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 struct bkey_alloc_unpacked {
  39         u64             journal_seq;
  40         u64             bucket;
  41         u8              dev;
  42         u8              gen;
  43         u8              oldest_gen;
  44         u8              data_type;
  45         bool            need_discard:1;
  46         bool            need_inc_gen:1;
  47 #define x(_name, _bits) u##_bits _name;
  48         BCH_ALLOC_FIELDS_V2()
  49 #undef  x
  50 };
  51
  52 static inline u64 alloc_field_v1_get(const struct bch_alloc *a,
  53                                      const void **p, unsigned field)
  54 {
  55         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  56         u64 v;
  57
  58         if (!(a->fields & (1 << field)))
  59                 return 0;
  60
  61         switch (bytes) {
  62         case 1:
  63                 v = *((const u8 *) *p);
  64                 break;
  65         case 2:
  66                 v = le16_to_cpup(*p);
  67                 break;
  68         case 4:
  69                 v = le32_to_cpup(*p);
  70                 break;
  71         case 8:
  72                 v = le64_to_cpup(*p);
  73                 break;
  74         default:
  75                 BUG();
  76         }
  77
  78         *p += bytes;
  79         return v;
  80 }
  81
  82 static inline void alloc_field_v1_put(struct bkey_i_alloc *a, void **p,
  83                                       unsigned field, u64 v)
  84 {
  85         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  86
  87         if (!v)
  88                 return;
  89
  90         a->v.fields |= 1 << field;
  91
  92         switch (bytes) {
  93         case 1:
  94                 *((u8 *) *p) = v;
  95                 break;
  96         case 2:
  97                 *((__le16 *) *p) = cpu_to_le16(v);
  98                 break;
  99         case 4:
 100                 *((__le32 *) *p) = cpu_to_le32(v);
 101                 break;
 102         case 8:
 103                 *((__le64 *) *p) = cpu_to_le64(v);
 104                 break;
 105         default:
 106                 BUG();
 107         }
 108
 109         *p += bytes;
 110 }
 111
 112 static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out,
 113                                  struct bkey_s_c k)
 114 {
 115         const struct bch_alloc *in = bkey_s_c_to_alloc(k).v;
 116         const void *d = in->data;
 117         unsigned idx = 0;
 118
 119         out->gen = in->gen;
 120
 121 #define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++);
 122         BCH_ALLOC_FIELDS_V1()
 123 #undef  x
 124 }
 125
 126 static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out,
 127                                 struct bkey_s_c k)
 128 {
 129         struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k);
 130         const u8 *in = a.v->data;
 131         const u8 *end = bkey_val_end(a);
 132         unsigned fieldnr = 0;
 133         int ret;
 134         u64 v;
 135
 136         out->gen        = a.v->gen;
 137         out->oldest_gen = a.v->oldest_gen;
 138         out->data_type  = a.v->data_type;
 139
 140 #define x(_name, _bits)                                                 \
 141         if (fieldnr < a.v->nr_fields) {                                 \
 142                 ret = bch2_varint_decode_fast(in, end, &v);             \
 143                 if (ret < 0)                                            \
 144                         return ret;                                     \
 145                 in += ret;                                              \
 146         } else {                                                        \
 147                 v = 0;                                                  \
 148         }                                                               \
 149         out->_name = v;                                                 \
 150         if (v != out->_name)                                            \
 151                 return -1;                                              \
 152         fieldnr++;
 153
 154         BCH_ALLOC_FIELDS_V2()
 155 #undef  x
 156         return 0;
 157 }
 158
 159 static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out,
 160                                 struct bkey_s_c k)
 161 {
 162         struct bkey_s_c_alloc_v3 a = bkey_s_c_to_alloc_v3(k);
 163         const u8 *in = a.v->data;
 164         const u8 *end = bkey_val_end(a);
 165         unsigned fieldnr = 0;
 166         int ret;
 167         u64 v;
 168
 169         out->gen        = a.v->gen;
 170         out->oldest_gen = a.v->oldest_gen;
 171         out->data_type  = a.v->data_type;
 172         out->need_discard = BCH_ALLOC_V3_NEED_DISCARD(a.v);
 173         out->need_inc_gen = BCH_ALLOC_V3_NEED_INC_GEN(a.v);
 174         out->journal_seq = le64_to_cpu(a.v->journal_seq);
 175
 176 #define x(_name, _bits)                                                 \
 177         if (fieldnr < a.v->nr_fields) {                                 \
 178                 ret = bch2_varint_decode_fast(in, end, &v);             \
 179                 if (ret < 0)                                            \
 180                         return ret;                                     \
 181                 in += ret;                                              \
 182         } else {                                                        \
 183                 v = 0;                                                  \
 184         }                                                               \
 185         out->_name = v;                                                 \
 186         if (v != out->_name)                                            \
 187                 return -1;                                              \
 188         fieldnr++;
 189
 190         BCH_ALLOC_FIELDS_V2()
 191 #undef  x
 192         return 0;
 193 }
 194
 195 static struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 196 {
 197         struct bkey_alloc_unpacked ret = {
 198                 .dev    = k.k->p.inode,
 199                 .bucket = k.k->p.offset,
 200                 .gen    = 0,
 201         };
 202
 203         switch (k.k->type) {
 204         case KEY_TYPE_alloc:
 205                 bch2_alloc_unpack_v1(&ret, k);
 206                 break;
 207         case KEY_TYPE_alloc_v2:
 208                 bch2_alloc_unpack_v2(&ret, k);
 209                 break;
 210         case KEY_TYPE_alloc_v3:
 211                 bch2_alloc_unpack_v3(&ret, k);
 212                 break;
 213         }
 214
 215         return ret;
 216 }
 217
 218 void bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out)
 219 {
 220         if (k.k->type == KEY_TYPE_alloc_v4) {
 221                 *out = *bkey_s_c_to_alloc_v4(k).v;
 222         } else {
 223                 struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 224
