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 "backpointers.h"
   6 #include "btree_cache.h"
   7 #include "btree_io.h"
   8 #include "btree_key_cache.h"
   9 #include "btree_update.h"
  10 #include "btree_update_interior.h"
  11 #include "btree_gc.h"
  12 #include "btree_write_buffer.h"
  13 #include "buckets.h"
  14 #include "buckets_waiting_for_journal.h"
  15 #include "clock.h"
  16 #include "debug.h"
  17 #include "ec.h"
  18 #include "error.h"
  19 #include "lru.h"
  20 #include "recovery.h"
  21 #include "trace.h"
  22 #include "varint.h"
  23
  24 #include <linux/kthread.h>
  25 #include <linux/math64.h>
  26 #include <linux/random.h>
  27 #include <linux/rculist.h>
  28 #include <linux/rcupdate.h>
  29 #include <linux/sched/task.h>
  30 #include <linux/sort.h>
  31
  32 /* Persistent alloc info: */
  33
  34 static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
  35 #define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8,
  36         BCH_ALLOC_FIELDS_V1()
  37 #undef x
  38 };
  39
  40 struct bkey_alloc_unpacked {
  41         u64             journal_seq;
  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 void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out,
  83                                  struct bkey_s_c k)
  84 {
  85         const struct bch_alloc *in = bkey_s_c_to_alloc(k).v;
  86         const void *d = in->data;
  87         unsigned idx = 0;
  88
  89         out->gen = in->gen;
  90
  91 #define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++);
  92         BCH_ALLOC_FIELDS_V1()
  93 #undef  x
  94 }
  95
  96 static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out,
  97                                 struct bkey_s_c k)
  98 {
  99         struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k);
 100         const u8 *in = a.v->data;
 101         const u8 *end = bkey_val_end(a);
 102         unsigned fieldnr = 0;
 103         int ret;
 104         u64 v;
 105
 106         out->gen        = a.v->gen;
 107         out->oldest_gen = a.v->oldest_gen;
 108         out->data_type  = a.v->data_type;
 109
 110 #define x(_name, _bits)                                                 \
 111         if (fieldnr < a.v->nr_fields) {                                 \
 112                 ret = bch2_varint_decode_fast(in, end, &v);             \
 113                 if (ret < 0)                                            \
 114                         return ret;                                     \
 115                 in += ret;                                              \
 116         } else {                                                        \
 117                 v = 0;                                                  \
 118         }                                                               \
 119         out->_name = v;                                                 \
 120         if (v != out->_name)                                            \
 121                 return -1;                                              \
 122         fieldnr++;
 123
 124         BCH_ALLOC_FIELDS_V2()
 125 #undef  x
 126         return 0;
 127 }
 128
 129 static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out,
 130                                 struct bkey_s_c k)
 131 {
 132         struct bkey_s_c_alloc_v3 a = bkey_s_c_to_alloc_v3(k);
 133         const u8 *in = a.v->data;
 134         const u8 *end = bkey_val_end(a);
 135         unsigned fieldnr = 0;
 136         int ret;
 137         u64 v;
 138
 139         out->gen        = a.v->gen;
 140         out->oldest_gen = a.v->oldest_gen;
 141         out->data_type  = a.v->data_type;
 142         out->need_discard = BCH_ALLOC_V3_NEED_DISCARD(a.v);
 143         out->need_inc_gen = BCH_ALLOC_V3_NEED_INC_GEN(a.v);
 144         out->journal_seq = le64_to_cpu(a.v->journal_seq);
 145
 146 #define x(_name, _bits)                                                 \
 147         if (fieldnr < a.v->nr_fields) {                                 \
 148                 ret = bch2_varint_decode_fast(in, end, &v);             \
 149                 if (ret < 0)                                            \
 150                         return ret;                                     \
 151                 in += ret;                                              \
 152         } else {                                                        \
 153                 v = 0;                                                  \
 154         }                                                               \
 155         out->_name = v;                                                 \
 156         if (v != out->_name)                                            \
 157                 return -1;                                              \
 158         fieldnr++;
 159
 160         BCH_ALLOC_FIELDS_V2()
 161 #undef  x
 162         return 0;
 163 }
 164
 165 static struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 166 {
 167         struct bkey_alloc_unpacked ret = { .gen = 0 };
 168
 169         switch (k.k->type) {
 170         case KEY_TYPE_alloc:
 171                 bch2_alloc_unpack_v1(&ret, k);
 172                 break;
 173         case KEY_TYPE_alloc_v2:
 174                 bch2_alloc_unpack_v2(&ret, k);
 175                 break;
 176         case KEY_TYPE_alloc_v3:
 177                 bch2_alloc_unpack_v3(&ret, k);
 178                 break;
 179         }
 180
 181         return ret;
 182 }
 183
 184 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
 185 {
 186         unsigned i, bytes = offsetof(struct bch_alloc, data);
 187
 188         for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++)
 189                 if (a->fields & (1 << i))
 190                         bytes += BCH_ALLOC_V1_FIELD_BYTES[i];
 191
 192         return DIV_ROUND_UP(bytes, sizeof(u64));
 193 }
 194
 195 int bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k,
 196                           enum bkey_invalid_flags flags,
 197                           struct printbuf *err)
 198 {
 199         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 200
 201         /* allow for unknown fields */
