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                             test_bit(BCH_FS_CHECK_ALLOC_TO_LRU_REFS_DONE, &c->flags)) {
 387                                 pr_buf(err, "cached bucket with read_time == 0");
 388                                 return -EINVAL;
 389                         }
 390                         break;
 391                 case BCH_DATA_stripe:
 392                         if (!a.v->stripe) {
 393                                 pr_buf(err, "data_type %s but stripe==0",
 394                                        bch2_data_types[a.v->data_type]);
 395                                 return -EINVAL;
 396                         }
 397                         break;
 398                 }
 399         }
 400
 401         return 0;
 402 }
 403
 404 void bch2_alloc_v4_swab(struct bkey_s k)
 405 {
 406         struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v;
 407
 408         a->journal_seq          = swab64(a->journal_seq);
 409         a->flags                = swab32(a->flags);
 410         a->dirty_sectors        = swab32(a->dirty_sectors);
 411         a->cached_sectors       = swab32(a->cached_sectors);
 412         a->io_time[0]           = swab64(a->io_time[0]);
 413         a->io_time[1]           = swab64(a->io_time[1]);
 414         a->stripe               = swab32(a->stripe);
 415         a->nr_external_backpointers = swab32(a->nr_external_backpointers);
 416 }
 417
 418 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 419 {
 420         struct bch_alloc_v4 a;
 421
 422         bch2_alloc_to_v4(k, &a);
 423
 424         pr_buf(out, "gen %u oldest_gen %u data_type %s journal_seq %llu need_discard %llu need_inc_gen %llu",
 425                a.gen, a.oldest_gen, bch2_data_types[a.data_type],
 426                a.journal_seq,
 427                BCH_ALLOC_V4_NEED_DISCARD(&a),
 428                BCH_ALLOC_V4_NEED_INC_GEN(&a));
 429         pr_buf(out, " dirty_sectors %u",        a.dirty_sectors);
 430         pr_buf(out, " cached_sectors %u",       a.cached_sectors);
 431         pr_buf(out, " stripe %u",               a.stripe);
 432         pr_buf(out, " stripe_redundancy %u",    a.stripe_redundancy);
 433         pr_buf(out, " read_time %llu",          a.io_time[READ]);
 434         pr_buf(out, " write_time %llu",         a.io_time[WRITE]);
 435 }
 436
 437 int bch2_alloc_read(struct bch_fs *c)
 438 {
 439         struct btree_trans trans;
 440         struct btree_iter iter;
 441         struct bkey_s_c k;
 442         struct bch_alloc_v4 a;
 443         struct bch_dev *ca;
 444         int ret;
 445
 446         bch2_trans_init(&trans, c, 0, 0);
 447
 448         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 449                            BTREE_ITER_PREFETCH, k, ret) {
 450                 /*
 451                  * Not a fsck error because this is checked/repaired by
 452                  * bch2_check_alloc_key() which runs later:
 453                  */
 454                 if (!bch2_dev_bucket_exists(c, k.k->p))
 455                         continue;
 456
 457                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 458                 bch2_alloc_to_v4(k, &a);
 459
 460                 *bucket_gen(ca, k.k->p.offset) = a.gen;
 461         }
 462         bch2_trans_iter_exit(&trans, &iter);
 463
 464         bch2_trans_exit(&trans);
 465
 466         if (ret)
 467                 bch_err(c, "error reading alloc info: %i", ret);
 468
 469         return ret;
 470 }
 471
 472 /* Free space/discard btree: */
 473
 474 static int bch2_bucket_do_index(struct btree_trans *trans,
 475                                 struct bkey_s_c alloc_k,
 476                                 const struct bch_alloc_v4 *a,
 477                                 bool set)
 478 {
 479         struct bch_fs *c = trans->c;
 480         struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 481         struct btree_iter iter;
 482         struct bkey_s_c old;
 483         struct bkey_i *k;
 484         enum btree_id btree;
 485         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 486         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 487         struct printbuf buf = PRINTBUF;
 488         int ret;
 489
 490         if (a->data_type != BCH_DATA_free &&
 491             a->data_type != BCH_DATA_need_discard)
 492                 return 0;
 493
 494         k = bch2_trans_kmalloc(trans, sizeof(*k));
 495         if (IS_ERR(k))
 496                 return PTR_ERR(k);
 497
 498         bkey_init(&k->k);
 499         k->k.type = new_type;
 500
 501         switch (a->data_type) {
 502         case BCH_DATA_free:
 503                 btree = BTREE_ID_freespace;
 504                 k->k.p = alloc_freespace_pos(alloc_k.k->p, *a);
 505                 bch2_key_resize(&k->k, 1);
 506                 break;
 507         case BCH_DATA_need_discard:
 508                 btree = BTREE_ID_need_discard;
 509                 k->k.p = alloc_k.k->p;
 510                 break;
 511         default:
 512                 return 0;
 513         }
 514
 515         bch2_trans_iter_init(trans, &iter, btree,
 516                              bkey_start_pos(&k->k),
 517                              BTREE_ITER_INTENT);