 225                 *out = (struct bch_alloc_v4) {
 226                         .journal_seq            = u.journal_seq,
 227                         .flags                  = u.need_discard,
 228                         .gen                    = u.gen,
 229                         .oldest_gen             = u.oldest_gen,
 230                         .data_type              = u.data_type,
 231                         .stripe_redundancy      = u.stripe_redundancy,
 232                         .dirty_sectors          = u.dirty_sectors,
 233                         .cached_sectors         = u.cached_sectors,
 234                         .io_time[READ]          = u.read_time,
 235                         .io_time[WRITE]         = u.write_time,
 236                         .stripe                 = u.stripe,
 237                 };
 238         }
 239 }
 240
 241 struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 242 {
 243         struct bkey_i_alloc_v4 *ret;
 244
 245         if (k.k->type == KEY_TYPE_alloc_v4) {
 246                 ret = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 247                 if (!IS_ERR(ret))
 248                         bkey_reassemble(&ret->k_i, k);
 249         } else {
 250                 ret = bch2_trans_kmalloc(trans, sizeof(*ret));
 251                 if (!IS_ERR(ret)) {
 252                         bkey_alloc_v4_init(&ret->k_i);
 253                         ret->k.p = k.k->p;
 254                         bch2_alloc_to_v4(k, &ret->v);
 255                 }
 256         }
 257         return ret;
 258 }
 259
 260 struct bkey_i_alloc_v4 *
 261 bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
 262                               struct bpos pos)
 263 {
 264         struct bkey_s_c k;
 265         struct bkey_i_alloc_v4 *a;
 266         int ret;
 267
 268         bch2_trans_iter_init(trans, iter, BTREE_ID_alloc, pos,
 269                              BTREE_ITER_WITH_UPDATES|
 270                              BTREE_ITER_CACHED|
 271                              BTREE_ITER_INTENT);
 272         k = bch2_btree_iter_peek_slot(iter);
 273         ret = bkey_err(k);
 274         if (ret) {
 275                 bch2_trans_iter_exit(trans, iter);
 276                 return ERR_PTR(ret);
 277         }
 278
 279         a = bch2_alloc_to_v4_mut(trans, k);
 280         if (IS_ERR(a))
 281                 bch2_trans_iter_exit(trans, iter);
 282         return a;
 283 }
 284
 285 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
 286 {
 287         unsigned i, bytes = offsetof(struct bch_alloc, data);
 288
 289         for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++)
 290                 if (a->fields & (1 << i))
 291                         bytes += BCH_ALLOC_V1_FIELD_BYTES[i];
 292
 293         return DIV_ROUND_UP(bytes, sizeof(u64));
 294 }
 295
 296 int bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k,
 297                           int rw, struct printbuf *err)
 298 {
 299         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 300
 301         /* allow for unknown fields */
 302         if (bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v)) {
 303                 pr_buf(err, "incorrect value size (%zu < %u)",
 304                        bkey_val_u64s(a.k), bch_alloc_v1_val_u64s(a.v));
 305                 return -EINVAL;
 306         }
 307
 308         return 0;
 309 }
 310
 311 int bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k,
 312                           int rw, struct printbuf *err)
 313 {
 314         struct bkey_alloc_unpacked u;
 315
 316         if (bch2_alloc_unpack_v2(&u, k)) {
 317                 pr_buf(err, "unpack error");
 318                 return -EINVAL;
 319         }
 320
 321         return 0;
 322 }
 323
 324 int bch2_alloc_v3_invalid(const struct bch_fs *c, struct bkey_s_c k,
 325                           int rw, struct printbuf *err)
 326 {
 327         struct bkey_alloc_unpacked u;
 328
 329         if (bch2_alloc_unpack_v3(&u, k)) {
 330                 pr_buf(err, "unpack error");
 331                 return -EINVAL;
 332         }
 333
 334         return 0;
 335 }
 336
 337 int bch2_alloc_v4_invalid(const struct bch_fs *c, struct bkey_s_c k,
 338                           int rw, struct printbuf *err)
 339 {
 340         struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 341
 342         if (bkey_val_bytes(k.k) != sizeof(struct bch_alloc_v4)) {
 343                 pr_buf(err, "bad val size (%zu != %zu)",
 344                        bkey_val_bytes(k.k), sizeof(struct bch_alloc_v4));
 345                 return -EINVAL;
 346         }
 347
 348         if (rw == WRITE) {
 349                 if (alloc_data_type(*a.v, a.v->data_type) != a.v->data_type) {
 350                         pr_buf(err, "invalid data type (got %u should be %u)",
 351                                a.v->data_type, alloc_data_type(*a.v, a.v->data_type));
 352                         return -EINVAL;
 353                 }
 354
 355                 switch (a.v->data_type) {
 356                 case BCH_DATA_free:
 357                 case BCH_DATA_need_gc_gens:
 358                 case BCH_DATA_need_discard:
 359                         if (a.v->dirty_sectors ||
 360                             a.v->cached_sectors ||
 361                             a.v->stripe) {
 362                                 pr_buf(err, "empty data type free but have data");
 363                                 return -EINVAL;
 364                         }
 365                         break;
 366                 case BCH_DATA_sb:
 367                 case BCH_DATA_journal:
 368                 case BCH_DATA_btree:
 369                 case BCH_DATA_user:
 370                 case BCH_DATA_parity:
 371                         if (!a.v->dirty_sectors) {
 372                                 pr_buf(err, "data_type %s but dirty_sectors==0",
 373                                        bch2_data_types[a.v->data_type]);
 374                                 return -EINVAL;
 375                         }
 376                         break;
 377                 case BCH_DATA_cached:
 378                         if (!a.v->cached_sectors ||
 379                             a.v->dirty_sectors ||
 380                             a.v->stripe) {
 381                                 pr_buf(err, "data type inconsistency");
 382                                 return -EINVAL;
 383                         }
 384
 385                         if (!a.v->io_time[READ]) {
 386                                 pr_buf(err, "cached bucket with read_time == 0");
 387                                 return -EINVAL;
 388                         }
 389                         break;
 390                 case BCH_DATA_stripe:
 391                         if (!a.v->stripe) {
 392                                 pr_buf(err, "data_type %s but stripe==0",
 393                                        bch2_data_types[a.v->data_type]);
