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 static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
  31 #define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8,
  32         BCH_ALLOC_FIELDS_V1()
  33 #undef x
  34 };
  35
  36 const char * const bch2_bucket_states[] = {
  37         "free",
  38         "need gc gens",
  39         "need discard",
  40         "cached",
  41         "dirty",
  42         NULL
  43 };
  44
  45 /* Persistent alloc info: */
  46
  47 static inline u64 alloc_field_v1_get(const struct bch_alloc *a,
  48                                      const void **p, unsigned field)
  49 {
  50         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  51         u64 v;
  52
  53         if (!(a->fields & (1 << field)))
  54                 return 0;
  55
  56         switch (bytes) {
  57         case 1:
  58                 v = *((const u8 *) *p);
  59                 break;
  60         case 2:
  61                 v = le16_to_cpup(*p);
  62                 break;
  63         case 4:
  64                 v = le32_to_cpup(*p);
  65                 break;
  66         case 8:
  67                 v = le64_to_cpup(*p);
  68                 break;
  69         default:
  70                 BUG();
  71         }
  72
  73         *p += bytes;
  74         return v;
  75 }
  76
  77 static inline void alloc_field_v1_put(struct bkey_i_alloc *a, void **p,
  78                                       unsigned field, u64 v)
  79 {
  80         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  81
  82         if (!v)
  83                 return;
  84
  85         a->v.fields |= 1 << field;
  86
  87         switch (bytes) {
  88         case 1:
  89                 *((u8 *) *p) = v;
  90                 break;
  91         case 2:
  92                 *((__le16 *) *p) = cpu_to_le16(v);
  93                 break;
  94         case 4:
  95                 *((__le32 *) *p) = cpu_to_le32(v);
  96                 break;
  97         case 8:
  98                 *((__le64 *) *p) = cpu_to_le64(v);
  99                 break;
 100         default:
 101                 BUG();
 102         }
 103
 104         *p += bytes;
 105 }
 106
 107 static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out,
 108                                  struct bkey_s_c k)
 109 {
 110         const struct bch_alloc *in = bkey_s_c_to_alloc(k).v;
 111         const void *d = in->data;
 112         unsigned idx = 0;
 113
 114         out->gen = in->gen;
 115
 116 #define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++);
 117         BCH_ALLOC_FIELDS_V1()
 118 #undef  x
 119 }
 120
 121 static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out,
 122                                 struct bkey_s_c k)
 123 {
 124         struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k);
 125         const u8 *in = a.v->data;
 126         const u8 *end = bkey_val_end(a);
 127         unsigned fieldnr = 0;
 128         int ret;
 129         u64 v;
 130
 131         out->gen        = a.v->gen;
 132         out->oldest_gen = a.v->oldest_gen;
 133         out->data_type  = a.v->data_type;
 134
 135 #define x(_name, _bits)                                                 \
 136         if (fieldnr < a.v->nr_fields) {                                 \
 137                 ret = bch2_varint_decode_fast(in, end, &v);             \
 138                 if (ret < 0)                                            \
 139                         return ret;                                     \
 140                 in += ret;                                              \
 141         } else {                                                        \
 142                 v = 0;                                                  \
 143         }                                                               \
 144         out->_name = v;                                                 \
 145         if (v != out->_name)                                            \
 146                 return -1;                                              \
 147         fieldnr++;
 148
 149         BCH_ALLOC_FIELDS_V2()
 150 #undef  x
 151         return 0;
 152 }
 153
 154 static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out,
 155                                 struct bkey_s_c k)
 156 {
 157         struct bkey_s_c_alloc_v3 a = bkey_s_c_to_alloc_v3(k);
 158         const u8 *in = a.v->data;
 159         const u8 *end = bkey_val_end(a);
 160         unsigned fieldnr = 0;
 161         int ret;
 162         u64 v;
 163
 164         out->gen        = a.v->gen;
 165         out->oldest_gen = a.v->oldest_gen;
 166         out->data_type  = a.v->data_type;
 167         out->need_discard = BCH_ALLOC_NEED_DISCARD(a.v);
 168         out->need_inc_gen = BCH_ALLOC_NEED_INC_GEN(a.v);
 169         out->journal_seq = le64_to_cpu(a.v->journal_seq);
 170
 171 #define x(_name, _bits)                                                 \
 172         if (fieldnr < a.v->nr_fields) {                                 \
 173                 ret = bch2_varint_decode_fast(in, end, &v);             \
 174                 if (ret < 0)                                            \
 175                         return ret;                                     \
 176                 in += ret;                                              \
 177         } else {                                                        \
 178                 v = 0;                                                  \
 179         }                                                               \
 180         out->_name = v;                                                 \
 181         if (v != out->_name)                                            \
 182                 return -1;                                              \
 183         fieldnr++;
 184
 185         BCH_ALLOC_FIELDS_V2()