 518         old = bch2_btree_iter_peek_slot(&iter);
 519         ret = bkey_err(old);
 520         if (ret)
 521                 goto err;
 522
 523         if (ca->mi.freespace_initialized &&
 524             bch2_trans_inconsistent_on(old.k->type != old_type, trans,
 525                         "incorrect key when %s %s btree (got %s should be %s)\n"
 526                         "  for %s",
 527                         set ? "setting" : "clearing",
 528                         bch2_btree_ids[btree],
 529                         bch2_bkey_types[old.k->type],
 530                         bch2_bkey_types[old_type],
 531                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 532                 ret = -EIO;
 533                 goto err;
 534         }
 535
 536         ret = bch2_trans_update(trans, &iter, k, 0);
 537 err:
 538         bch2_trans_iter_exit(trans, &iter);
 539         printbuf_exit(&buf);
 540         return ret;
 541 }
 542
 543 int bch2_trans_mark_alloc(struct btree_trans *trans,
 544                           enum btree_id btree_id, unsigned level,
 545                           struct bkey_s_c old, struct bkey_i *new,
 546                           unsigned flags)
 547 {
 548         struct bch_fs *c = trans->c;
 549         struct bch_alloc_v4 old_a, *new_a;
 550         u64 old_lru, new_lru;
 551         int ret = 0;
 552
 553         /*
 554          * Deletion only happens in the device removal path, with
 555          * BTREE_TRIGGER_NORUN:
 556          */
 557         BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
 558
 559         bch2_alloc_to_v4(old, &old_a);
 560         new_a = &bkey_i_to_alloc_v4(new)->v;
 561
 562         new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
 563
 564         if (new_a->dirty_sectors > old_a.dirty_sectors ||
 565             new_a->cached_sectors > old_a.cached_sectors) {
 566                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 567                 new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 568                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
 569                 SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
 570         }
 571
 572         if (data_type_is_empty(new_a->data_type) &&
 573             BCH_ALLOC_V4_NEED_INC_GEN(new_a) &&
 574             !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
 575                 new_a->gen++;
 576                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
 577         }
 578
 579         if (old_a.data_type != new_a->data_type ||
 580             (new_a->data_type == BCH_DATA_free &&
 581              alloc_freespace_genbits(old_a) != alloc_freespace_genbits(*new_a))) {
 582                 ret =   bch2_bucket_do_index(trans, old, &old_a, false) ?:
 583                         bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_a, true);
 584                 if (ret)
 585                         return ret;
 586         }
 587
 588         if (new_a->data_type == BCH_DATA_cached &&
 589             !new_a->io_time[READ])
 590                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 591
 592         old_lru = alloc_lru_idx(old_a);
 593         new_lru = alloc_lru_idx(*new_a);
 594
 595         if (old_lru != new_lru) {
 596                 ret = bch2_lru_change(trans, new->k.p.inode, new->k.p.offset,
 597                                       old_lru, &new_lru, old);
 598                 if (ret)
 599                         return ret;
 600
 601                 if (new_a->data_type == BCH_DATA_cached)
 602                         new_a->io_time[READ] = new_lru;
 603         }
 604
 605         return 0;
 606 }
 607
 608 static int bch2_check_alloc_key(struct btree_trans *trans,
 609                                 struct btree_iter *alloc_iter)
 610 {
 611         struct bch_fs *c = trans->c;
 612         struct bch_dev *ca;
 613         struct btree_iter discard_iter, freespace_iter;
 614         struct bch_alloc_v4 a;
 615         unsigned discard_key_type, freespace_key_type;
 616         struct bkey_s_c alloc_k, k;
 617         struct printbuf buf = PRINTBUF;
 618         struct printbuf buf2 = PRINTBUF;
 619         int ret;
 620
 621         alloc_k = bch2_btree_iter_peek(alloc_iter);
 622         if (!alloc_k.k)
 623                 return 0;
 624
 625         ret = bkey_err(alloc_k);
 626         if (ret)
 627                 return ret;
 628
 629         if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
 630                         "alloc key for invalid device:bucket %llu:%llu",
 631                         alloc_k.k->p.inode, alloc_k.k->p.offset))
 632                 return bch2_btree_delete_at(trans, alloc_iter, 0);
 633
 634         ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 635         if (!ca->mi.freespace_initialized)
 636                 return 0;
 637
 638         bch2_alloc_to_v4(alloc_k, &a);
 639
 640         discard_key_type = a.data_type == BCH_DATA_need_discard
 641                 ? KEY_TYPE_set : 0;
 642         freespace_key_type = a.data_type == BCH_DATA_free
 643                 ? KEY_TYPE_set : 0;
 644
 645         bch2_trans_iter_init(trans, &discard_iter, BTREE_ID_need_discard,