 394                                 return -EINVAL;
 395                         }
 396                         break;
 397                 }
 398         }
 399
 400         return 0;
 401 }
 402
 403 void bch2_alloc_v4_swab(struct bkey_s k)
 404 {
 405         struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v;
 406
 407         a->journal_seq          = swab64(a->journal_seq);
 408         a->flags                = swab32(a->flags);
 409         a->dirty_sectors        = swab32(a->dirty_sectors);
 410         a->cached_sectors       = swab32(a->cached_sectors);
 411         a->io_time[0]           = swab64(a->io_time[0]);
 412         a->io_time[1]           = swab64(a->io_time[1]);
 413         a->stripe               = swab32(a->stripe);
 414         a->nr_external_backpointers = swab32(a->nr_external_backpointers);
 415 }
 416
 417 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 418 {
 419         struct bch_alloc_v4 a;
 420
 421         bch2_alloc_to_v4(k, &a);
 422
 423         pr_buf(out, "gen %u oldest_gen %u data_type %s journal_seq %llu need_discard %llu need_inc_gen %llu",
 424                a.gen, a.oldest_gen, bch2_data_types[a.data_type],
 425                a.journal_seq,
 426                BCH_ALLOC_V4_NEED_DISCARD(&a),
 427                BCH_ALLOC_V4_NEED_INC_GEN(&a));
 428         pr_buf(out, " dirty_sectors %u",        a.dirty_sectors);
 429         pr_buf(out, " cached_sectors %u",       a.cached_sectors);
 430         pr_buf(out, " stripe %u",               a.stripe);
 431         pr_buf(out, " stripe_redundancy %u",    a.stripe_redundancy);
 432         pr_buf(out, " read_time %llu",          a.io_time[READ]);
 433         pr_buf(out, " write_time %llu",         a.io_time[WRITE]);
 434 }
 435
 436 int bch2_alloc_read(struct bch_fs *c)
 437 {
 438         struct btree_trans trans;
 439         struct btree_iter iter;
 440         struct bkey_s_c k;
 441         struct bch_alloc_v4 a;
 442         struct bch_dev *ca;
 443         int ret;
 444
 445         bch2_trans_init(&trans, c, 0, 0);
 446
 447         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 448                            BTREE_ITER_PREFETCH, k, ret) {
 449                 /*
 450                  * Not a fsck error because this is checked/repaired by
 451                  * bch2_check_alloc_key() which runs later:
 452                  */
 453                 if (!bch2_dev_bucket_exists(c, k.k->p))
 454                         continue;
 455
 456                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 457                 bch2_alloc_to_v4(k, &a);
 458
 459                 *bucket_gen(ca, k.k->p.offset) = a.gen;
 460         }
 461         bch2_trans_iter_exit(&trans, &iter);
 462
 463         bch2_trans_exit(&trans);
 464
 465         if (ret)
 466                 bch_err(c, "error reading alloc info: %i", ret);
 467
 468         return ret;
 469 }
 470
 471 /* Free space/discard btree: */
 472
 473 static int bch2_bucket_do_index(struct btree_trans *trans,
 474                                 struct bkey_s_c alloc_k,
 475                                 const struct bch_alloc_v4 *a,
 476                                 bool set)
 477 {
 478         struct bch_fs *c = trans->c;
 479         struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 480         struct btree_iter iter;
 481         struct bkey_s_c old;
 482         struct bkey_i *k;
 483         enum btree_id btree;
 484         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 485         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 486         struct printbuf buf = PRINTBUF;
 487         int ret;
 488
 489         if (a->data_type != BCH_DATA_free &&
 490             a->data_type != BCH_DATA_need_discard)
 491                 return 0;
 492
 493         k = bch2_trans_kmalloc(trans, sizeof(*k));
 494         if (IS_ERR(k))
 495                 return PTR_ERR(k);
 496
 497         bkey_init(&k->k);
 498         k->k.type = new_type;
 499
 500         switch (a->data_type) {
 501         case BCH_DATA_free:
 502                 btree = BTREE_ID_freespace;
 503                 k->k.p = alloc_freespace_pos(alloc_k.k->p, *a);
 504                 bch2_key_resize(&k->k, 1);
 505                 break;
 506         case BCH_DATA_need_discard:
 507                 btree = BTREE_ID_need_discard;
 508                 k->k.p = alloc_k.k->p;
 509                 break;
 510         default:
 511                 return 0;
 512         }
 513
 514         bch2_trans_iter_init(trans, &iter, btree,
 515                              bkey_start_pos(&k->k),
 516                              BTREE_ITER_INTENT);
 517         old = bch2_btree_iter_peek_slot(&iter);
 518         ret = bkey_err(old);
 519         if (ret)
 520                 goto err;
 521
 522         if (ca->mi.freespace_initialized &&
 523             bch2_trans_inconsistent_on(old.k->type != old_type, trans,
 524                         "incorrect key when %s %s btree (got %s should be %s)\n"
 525                         "  for %s",
 526                         set ? "setting" : "clearing",
 527                         bch2_btree_ids[btree],
 528                         bch2_bkey_types[old.k->type],
 529                         bch2_bkey_types[old_type],
 530                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 531                 ret = -EIO;
 532                 goto err;
 533         }
 534
 535         ret = bch2_trans_update(trans, &iter, k, 0);
 536 err:
 537         bch2_trans_iter_exit(trans, &iter);
 538         printbuf_exit(&buf);
 539         return ret;
 540 }
 541
 542 int bch2_trans_mark_alloc(struct btree_trans *trans,
 543                           struct bkey_s_c old, struct bkey_i *new,
 544                           unsigned flags)
 545 {
 546         struct bch_fs *c = trans->c;
 547         struct bch_alloc_v4 old_a, *new_a;
 548         int ret = 0;
 549
 550         /*
 551          * Deletion only happens in the device removal path, with
 552          * BTREE_TRIGGER_NORUN:
 553          */
 554         BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
 555
 556         bch2_alloc_to_v4(old, &old_a);
 557         new_a = &bkey_i_to_alloc_v4(new)->v;
 558
 559         new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
 560
 561         if (new_a->dirty_sectors > old_a.dirty_sectors ||
 562             new_a->cached_sectors > old_a.cached_sectors) {
 563                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 564                 new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 565                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
 566                 SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
 567         }
 568
 569         if (data_type_is_empty(new_a->data_type) &&
 570             BCH_ALLOC_V4_NEED_INC_GEN(new_a) &&
 571             !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
 572                 new_a->gen++;