 202         if (bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v)) {
 203                 prt_printf(err, "incorrect value size (%zu < %u)",
 204                        bkey_val_u64s(a.k), bch_alloc_v1_val_u64s(a.v));
 205                 return -BCH_ERR_invalid_bkey;
 206         }
 207
 208         return 0;
 209 }
 210
 211 int bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k,
 212                           enum bkey_invalid_flags flags,
 213                           struct printbuf *err)
 214 {
 215         struct bkey_alloc_unpacked u;
 216
 217         if (bch2_alloc_unpack_v2(&u, k)) {
 218                 prt_printf(err, "unpack error");
 219                 return -BCH_ERR_invalid_bkey;
 220         }
 221
 222         return 0;
 223 }
 224
 225 int bch2_alloc_v3_invalid(const struct bch_fs *c, struct bkey_s_c k,
 226                           enum bkey_invalid_flags flags,
 227                           struct printbuf *err)
 228 {
 229         struct bkey_alloc_unpacked u;
 230
 231         if (bch2_alloc_unpack_v3(&u, k)) {
 232                 prt_printf(err, "unpack error");
 233                 return -BCH_ERR_invalid_bkey;
 234         }
 235
 236         return 0;
 237 }
 238
 239 int bch2_alloc_v4_invalid(const struct bch_fs *c, struct bkey_s_c k,
 240                           enum bkey_invalid_flags flags,
 241                           struct printbuf *err)
 242 {
 243         struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 244
 245         if (alloc_v4_u64s(a.v) > bkey_val_u64s(k.k)) {
 246                 prt_printf(err, "bad val size (%u > %lu)",
 247                        alloc_v4_u64s(a.v), bkey_val_u64s(k.k));
 248                 return -BCH_ERR_invalid_bkey;
 249         }
 250
 251         if (!BCH_ALLOC_V4_BACKPOINTERS_START(a.v) &&
 252             BCH_ALLOC_V4_NR_BACKPOINTERS(a.v)) {
 253                 prt_printf(err, "invalid backpointers_start");
 254                 return -BCH_ERR_invalid_bkey;
 255         }
 256
 257         if (alloc_data_type(*a.v, a.v->data_type) != a.v->data_type) {
 258                 prt_printf(err, "invalid data type (got %u should be %u)",
 259                        a.v->data_type, alloc_data_type(*a.v, a.v->data_type));
 260                 return -BCH_ERR_invalid_bkey;
 261         }
 262
 263         switch (a.v->data_type) {
 264         case BCH_DATA_free:
 265         case BCH_DATA_need_gc_gens:
 266         case BCH_DATA_need_discard:
 267                 if (a.v->dirty_sectors ||
 268                     a.v->cached_sectors ||
 269                     a.v->stripe) {
 270                         prt_printf(err, "empty data type free but have data");
 271                         return -BCH_ERR_invalid_bkey;
 272                 }
 273                 break;
 274         case BCH_DATA_sb:
 275         case BCH_DATA_journal:
 276         case BCH_DATA_btree:
 277         case BCH_DATA_user:
 278         case BCH_DATA_parity:
 279                 if (!a.v->dirty_sectors) {
 280                         prt_printf(err, "data_type %s but dirty_sectors==0",
 281                                bch2_data_types[a.v->data_type]);
 282                         return -BCH_ERR_invalid_bkey;
 283                 }
 284                 break;
 285         case BCH_DATA_cached:
 286                 if (!a.v->cached_sectors ||
 287                     a.v->dirty_sectors ||
 288                     a.v->stripe) {
 289                         prt_printf(err, "data type inconsistency");
 290                         return -BCH_ERR_invalid_bkey;
 291                 }
 292
 293                 if (!a.v->io_time[READ] &&
 294                     c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_to_lru_refs) {
 295                         prt_printf(err, "cached bucket with read_time == 0");
 296                         return -BCH_ERR_invalid_bkey;
 297                 }
 298                 break;
 299         case BCH_DATA_stripe:
 300                 break;
 301         }
 302
 303         return 0;
 304 }
 305
 306 static inline u64 swab40(u64 x)
 307 {
 308         return (((x & 0x00000000ffULL) << 32)|
 309                 ((x & 0x000000ff00ULL) << 16)|
 310                 ((x & 0x0000ff0000ULL) >>  0)|
 311                 ((x & 0x00ff000000ULL) >> 16)|
 312                 ((x & 0xff00000000ULL) >> 32));
 313 }
 314
 315 void bch2_alloc_v4_swab(struct bkey_s k)
 316 {
 317         struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v;
 318         struct bch_backpointer *bp, *bps;
 319
 320         a->journal_seq          = swab64(a->journal_seq);
 321         a->flags                = swab32(a->flags);
 322         a->dirty_sectors        = swab32(a->dirty_sectors);
 323         a->cached_sectors       = swab32(a->cached_sectors);
 324         a->io_time[0]           = swab64(a->io_time[0]);
 325         a->io_time[1]           = swab64(a->io_time[1]);
 326         a->stripe               = swab32(a->stripe);
 327         a->nr_external_backpointers = swab32(a->nr_external_backpointers);
 328
 329         bps = alloc_v4_backpointers(a);
 330         for (bp = bps; bp < bps + BCH_ALLOC_V4_NR_BACKPOINTERS(a); bp++) {
 331                 bp->bucket_offset       = swab40(bp->bucket_offset);
 332                 bp->bucket_len          = swab32(bp->bucket_len);
 333                 bch2_bpos_swab(&bp->pos);
 334         }
 335 }
 336
 337 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 338 {
 339         struct bch_alloc_v4 _a;
 340         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &_a);
 341         unsigned i;
 342
 343         prt_newline(out);
 344         printbuf_indent_add(out, 2);
 345
 346         prt_printf(out, "gen %u oldest_gen %u data_type %s",
 347                a->gen, a->oldest_gen,
 348                a->data_type < BCH_DATA_NR
 349                ? bch2_data_types[a->data_type]
 350                : "(invalid data type)");
 351         prt_newline(out);
 352         prt_printf(out, "journal_seq       %llu",       a->journal_seq);
 353         prt_newline(out);
 354         prt_printf(out, "need_discard      %llu",       BCH_ALLOC_V4_NEED_DISCARD(a));
 355         prt_newline(out);