 646                              alloc_k.k->p, 0);
 647         bch2_trans_iter_init(trans, &freespace_iter, BTREE_ID_freespace,
 648                              alloc_freespace_pos(alloc_k.k->p, a), 0);
 649
 650         k = bch2_btree_iter_peek_slot(&discard_iter);
 651         ret = bkey_err(k);
 652         if (ret)
 653                 goto err;
 654
 655         if (fsck_err_on(k.k->type != discard_key_type, c,
 656                         "incorrect key in need_discard btree (got %s should be %s)\n"
 657                         "  %s",
 658                         bch2_bkey_types[k.k->type],
 659                         bch2_bkey_types[discard_key_type],
 660                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 661                 struct bkey_i *update =
 662                         bch2_trans_kmalloc(trans, sizeof(*update));
 663
 664                 ret = PTR_ERR_OR_ZERO(update);
 665                 if (ret)
 666                         goto err;
 667
 668                 bkey_init(&update->k);
 669                 update->k.type  = discard_key_type;
 670                 update->k.p     = discard_iter.pos;
 671
 672                 ret = bch2_trans_update(trans, &discard_iter, update, 0);
 673                 if (ret)
 674                         goto err;
 675         }
 676
 677         k = bch2_btree_iter_peek_slot(&freespace_iter);
 678         ret = bkey_err(k);
 679         if (ret)
 680                 goto err;
 681
 682         if (fsck_err_on(k.k->type != freespace_key_type, c,
 683                         "incorrect key in freespace btree (got %s should be %s)\n"
 684                         "  %s",
 685                         bch2_bkey_types[k.k->type],
 686                         bch2_bkey_types[freespace_key_type],
 687                         (printbuf_reset(&buf),
 688                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 689                 struct bkey_i *update =
 690                         bch2_trans_kmalloc(trans, sizeof(*update));
 691
 692                 ret = PTR_ERR_OR_ZERO(update);
 693                 if (ret)
 694                         goto err;
 695
 696                 bkey_init(&update->k);
 697                 update->k.type  = freespace_key_type;
 698                 update->k.p     = freespace_iter.pos;
 699                 bch2_key_resize(&update->k, 1);
 700
 701                 ret = bch2_trans_update(trans, &freespace_iter, update, 0);
 702                 if (ret)
 703                         goto err;
 704         }
 705 err:
 706 fsck_err:
 707         bch2_trans_iter_exit(trans, &freespace_iter);
 708         bch2_trans_iter_exit(trans, &discard_iter);
 709         printbuf_exit(&buf2);
 710         printbuf_exit(&buf);
 711         return ret;
 712 }
 713
 714 static int bch2_check_discard_freespace_key(struct btree_trans *trans,
 715                                             struct btree_iter *iter)
 716 {
 717         struct bch_fs *c = trans->c;
 718         struct btree_iter alloc_iter;
 719         struct bkey_s_c k, freespace_k;
 720         struct bch_alloc_v4 a;
 721         u64 genbits;
 722         struct bpos pos;
 723         enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
 724                 ? BCH_DATA_need_discard
 725                 : BCH_DATA_free;
 726         struct printbuf buf = PRINTBUF;
 727         int ret;
 728
 729         freespace_k = bch2_btree_iter_peek(iter);
 730         if (!freespace_k.k)
 731                 return 1;
 732
 733         ret = bkey_err(freespace_k);
 734         if (ret)
 735                 return ret;
 736
 737         pos = iter->pos;
 738         pos.offset &= ~(~0ULL << 56);
 739         genbits = iter->pos.offset & (~0ULL << 56);
 740
 741         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
 742
 743         if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
 744                         "entry in %s btree for nonexistant dev:bucket %llu:%llu",
 745                         bch2_btree_ids[iter->btree_id], pos.inode, pos.offset))
 746                 goto delete;
 747
 748         k = bch2_btree_iter_peek_slot(&alloc_iter);
 749         ret = bkey_err(k);
 750         if (ret)
 751                 goto err;
 752
 753         bch2_alloc_to_v4(k, &a);
 754
 755         if (fsck_err_on(a.data_type != state ||
 756                         (state == BCH_DATA_free &&
 757                          genbits != alloc_freespace_genbits(a)), c,
 758                         "%s\n  incorrectly set in %s index (free %u, genbits %llu should be %llu)",
 759                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
 760                         bch2_btree_ids[iter->btree_id],
 761                         a.data_type == state,
 762                         genbits >> 56, alloc_freespace_genbits(a) >> 56))
 763                 goto delete;
 764 out:
 765 err:
 766 fsck_err:
 767         bch2_trans_iter_exit(trans, &alloc_iter);
 768         printbuf_exit(&buf);
 769         return ret;
 770 delete:
 771         ret = bch2_btree_delete_extent_at(trans, iter,
 772                         iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0);
 773         goto out;
 774 }
 775