 573                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
 574         }
 575
 576         if (old_a.data_type != new_a->data_type ||
 577             (new_a->data_type == BCH_DATA_free &&
 578              alloc_freespace_genbits(old_a) != alloc_freespace_genbits(*new_a))) {
 579                 ret =   bch2_bucket_do_index(trans, old, &old_a, false) ?:
 580                         bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_a, true);
 581                 if (ret)
 582                         return ret;
 583         }
 584
 585         if (new_a->data_type == BCH_DATA_cached &&
 586             !new_a->io_time[READ])
 587                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 588
 589         if ((old_a.data_type == BCH_DATA_cached) !=
 590             (new_a->data_type == BCH_DATA_cached)) {
 591                 u64 old_lru = alloc_lru_idx(old_a);
 592                 u64 new_lru = alloc_lru_idx(*new_a);
 593
 594                 ret = bch2_lru_change(trans, new->k.p.inode, new->k.p.offset,
 595                                       old_lru, &new_lru);
 596                 if (ret)
 597                         return ret;
 598
 599                 if (new_a->data_type == BCH_DATA_cached)
 600                         new_a->io_time[READ] = new_lru;
 601         }
 602
 603         return 0;
 604 }
 605
 606 static int bch2_check_alloc_key(struct btree_trans *trans,
 607                                 struct btree_iter *alloc_iter)
 608 {
 609         struct bch_fs *c = trans->c;
 610         struct bch_dev *ca;
 611         struct btree_iter discard_iter, freespace_iter;
 612         struct bch_alloc_v4 a;
 613         unsigned discard_key_type, freespace_key_type;
 614         struct bkey_s_c alloc_k, k;
 615         struct printbuf buf = PRINTBUF;
 616         struct printbuf buf2 = PRINTBUF;
 617         int ret;
 618
 619         alloc_k = bch2_btree_iter_peek(alloc_iter);
 620         if (!alloc_k.k)
 621                 return 0;
 622
 623         ret = bkey_err(alloc_k);
 624         if (ret)
 625                 return ret;
 626
 627         if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
 628                         "alloc key for invalid device:bucket %llu:%llu",
 629                         alloc_k.k->p.inode, alloc_k.k->p.offset))
 630                 return bch2_btree_delete_at(trans, alloc_iter, 0);
 631
 632         ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 633         if (!ca->mi.freespace_initialized)
 634                 return 0;
 635
 636         bch2_alloc_to_v4(alloc_k, &a);
 637
 638         discard_key_type = a.data_type == BCH_DATA_need_discard
 639                 ? KEY_TYPE_set : 0;
 640         freespace_key_type = a.data_type == BCH_DATA_free
 641                 ? KEY_TYPE_set : 0;
 642
 643         bch2_trans_iter_init(trans, &discard_iter, BTREE_ID_need_discard,
 644                              alloc_k.k->p, 0);
 645         bch2_trans_iter_init(trans, &freespace_iter, BTREE_ID_freespace,
 646                              alloc_freespace_pos(alloc_k.k->p, a), 0);
 647
 648         k = bch2_btree_iter_peek_slot(&discard_iter);
 649         ret = bkey_err(k);
 650         if (ret)
 651                 goto err;
 652
 653         if (fsck_err_on(k.k->type != discard_key_type, c,
 654                         "incorrect key in need_discard btree (got %s should be %s)\n"
 655                         "  %s",
 656                         bch2_bkey_types[k.k->type],
 657                         bch2_bkey_types[discard_key_type],
 658                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 659                 struct bkey_i *update =
 660                         bch2_trans_kmalloc(trans, sizeof(*update));
 661
 662                 ret = PTR_ERR_OR_ZERO(update);
 663                 if (ret)
 664                         goto err;
 665
 666                 bkey_init(&update->k);
 667                 update->k.type  = discard_key_type;
 668                 update->k.p     = discard_iter.pos;
 669
 670                 ret = bch2_trans_update(trans, &discard_iter, update, 0);
 671                 if (ret)
 672                         goto err;
 673         }
 674
 675         k = bch2_btree_iter_peek_slot(&freespace_iter);
 676         ret = bkey_err(k);
 677         if (ret)
 678                 goto err;
 679
 680         if (fsck_err_on(k.k->type != freespace_key_type, c,
 681                         "incorrect key in freespace btree (got %s should be %s)\n"
 682                         "  %s",
 683                         bch2_bkey_types[k.k->type],
 684                         bch2_bkey_types[freespace_key_type],
 685                         (printbuf_reset(&buf),
 686                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 687                 struct bkey_i *update =
 688                         bch2_trans_kmalloc(trans, sizeof(*update));
 689
 690                 ret = PTR_ERR_OR_ZERO(update);
 691                 if (ret)
 692                         goto err;
 693
 694                 bkey_init(&update->k);
 695                 update->k.type  = freespace_key_type;
 696                 update->k.p     = freespace_iter.pos;
 697                 bch2_key_resize(&update->k, 1);
 698
 699                 ret = bch2_trans_update(trans, &freespace_iter, update, 0);
 700                 if (ret)
 701                         goto err;
 702         }
 703 err:
 704 fsck_err:
 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 int bch2_check_discard_freespace_key(struct btree_trans *trans,
 713                                             struct btree_iter *iter)
 714 {
 715         struct bch_fs *c = trans->c;
 716         struct btree_iter alloc_iter;
 717         struct bkey_s_c k, freespace_k;
 718         struct bch_alloc_v4 a;
 719         u64 genbits;
 720         struct bpos pos;
 721         enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
 722                 ? BCH_DATA_need_discard
 723                 : BCH_DATA_free;
 724         struct printbuf buf = PRINTBUF;
 725         int ret;
 726
 727         freespace_k = bch2_btree_iter_peek(iter);
 728         if (!freespace_k.k)
 729                 return 1;
 730
 731         ret = bkey_err(freespace_k);
 732         if (ret)
 733                 return ret;
 734
 735         pos = iter->pos;
 736         pos.offset &= ~(~0ULL << 56);
 737         genbits = iter->pos.offset & (~0ULL << 56);
 738
 739         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
 740
 741         if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
 742                         "entry in %s btree for nonexistant dev:bucket %llu:%llu",
 743                         bch2_btree_ids[iter->btree_id], pos.inode, pos.offset))
 744                 goto delete;
 745
 746         k = bch2_btree_iter_peek_slot(&alloc_iter);
 747         ret = bkey_err(k);
 748         if (ret)
 749                 goto err;
 750
 751         bch2_alloc_to_v4(k, &a);