 356         prt_printf(out, "need_inc_gen      %llu",       BCH_ALLOC_V4_NEED_INC_GEN(a));
 357         prt_newline(out);
 358         prt_printf(out, "dirty_sectors     %u", a->dirty_sectors);
 359         prt_newline(out);
 360         prt_printf(out, "cached_sectors    %u", a->cached_sectors);
 361         prt_newline(out);
 362         prt_printf(out, "stripe            %u", a->stripe);
 363         prt_newline(out);
 364         prt_printf(out, "stripe_redundancy %u", a->stripe_redundancy);
 365         prt_newline(out);
 366         prt_printf(out, "io_time[READ]     %llu",       a->io_time[READ]);
 367         prt_newline(out);
 368         prt_printf(out, "io_time[WRITE]    %llu",       a->io_time[WRITE]);
 369         prt_newline(out);
 370         prt_printf(out, "fragmentation     %llu",       a->fragmentation_lru);
 371         prt_newline(out);
 372         prt_printf(out, "bp_start          %llu", BCH_ALLOC_V4_BACKPOINTERS_START(a));
 373         prt_newline(out);
 374
 375         if (BCH_ALLOC_V4_NR_BACKPOINTERS(a)) {
 376                 struct bkey_s_c_alloc_v4 a_raw = bkey_s_c_to_alloc_v4(k);
 377                 const struct bch_backpointer *bps = alloc_v4_backpointers_c(a_raw.v);
 378
 379                 prt_printf(out, "backpointers:     %llu", BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v));
 380                 printbuf_indent_add(out, 2);
 381
 382                 for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v); i++) {
 383                         prt_newline(out);
 384                         bch2_backpointer_to_text(out, &bps[i]);
 385                 }
 386
 387                 printbuf_indent_sub(out, 2);
 388         }
 389
 390         printbuf_indent_sub(out, 2);
 391 }
 392
 393 void __bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out)
 394 {
 395         if (k.k->type == KEY_TYPE_alloc_v4) {
 396                 void *src, *dst;
 397
 398                 *out = *bkey_s_c_to_alloc_v4(k).v;
 399
 400                 src = alloc_v4_backpointers(out);
 401                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 402                 dst = alloc_v4_backpointers(out);
 403
 404                 if (src < dst)
 405                         memset(src, 0, dst - src);
 406
 407                 SET_BCH_ALLOC_V4_NR_BACKPOINTERS(out, 0);
 408         } else {
 409                 struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 410
 411                 *out = (struct bch_alloc_v4) {
 412                         .journal_seq            = u.journal_seq,
 413                         .flags                  = u.need_discard,
 414                         .gen                    = u.gen,
 415                         .oldest_gen             = u.oldest_gen,
 416                         .data_type              = u.data_type,
 417                         .stripe_redundancy      = u.stripe_redundancy,
 418                         .dirty_sectors          = u.dirty_sectors,
 419                         .cached_sectors         = u.cached_sectors,
 420                         .io_time[READ]          = u.read_time,
 421                         .io_time[WRITE]         = u.write_time,
 422                         .stripe                 = u.stripe,
 423                 };
 424
 425                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 426         }
 427 }
 428
 429 static noinline struct bkey_i_alloc_v4 *
 430 __bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 431 {
 432         struct bkey_i_alloc_v4 *ret;
 433
 434         ret = bch2_trans_kmalloc(trans, max(bkey_bytes(k.k), sizeof(struct bkey_i_alloc_v4)));
 435         if (IS_ERR(ret))
 436                 return ret;
 437
 438         if (k.k->type == KEY_TYPE_alloc_v4) {
 439                 void *src, *dst;
 440
 441                 bkey_reassemble(&ret->k_i, k);
 442
 443                 src = alloc_v4_backpointers(&ret->v);
 444                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(&ret->v, BCH_ALLOC_V4_U64s);
 445                 dst = alloc_v4_backpointers(&ret->v);
 446
 447                 if (src < dst)
 448                         memset(src, 0, dst - src);
 449
 450                 SET_BCH_ALLOC_V4_NR_BACKPOINTERS(&ret->v, 0);
 451                 set_alloc_v4_u64s(ret);
 452         } else {
 453                 bkey_alloc_v4_init(&ret->k_i);
 454                 ret->k.p = k.k->p;
 455                 bch2_alloc_to_v4(k, &ret->v);
 456         }
 457         return ret;
 458 }
 459
 460 static inline struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut_inlined(struct btree_trans *trans, struct bkey_s_c k)
 461 {
 462         struct bkey_s_c_alloc_v4 a;
 463
 464         if (likely(k.k->type == KEY_TYPE_alloc_v4) &&
 465             ((a = bkey_s_c_to_alloc_v4(k), true) &&
 466              BCH_ALLOC_V4_NR_BACKPOINTERS(a.v) == 0))
 467                 return bch2_bkey_make_mut_noupdate_typed(trans, k, alloc_v4);
 468
 469         return __bch2_alloc_to_v4_mut(trans, k);
 470 }
 471
 472 struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 473 {
 474         return bch2_alloc_to_v4_mut_inlined(trans, k);
 475 }
 476
 477 struct bkey_i_alloc_v4 *
 478 bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
 479                               struct bpos pos)
 480 {
 481         struct bkey_s_c k;
 482         struct bkey_i_alloc_v4 *a;
 483         int ret;
 484
 485         k = bch2_bkey_get_iter(trans, iter, BTREE_ID_alloc, pos,
 486                              BTREE_ITER_WITH_UPDATES|
 487                              BTREE_ITER_CACHED|
 488                              BTREE_ITER_INTENT);
 489         ret = bkey_err(k);
 490         if (unlikely(ret))
 491                 return ERR_PTR(ret);
 492
 493         a = bch2_alloc_to_v4_mut_inlined(trans, k);
 494         ret = PTR_ERR_OR_ZERO(a);
 495         if (unlikely(ret))
 496                 goto err;
 497         return a;
 498 err:
 499         bch2_trans_iter_exit(trans, iter);
 500         return ERR_PTR(ret);
 501 }
 502
 503 static struct bpos alloc_gens_pos(struct bpos pos, unsigned *offset)
 504 {
 505         *offset = pos.offset & KEY_TYPE_BUCKET_GENS_MASK;
 506
 507         pos.offset >>= KEY_TYPE_BUCKET_GENS_BITS;