 776 int bch2_check_alloc_info(struct bch_fs *c)
 777 {
 778         struct btree_trans trans;
 779         struct btree_iter iter;
 780         struct bkey_s_c k;
 781         int ret = 0;
 782
 783         bch2_trans_init(&trans, c, 0, 0);
 784
 785         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 786                            BTREE_ITER_PREFETCH, k, ret) {
 787                 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
 788                         bch2_check_alloc_key(&trans, &iter));
 789                 if (ret)
 790                         break;
 791         }
 792         bch2_trans_iter_exit(&trans, &iter);
 793
 794         if (ret)
 795                 goto err;
 796
 797         bch2_trans_iter_init(&trans, &iter, BTREE_ID_need_discard, POS_MIN,
 798                              BTREE_ITER_PREFETCH);
 799         while (1) {
 800                 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
 801                         bch2_check_discard_freespace_key(&trans, &iter));
 802                 if (ret)
 803                         break;
 804
 805                 bch2_btree_iter_set_pos(&iter, bpos_nosnap_successor(iter.pos));
 806         }
 807         bch2_trans_iter_exit(&trans, &iter);
 808
 809         if (ret)
 810                 goto err;
 811
 812         bch2_trans_iter_init(&trans, &iter, BTREE_ID_freespace, POS_MIN,
 813                              BTREE_ITER_PREFETCH);
 814         while (1) {
 815                 ret = __bch2_trans_do(&trans, NULL, NULL, 0,
 816                         bch2_check_discard_freespace_key(&trans, &iter));
 817                 if (ret)
 818                         break;
 819
 820                 bch2_btree_iter_set_pos(&iter, bpos_nosnap_successor(iter.pos));
 821         }
 822         bch2_trans_iter_exit(&trans, &iter);
 823 err:
 824         bch2_trans_exit(&trans);
 825         return ret < 0 ? ret : 0;
 826 }
 827
 828 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
 829                                        struct btree_iter *alloc_iter)
 830 {
 831         struct bch_fs *c = trans->c;
 832         struct btree_iter lru_iter;
 833         struct bch_alloc_v4 a;
 834         struct bkey_s_c alloc_k, k;
 835         struct printbuf buf = PRINTBUF;
 836         struct printbuf buf2 = PRINTBUF;
 837         int ret;
 838
 839         alloc_k = bch2_btree_iter_peek(alloc_iter);
 840         if (!alloc_k.k)
 841                 return 0;
 842
 843         ret = bkey_err(alloc_k);
 844         if (ret)
 845                 return ret;
 846
 847         bch2_alloc_to_v4(alloc_k, &a);
 848
 849         if (a.data_type != BCH_DATA_cached)
 850                 return 0;
 851
 852         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
 853                              POS(alloc_k.k->p.inode, a.io_time[READ]), 0);
 854
 855         k = bch2_btree_iter_peek_slot(&lru_iter);
 856         ret = bkey_err(k);
 857         if (ret)
 858                 goto err;
 859
 860         if (fsck_err_on(!a.io_time[READ], c,
 861                         "cached bucket with read_time 0\n"
 862                         "  %s",
 863                 (printbuf_reset(&buf),
 864                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) ||
 865             fsck_err_on(k.k->type != KEY_TYPE_lru ||
 866                         le64_to_cpu(bkey_s_c_to_lru(k).v->idx) != alloc_k.k->p.offset, c,
 867                         "incorrect/missing lru entry\n"
 868                         "  %s\n"
 869                         "  %s",
 870                         (printbuf_reset(&buf),
 871                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
 872                         (bch2_bkey_val_to_text(&buf2, c, k), buf2.buf))) {
 873                 u64 read_time = a.io_time[READ];
 874
 875                 if (!a.io_time[READ])
 876                         a.io_time[READ] = atomic64_read(&c->io_clock[READ].now);
 877
 878                 ret = bch2_lru_set(trans,
 879                                    alloc_k.k->p.inode,
 880                                    alloc_k.k->p.offset,
 881                                    &a.io_time[READ]);
 882                 if (ret)
 883                         goto err;
 884
 885                 if (a.io_time[READ] != read_time) {
 886                         struct bkey_i_alloc_v4 *a_mut =
 887                                 bch2_alloc_to_v4_mut(trans, alloc_k);
 888                         ret = PTR_ERR_OR_ZERO(a_mut);
 889                         if (ret)
 890                                 goto err;
 891
 892                         a_mut->v.io_time[READ] = a.io_time[READ];
 893                         ret = bch2_trans_update(trans, alloc_iter,
 894                                                 &a_mut->k_i, BTREE_TRIGGER_NORUN);
 895                         if (ret)
 896                                 goto err;
 897                 }
 898         }
 899 err:
 900 fsck_err:
 901         bch2_trans_iter_exit(trans, &lru_iter);
 902         printbuf_exit(&buf2);
 903         printbuf_exit(&buf);
 904         return ret;
 905 }
 906
 907 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
 908 {
 909         struct btree_trans trans;