 752
 753         if (fsck_err_on(a.data_type != state ||
 754                         (state == BCH_DATA_free &&
 755                          genbits != alloc_freespace_genbits(a)), c,
 756                         "%s\n  incorrectly set in %s index (free %u, genbits %llu should be %llu)",
 757                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
 758                         bch2_btree_ids[iter->btree_id],
 759                         a.data_type == state,
 760                         genbits >> 56, alloc_freespace_genbits(a) >> 56))
 761                 goto delete;
 762 out:
 763 err:
 764 fsck_err:
 765         bch2_trans_iter_exit(trans, &alloc_iter);
 766         printbuf_exit(&buf);
 767         return ret;
 768 delete:
 769         ret = bch2_btree_delete_extent_at(trans, iter,
 770                         iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0);
 771         goto out;
 772 }
 773
 774 int bch2_check_alloc_info(struct bch_fs *c)
 775 {
 776         struct btree_trans trans;
 777         struct btree_iter iter;
 778         struct bkey_s_c k;
 779         int ret = 0;
 780
 781         bch2_trans_init(&trans, c, 0, 0);
 782
 783         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 784                            BTREE_ITER_PREFETCH, k, ret) {
 785                 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
 786                         bch2_check_alloc_key(&trans, &iter));
 787                 if (ret)
 788                         break;
 789         }
 790         bch2_trans_iter_exit(&trans, &iter);
 791
 792         if (ret)
 793                 goto err;
 794
 795         bch2_trans_iter_init(&trans, &iter, BTREE_ID_need_discard, POS_MIN,
 796                              BTREE_ITER_PREFETCH);
 797         while (1) {
 798                 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
 799                         bch2_check_discard_freespace_key(&trans, &iter));
 800                 if (ret)
 801                         break;
 802
 803                 bch2_btree_iter_set_pos(&iter, bpos_nosnap_successor(iter.pos));
 804         }
 805         bch2_trans_iter_exit(&trans, &iter);
 806
 807         if (ret)
 808                 goto err;
 809
 810         bch2_trans_iter_init(&trans, &iter, BTREE_ID_freespace, POS_MIN,
 811                              BTREE_ITER_PREFETCH);
 812         while (1) {
 813                 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
 814                         bch2_check_discard_freespace_key(&trans, &iter));
 815                 if (ret)
 816                         break;
 817
 818                 bch2_btree_iter_set_pos(&iter, bpos_nosnap_successor(iter.pos));
 819         }
 820         bch2_trans_iter_exit(&trans, &iter);
 821 err:
 822         bch2_trans_exit(&trans);
 823         return ret < 0 ? ret : 0;
 824 }
 825
 826 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
 827                                        struct btree_iter *alloc_iter)
 828 {
 829         struct bch_fs *c = trans->c;
 830         struct btree_iter lru_iter;
 831         struct bch_alloc_v4 a;
 832         struct bkey_s_c alloc_k, k;
 833         struct printbuf buf = PRINTBUF;
 834         struct printbuf buf2 = PRINTBUF;
 835         int ret;
 836
 837         alloc_k = bch2_btree_iter_peek(alloc_iter);
 838         if (!alloc_k.k)
 839                 return 0;
 840
 841         ret = bkey_err(alloc_k);
 842         if (ret)
 843                 return ret;
 844
 845         bch2_alloc_to_v4(alloc_k, &a);
 846
 847         if (a.data_type != BCH_DATA_cached)
 848                 return 0;
 849
 850         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
 851                              POS(alloc_k.k->p.inode, a.io_time[READ]), 0);
 852
 853         k = bch2_btree_iter_peek_slot(&lru_iter);
 854         ret = bkey_err(k);
 855         if (ret)
 856                 goto err;
 857
 858         if (fsck_err_on(!a.io_time[READ], c,
 859                         "cached bucket with read_time 0\n"
 860                         "  %s",
 861                 (printbuf_reset(&buf),
 862                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) ||
 863             fsck_err_on(k.k->type != KEY_TYPE_lru ||
 864                         le64_to_cpu(bkey_s_c_to_lru(k).v->idx) != alloc_k.k->p.offset, c,
 865                         "incorrect/missing lru entry\n"
 866                         "  %s\n"
 867                         "  %s",
 868                         (printbuf_reset(&buf),
 869                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
 870                         (bch2_bkey_val_to_text(&buf2, c, k), buf2.buf))) {
 871                 u64 read_time = a.io_time[READ];
 872
 873                 if (!a.io_time[READ])
 874                         a.io_time[READ] = atomic64_read(&c->io_clock[READ].now);
 875
 876                 ret = bch2_lru_set(trans,
 877                                    alloc_k.k->p.inode,
 878                                    alloc_k.k->p.offset,
 879                                    &a.io_time[READ]);
 880                 if (ret)
 881                         goto err;
 882
 883                 if (a.io_time[READ] != read_time) {
 884                         struct bkey_i_alloc_v4 *a_mut =
 885                                 bch2_alloc_to_v4_mut(trans, alloc_k);
 886                         ret = PTR_ERR_OR_ZERO(a_mut);
 887                         if (ret)
 888                                 goto err;
 889
 890                         a_mut->v.io_time[READ] = a.io_time[READ];
 891                         ret = bch2_trans_update(trans, alloc_iter,
 892                                                 &a_mut->k_i, BTREE_TRIGGER_NORUN);
 893                         if (ret)
 894                                 goto err;
 895                 }
 896         }
 897 err:
 898 fsck_err:
 899         bch2_trans_iter_exit(trans, &lru_iter);
 900         printbuf_exit(&buf2);
 901         printbuf_exit(&buf);
 902         return ret;
 903 }
 904
 905 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
 906 {
 907         struct btree_trans trans;
 908         struct btree_iter iter;
 909         struct bkey_s_c k;
 910         int ret = 0;
 911
 912         bch2_trans_init(&trans, c, 0, 0);
 913
 914         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 915                            BTREE_ITER_PREFETCH, k, ret) {
 916                 ret = __bch2_trans_do(&trans, NULL, NULL,
 917                                       BTREE_INSERT_NOFAIL|
 918                                       BTREE_INSERT_LAZY_RW,
 919                         bch2_check_alloc_to_lru_ref(&trans, &iter));
 920                 if (ret)
 921                         break;
 922         }
 923         bch2_trans_iter_exit(&trans, &iter);
 924
 925         bch2_trans_exit(&trans);
 926         return ret < 0 ? ret : 0;
 927 }
 928