 508         return pos;
 509 }
 510
 511 static struct bpos bucket_gens_pos_to_alloc(struct bpos pos, unsigned offset)
 512 {
 513         pos.offset <<= KEY_TYPE_BUCKET_GENS_BITS;
 514         pos.offset += offset;
 515         return pos;
 516 }
 517
 518 static unsigned alloc_gen(struct bkey_s_c k, unsigned offset)
 519 {
 520         return k.k->type == KEY_TYPE_bucket_gens
 521                 ? bkey_s_c_to_bucket_gens(k).v->gens[offset]
 522                 : 0;
 523 }
 524
 525 int bch2_bucket_gens_invalid(const struct bch_fs *c, struct bkey_s_c k,
 526                              enum bkey_invalid_flags flags,
 527                              struct printbuf *err)
 528 {
 529         if (bkey_val_bytes(k.k) != sizeof(struct bch_bucket_gens)) {
 530                 prt_printf(err, "bad val size (%lu != %zu)",
 531                        bkey_val_bytes(k.k), sizeof(struct bch_bucket_gens));
 532                 return -BCH_ERR_invalid_bkey;
 533         }
 534
 535         return 0;
 536 }
 537
 538 void bch2_bucket_gens_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 539 {
 540         struct bkey_s_c_bucket_gens g = bkey_s_c_to_bucket_gens(k);
 541         unsigned i;
 542
 543         for (i = 0; i < ARRAY_SIZE(g.v->gens); i++) {
 544                 if (i)
 545                         prt_char(out, ' ');
 546                 prt_printf(out, "%u", g.v->gens[i]);
 547         }
 548 }
 549
 550 int bch2_bucket_gens_init(struct bch_fs *c)
 551 {
 552         struct btree_trans trans;
 553         struct btree_iter iter;
 554         struct bkey_s_c k;
 555         struct bch_alloc_v4 a;
 556         struct bkey_i_bucket_gens g;
 557         bool have_bucket_gens_key = false;
 558         unsigned offset;
 559         struct bpos pos;
 560         u8 gen;
 561         int ret;
 562
 563         bch2_trans_init(&trans, c, 0, 0);
 564
 565         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 566                            BTREE_ITER_PREFETCH, k, ret) {
 567                 /*
 568                  * Not a fsck error because this is checked/repaired by
 569                  * bch2_check_alloc_key() which runs later:
 570                  */
 571                 if (!bch2_dev_bucket_exists(c, k.k->p))
 572                         continue;
 573
 574                 gen = bch2_alloc_to_v4(k, &a)->gen;
 575                 pos = alloc_gens_pos(iter.pos, &offset);
 576
 577                 if (have_bucket_gens_key && bkey_cmp(iter.pos, pos)) {
 578                         ret = commit_do(&trans, NULL, NULL,
 579                                         BTREE_INSERT_NOFAIL|
 580                                         BTREE_INSERT_LAZY_RW,
 581                                 __bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i, 0));
 582                         if (ret)
 583                                 break;
 584                         have_bucket_gens_key = false;
 585                 }
 586
 587                 if (!have_bucket_gens_key) {
 588                         bkey_bucket_gens_init(&g.k_i);
 589                         g.k.p = pos;
 590                         have_bucket_gens_key = true;
 591                 }
 592
 593                 g.v.gens[offset] = gen;
 594         }
 595         bch2_trans_iter_exit(&trans, &iter);
 596
 597         if (have_bucket_gens_key && !ret)
 598                 ret = commit_do(&trans, NULL, NULL,
 599                                 BTREE_INSERT_NOFAIL|
 600                                 BTREE_INSERT_LAZY_RW,
 601                         __bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i, 0));
 602
 603         bch2_trans_exit(&trans);
 604
 605         if (ret)
 606                 bch_err_fn(c, ret);
 607         return ret;
 608 }
 609
 610 int bch2_alloc_read(struct bch_fs *c)
 611 {
 612         struct btree_trans trans;
 613         struct btree_iter iter;
 614         struct bkey_s_c k;
 615         struct bch_dev *ca;
 616         int ret;
 617
 618         down_read(&c->gc_lock);
 619         bch2_trans_init(&trans, c, 0, 0);
 620
 621         if (c->sb.version_upgrade_complete >= bcachefs_metadata_version_bucket_gens) {
 622                 const struct bch_bucket_gens *g;
 623                 u64 b;
 624
 625                 for_each_btree_key(&trans, iter, BTREE_ID_bucket_gens, POS_MIN,
 626                                    BTREE_ITER_PREFETCH, k, ret) {
 627                         u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
 628                         u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
 629
 630                         if (k.k->type != KEY_TYPE_bucket_gens)
 631                                 continue;
 632
 633                         g = bkey_s_c_to_bucket_gens(k).v;
 634
 635                         /*
 636                          * Not a fsck error because this is checked/repaired by
 637                          * bch2_check_alloc_key() which runs later:
 638                          */
 639                         if (!bch2_dev_exists2(c, k.k->p.inode))
 640                                 continue;
 641
 642                         ca = bch_dev_bkey_exists(c, k.k->p.inode);
 643
 644                         for (b = max_t(u64, ca->mi.first_bucket, start);
 645                              b < min_t(u64, ca->mi.nbuckets, end);
 646                              b++)
 647                                 *bucket_gen(ca, b) = g->gens[b & KEY_TYPE_BUCKET_GENS_MASK];
 648                 }
 649                 bch2_trans_iter_exit(&trans, &iter);
 650         } else {
 651                 struct bch_alloc_v4 a;
 652
 653                 for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 654                                    BTREE_ITER_PREFETCH, k, ret) {
 655                         /*
 656                          * Not a fsck error because this is checked/repaired by
 657                          * bch2_check_alloc_key() which runs later:
 658                          */
 659                         if (!bch2_dev_bucket_exists(c, k.k->p))
 660                                 continue;
 661
 662                         ca = bch_dev_bkey_exists(c, k.k->p.inode);
 663
