&c->io_clock[rw].now);
1156         if (a->v.io_time[rw] == now)
1157                 goto out;
1158
1159         a->v.io_time[rw] = now;
1160
1161         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1162                 bch2_trans_commit(trans, NULL, NULL, 0);
1163 out:
1164         bch2_trans_iter_exit(trans, &iter);
1165         return ret;
1166 }
1167
1168 /* Startup/shutdown (ro/rw): */
1169
1170 void bch2_recalc_capacity(struct bch_fs *c)
1171 {
1172         struct bch_dev *ca;
1173         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
1174         unsigned bucket_size_max = 0;
1175         unsigned long ra_pages = 0;
1176         unsigned i;
1177
1178         lockdep_assert_held(&c->state_lock);
1179
1180         for_each_online_member(ca, c, i) {
1181                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
1182
1183                 ra_pages += bdi->ra_pages;
1184         }
1185
1186         bch2_set_ra_pages(c, ra_pages);
1187
1188         for_each_rw_member(ca, c, i) {
1189                 u64 dev_reserve = 0;
1190
1191                 /*
1192                  * We need to reserve buckets (from the number
1193                  * of currently available buckets) against
1194                  * foreground writes so that mainly copygc can
1195                  * make forward progress.
1196                  *
1197                  * We need enough to refill the various reserves
1198                  * from scratch - copygc will use its entire
1199                  * reserve all at once, then run against when
1200                  * its reserve is refilled (from the formerly
1201                  * available buckets).
1202                  *
1203                  * This reserve is just used when considering if
1204                  * allocations for foreground writes must wait -
1205                  * not -ENOSPC calculations.
1206                  */
1207
1208                 dev_reserve += ca->nr_btree_reserve * 2;
1209                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
1210
1211                 dev_reserve += 1;       /* btree write point */
1212                 dev_reserve += 1;       /* copygc write point */
1213                 dev_reserve += 1;       /* rebalance write point */
1214
1215                 dev_reserve *= ca->mi.bucket_size;
1216
1217                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
1218                                              ca->mi.first_bucket);
1219
1220                 reserved_sectors += dev_reserve * 2;
1221
1222                 bucket_size_max = max_t(unsigned, bucket_size_max,
1223                                         ca->mi.bucket_size);
1224         }
1225
1226         gc_reserve = c->opts.gc_reserve_bytes
1227                 ? c->opts.gc_reserve_bytes >> 9
1228                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
1229
1230         reserved_sectors = max(gc_reserve, reserved_sectors);
1231
1232         reserved_sectors = min(reserved_sectors, capacity);
1233
1234         c->capacity = capacity - reserved_sectors;
1235
1236         c->bucket_size_max = bucket_size_max;
1237
1238         /* Wake up case someone was waiting for buckets */
1239         closure_wake_up(&c->freelist_wait);
1240 }
1241
1242 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
1243 {
1244         struct open_bucket *ob;
1245         bool ret = false;
1246
1247         for (ob = c->open_buckets;
1248              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1249              ob++) {
1250                 spin_lock(&ob->lock);
1251                 if (ob->valid && !ob->on_partial_list &&
1252                     ob->dev == ca->dev_idx)
1253                         ret = true;
1254                 spin_unlock(&ob->lock);
1255         }
1256
1257         return ret;
1258 }
1259
1260 /* device goes ro: */
1261 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
1262 {
1263         unsigned i;
1264
1265         /* First, remove device from allocation groups: */
1266
1267         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1268                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
1269
1270         /*
1271          * Capacity is calculated based off of devices in allocation groups:
1272          */
1273         bch2_recalc_capacity(c);
1274
1275         /* Next, close write points that point to this device... */
1276         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1277                 bch2_writepoint_stop(c, ca, &c->write_points[i]);
1278
1279         bch2_writepoint_stop(c, ca, &c->copygc_write_point);
1280         bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
1281         bch2_writepoint_stop(c, ca, &c->btree_write_point);
1282
1283         mutex_lock(&c->btree_reserve_cache_lock);
1284         while (c->btree_reserve_cache_nr) {
1285                 struct btree_alloc *a =
1286                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1287
1288                 bch2_open_buckets_put(c, &a->ob);
1289         }
1290         mutex_unlock(&c->btree_reserve_cache_lock);
1291
1292         while (1) {
1293                 struct open_bucket *ob;
1294
1295                 spin_lock(&c->freelist_lock);
1296                 if (!ca->open_buckets_partial_nr) {
1297                         spin_unlock(&c->freelist_lock);
1298                         break;
1299                 }
1300                 ob = c->open_buckets +
1301                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
1302                 ob->on_partial_list = false;
1303                 spin_unlock(&c->freelist_lock);
1304
1305                 bch2_open_bucket_put(c, ob);
1306         }
1307
1308         bch2_ec_stop_dev(c, ca);
1309
1310         /*
1311          * Wake up threads that were blocked on allocation, so they can notice
1312          * the device can no longer be removed and the capacity has changed:
1313          */
1314         closure_wake_up(&c->freelist_wait);
1315
1316         /*
1317          * journal_res_get() can block waiting for free space in the journal -
1318          * it needs to notice there may not be devices to allocate from anymore:
1319          */
1320         wake_up(&c->journal.wait);
1321
1322         /* Now wait for any in flight writes: */
1323
1324         closure_wait_event(&c->open_buckets_wait,
1325                            !bch2_dev_has_open_write_point(c, ca));
1326 }
1327
1328 /* device goes rw: */
1329 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
1330 {
1331         unsigned i;
1332
1333         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1334                 if (ca->mi.data_allowed & (1 << i))
1335                         set_bit(ca->dev_idx, c->rw_devs[i].d);
1336 }
1337
1338 void bch2_fs_allocator_background_init(struct bch_fs *c)
1339 {
1340         spin_lock_init(&c->freelist_lock);
1341         INIT_WORK(&c->discard_work, bch2_do_discards_work);
1342         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
1343 }