 929 static int bch2_clear_need_discard(struct btree_trans *trans, struct bpos pos,
 930                                    struct bch_dev *ca, bool *discard_done)
 931 {
 932         struct bch_fs *c = trans->c;
 933         struct btree_iter iter;
 934         struct bkey_s_c k;
 935         struct bkey_i_alloc_v4 *a;
 936         struct printbuf buf = PRINTBUF;
 937         int ret;
 938
 939         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, pos,
 940                              BTREE_ITER_CACHED);
 941         k = bch2_btree_iter_peek_slot(&iter);
 942         ret = bkey_err(k);
 943         if (ret)
 944                 goto out;
 945
 946         a = bch2_alloc_to_v4_mut(trans, k);
 947         ret = PTR_ERR_OR_ZERO(a);
 948         if (ret)
 949                 goto out;
 950
 951         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
 952                 a->v.gen++;
 953                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
 954                 goto write;
 955         }
 956
 957         if (bch2_trans_inconsistent_on(a->v.journal_seq > c->journal.flushed_seq_ondisk, trans,
 958                         "clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
 959                         "%s",
 960                         a->v.journal_seq,
 961                         c->journal.flushed_seq_ondisk,
 962                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 963                 ret = -EIO;
 964                 goto out;
 965         }
 966
 967         if (bch2_trans_inconsistent_on(a->v.data_type != BCH_DATA_need_discard, trans,
 968                         "bucket incorrectly set in need_discard btree\n"
 969                         "%s",
 970                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 971                 ret = -EIO;
 972                 goto out;
 973         }
 974
 975         if (!*discard_done && ca->mi.discard && !c->opts.nochanges) {
 976                 /*
 977                  * This works without any other locks because this is the only
 978                  * thread that removes items from the need_discard tree
 979                  */
 980                 bch2_trans_unlock(trans);
 981                 blkdev_issue_discard(ca->disk_sb.bdev,
 982                                      k.k->p.offset * ca->mi.bucket_size,
 983                                      ca->mi.bucket_size,
 984                                      GFP_KERNEL, 0);
 985                 *discard_done = true;
 986
 987                 ret = bch2_trans_relock(trans) ? 0 : -EINTR;
 988                 if (ret)
 989                         goto out;
 990         }
 991
 992         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
 993         a->v.data_type = alloc_data_type(a->v, a->v.data_type);
 994 write:
 995         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
 996 out:
 997         bch2_trans_iter_exit(trans, &iter);
 998         printbuf_exit(&buf);
 999         return ret;
1000 }
1001
1002 static void bch2_do_discards_work(struct work_struct *work)
1003 {
1004         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
1005         struct bch_dev *ca = NULL;
1006         struct btree_trans trans;
1007         struct btree_iter iter;
1008         struct bkey_s_c k;
1009         u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
1010         int ret;
1011
1012         bch2_trans_init(&trans, c, 0, 0);
1013
1014         for_each_btree_key(&trans, iter, BTREE_ID_need_discard,
1015                            POS_MIN, 0, k, ret) {
1016                 bool discard_done = false;
1017
1018                 if (ca && k.k->p.inode != ca->dev_idx) {
1019                         percpu_ref_put(&ca->io_ref);
1020                         ca = NULL;
1021                 }
1022
1023                 if (!ca) {
1024                         ca = bch_dev_bkey_exists(c, k.k->p.inode);
1025                         if (!percpu_ref_tryget(&ca->io_ref)) {
1026                                 ca = NULL;
1027                                 bch2_btree_iter_set_pos(&iter, POS(k.k->p.inode + 1, 0));
1028                                 continue;
1029                         }
1030                 }
1031
1032                 seen++;
1033
1034                 if (bch2_bucket_is_open_safe(c, k.k->p.inode, k.k->p.offset)) {
1035                         open++;
1036                         continue;
1037                 }
1038
1039                 if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
1040                                 c->journal.flushed_seq_ondisk,
1041                                 k.k->p.inode, k.k->p.offset)) {
1042                         need_journal_commit++;
1043                         continue;
1044                 }
1045
1046                 ret = __bch2_trans_do(&trans, NULL, NULL,
1047                                       BTREE_INSERT_USE_RESERVE|
1048                                       BTREE_INSERT_NOFAIL,
1049                                 bch2_clear_need_discard(&trans, k.k->p, ca, &discard_done));
1050                 if (ret)
1051                         break;
1052
1053                 discarded++;
1054         }
1055         bch2_trans_iter_exit(&trans, &iter);
1056
1057         if (ca)
1058                 percpu_ref_put(&ca->io_ref);
1059
1060         bch2_trans_exit(&trans);
1061
1062         if (need_journal_commit * 2 > seen)
1063                 bch2_journal_flush_async(&c->journal, NULL);
1064
1065         percpu_ref_put(&c->writes);
1066
1067         trace_do_discards(c, seen, open, need_journal_commit, discarded, ret);
1068 }
1069
1070 void bch2_do_discards(struct bch_fs *c)
1071 {
1072         if (percpu_ref_tryget(&c->writes) &&
1073             !queue_work(system_long_wq, &c->discard_work))
1074                 percpu_ref_put(&c->writes);
1075 }
1076
1077 static int invalidate_one_bucket(struct btree_trans *trans, struct bch_dev *ca)
1078 {
1079         struct bch_fs *c = trans->c;
1080         struct btree_iter lru_iter, alloc_iter = { NULL };
1081         struct bkey_s_c k;
1082         struct bkey_i_alloc_v4 *a;
1083         u64 bucket, idx;
1084         struct printbuf buf = PRINTBUF;
1085         int ret;
1086
1087         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
1088                              POS(ca->dev_idx, 0), 0);
1089         k = bch2_btree_iter_peek(&lru_iter);
1090         ret = bkey_err(k);
1091         if (ret)
1092                 goto out;
1093
1094         if (!k.k || k.k->p.inode != ca->dev_idx)
1095                 goto out;
1096
1097         if (bch2_trans_inconsistent_on(k.k->type != KEY_TYPE_lru, trans,
1098                                        "non lru key in lru btree"))
1099                 goto out;
1100
1101         idx     = k.k->p.offset;
1102         bucket  = le64_to_cpu(bkey_s_c_to_lru(k).v->idx);
1103
1104         a = bch2_trans_start_alloc_update(trans, &alloc_iter,
1105                                           POS(ca->dev_idx, bucket));