 186 #undef  x
 187         return 0;
 188 }
 189
 190 static void bch2_alloc_pack_v3(struct bkey_alloc_buf *dst,
 191                                const struct bkey_alloc_unpacked src)
 192 {
 193         struct bkey_i_alloc_v3 *a = bkey_alloc_v3_init(&dst->k);
 194         unsigned nr_fields = 0, last_nonzero_fieldnr = 0;
 195         u8 *out = a->v.data;
 196         u8 *end = (void *) &dst[1];
 197         u8 *last_nonzero_field = out;
 198         unsigned bytes;
 199
 200         a->k.p          = POS(src.dev, src.bucket);
 201         a->v.gen        = src.gen;
 202         a->v.oldest_gen = src.oldest_gen;
 203         a->v.data_type  = src.data_type;
 204         a->v.journal_seq = cpu_to_le64(src.journal_seq);
 205         SET_BCH_ALLOC_NEED_DISCARD(&a->v, src.need_discard);
 206         SET_BCH_ALLOC_NEED_INC_GEN(&a->v, src.need_inc_gen);
 207
 208 #define x(_name, _bits)                                                 \
 209         nr_fields++;                                                    \
 210                                                                         \
 211         if (src._name) {                                                \
 212                 out += bch2_varint_encode_fast(out, src._name);         \
 213                                                                         \
 214                 last_nonzero_field = out;                               \
 215                 last_nonzero_fieldnr = nr_fields;                       \
 216         } else {                                                        \
 217                 *out++ = 0;                                             \
 218         }
 219
 220         BCH_ALLOC_FIELDS_V2()
 221 #undef  x
 222         BUG_ON(out > end);
 223
 224         out = last_nonzero_field;
 225         a->v.nr_fields = last_nonzero_fieldnr;
 226
 227         bytes = (u8 *) out - (u8 *) &a->v;
 228         set_bkey_val_bytes(&a->k, bytes);
 229         memset_u64s_tail(&a->v, 0, bytes);
 230 }
 231
 232 struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 233 {
 234         struct bkey_alloc_unpacked ret = {
 235                 .dev    = k.k->p.inode,
 236                 .bucket = k.k->p.offset,
 237                 .gen    = 0,
 238         };
 239
 240         switch (k.k->type) {
 241         case KEY_TYPE_alloc:
 242                 bch2_alloc_unpack_v1(&ret, k);
 243                 break;
 244         case KEY_TYPE_alloc_v2:
 245                 bch2_alloc_unpack_v2(&ret, k);
 246                 break;
 247         case KEY_TYPE_alloc_v3:
 248                 bch2_alloc_unpack_v3(&ret, k);
 249                 break;
 250         }
 251
 252         return ret;
 253 }
 254
 255 struct bkey_alloc_buf *bch2_alloc_pack(struct btree_trans *trans,
 256                                        const struct bkey_alloc_unpacked src)
 257 {
 258         struct bkey_alloc_buf *dst;
 259
 260         dst = bch2_trans_kmalloc(trans, sizeof(struct bkey_alloc_buf));
 261         if (!IS_ERR(dst))
 262                 bch2_alloc_pack_v3(dst, src);
 263
 264         return dst;
 265 }
 266
 267 int bch2_alloc_write(struct btree_trans *trans, struct btree_iter *iter,
 268                      struct bkey_alloc_unpacked *u, unsigned trigger_flags)
 269 {
 270         struct bkey_alloc_buf *a = bch2_alloc_pack(trans, *u);
 271
 272         return PTR_ERR_OR_ZERO(a) ?:
 273                 bch2_trans_update(trans, iter, &a->k, trigger_flags);
 274 }
 275
 276 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
 277 {
 278         unsigned i, bytes = offsetof(struct bch_alloc, data);
 279
 280         for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++)
 281                 if (a->fields & (1 << i))
 282                         bytes += BCH_ALLOC_V1_FIELD_BYTES[i];
 283
 284         return DIV_ROUND_UP(bytes, sizeof(u64));
 285 }
 286
 287 const char *bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k)
 288 {
 289         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 290
 291         if (k.k->p.inode >= c->sb.nr_devices ||
 292             !c->devs[k.k->p.inode])
 293                 return "invalid device";
 294
 295         /* allow for unknown fields */
 296         if (bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v))
 297                 return "incorrect value size";
 298
 299         return NULL;
 300 }
 301
 302 const char *bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k)
 303 {
 304         struct bkey_alloc_unpacked u;
 305
 306         if (k.k->p.inode >= c->sb.nr_devices ||
 307             !c->devs[k.k->p.inode])
 308                 return "invalid device";
 309
 310         if (bch2_alloc_unpack_v2(&u, k))
 311                 return "unpack error";
 312
 313         return NULL;
 314 }
 315
 316 const char *bch2_alloc_v3_invalid(const struct bch_fs *c, struct bkey_s_c k)
 317 {
 318         struct bkey_alloc_unpacked u;
 319
 320         if (k.k->p.inode >= c->sb.nr_devices ||
 321             !c->devs[k.k->p.inode])
 322                 return "invalid device";
 323
 324         if (bch2_alloc_unpack_v3(&u, k))
 325                 return "unpack error";
 326
 327         return NULL;
 328 }
 329
 330 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c,
 331                            struct bkey_s_c k)
 332 {
 333         struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 334
 335         pr_buf(out, "gen %u oldest_gen %u data_type %s journal_seq %llu need_discard %u",
 336                u.gen, u.oldest_gen, bch2_data_types[u.data_type],
 337                u.journal_seq, u.need_discard);
 338 #define x(_name, ...)   pr_buf(out, " " #_name " %llu", (u64) u._name);
 339         BCH_ALLOC_FIELDS_V2()
 340 #undef  x
 341 }
 342
 343 int bch2_alloc_read(struct bch_fs *c)
 344 {
 345         struct btree_trans trans;
 346         struct btree_iter iter;