 664                         *bucket_gen(ca, k.k->p.offset) = bch2_alloc_to_v4(k, &a)->gen;
 665                 }
 666                 bch2_trans_iter_exit(&trans, &iter);
 667         }
 668
 669         bch2_trans_exit(&trans);
 670         up_read(&c->gc_lock);
 671
 672         if (ret)
 673                 bch_err_fn(c, ret);
 674
 675         return ret;
 676 }
 677
 678 /* Free space/discard btree: */
 679
 680 static int bch2_bucket_do_index(struct btree_trans *trans,
 681                                 struct bkey_s_c alloc_k,
 682                                 const struct bch_alloc_v4 *a,
 683                                 bool set)
 684 {
 685         struct bch_fs *c = trans->c;
 686         struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 687         struct btree_iter iter;
 688         struct bkey_s_c old;
 689         struct bkey_i *k;
 690         enum btree_id btree;
 691         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 692         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 693         struct printbuf buf = PRINTBUF;
 694         int ret;
 695
 696         if (a->data_type != BCH_DATA_free &&
 697             a->data_type != BCH_DATA_need_discard)
 698                 return 0;
 699
 700         k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
 701         if (IS_ERR(k))
 702                 return PTR_ERR(k);
 703
 704         bkey_init(&k->k);
 705         k->k.type = new_type;
 706
 707         switch (a->data_type) {
 708         case BCH_DATA_free:
 709                 btree = BTREE_ID_freespace;
 710                 k->k.p = alloc_freespace_pos(alloc_k.k->p, *a);
 711                 bch2_key_resize(&k->k, 1);
 712                 break;
 713         case BCH_DATA_need_discard:
 714                 btree = BTREE_ID_need_discard;
 715                 k->k.p = alloc_k.k->p;
 716                 break;
 717         default:
 718                 return 0;
 719         }
 720
 721         old = bch2_bkey_get_iter(trans, &iter, btree,
 722                              bkey_start_pos(&k->k),
 723                              BTREE_ITER_INTENT);
 724         ret = bkey_err(old);
 725         if (ret)
 726                 return ret;
 727
 728         if (ca->mi.freespace_initialized &&
 729             c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info &&
 730             bch2_trans_inconsistent_on(old.k->type != old_type, trans,
 731                         "incorrect key when %s %s:%llu:%llu:0 (got %s should be %s)\n"
 732                         "  for %s",
 733                         set ? "setting" : "clearing",
 734                         bch2_btree_ids[btree],
 735                         iter.pos.inode,
 736                         iter.pos.offset,
 737                         bch2_bkey_types[old.k->type],
 738                         bch2_bkey_types[old_type],
 739                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 740                 ret = -EIO;
 741                 goto err;
 742         }
 743
 744         ret = bch2_trans_update(trans, &iter, k, 0);
 745 err:
 746         bch2_trans_iter_exit(trans, &iter);
 747         printbuf_exit(&buf);
 748         return ret;
 749 }
 750
 751 static noinline int bch2_bucket_gen_update(struct btree_trans *trans,
 752                                            struct bpos bucket, u8 gen)
 753 {
 754         struct btree_iter iter;
 755         unsigned offset;
 756         struct bpos pos = alloc_gens_pos(bucket, &offset);
 757         struct bkey_i_bucket_gens *g;
 758         struct bkey_s_c k;
 759         int ret;
 760
 761         g = bch2_trans_kmalloc(trans, sizeof(*g));
 762         ret = PTR_ERR_OR_ZERO(g);
 763         if (ret)
 764                 return ret;
 765
 766         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_bucket_gens, pos,
 767                                BTREE_ITER_INTENT|
 768                                BTREE_ITER_WITH_UPDATES);
 769         ret = bkey_err(k);
 770         if (ret)
 771                 return ret;
 772
 773         if (k.k->type != KEY_TYPE_bucket_gens) {
 774                 bkey_bucket_gens_init(&g->k_i);
 775                 g->k.p = iter.pos;
 776         } else {
 777                 bkey_reassemble(&g->k_i, k);
 778         }
 779
 780         g->v.gens[offset] = gen;
 781
 782         ret = bch2_trans_update(trans, &iter, &g->k_i, 0);
 783         bch2_trans_iter_exit(trans, &iter);
 784         return ret;
 785 }
 786
 787 int bch2_trans_mark_alloc(struct btree_trans *trans,
 788                           enum btree_id btree_id, unsigned level,
 789                           struct bkey_s_c old, struct bkey_i *new,
 790                           unsigned flags)
 791 {
 792         struct bch_fs *c = trans->c;
 793         struct bch_alloc_v4 old_a_convert, *new_a;
 794         const struct bch_alloc_v4 *old_a;
 795         u64 old_lru, new_lru;
 796         int ret = 0;
 797
 798         /*
 799          * Deletion only happens in the device removal path, with
 800          * BTREE_TRIGGER_NORUN:
 801          */
 802         BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
 803
 804         old_a = bch2_alloc_to_v4(old, &old_a_convert);
 805         new_a = &bkey_i_to_alloc_v4(new)->v;
 806
 807         new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
 808
 809         if (new_a->dirty_sectors > old_a->dirty_sectors ||
 810             new_a->cached_sectors > old_a->cached_sectors) {
 811                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 812                 new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 813                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
 814                 SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
 815         }
 816
 817         if (data_type_is_empty(new_a->data_type) &&
 818             BCH_ALLOC_V4_NEED_INC_GEN(new_a) &&
 819             !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
 820                 new_a->gen++;
 821                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
 822         }
 823
 824         if (old_a->data_type != new_a->data_type ||
 825             (new_a->data_type == BCH_DATA_free &&