 910         struct btree_iter iter;
 911         struct bkey_s_c k;
 912         int ret = 0;
 913
 914         bch2_trans_init(&trans, c, 0, 0);
 915
 916         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 917                            BTREE_ITER_PREFETCH, k, ret) {
 918                 ret = __bch2_trans_do(&trans, NULL, NULL,
 919                                       BTREE_INSERT_NOFAIL|
 920                                       BTREE_INSERT_LAZY_RW,
 921                         bch2_check_alloc_to_lru_ref(&trans, &iter));
 922                 if (ret)
 923                         break;
 924         }
 925         bch2_trans_iter_exit(&trans, &iter);
 926
 927         bch2_trans_exit(&trans);
 928         return ret < 0 ? ret : 0;
 929 }
 930
 931 static int bch2_clear_need_discard(struct btree_trans *trans, struct bpos pos,
 932                                    struct bch_dev *ca, bool *discard_done)
 933 {
 934         struct bch_fs *c = trans->c;
 935         struct btree_iter iter;
 936         struct bkey_s_c k;
 937         struct bkey_i_alloc_v4 *a;
 938         struct printbuf buf = PRINTBUF;
 939         int ret;
 940
 941         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, pos,
 942                              BTREE_ITER_CACHED);
 943         k = bch2_btree_iter_peek_slot(&iter);
 944         ret = bkey_err(k);
 945         if (ret)
 946                 goto out;
 947
 948         a = bch2_alloc_to_v4_mut(trans, k);
 949         ret = PTR_ERR_OR_ZERO(a);
 950         if (ret)
 951                 goto out;
 952
 953         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
 954                 a->v.gen++;
 955                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
 956                 goto write;
 957         }
 958
 959         if (bch2_trans_inconsistent_on(a->v.journal_seq > c->journal.flushed_seq_ondisk, trans,
 960                         "clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
 961                         "%s",
 962                         a->v.journal_seq,
 963                         c->journal.flushed_seq_ondisk,
 964                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 965                 ret = -EIO;
 966                 goto out;
 967         }
 968
 969         if (bch2_trans_inconsistent_on(a->v.data_type != BCH_DATA_need_discard, trans,
 970                         "bucket incorrectly set in need_discard btree\n"
 971                         "%s",
 972                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
 973                 ret = -EIO;
 974                 goto out;
 975         }
 976
 977         if (!*discard_done && ca->mi.discard && !c->opts.nochanges) {
 978                 /*
 979                  * This works without any other locks because this is the only
 980                  * thread that removes items from the need_discard tree
 981                  */
 982                 bch2_trans_unlock(trans);
 983                 blkdev_issue_discard(ca->disk_sb.bdev,
 984                                      k.k->p.offset * ca->mi.bucket_size,
 985                                      ca->mi.bucket_size,
 986                                      GFP_KERNEL, 0);
 987                 *discard_done = true;
 988
 989                 ret = bch2_trans_relock(trans) ? 0 : -EINTR;
 990                 if (ret)
 991                         goto out;
 992         }
 993
 994         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
 995         a->v.data_type = alloc_data_type(a->v, a->v.data_type);
 996 write:
 997         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
 998 out:
 999         bch2_trans_iter_exit(trans, &iter);
1000         printbuf_exit(&buf);
1001         return ret;
1002 }
1003
1004 static void bch2_do_discards_work(struct work_struct *work)
1005 {
1006         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
1007         struct bch_dev *ca = NULL;
1008         struct btree_trans trans;
1009         struct btree_iter iter;
1010         struct bkey_s_c k;
1011         u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
1012         int ret;
1013
1014         bch2_trans_init(&trans, c, 0, 0);
1015
1016         for_each_btree_key(&trans, iter, BTREE_ID_need_discard,
1017                            POS_MIN, 0, k, ret) {
1018                 bool discard_done = false;
1019
1020                 if (ca && k.k->p.inode != ca->dev_idx) {
1021                         percpu_ref_put(&ca->io_ref);
1022                         ca = NULL;
1023                 }
1024
1025                 if (!ca) {
1026                         ca = bch_dev_bkey_exists(c, k.k->p.inode);
1027                         if (!percpu_ref_tryget(&ca->io_ref)) {
1028                                 ca = NULL;
1029                                 bch2_btree_iter_set_pos(&iter, POS(k.k->p.inode + 1, 0));
1030                                 continue;
1031                         }
1032                 }
1033
1034                 seen++;
1035
1036                 if (bch2_bucket_is_open_safe(c, k.k->p.inode, k.k->p.offset)) {
1037                         open++;
1038                         continue;
1039                 }
1040