 347         struct bkey_s_c k;
 348         struct bch_dev *ca;
 349         int ret;
 350
 351         bch2_trans_init(&trans, c, 0, 0);
 352
 353         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 354                            BTREE_ITER_PREFETCH, k, ret) {
 355                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 356
 357                 *bucket_gen(ca, k.k->p.offset) = bch2_alloc_unpack(k).gen;
 358         }
 359         bch2_trans_iter_exit(&trans, &iter);
 360
 361         bch2_trans_exit(&trans);
 362
 363         if (ret)
 364                 bch_err(c, "error reading alloc info: %i", ret);
 365
 366         return ret;
 367 }
 368
 369 /* Free space/discard btree: */
 370
 371 static int bch2_bucket_do_index(struct btree_trans *trans,
 372                                 struct bkey_s_c alloc_k,
 373                                 struct bkey_alloc_unpacked a,
 374                                 bool set)
 375 {
 376         struct bch_fs *c = trans->c;
 377         struct bch_dev *ca = bch_dev_bkey_exists(c, a.dev);
 378         struct btree_iter iter;
 379         struct bkey_s_c old;
 380         struct bkey_i *k;
 381         enum bucket_state state = bucket_state(a);
 382         enum btree_id btree;
 383         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 384         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 385         struct printbuf buf = PRINTBUF;
 386         int ret;
 387
 388         if (state != BUCKET_free &&
 389             state != BUCKET_need_discard)
 390                 return 0;
 391
 392         k = bch2_trans_kmalloc(trans, sizeof(*k));
 393         if (IS_ERR(k))
 394                 return PTR_ERR(k);
 395
 396         bkey_init(&k->k);
 397         k->k.type = new_type;
 398
 399         switch (state) {
 400         case BUCKET_free:
 401                 btree = BTREE_ID_freespace;
 402                 k->k.p = alloc_freespace_pos(a);
 403                 bch2_key_resize(&k->k, 1);
 404                 break;
 405         case BUCKET_need_discard:
 406                 btree = BTREE_ID_need_discard;
 407                 k->k.p = POS(a.dev, a.bucket);
 408                 break;
 409         default:
 410                 return 0;
 411         }
 412
 413         bch2_trans_iter_init(trans, &iter, btree,
 414                              bkey_start_pos(&k->k),
 415                              BTREE_ITER_INTENT);
 416         old = bch2_btree_iter_peek_slot(&iter);
 417         ret = bkey_err(old);
 418         if (ret)
 419                 goto err;
 420
 421         if (ca->mi.freespace_initialized &&
 422             bch2_fs_inconsistent_on(old.k->type != old_type, c,
 423                         "incorrect key when %s %s btree (got %s should be %s)\n"
 424                         "  for %s",
 425                         set ? "setting" : "clearing",
 426                         bch2_btree_ids[btree],
 427                         bch2_bkey_types[old.k->type],
 428                         bch2_bkey_types[old_type],
 429                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 430                 ret = -EIO;
 431                 goto err;
 432         }
 433
 434         ret = bch2_trans_update(trans, &iter, k, 0);
 435 err:
 436         bch2_trans_iter_exit(trans, &iter);
 437         printbuf_exit(&buf);
 438         return ret;
 439 }
 440
 441 int bch2_trans_mark_alloc(struct btree_trans *trans,
 442                           struct bkey_s_c old, struct bkey_i *new,
 443                           unsigned flags)
 444 {
 445         struct bch_fs *c = trans->c;
 446         struct bkey_alloc_unpacked old_u = bch2_alloc_unpack(old);
 447         struct bkey_alloc_unpacked new_u = bch2_alloc_unpack(bkey_i_to_s_c(new));
 448         u64 old_lru, new_lru;
 449         bool need_repack = false;
 450         int ret = 0;
 451
 452         if (new_u.dirty_sectors > old_u.dirty_sectors ||
 453             new_u.cached_sectors > old_u.cached_sectors) {
 454                 new_u.read_time = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 455                 new_u.write_time = max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 456                 new_u.need_inc_gen = true;
 457                 new_u.need_discard = true;
 458                 need_repack = true;
 459         }
 460
 461         if (old_u.data_type && !new_u.data_type &&
 462             old_u.gen == new_u.gen &&
 463             !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
 464                 new_u.gen++;
 465                 new_u.need_inc_gen = false;
 466                 need_repack = true;
 467         }
 468
 469         if (bucket_state(old_u) != bucket_state(new_u) ||
 470             (bucket_state(new_u) == BUCKET_free &&
 471              alloc_freespace_genbits(old_u) != alloc_freespace_genbits(new_u))) {
 472                 ret =   bch2_bucket_do_index(trans, old, old_u, false) ?:
 473                         bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_u, true);
 474                 if (ret)
 475                         return ret;
 476         }
 477
 478         old_lru = alloc_lru_idx(old_u);
 479         new_lru = alloc_lru_idx(new_u);
 480
 481         if (old_lru != new_lru) {
 482                 ret = bch2_lru_change(trans, new->k.p.inode, new->k.p.offset,
 483                                       old_lru, &new_lru);
 484                 if (ret)
 485                         return ret;
 486
 487                 if (new_lru && new_u.read_time != new_lru) {
 488                         new_u.read_time = new_lru;
 489                         need_repack = true;
 490                 }
 491         }
 492
 493         if (need_repack && !bkey_deleted(&new->k))
 494                 bch2_alloc_pack_v3((void *) new, new_u);
 495
 496         return 0;
 497 }
 498