 826              alloc_freespace_genbits(*old_a) != alloc_freespace_genbits(*new_a))) {
 827                 ret =   bch2_bucket_do_index(trans, old, old_a, false) ?:
 828                         bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_a, true);
 829                 if (ret)
 830                         return ret;
 831         }
 832
 833         if (new_a->data_type == BCH_DATA_cached &&
 834             !new_a->io_time[READ])
 835                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 836
 837         old_lru = alloc_lru_idx_read(*old_a);
 838         new_lru = alloc_lru_idx_read(*new_a);
 839
 840         if (old_lru != new_lru) {
 841                 ret = bch2_lru_change(trans, new->k.p.inode,
 842                                       bucket_to_u64(new->k.p),
 843                                       old_lru, new_lru);
 844                 if (ret)
 845                         return ret;
 846         }
 847
 848         new_a->fragmentation_lru = alloc_lru_idx_fragmentation(*new_a,
 849                                         bch_dev_bkey_exists(c, new->k.p.inode));
 850
 851         if (old_a->fragmentation_lru != new_a->fragmentation_lru) {
 852                 ret = bch2_lru_change(trans,
 853                                 BCH_LRU_FRAGMENTATION_START,
 854                                 bucket_to_u64(new->k.p),
 855                                 old_a->fragmentation_lru, new_a->fragmentation_lru);
 856                 if (ret)
 857                         return ret;
 858         }
 859
 860         if (old_a->gen != new_a->gen) {
 861                 ret = bch2_bucket_gen_update(trans, new->k.p, new_a->gen);
 862                 if (ret)
 863                         return ret;
 864         }
 865
 866         return 0;
 867 }
 868
 869 /*
 870  * This synthesizes deleted extents for holes, similar to BTREE_ITER_SLOTS for
 871  * extents style btrees, but works on non-extents btrees:
 872  */
 873 static struct bkey_s_c bch2_get_key_or_hole(struct btree_iter *iter, struct bpos end, struct bkey *hole)
 874 {
 875         struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
 876
 877         if (bkey_err(k))
 878                 return k;
 879
 880         if (k.k->type) {
 881                 return k;
 882         } else {
 883                 struct btree_iter iter2;
 884                 struct bpos next;
 885
 886                 bch2_trans_copy_iter(&iter2, iter);
 887
 888                 if (!bpos_eq(iter->path->l[0].b->key.k.p, SPOS_MAX))
 889                         end = bkey_min(end, bpos_nosnap_successor(iter->path->l[0].b->key.k.p));
 890
 891                 end = bkey_min(end, POS(iter->pos.inode, iter->pos.offset + U32_MAX - 1));
 892
 893                 /*
 894                  * btree node min/max is a closed interval, upto takes a half
 895                  * open interval:
 896                  */
 897                 k = bch2_btree_iter_peek_upto(&iter2, end);
 898                 next = iter2.pos;
 899                 bch2_trans_iter_exit(iter->trans, &iter2);
 900
 901                 BUG_ON(next.offset >= iter->pos.offset + U32_MAX);
 902
 903                 if (bkey_err(k))
 904                         return k;
 905
 906                 bkey_init(hole);
 907                 hole->p = iter->pos;
 908
 909                 bch2_key_resize(hole, next.offset - iter->pos.offset);
 910                 return (struct bkey_s_c) { hole, NULL };
 911         }
 912 }
 913
 914 static bool next_bucket(struct bch_fs *c, struct bpos *bucket)
 915 {
 916         struct bch_dev *ca;
 917         unsigned iter;
 918
 919         if (bch2_dev_bucket_exists(c, *bucket))
 920                 return true;
 921
 922         if (bch2_dev_exists2(c, bucket->inode)) {
 923                 ca = bch_dev_bkey_exists(c, bucket->inode);
 924
 925                 if (bucket->offset < ca->mi.first_bucket) {
 926                         bucket->offset = ca->mi.first_bucket;
 927                         return true;
 928                 }
 929
 930                 bucket->inode++;
 931                 bucket->offset = 0;
 932         }
 933
 934         rcu_read_lock();
 935         iter = bucket->inode;
 936         ca = __bch2_next_dev(c, &iter, NULL);
 937         if (ca)
 938                 *bucket = POS(ca->dev_idx, ca->mi.first_bucket);
 939         rcu_read_unlock();
 940
 941         return ca != NULL;
 942 }
 943
 944 static struct bkey_s_c bch2_get_key_or_real_bucket_hole(struct btree_iter *iter, struct bkey *hole)
 945 {
 946         struct bch_fs *c = iter->trans->c;
 947         struct bkey_s_c k;
 948 again:
 949         k = bch2_get_key_or_hole(iter, POS_MAX, hole);
 950         if (bkey_err(k))
 951                 return k;
 952
 953         if (!k.k->type) {
 954                 struct bpos bucket = bkey_start_pos(k.k);
 955
 956                 if (!bch2_dev_bucket_exists(c, bucket)) {
 957                         if (!next_bucket(c, &bucket))
 958                                 return bkey_s_c_null;
 959
 960                         bch2_btree_iter_set_pos(iter, bucket);
 961                         goto again;
 962                 }
 963
 964                 if (!bch2_dev_bucket_exists(c, k.k->p)) {
 965                         struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
 966
 967                         bch2_key_resize(hole, ca->mi.nbuckets - bucket.offset);
 968                 }
 969         }
 970
 971         return k;
 972 }
 973
 974 static noinline_for_stack
 975 int bch2_check_alloc_key(struct btree_trans *trans,
 976                          struct bkey_s_c alloc_k,
 977                          struct btree_iter *alloc_iter,
 978                          struct btree_iter *discard_iter,
 979                          struct btree_iter *freespace_iter,
 980                          struct btree_iter *bucket_gens_iter)
 981 {
 982         struct bch_fs *c = trans->c;
 983         struct bch_dev *ca;
 984         struct bch_alloc_v4 a_convert;
 985         const struct bch_alloc_v4 *a;
 986         unsigned discard_key_type, freespace_key_type;
 987         unsigned gens_offset;
 988         struct bkey_s_c k;