1041                 if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
1042                                 c->journal.flushed_seq_ondisk,
1043                                 k.k->p.inode, k.k->p.offset)) {
1044                         need_journal_commit++;
1045                         continue;
1046                 }
1047
1048                 ret = __bch2_trans_do(&trans, NULL, NULL,
1049                                       BTREE_INSERT_USE_RESERVE|
1050                                       BTREE_INSERT_NOFAIL,
1051                                 bch2_clear_need_discard(&trans, k.k->p, ca, &discard_done));
1052                 if (ret)
1053                         break;
1054
1055                 discarded++;
1056         }
1057         bch2_trans_iter_exit(&trans, &iter);
1058
1059         if (ca)
1060                 percpu_ref_put(&ca->io_ref);
1061
1062         bch2_trans_exit(&trans);
1063
1064         if (need_journal_commit * 2 > seen)
1065                 bch2_journal_flush_async(&c->journal, NULL);
1066
1067         percpu_ref_put(&c->writes);
1068
1069         trace_discard_buckets(c, seen, open, need_journal_commit, discarded, ret);
1070 }
1071
1072 void bch2_do_discards(struct bch_fs *c)
1073 {
1074         if (percpu_ref_tryget(&c->writes) &&
1075             !queue_work(system_long_wq, &c->discard_work))
1076                 percpu_ref_put(&c->writes);
1077 }
1078
1079 static int invalidate_one_bucket(struct btree_trans *trans, struct bch_dev *ca)
1080 {
1081         struct bch_fs *c = trans->c;
1082         struct btree_iter lru_iter, alloc_iter = { NULL };
1083         struct bkey_s_c k;
1084         struct bkey_i_alloc_v4 *a;
1085         u64 bucket, idx;
1086         struct printbuf buf = PRINTBUF;
1087         int ret;
1088
1089         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
1090                              POS(ca->dev_idx, 0), 0);
1091 next_lru:
1092         k = bch2_btree_iter_peek(&lru_iter);
1093         ret = bkey_err(k);
1094         if (ret)
1095                 goto out;
1096
1097         if (!k.k || k.k->p.inode != ca->dev_idx)
1098                 goto out;
1099
1100         if (k.k->type != KEY_TYPE_lru) {
1101                 pr_buf(&buf, "non lru key in lru btree:\n  ");
1102                 bch2_bkey_val_to_text(&buf, c, k);
1103
1104                 if (!test_bit(BCH_FS_CHECK_LRUS_DONE, &c->flags)) {
1105                         bch_err(c, "%s", buf.buf);
1106                         bch2_btree_iter_advance(&lru_iter);
1107                         goto next_lru;
1108                 } else {
1109                         bch2_trans_inconsistent(trans, "%s", buf.buf);
1110                         ret = -EINVAL;
1111                         goto out;
1112                 }
1113         }
1114
1115         idx     = k.k->p.offset;
1116         bucket  = le64_to_cpu(bkey_s_c_to_lru(k).v->idx);
1117
1118         a = bch2_trans_start_alloc_update(trans, &alloc_iter,
1119                                           POS(ca->dev_idx, bucket));
1120         ret = PTR_ERR_OR_ZERO(a);
1121         if (ret)
1122                 goto out;
1123
1124         if (idx != alloc_lru_idx(a->v)) {
1125                 pr_buf(&buf, "alloc key does not point back to lru entry when invalidating bucket:\n  ");
1126                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
1127                 pr_buf(&buf, "\n  ");
1128                 bch2_bkey_val_to_text(&buf, c, k);
1129
1130                 if (!test_bit(BCH_FS_CHECK_LRUS_DONE, &c->flags)) {
1131                         bch_err(c, "%s", buf.buf);
1132                         bch2_btree_iter_advance(&lru_iter);
1133                         goto next_lru;
1134                 } else {
1135                         bch2_trans_inconsistent(trans, "%s", buf.buf);
1136                         ret = -EINVAL;
1137                         goto out;
1138                 }
1139         }
1140
1141         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1142         a->v.gen++;
1143         a->v.data_type          = 0;
1144         a->v.dirty_sectors      = 0;
1145         a->v.cached_sectors     = 0;
1146         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1147         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1148
1149         ret = bch2_trans_update(trans, &alloc_iter, &a->k_i,
1150                                 BTREE_TRIGGER_BUCKET_INVALIDATE);
1151         if (ret)
1152                 goto out;
1153
1154         trace_invalidate_bucket(c, a->k.p.inode, a->k.p.offset);
1155 out:
1156         bch2_trans_iter_exit(trans, &alloc_iter);
1157         bch2_trans_iter_exit(trans, &lru_iter);
1158         printbuf_exit(&buf);
1159         return ret;
1160 }
1161
1162 static void bch2_do_invalidates_work(struct work_struct *work)
1163 {
1164         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1165         struct bch_dev *ca;
1166         struct btree_trans trans;
1167         unsigned i;
1168         int ret = 0;
1169
1170         bch2_trans_init(&trans, c, 0, 0);
1171
1172         for_each_member_device(ca, c, i) {
1173                 s64 nr_to_invalidate =
1174                         should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
1175