 499 static int bch2_check_alloc_key(struct btree_trans *trans,
 500                                 struct btree_iter *alloc_iter)
 501 {
 502         struct bch_fs *c = trans->c;
 503         struct btree_iter discard_iter, freespace_iter, lru_iter;
 504         struct bkey_alloc_unpacked a;
 505         unsigned discard_key_type, freespace_key_type;
 506         struct bkey_s_c alloc_k, k;
 507         struct printbuf buf = PRINTBUF;
 508         struct printbuf buf2 = PRINTBUF;
 509         int ret;
 510
 511         alloc_k = bch2_btree_iter_peek(alloc_iter);
 512         if (!alloc_k.k)
 513                 return 0;
 514
 515         ret = bkey_err(alloc_k);
 516         if (ret)
 517                 return ret;
 518
 519         a = bch2_alloc_unpack(alloc_k);
 520         discard_key_type = bucket_state(a) == BUCKET_need_discard
 521                 ? KEY_TYPE_set : 0;
 522         freespace_key_type = bucket_state(a) == BUCKET_free
 523                 ? KEY_TYPE_set : 0;
 524
 525         bch2_trans_iter_init(trans, &discard_iter, BTREE_ID_need_discard,
 526                              alloc_k.k->p, 0);
 527         bch2_trans_iter_init(trans, &freespace_iter, BTREE_ID_freespace,
 528                              alloc_freespace_pos(a), 0);
 529         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
 530                              POS(a.dev, a.read_time), 0);
 531
 532         k = bch2_btree_iter_peek_slot(&discard_iter);
 533         ret = bkey_err(k);
 534         if (ret)
 535                 goto err;
 536
 537         if (fsck_err_on(k.k->type != discard_key_type, c,
 538                         "incorrect key in need_discard btree (got %s should be %s)\n"
 539                         "  %s",
 540                         bch2_bkey_types[k.k->type],
 541                         bch2_bkey_types[discard_key_type],
 542                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 543                 struct bkey_i *update =
 544                         bch2_trans_kmalloc(trans, sizeof(*update));
 545
 546                 ret = PTR_ERR_OR_ZERO(update);
 547                 if (ret)
 548                         goto err;
 549
 550                 bkey_init(&update->k);
 551                 update->k.type  = discard_key_type;
 552                 update->k.p     = discard_iter.pos;
 553
 554                 ret =   bch2_trans_update(trans, &discard_iter, update, 0) ?:
 555                         bch2_trans_commit(trans, NULL, NULL, 0);
 556                 if (ret)
 557                         goto err;
 558         }
 559
 560         k = bch2_btree_iter_peek_slot(&freespace_iter);
 561         ret = bkey_err(k);
 562         if (ret)
 563                 goto err;
 564
 565         if (fsck_err_on(k.k->type != freespace_key_type, c,
 566                         "incorrect key in freespace btree (got %s should be %s)\n"
 567                         "  %s",
 568                         bch2_bkey_types[k.k->type],
 569                         bch2_bkey_types[freespace_key_type],
 570                         (printbuf_reset(&buf),
 571                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 572                 struct bkey_i *update =
 573                         bch2_trans_kmalloc(trans, sizeof(*update));
 574
 575                 ret = PTR_ERR_OR_ZERO(update);
 576                 if (ret)
 577                         goto err;
 578
 579                 bkey_init(&update->k);
 580                 update->k.type  = freespace_key_type;
 581                 update->k.p     = freespace_iter.pos;
 582                 bch2_key_resize(&update->k, 1);
 583
 584                 ret   = bch2_trans_update(trans, &freespace_iter, update, 0) ?:
 585                         bch2_trans_commit(trans, NULL, NULL, 0);
 586                 if (ret)
 587                         goto err;
 588         }
 589
 590         if (bucket_state(a) == BUCKET_cached) {
 591                 if (fsck_err_on(!a.read_time, c,
 592                                 "cached bucket with read_time 0\n"
 593                                 "  %s",
 594                         (printbuf_reset(&buf),
 595                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 596
 597                         a.read_time = atomic64_read(&c->io_clock[READ].now);
 598
 599                         ret   = bch2_lru_change(trans, a.dev, a.bucket,
 600                                                 0, &a.read_time) ?:
 601                                 bch2_alloc_write(trans, alloc_iter, &a, BTREE_TRIGGER_NORUN);
 602                                 bch2_trans_commit(trans, NULL, NULL, 0);
 603                         if (ret)
 604                                 goto err;
 605                 }
 606
 607                 k = bch2_btree_iter_peek_slot(&lru_iter);
 608                 ret = bkey_err(k);
 609                 if (ret)
 610                         goto err;
 611
 612                 if (fsck_err_on(k.k->type != KEY_TYPE_lru ||
 613                                 le64_to_cpu(bkey_s_c_to_lru(k).v->idx) != a.bucket, c,
 614                         "incorrect/missing lru entry\n"
 615                         "  %s\n"
 616                         "  %s",
 617                         (printbuf_reset(&buf),
 618                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
 619                         (bch2_bkey_val_to_text(&buf2, c, k), buf2.buf))) {
 620                         u64 read_time = a.read_time;
 621
 622                         ret   = bch2_lru_change(trans, a.dev, a.bucket,
 623                                                 0, &a.read_time) ?:
 624                                 (a.read_time != read_time
 625                                  ? bch2_alloc_write(trans, alloc_iter, &a, BTREE_TRIGGER_NORUN)
 626                                  : 0) ?:
 627                                 bch2_trans_commit(trans, NULL, NULL, 0);