 989         struct printbuf buf = PRINTBUF;
 990         int ret;
 991
 992         if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
 993                         "alloc key for invalid device:bucket %llu:%llu",
 994                         alloc_k.k->p.inode, alloc_k.k->p.offset))
 995                 return bch2_btree_delete_at(trans, alloc_iter, 0);
 996
 997         ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 998         if (!ca->mi.freespace_initialized)
 999                 return 0;
1000
1001         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1002
1003         discard_key_type = a->data_type == BCH_DATA_need_discard ? KEY_TYPE_set : 0;
1004         bch2_btree_iter_set_pos(discard_iter, alloc_k.k->p);
1005         k = bch2_btree_iter_peek_slot(discard_iter);
1006         ret = bkey_err(k);
1007         if (ret)
1008                 goto err;
1009
1010         if (k.k->type != discard_key_type &&
1011             (c->opts.reconstruct_alloc ||
1012              fsck_err(c, "incorrect key in need_discard btree (got %s should be %s)\n"
1013                       "  %s",
1014                       bch2_bkey_types[k.k->type],
1015                       bch2_bkey_types[discard_key_type],
1016                       (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1017                 struct bkey_i *update =
1018                         bch2_trans_kmalloc(trans, sizeof(*update));
1019
1020                 ret = PTR_ERR_OR_ZERO(update);
1021                 if (ret)
1022                         goto err;
1023
1024                 bkey_init(&update->k);
1025                 update->k.type  = discard_key_type;
1026                 update->k.p     = discard_iter->pos;
1027
1028                 ret = bch2_trans_update(trans, discard_iter, update, 0);
1029                 if (ret)
1030                         goto err;
1031         }
1032
1033         freespace_key_type = a->data_type == BCH_DATA_free ? KEY_TYPE_set : 0;
1034         bch2_btree_iter_set_pos(freespace_iter, alloc_freespace_pos(alloc_k.k->p, *a));
1035         k = bch2_btree_iter_peek_slot(freespace_iter);
1036         ret = bkey_err(k);
1037         if (ret)
1038                 goto err;
1039
1040         if (k.k->type != freespace_key_type &&
1041             (c->opts.reconstruct_alloc ||
1042              fsck_err(c, "incorrect key in freespace btree (got %s should be %s)\n"
1043                       "  %s",
1044                       bch2_bkey_types[k.k->type],
1045                       bch2_bkey_types[freespace_key_type],
1046                       (printbuf_reset(&buf),
1047                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1048                 struct bkey_i *update =
1049                         bch2_trans_kmalloc(trans, sizeof(*update));
1050
1051                 ret = PTR_ERR_OR_ZERO(update);
1052                 if (ret)
1053                         goto err;
1054
1055                 bkey_init(&update->k);
1056                 update->k.type  = freespace_key_type;
1057                 update->k.p     = freespace_iter->pos;
1058                 bch2_key_resize(&update->k, 1);
1059
1060                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1061                 if (ret)
1062                         goto err;
1063         }
1064
1065         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(alloc_k.k->p, &gens_offset));
1066         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1067         ret = bkey_err(k);
1068         if (ret)
1069                 goto err;
1070
1071         if (a->gen != alloc_gen(k, gens_offset) &&
1072             (c->opts.reconstruct_alloc ||
1073              fsck_err(c, "incorrect gen in bucket_gens btree (got %u should be %u)\n"
1074                       "  %s",
1075                       alloc_gen(k, gens_offset), a->gen,
1076                       (printbuf_reset(&buf),
1077                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1078                 struct bkey_i_bucket_gens *g =
1079                         bch2_trans_kmalloc(trans, sizeof(*g));
1080
1081                 ret = PTR_ERR_OR_ZERO(g);
1082                 if (ret)
1083                         goto err;
1084
1085                 if (k.k->type == KEY_TYPE_bucket_gens) {
1086                         bkey_reassemble(&g->k_i, k);
1087                 } else {
1088                         bkey_bucket_gens_init(&g->k_i);
1089                         g->k.p = alloc_gens_pos(alloc_k.k->p, &gens_offset);
1090                 }
1091
1092                 g->v.gens[gens_offset] = a->gen;
1093
1094                 ret = bch2_trans_update(trans, bucket_gens_iter, &g->k_i, 0);
1095                 if (ret)
1096                         goto err;
1097         }
1098 err:
1099 fsck_err:
1100         printbuf_exit(&buf);
1101         return ret;
1102 }
1103
1104 static noinline_for_stack
1105 int bch2_check_alloc_hole_freespace(struct btree_trans *trans,
1106                                     struct bpos start,
1107                                     struct bpos *end,
1108                                     struct btree_iter *freespace_iter)
1109 {
1110         struct bch_fs *c = trans->c;
1111         struct bch_dev *ca;
1112         struct bkey_s_c k;
1113         struct printbuf buf = PRINTBUF;
1114         int ret;
1115
1116         ca = bch_dev_bkey_exists(c, start.inode);
1117         if (!ca->mi.freespace_initialized)
1118                 return 0;
1119
1120         bch2_btree_iter_set_pos(freespace_iter, start);
1121
1122         k = bch2_btree_iter_peek_slot(freespace_iter);
1123         ret = bkey_err(k);
1124         if (ret)
1125                 goto err;
1126
1127         *end = bkey_min(k.k->p, *end);
1128
1129         if (k.k->type != KEY_TYPE_set &&
1130             (c->opts.reconstruct_alloc ||
1131              fsck_err(c, "hole in alloc btree missing in freespace btree\n"
1132                       "  device %llu buckets %llu-%llu",
1133                       freespace_iter->pos.inode,
1134                       freespace_iter->pos.offset,
1135                       end->offset))) {
1136                 struct bkey_i *update =
1137                         bch2_trans_kmalloc(trans, sizeof(*update));