1106         ret = PTR_ERR_OR_ZERO(a);
1107         if (ret)
1108                 goto out;
1109
1110         if (idx != alloc_lru_idx(a->v)) {
1111                 pr_buf(&buf, "alloc key does not point back to lru entry when invalidating bucket:\n  ");
1112
1113                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
1114                 pr_buf(&buf, "\n  ");
1115                 bch2_bkey_val_to_text(&buf, c, k);
1116                 bch2_trans_inconsistent(trans, "%s", buf.buf);
1117                 ret = -EINVAL;
1118                 goto out;
1119         }
1120
1121         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1122         a->v.gen++;
1123         a->v.data_type          = 0;
1124         a->v.dirty_sectors      = 0;
1125         a->v.cached_sectors     = 0;
1126         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1127         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1128
1129         ret = bch2_trans_update(trans, &alloc_iter, &a->k_i,
1130                                 BTREE_TRIGGER_BUCKET_INVALIDATE);
1131 out:
1132         bch2_trans_iter_exit(trans, &alloc_iter);
1133         bch2_trans_iter_exit(trans, &lru_iter);
1134         printbuf_exit(&buf);
1135         return ret;
1136 }
1137
1138 static void bch2_do_invalidates_work(struct work_struct *work)
1139 {
1140         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1141         struct bch_dev *ca;
1142         struct btree_trans trans;
1143         unsigned i;
1144         int ret = 0;
1145
1146         bch2_trans_init(&trans, c, 0, 0);
1147
1148         for_each_member_device(ca, c, i) {
1149                 s64 nr_to_invalidate =
1150                         should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
1151
1152                 while (!ret && nr_to_invalidate-- >= 0)
1153                         ret = __bch2_trans_do(&trans, NULL, NULL,
1154                                               BTREE_INSERT_USE_RESERVE|
1155                                               BTREE_INSERT_NOFAIL,
1156                                         invalidate_one_bucket(&trans, ca));
1157         }
1158
1159         bch2_trans_exit(&trans);
1160         percpu_ref_put(&c->writes);
1161 }
1162
1163 void bch2_do_invalidates(struct bch_fs *c)
1164 {
1165         if (percpu_ref_tryget(&c->writes))
1166                 queue_work(system_long_wq, &c->invalidate_work);
1167 }
1168
1169 static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca)
1170 {
1171         struct btree_trans trans;
1172         struct btree_iter iter;
1173         struct bkey_s_c k;
1174         struct bch_alloc_v4 a;
1175         struct bch_member *m;
1176         int ret;
1177
1178         bch2_trans_init(&trans, c, 0, 0);
1179
1180         for_each_btree_key(&trans, iter, BTREE_ID_alloc,
1181                            POS(ca->dev_idx, ca->mi.first_bucket),
1182                            BTREE_ITER_SLOTS|
1183                            BTREE_ITER_PREFETCH, k, ret) {
1184                 if (iter.pos.offset >= ca->mi.nbuckets)
1185                         break;
1186
1187                 bch2_alloc_to_v4(k, &a);
1188                 ret = __bch2_trans_do(&trans, NULL, NULL,
1189                                       BTREE_INSERT_LAZY_RW,
1190                                  bch2_bucket_do_index(&trans, k, &a, true));
1191                 if (ret)
1192                         break;
1193         }
1194         bch2_trans_iter_exit(&trans, &iter);
1195
1196         bch2_trans_exit(&trans);
1197
1198         if (ret) {
1199                 bch_err(ca, "error initializing free space: %i", ret);
1200                 return ret;
1201         }
1202
1203         mutex_lock(&c->sb_lock);
1204         m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx;
1205         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1206         mutex_unlock(&c->sb_lock);
1207
1208         return ret;
1209 }
1210
1211 int bch2_fs_freespace_init(struct bch_fs *c)
1212 {
1213         struct bch_dev *ca;
1214         unsigned i;
1215         int ret = 0;
1216         bool doing_init = false;
1217
1218         /*
1219          * We can crash during the device add path, so we need to check this on
1220          * every mount:
1221          */
1222
1223         for_each_member_device(ca, c, i) {
1224                 if (ca->mi.freespace_initialized)
1225                         continue;
1226
1227                 if (!doing_init) {
1228                         bch_info(c, "initializing freespace");
1229                         doing_init = true;
1230                 }
1231
1232                 ret = bch2_dev_freespace_init(c, ca);
1233                 if (ret) {
1234                         percpu_ref_put(&ca->ref);
1235                         return ret;
1236                 }
1237         }
1238
1239         if (doing_init) {
1240                 mutex_lock(&c->sb_lock);
1241                 bch2_write_super(c);
1242                 mutex_unlock(&c->sb_lock);
1243
1244                 bch_verbose(c, "done initializing freespace");
1245         }
1246
1247         return ret;
1248 }
1249
1250 /* Bucket IO clocks: */
1251
1252 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
1253                               size_t bucket_nr, int rw)
1254 {
1255         struct bch_fs *c = trans->c;
1256         struct btree_iter iter;
1257         struct bkey_i_alloc_v4 *a;
1258         u64 now;
1259         int ret = 0;
1260
1261         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
1262         ret = PTR_ERR_OR_ZERO(a);
1263         if (ret)
1264                 return ret;
1265
1266         now = atomic64_read(&c->io_clock[rw].now);
1267         if (a->v.io_time[rw] == now)
1268                 goto out;
1269
1270         a->v.io_time[rw] = now;
1271
1272         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1273                 bch2_trans_commit(trans, NULL, NULL, 0);
1274 out:
1275         bch2_trans_iter_exit(trans, &iter);
1276         return ret;
1277 }
1278
1279 /* Startup/shutdown (ro/rw): */
1280
1281 void bch2_recalc_capacity(struct bch_fs *c)
1282 {
1283         struct bch_dev *ca;
1284         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
1285         unsigned bucket_size_max = 0;
1286         unsigned long ra_pages = 0;
1287         unsigned i;
1288
1289         lockdep_assert_held(&c->state_lock);
1290
1291         for_each_online_member(ca, c, i) {
1292                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
1293
1294                 ra_pages += bdi->ra_pages;
1295         }
1296
1297         bch2_set_ra_pages(c, ra_pages);
1298
1299         for_each_rw_member(ca, c, i) {
1300                 u64 dev_reserve = 0;
1301
1302                 /*
1303                  * We need to reserve buckets (from the number
1304                  * of currently available buckets) against
1305                  * foreground writes so that mainly copygc can
1306                  * make forward progress.