 628                         if (ret)
 629                                 goto err;
 630                 }
 631         }
 632 err:
 633 fsck_err:
 634         bch2_trans_iter_exit(trans, &lru_iter);
 635         bch2_trans_iter_exit(trans, &freespace_iter);
 636         bch2_trans_iter_exit(trans, &discard_iter);
 637         printbuf_exit(&buf2);
 638         printbuf_exit(&buf);
 639         return ret;
 640 }
 641
 642 static inline bool bch2_dev_bucket_exists(struct bch_fs *c, struct bpos pos)
 643 {
 644         struct bch_dev *ca;
 645
 646         if (pos.inode >= c->sb.nr_devices || !c->devs[pos.inode])
 647                 return false;
 648
 649         ca = bch_dev_bkey_exists(c, pos.inode);
 650         return pos.offset >= ca->mi.first_bucket &&
 651                 pos.offset < ca->mi.nbuckets;
 652 }
 653
 654 static int bch2_check_freespace_key(struct btree_trans *trans,
 655                                     struct btree_iter *freespace_iter,
 656                                     bool initial)
 657 {
 658         struct bch_fs *c = trans->c;
 659         struct btree_iter alloc_iter;
 660         struct bkey_s_c k, freespace_k;
 661         struct bkey_alloc_unpacked a;
 662         u64 genbits;
 663         struct bpos pos;
 664         struct bkey_i *update;
 665         struct printbuf buf = PRINTBUF;
 666         int ret;
 667
 668         freespace_k = bch2_btree_iter_peek(freespace_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 = freespace_iter->pos;
 677         pos.offset &= ~(~0ULL << 56);
 678         genbits = freespace_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 freespace btree but device or bucket does not exist",
 684                         pos.inode, pos.offset))
 685                 goto delete;
 686
 687         k = bch2_btree_iter_peek_slot(&alloc_iter);
 688         ret = bkey_err(k);
 689         if (ret)
 690                 goto err;
 691
 692         a = bch2_alloc_unpack(k);
 693
 694         if (fsck_err_on(bucket_state(a) != BUCKET_free ||
 695                         genbits != alloc_freespace_genbits(a), c,
 696                         "%s\n  incorrectly set in freespace index (free %u, genbits %llu should be %llu)",
 697                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
 698                         bucket_state(a) == BUCKET_free,
 699                         genbits >> 56, alloc_freespace_genbits(a) >> 56))
 700                 goto delete;
 701 out:
 702 err:
 703 fsck_err:
 704         bch2_trans_iter_exit(trans, &alloc_iter);
 705         printbuf_exit(&buf);
 706         return ret;
 707 delete:
 708         update = bch2_trans_kmalloc(trans, sizeof(*update));
 709         ret = PTR_ERR_OR_ZERO(update);
 710         if (ret)
 711                 goto err;
 712
 713         bkey_init(&update->k);
 714         update->k.p = freespace_iter->pos;
 715         bch2_key_resize(&update->k, 1);
 716
 717         ret   = bch2_trans_update(trans, freespace_iter, update, 0) ?:
 718                 bch2_trans_commit(trans, NULL, NULL, 0);
 719         goto out;
 720 }
 721
 722 int bch2_check_alloc_info(struct bch_fs *c, bool initial)
 723 {
 724         struct btree_trans trans;
 725         struct btree_iter iter;
 726         struct bkey_s_c k;
 727         int ret = 0, last_dev = -1;
 728
 729         bch2_trans_init(&trans, c, 0, 0);
 730
 731         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 732                            BTREE_ITER_PREFETCH, k, ret) {
 733                 if (k.k->p.inode != last_dev) {
 734                         struct bch_dev *ca = bch_dev_bkey_exists(c, k.k->p.inode);
 735
 736                         if (!ca->mi.freespace_initialized) {
 737                                 bch2_btree_iter_set_pos(&iter, POS(k.k->p.inode + 1, 0));
 738                                 continue;
 739                         }
 740
 741                         last_dev = k.k->p.inode;
 742                 }
 743
 744                 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
 745                         bch2_check_alloc_key(&trans, &iter));
 746                 if (ret)
 747                         break;
 748         }
 749         bch2_trans_iter_exit(&trans, &iter);
 750
 751         if (ret)
 752                 goto err;
 753
 754         bch2_trans_iter_init(&trans, &iter, BTREE_ID_freespace, POS_MIN,
 755                              BTREE_ITER_PREFETCH);
 756         while (1) {
 757                 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
 758                         bch2_check_freespace_key(&trans, &iter, initial));
 759                 if (ret)
 760                         break;
 761
 762                 bch2_btree_iter_set_pos(&iter, bpos_nosnap_successor(iter.pos));
 763         }
 764         bch2_trans_iter_exit(&trans, &iter);
 765 err:
 766         bch2_trans_exit(&trans);
 767         return ret < 0 ? ret : 0;
 768 }
 769
 770 static int bch2_clear_need_discard(struct btree_trans *trans, struct bpos pos,
 771                                    struct bch_dev *ca, bool *discard_done)
 772 {
 773         struct bch_fs *c = trans->c;
 774         struct btree_iter iter;
 775         struct bkey_s_c k;
 776         struct bkey_alloc_unpacked a;
 777         struct printbuf buf = PRINTBUF;
 778         int ret;
 779
 780         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, pos,
 781                              BTREE_ITER_CACHED);
 782         k = bch2_btree_iter_peek_slot(&iter);
 783         ret = bkey_err(k);
 784         if (ret)
 785                 goto out;
 786
 787         a = bch2_alloc_unpack(k);
 788
 789         if (a.need_inc_gen) {
 790                 a.gen++;
 791                 a.need_inc_gen = false;
 792                 goto write;
 793         }
 794
 795         BUG_ON(a.journal_seq > c->journal.flushed_seq_ondisk);