1176                 while (!ret && nr_to_invalidate-- >= 0)
1177                         ret = __bch2_trans_do(&trans, NULL, NULL,
1178                                               BTREE_INSERT_USE_RESERVE|
1179                                               BTREE_INSERT_NOFAIL,
1180                                         invalidate_one_bucket(&trans, ca));
1181         }
1182
1183         bch2_trans_exit(&trans);
1184         percpu_ref_put(&c->writes);
1185 }
1186
1187 void bch2_do_invalidates(struct bch_fs *c)
1188 {
1189         if (percpu_ref_tryget(&c->writes))
1190                 queue_work(system_long_wq, &c->invalidate_work);
1191 }
1192
1193 static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca)
1194 {
1195         struct btree_trans trans;
1196         struct btree_iter iter;
1197         struct bkey_s_c k;
1198         struct bch_alloc_v4 a;
1199         struct bch_member *m;
1200         int ret;
1201
1202         bch2_trans_init(&trans, c, 0, 0);
1203
1204         for_each_btree_key(&trans, iter, BTREE_ID_alloc,
1205                            POS(ca->dev_idx, ca->mi.first_bucket),
1206                            BTREE_ITER_SLOTS|
1207                            BTREE_ITER_PREFETCH, k, ret) {
1208                 if (iter.pos.offset >= ca->mi.nbuckets)
1209                         break;
1210
1211                 bch2_alloc_to_v4(k, &a);
1212                 ret = __bch2_trans_do(&trans, NULL, NULL,
1213                                       BTREE_INSERT_LAZY_RW,
1214                                  bch2_bucket_do_index(&trans, k, &a, true));
1215                 if (ret)
1216                         break;
1217         }
1218         bch2_trans_iter_exit(&trans, &iter);
1219
1220         bch2_trans_exit(&trans);
1221
1222         if (ret) {
1223                 bch_err(ca, "error initializing free space: %i", ret);
1224                 return ret;
1225         }
1226
1227         mutex_lock(&c->sb_lock);
1228         m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx;
1229         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1230         mutex_unlock(&c->sb_lock);
1231
1232         return ret;
1233 }
1234
1235 int bch2_fs_freespace_init(struct bch_fs *c)
1236 {
1237         struct bch_dev *ca;
1238         unsigned i;
1239         int ret = 0;
1240         bool doing_init = false;
1241
1242         /*
1243          * We can crash during the device add path, so we need to check this on
1244          * every mount:
1245          */
1246
1247         for_each_member_device(ca, c, i) {
1248                 if (ca->mi.freespace_initialized)
1249                         continue;
1250
1251                 if (!doing_init) {
1252                         bch_info(c, "initializing freespace");
1253                         doing_init = true;
1254                 }
1255
1256                 ret = bch2_dev_freespace_init(c, ca);
1257                 if (ret) {
1258                         percpu_ref_put(&ca->ref);
1259                         return ret;
1260                 }
1261         }
1262
1263         if (doing_init) {
1264                 mutex_lock(&c->sb_lock);
1265                 bch2_write_super(c);
1266                 mutex_unlock(&c->sb_lock);
1267
1268                 bch_verbose(c, "done initializing freespace");
1269         }
1270
1271         return ret;
1272 }
1273
1274 /* Bucket IO clocks: */
1275
1276 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
1277                               size_t bucket_nr, int rw)
1278 {
1279         struct bch_fs *c = trans->c;
1280         struct btree_iter iter;
1281         struct bkey_i_alloc_v4 *a;
1282         u64 now;
1283         int ret = 0;
1284
1285         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
1286         ret = PTR_ERR_OR_ZERO(a);
1287         if (ret)
1288                 return ret;
1289
1290         now = atomic64_read(&c->io_clock[rw].now);
1291         if (a->v.io_time[rw] == now)
1292                 goto out;
1293
1294         a->v.io_time[rw] = now;
1295
1296         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1297                 bch2_trans_commit(trans, NULL, NULL, 0);
1298 out:
1299         bch2_trans_iter_exit(trans, &iter);
1300         return ret;
1301 }
1302
1303 /* Startup/shutdown (ro/rw): */
1304
1305 void bch2_recalc_capacity(struct bch_fs *c)
1306 {
1307         struct bch_dev *ca;
1308         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
1309         unsigned bucket_size_max = 0;
1310         unsigned long ra_pages = 0;
1311         unsigned i;
1312
1313         lockdep_assert_held(&c->state_lock);
1314
1315         for_each_online_member(ca, c, i) {
1316                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
1317
1318                 ra_pages += bdi->ra_pages;
1319         }
1320
1321         bch2_set_ra_pages(c, ra_pages);
1322
1323         for_each_rw_member(ca, c, i) {
1324                 u64 dev_reserve = 0;
1325
1326                 /*
1327                  * We need to reserve buckets (from the number
1328                  * of currently available buckets) against
1329                  * foreground writes so that mainly copygc can
1330                  * make forward progress.