1307                  *
1308                  * We need enough to refill the various reserves
1309                  * from scratch - copygc will use its entire
1310                  * reserve all at once, then run against when
1311                  * its reserve is refilled (from the formerly
1312                  * available buckets).
1313                  *
1314                  * This reserve is just used when considering if
1315                  * allocations for foreground writes must wait -
1316                  * not -ENOSPC calculations.
1317                  */
1318
1319                 dev_reserve += ca->nr_btree_reserve * 2;
1320                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
1321
1322                 dev_reserve += 1;       /* btree write point */
1323                 dev_reserve += 1;       /* copygc write point */
1324                 dev_reserve += 1;       /* rebalance write point */
1325
1326                 dev_reserve *= ca->mi.bucket_size;
1327
1328                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
1329                                              ca->mi.first_bucket);
1330
1331                 reserved_sectors += dev_reserve * 2;
1332
1333                 bucket_size_max = max_t(unsigned, bucket_size_max,
1334                                         ca->mi.bucket_size);
1335         }
1336
1337         gc_reserve = c->opts.gc_reserve_bytes
1338                 ? c->opts.gc_reserve_bytes >> 9
1339                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
1340
1341         reserved_sectors = max(gc_reserve, reserved_sectors);
1342
1343         reserved_sectors = min(reserved_sectors, capacity);
1344
1345         c->capacity = capacity - reserved_sectors;
1346
1347         c->bucket_size_max = bucket_size_max;
1348
1349         /* Wake up case someone was waiting for buckets */
1350         closure_wake_up(&c->freelist_wait);
1351 }
1352
1353 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
1354 {
1355         struct open_bucket *ob;
1356         bool ret = false;
1357
1358         for (ob = c->open_buckets;
1359              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1360              ob++) {
1361                 spin_lock(&ob->lock);
1362                 if (ob->valid && !ob->on_partial_list &&
1363                     ob->dev == ca->dev_idx)
1364                         ret = true;
1365                 spin_unlock(&ob->lock);
1366         }
1367
1368         return ret;
1369 }
1370
1371 /* device goes ro: */
1372 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
1373 {
1374         unsigned i;
1375
1376         /* First, remove device from allocation groups: */
1377
1378         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1379                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
1380
1381         /*
1382          * Capacity is calculated based off of devices in allocation groups:
1383          */
1384         bch2_recalc_capacity(c);
1385
1386         /* Next, close write points that point to this device... */
1387         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1388                 bch2_writepoint_stop(c, ca, &c->write_points[i]);
1389
1390         bch2_writepoint_stop(c, ca, &c->copygc_write_point);
1391         bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
1392         bch2_writepoint_stop(c, ca, &c->btree_write_point);
1393
1394         mutex_lock(&c->btree_reserve_cache_lock);
1395         while (c->btree_reserve_cache_nr) {
1396                 struct btree_alloc *a =
1397                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1398
1399                 bch2_open_buckets_put(c, &a->ob);
1400         }
1401         mutex_unlock(&c->btree_reserve_cache_lock);
1402
1403         while (1) {
1404                 struct open_bucket *ob;
1405
1406                 spin_lock(&c->freelist_lock);
1407                 if (!ca->open_buckets_partial_nr) {
1408                         spin_unlock(&c->freelist_lock);
1409                         break;
1410                 }
1411                 ob = c->open_buckets +
1412                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
1413                 ob->on_partial_list = false;
1414                 spin_unlock(&c->freelist_lock);
1415
1416                 bch2_open_bucket_put(c, ob);
1417         }
1418
1419         bch2_ec_stop_dev(c, ca);
1420
1421         /*
1422          * Wake up threads that were blocked on allocation, so they can notice
1423          * the device can no longer be removed and the capacity has changed:
1424          */
1425         closure_wake_up(&c->freelist_wait);
1426
1427         /*
1428          * journal_res_get() can block waiting for free space in the journal -
1429          * it needs to notice there may not be devices to allocate from anymore:
1430          */
1431         wake_up(&c->journal.wait);
1432
1433         /* Now wait for any in flight writes: */
1434
1435         closure_wait_event(&c->open_buckets_wait,
1436                            !bch2_dev_has_open_write_point(c, ca));
1437 }
1438
1439 /* device goes rw: */
1440 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
1441 {
1442         unsigned i;
1443
1444         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1445                 if (ca->mi.data_allowed & (1 << i))
1446                         set_bit(ca->dev_idx, c->rw_devs[i].d);
1447 }
1448
1449 void bch2_fs_allocator_background_init(struct bch_fs *c)
1450 {
1451         spin_lock_init(&c->freelist_lock);
1452         INIT_WORK(&c->discard_work, bch2_do_discards_work);
1453         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
1454 }