 796
 797         if (bch2_fs_inconsistent_on(!a.need_discard, c,
 798                         "%s\n  incorrectly set in need_discard btree",
 799                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 800                 ret = -EIO;
 801                 goto out;
 802         }
 803
 804         if (!*discard_done && ca->mi.discard && !c->opts.nochanges) {
 805                 /*
 806                  * This works without any other locks because this is the only
 807                  * thread that removes items from the need_discard tree
 808                  */
 809                 bch2_trans_unlock(trans);
 810                 blkdev_issue_discard(ca->disk_sb.bdev,
 811                                      k.k->p.offset * ca->mi.bucket_size,
 812                                      ca->mi.bucket_size,
 813                                      GFP_KERNEL, 0);
 814                 *discard_done = true;
 815
 816                 ret = bch2_trans_relock(trans);
 817                 if (ret)
 818                         goto out;
 819         }
 820
 821         a.need_discard = false;
 822 write:
 823         ret = bch2_alloc_write(trans, &iter, &a, 0);
 824 out:
 825         bch2_trans_iter_exit(trans, &iter);
 826         printbuf_exit(&buf);
 827         return ret;
 828 }
 829
 830 static void bch2_do_discards_work(struct work_struct *work)
 831 {
 832         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
 833         struct bch_dev *ca = NULL;
 834         struct btree_trans trans;
 835         struct btree_iter iter;
 836         struct bkey_s_c k;
 837         int ret;
 838
 839         bch2_trans_init(&trans, c, 0, 0);
 840
 841         for_each_btree_key(&trans, iter, BTREE_ID_need_discard,
 842                            POS_MIN, 0, k, ret) {
 843                 bool discard_done = false;
 844
 845                 if (ca && k.k->p.inode != ca->dev_idx) {
 846                         percpu_ref_put(&ca->io_ref);
 847                         ca = NULL;
 848                 }
 849
 850                 if (!ca) {
 851                         ca = bch_dev_bkey_exists(c, k.k->p.inode);
 852                         if (!percpu_ref_tryget(&ca->io_ref)) {
 853                                 ca = NULL;
 854                                 bch2_btree_iter_set_pos(&iter, POS(k.k->p.inode + 1, 0));
 855                                 continue;
 856                         }
 857                 }
 858
 859                 if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 860                                 c->journal.flushed_seq_ondisk,
 861                                 k.k->p.inode, k.k->p.offset) ||
 862                     bch2_bucket_is_open_safe(c, k.k->p.inode, k.k->p.offset))
 863                         continue;
 864
 865                 ret = __bch2_trans_do(&trans, NULL, NULL,
 866                                       BTREE_INSERT_USE_RESERVE|
 867                                       BTREE_INSERT_NOFAIL,
 868                                 bch2_clear_need_discard(&trans, k.k->p, ca, &discard_done));
 869                 if (ret)
 870                         break;
 871         }
 872         bch2_trans_iter_exit(&trans, &iter);
 873
 874         if (ca)
 875                 percpu_ref_put(&ca->io_ref);
 876
 877         bch2_trans_exit(&trans);
 878         percpu_ref_put(&c->writes);
 879 }
 880
 881 void bch2_do_discards(struct bch_fs *c)
 882 {
 883         if (percpu_ref_tryget(&c->writes) &&
 884             !queue_work(system_long_wq, &c->discard_work))
 885                 percpu_ref_put(&c->writes);
 886 }
 887
 888 static int invalidate_one_bucket(struct btree_trans *trans, struct bch_dev *ca)
 889 {
 890         struct bch_fs *c = trans->c;
 891         struct btree_iter lru_iter, alloc_iter = { NULL };
 892         struct bkey_s_c k;
 893         struct bkey_alloc_unpacked a;
 894         u64 bucket, idx;
 895         int ret;
 896
 897         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
 898                              POS(ca->dev_idx, 0), 0);
 899         k = bch2_btree_iter_peek(&lru_iter);
 900         ret = bkey_err(k);
 901         if (ret)
 902                 goto out;
 903
 904         if (!k.k || k.k->p.inode != ca->dev_idx)
 905                 goto out;
 906
 907         if (bch2_fs_inconsistent_on(k.k->type != KEY_TYPE_lru, c,
 908                                     "non lru key in lru btree"))
 909                 goto out;
 910
 911         idx     = k.k->p.offset;
 912         bucket  = le64_to_cpu(bkey_s_c_to_lru(k).v->idx);
 913
 914         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc,
 915                              POS(ca->dev_idx, bucket),
 916                              BTREE_ITER_CACHED|
 917                              BTREE_ITER_INTENT);
 918         k = bch2_btree_iter_peek_slot(&alloc_iter);
 919         ret = bkey_err(k);
 920         if (ret)
 921                 goto out;
 922
 923         a = bch2_alloc_unpack(k);
 924
 925         if (bch2_fs_inconsistent_on(idx != alloc_lru_idx(a), c,
 926                         "invalidating bucket with wrong lru idx (got %llu should be %llu",
 927                         idx, alloc_lru_idx(a)))
 928                 goto out;
 929
 930         a.gen++;
 931         a.need_inc_gen          = false;
 932         a.data_type             = 0;
 933         a.dirty_sectors         = 0;
 934         a.cached_sectors        = 0;
 935         a.read_time             = atomic64_read(&c->io_clock[READ].now);
 936         a.write_time            = atomic64_read(&c->io_clock[WRITE].now);
 937
 938         ret = bch2_alloc_write(trans, &alloc_iter, &a,
 939                                BTREE_TRIGGER_BUCKET_INVALIDATE);
 940 out:
 941         bch2_trans_iter_exit(trans, &alloc_iter);
 942         bch2_trans_iter_exit(trans, &lru_iter);
 943         return ret;
 944 }