1331                  *
1332                  * We need enough to refill the various reserves
1333                  * from scratch - copygc will use its entire
1334                  * reserve all at once, then run against when
1335                  * its reserve is refilled (from the formerly
1336                  * available buckets).
1337                  *
1338                  * This reserve is just used when considering if
1339                  * allocations for foreground writes must wait -
1340                  * not -ENOSPC calculations.
1341                  */
1342
1343                 dev_reserve += ca->nr_btree_reserve * 2;
1344                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
1345
1346                 dev_reserve += 1;       /* btree write point */
1347                 dev_reserve += 1;       /* copygc write point */
1348                 dev_reserve += 1;       /* rebalance write point */
1349
1350                 dev_reserve *= ca->mi.bucket_size;
1351
1352                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
1353                                              ca->mi.first_bucket);
1354
1355                 reserved_sectors += dev_reserve * 2;
1356
1357                 bucket_size_max = max_t(unsigned, bucket_size_max,
1358                                         ca->mi.bucket_size);
1359         }
1360
1361         gc_reserve = c->opts.gc_reserve_bytes
1362                 ? c->opts.gc_reserve_bytes >> 9
1363                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
1364
1365         reserved_sectors = max(gc_reserve, reserved_sectors);
1366
1367         reserved_sectors = min(reserved_sectors, capacity);
1368
1369         c->capacity = capacity - reserved_sectors;
1370
1371         c->bucket_size_max = bucket_size_max;
1372
1373         /* Wake up case someone was waiting for buckets */
1374         closure_wake_up(&c->freelist_wait);
1375 }
1376
1377 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
1378 {
1379         struct open_bucket *ob;
1380         bool ret = false;
1381
1382         for (ob = c->open_buckets;
1383              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1384              ob++) {
1385                 spin_lock(&ob->lock);
1386                 if (ob->valid && !ob->on_partial_list &&
1387                     ob->dev == ca->dev_idx)
1388                         ret = true;
1389                 spin_unlock(&ob->lock);
1390         }
1391
1392         return ret;
1393 }
1394
1395 /* device goes ro: */
1396 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
1397 {
1398         unsigned i;
1399
1400         /* First, remove device from allocation groups: */
1401
1402         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1403                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
1404
1405         /*
1406          * Capacity is calculated based off of devices in allocation groups:
1407          */
1408         bch2_recalc_capacity(c);
1409
1410         /* Next, close write points that point to this device... */
1411         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1412                 bch2_writepoint_stop(c, ca, &c->write_points[i]);
1413
1414         bch2_writepoint_stop(c, ca, &c->copygc_write_point);
1415         bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
1416         bch2_writepoint_stop(c, ca, &c->btree_write_point);
1417
1418         mutex_lock(&c->btree_reserve_cache_lock);
1419         while (c->btree_reserve_cache_nr) {
1420                 struct btree_alloc *a =
1421                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1422
1423                 bch2_open_buckets_put(c, &a->ob);
1424         }
1425         mutex_unlock(&c->btree_reserve_cache_lock);
1426
1427         while (1) {
1428                 struct open_bucket *ob;
1429
1430                 spin_lock(&c->freelist_lock);
1431                 if (!ca->open_buckets_partial_nr) {
1432                         spin_unlock(&c->freelist_lock);
1433                         break;
1434                 }
1435                 ob = c->open_buckets +
1436                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
1437                 ob->on_partial_list = false;
1438                 spin_unlock(&c->freelist_lock);
1439
1440                 bch2_open_bucket_put(c, ob);
1441         }
1442
1443         bch2_ec_stop_dev(c, ca);
1444
1445         /*
1446          * Wake up threads that were blocked on allocation, so they can notice
1447          * the device can no longer be removed and the capacity has changed:
1448          */
1449         closure_wake_up(&c->freelist_wait);
1450
1451         /*
1452          * journal_res_get() can block waiting for free space in the journal -
1453          * it needs to notice there may not be devices to allocate from anymore:
1454          */
1455         wake_up(&c->journal.wait);
1456
1457         /* Now wait for any in flight writes: */
1458
1459         closure_wait_event(&c->open_buckets_wait,
1460                            !bch2_dev_has_open_write_point(c, ca));
1461 }
1462
1463 /* device goes rw: */
1464 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
1465 {
1466         unsigned i;
1467
1468         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1469                 if (ca->mi.data_allowed & (1 << i))
1470                         set_bit(ca->dev_idx, c->rw_devs[i].d);
1471 }
1472
1473 void bch2_fs_allocator_background_init(struct bch_fs *c)
1474 {
1475         spin_lock_init(&c->freelist_lock);
1476         INIT_WORK(&c->discard_work, bch2_do_discards_work);
1477         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
1478 }