 945
 946 static void bch2_do_invalidates_work(struct work_struct *work)
 947 {
 948         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
 949         struct bch_dev *ca;
 950         struct btree_trans trans;
 951         unsigned i;
 952         int ret = 0;
 953
 954         bch2_trans_init(&trans, c, 0, 0);
 955
 956         for_each_member_device(ca, c, i)
 957                 while (!ret && should_invalidate_buckets(ca))
 958                         ret = __bch2_trans_do(&trans, NULL, NULL,
 959                                               BTREE_INSERT_USE_RESERVE|
 960                                               BTREE_INSERT_NOFAIL,
 961                                         invalidate_one_bucket(&trans, ca));
 962
 963         bch2_trans_exit(&trans);
 964         percpu_ref_put(&c->writes);
 965 }
 966
 967 void bch2_do_invalidates(struct bch_fs *c)
 968 {
 969         if (percpu_ref_tryget(&c->writes))
 970                 queue_work(system_long_wq, &c->invalidate_work);
 971 }
 972
 973 static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca)
 974 {
 975         struct btree_trans trans;
 976         struct btree_iter iter;
 977         struct bkey_s_c k;
 978         struct bkey_alloc_unpacked a;
 979         struct bch_member *m;
 980         int ret;
 981
 982         bch2_trans_init(&trans, c, 0, 0);
 983
 984         for_each_btree_key(&trans, iter, BTREE_ID_alloc,
 985                            POS(ca->dev_idx, ca->mi.first_bucket),
 986                            BTREE_ITER_SLOTS|
 987                            BTREE_ITER_PREFETCH, k, ret) {
 988                 if (iter.pos.offset >= ca->mi.nbuckets)
 989                         break;
 990
 991                 a = bch2_alloc_unpack(k);
 992                 ret = __bch2_trans_do(&trans, NULL, NULL,
 993                                       BTREE_INSERT_LAZY_RW,
 994                                  bch2_bucket_do_index(&trans, k, a, true));
 995                 if (ret)
 996                         break;
 997         }
 998         bch2_trans_iter_exit(&trans, &iter);
 999
1000         bch2_trans_exit(&trans);
1001
1002         if (ret) {
1003                 bch_err(ca, "error initializing free space: %i", ret);
1004                 return ret;
1005         }
1006
1007         mutex_lock(&c->sb_lock);
1008         m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx;
1009         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1010         mutex_unlock(&c->sb_lock);
1011
1012         return ret;
1013 }
1014
1015 int bch2_fs_freespace_init(struct bch_fs *c)
1016 {
1017         struct bch_dev *ca;
1018         unsigned i;
1019         int ret = 0;
1020         bool doing_init = false;
1021
1022         /*
1023          * We can crash during the device add path, so we need to check this on
1024          * every mount:
1025          */
1026
1027         for_each_member_device(ca, c, i) {
1028                 if (ca->mi.freespace_initialized)
1029                         continue;
1030
1031                 if (!doing_init) {
1032                         bch_info(c, "initializing freespace");
1033                         doing_init = true;
1034                 }
1035
1036                 ret = bch2_dev_freespace_init(c, ca);
1037                 if (ret) {
1038                         percpu_ref_put(&ca->ref);
1039                         return ret;
1040                 }
1041         }
1042
1043         if (doing_init) {
1044                 mutex_lock(&c->sb_lock);
1045                 bch2_write_super(c);
1046                 mutex_unlock(&c->sb_lock);
1047
1048                 bch_verbose(c, "done initializing freespace");
1049         }
1050
1051         return ret;
1052 }
1053
1054 /* Bucket IO clocks: */
1055
1056 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
1057                               size_t bucket_nr, int rw)
1058 {
1059         struct bch_fs *c = trans->c;
1060         struct btree_iter iter;
1061         struct bkey_s_c k;
1062         struct bkey_alloc_unpacked u;
1063         u64 *time, now;
1064         int ret = 0;
1065
1066         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS(dev, bucket_nr),
1067                              BTREE_ITER_CACHED|
1068                              BTREE_ITER_INTENT);
1069         k = bch2_btree_iter_peek_slot(&iter);
1070         ret = bkey_err(k);
1071         if (ret)
1072                 goto out;
1073
1074         u = bch2_alloc_unpack(k);
1075
1076         time = rw == READ ? &u.read_time : &u.write_time;
1077         now = atomic64_read(&c->io_clock[rw].now);
1078         if (*time == now)
1079                 goto out;
1080
1081         *time = now;
1082
1083         ret   = bch2_alloc_write(trans, &iter, &u, 0) ?:
1084                 bch2_trans_commit(trans, NULL, NULL, 0);
1085 out:
1086         bch2_trans_iter_exit(trans, &iter);
1087         return ret;
1088 }
1089
1090 /* Startup/shutdown (ro/rw): */
1091
1092 void bch2_recalc_capacity(struct bch_fs *c)
1093 {
1094         struct bch_dev *ca;
1095         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
1096         unsigned bucket_size_max = 0;
1097         unsigned long ra_pages = 0;
1098         unsigned i;
1099
1100         lockdep_assert_held(&c->state_lock);
1101
1102         for_each_online_member(ca, c, i) {
1103                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
1104
1105                 ra_pages += bdi->ra_pages;
1106         }
1107
1108         bch2_set_ra_pages(c, ra_pages);
1109
1110         for_each_rw_member(ca, c, i) {
1111                 u64 dev_reserve = 0;
1112
1113                 /*
1114                  * We need to reserve buckets (from the number
1115                  * of currently available buckets) against
1116                  * foreground writes so that mainly copygc can
1117                  * make forward progress.
1118                  *
1119                  * We need enough to refill the various reserves
1120                  * from scratch - copygc will use its entire
1121                  * reserve all at once, then run against when
1122                  * its reserve is refilled (from the formerly
1123                  * available buckets).
1124                  *
1125                  * This reserve is just used when considering if
1126                  * allocations for foreground writes must wait -
1127                  * not -ENOSPC calculations.
1128                  */
1129
1130                 dev_reserve += ca->nr_btree_reserve * 2;
1131                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
1132
1133                 dev_reserve += 1;       /* btree write point */
1134                 dev_reserve += 1;       /* copygc write point */
1135                 dev_reserve += 1;       /* rebalance write point */
1136
1137                 dev_reserve *= ca->mi.bucket_size;
1138
1139                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
1140                                              ca->mi.first_bucket);
1141
1142                 reserved_sectors += dev_reserve * 2;
1143
1144                 bucket_size_max = max_t(unsigned, bucket_size_max,
1145                                         ca->mi.bucket_size);
1146         }
1147
1148         gc_reserve = c->opts.gc_reserve_bytes
1149                 ? c->opts.gc_reserve_bytes >> 9
1150                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
1151
1152         reserved_sectors = max(gc_reserve, reserved_sectors);
1153
1154         reserved_sectors = min(reserved_sectors, capacity);
1155
1156         c->capacity = capacity - reserved_sectors;
1157
1158         c->bucket_size_max = bucket_size_max;
1159
1160         /* Wake up case someone was waiting for buckets */
1161         closure_wake_up(&c->freelist_wait);
1162 }
1163
1164 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
1165 {
1166         struct open_bucket *ob;
1167         bool ret = false;
1168
1169         for (ob = c->open_buckets;
1170              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1171              ob++) {
1172                 spin_lock(&ob->lock);
1173                 if (ob->valid && !ob->on_partial_list &&
1174                     ob->dev == ca->dev_idx)
1175                         ret = true;
1176                 spin_unlock(&ob->lock);
1177         }
1178
1179         return ret;
1180 }
1181
1182 /* device goes ro: */
1183 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
1184 {
1185         unsigned i;
1186
1187         /* First, remove device from allocation groups: */
1188
1189         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1190                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
1191
1192         /*
1193          * Capacity is calculated based off of devices in allocation groups:
1194          */
1195         bch2_recalc_capacity(c);
1196
1197         /* Next, close write points that point to this device... */
1198         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1199                 bch2_writepoint_stop(c, ca, &c->write_points[i]);
1200
1201         bch2_writepoint_stop(c, ca, &c->copygc_write_point);
1202         bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
1203         bch2_writepoint_stop(c, ca, &c->btree_write_point);
1204
1205         mutex_lock(&c->btree_reserve_cache_lock);
1206         while (c->btree_reserve_cache_nr) {
1207                 struct btree_alloc *a =
1208                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1209
1210                 bch2_open_buckets_put(c, &a->ob);
1211         }
1212         mutex_unlock(&c->btree_reserve_cache_lock);
1213
1214         while (1) {
1215                 struct open_bucket *ob;
1216
1217                 spin_lock(&c->freelist_lock);
1218                 if (!ca->open_buckets_partial_nr) {
1219                         spin_unlock(&c->freelist_lock);
1220                         break;
1221                 }
1222                 ob = c->open_buckets +
1223                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
1224                 ob->on_partial_list = false;
1225                 spin_unlock(&c->freelist_lock);
1226
1227                 bch2_open_bucket_put(c, ob);
1228         }
1229
1230         bch2_ec_stop_dev(c, ca);
1231
1232         /*
1233          * Wake up threads that were blocked on allocation, so they can notice
1234          * the device can no longer be removed and the capacity has changed:
1235          */
1236         closure_wake_up(&c->freelist_wait);
1237
1238         /*
1239          * journal_res_get() can block waiting for free space in the journal -
1240          * it needs to notice there may not be devices to allocate from anymore:
1241          */
1242         wake_up(&c->journal.wait);
1243
1244         /* Now wait for any in flight writes: */
1245
1246         closure_wait_event(&c->open_buckets_wait,
1247                            !bch2_dev_has_open_write_point(c, ca));
1248 }
1249
1250 /* device goes rw: */
1251 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
1252 {
1253         unsigned i;
1254
1255         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1256                 if (ca->mi.data_allowed & (1 << i))
1257                         set_bit(ca->dev_idx, c->rw_devs[i].d);
1258 }
1259
1260 void bch2_fs_allocator_background_init(struct bch_fs *c)
1261 {
1262         spin_lock_init(&c->freelist_lock);
1263         INIT_WORK(&c->discard_work, bch2_do_discards_work);
1264         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
1265 }