1138
1139                 ret = PTR_ERR_OR_ZERO(update);
1140                 if (ret)
1141                         goto err;
1142
1143                 bkey_init(&update->k);
1144                 update->k.type  = KEY_TYPE_set;
1145                 update->k.p     = freespace_iter->pos;
1146                 bch2_key_resize(&update->k,
1147                                 min_t(u64, U32_MAX, end->offset -
1148                                       freespace_iter->pos.offset));
1149
1150                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1151                 if (ret)
1152                         goto err;
1153         }
1154 err:
1155 fsck_err:
1156         printbuf_exit(&buf);
1157         return ret;
1158 }
1159
1160 static noinline_for_stack
1161 int bch2_check_alloc_hole_bucket_gens(struct btree_trans *trans,
1162                                       struct bpos start,
1163                                       struct bpos *end,
1164                                       struct btree_iter *bucket_gens_iter)
1165 {
1166         struct bch_fs *c = trans->c;
1167         struct bkey_s_c k;
1168         struct printbuf buf = PRINTBUF;
1169         unsigned i, gens_offset, gens_end_offset;
1170         int ret;
1171
1172         if (c->sb.version < bcachefs_metadata_version_bucket_gens)
1173                 return 0;
1174
1175         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(start, &gens_offset));
1176
1177         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1178         ret = bkey_err(k);
1179         if (ret)
1180                 goto err;
1181
1182         if (bkey_cmp(alloc_gens_pos(start, &gens_offset),
1183                      alloc_gens_pos(*end,  &gens_end_offset)))
1184                 gens_end_offset = KEY_TYPE_BUCKET_GENS_NR;
1185
1186         if (k.k->type == KEY_TYPE_bucket_gens) {
1187                 struct bkey_i_bucket_gens g;
1188                 bool need_update = false;
1189
1190                 bkey_reassemble(&g.k_i, k);
1191
1192                 for (i = gens_offset; i < gens_end_offset; i++) {
1193                         if (fsck_err_on(g.v.gens[i], c,
1194                                         "hole in alloc btree at %llu:%llu with nonzero gen in bucket_gens btree (%u)",
1195                                         bucket_gens_pos_to_alloc(k.k->p, i).inode,
1196                                         bucket_gens_pos_to_alloc(k.k->p, i).offset,
1197                                         g.v.gens[i])) {
1198                                 g.v.gens[i] = 0;
1199                                 need_update = true;
1200                         }
1201                 }
1202
1203                 if (need_update) {
1204                         struct bkey_i *k = bch2_trans_kmalloc(trans, sizeof(g));
1205
1206                         ret = PTR_ERR_OR_ZERO(k);
1207                         if (ret)
1208                                 goto err;
1209
1210                         memcpy(k, &g, sizeof(g));
1211
1212                         ret = bch2_trans_update(trans, bucket_gens_iter, k, 0);
1213                         if (ret)
1214                                 goto err;
1215                 }
1216         }
1217
1218         *end = bkey_min(*end, bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0));
1219 err:
1220 fsck_err:
1221         printbuf_exit(&buf);
1222         return ret;
1223 }
1224
1225 static noinline_for_stack int __bch2_check_discard_freespace_key(struct btree_trans *trans,
1226                                               struct btree_iter *iter)
1227 {
1228         struct bch_fs *c = trans->c;
1229         struct btree_iter alloc_iter;
1230         struct bkey_s_c alloc_k;
1231         struct bch_alloc_v4 a_convert;
1232         const struct bch_alloc_v4 *a;
1233         u64 genbits;
1234         struct bpos pos;
1235         enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
1236                 ? BCH_DATA_need_discard
1237                 : BCH_DATA_free;
1238         struct printbuf buf = PRINTBUF;
1239         int ret;
1240
1241         pos = iter->pos;
1242         pos.offset &= ~(~0ULL << 56);
1243         genbits = iter->pos.offset & (~0ULL << 56);
1244
1245         alloc_k = bch2_bkey_get_iter(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
1246         ret = bkey_err(alloc_k);
1247         if (ret)
1248                 return ret;
1249
1250         if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
1251                         "entry in %s btree for nonexistant dev:bucket %llu:%llu",
1252                         bch2_btree_ids[iter->btree_id], pos.inode, pos.offset))
1253                 goto delete;
1254
1255         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1256
1257         if (fsck_err_on(a->data_type != state ||
1258                         (state == BCH_DATA_free &&
1259                          genbits != alloc_freespace_genbits(*a)), c,
1260                         "%s\n  incorrectly set at %s:%llu:%llu:0 (free %u, genbits %llu should be %llu)",
1261                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
1262                         bch2_btree_ids[iter->btree_id],
1263                         iter->pos.inode,
1264                         iter->pos.offset,
1265                         a->data_type == state,
1266                         genbits >> 56, alloc_freespace_genbits(*a) >> 56))
1267                 goto delete;
1268 out:
1269 fsck_err:
1270         set_btree_iter_dontneed(&alloc_iter);
1271         bch2_trans_iter_exit(trans, &alloc_iter);
1272         printbuf_exit(&buf);
1273         return ret;
1274 delete:
1275         ret =   bch2_btree_delete_extent_at(trans, iter,
1276                         iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0) ?:
1277                 bch2_trans_commit(trans, NULL, NULL,
1278                         BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW);
1279         goto out;
1280 }
1281
1282 static int bch2_check_discard_freespace_key(struct btree_trans *trans,
1283                                             struct btree_iter *iter,
1284                                             struct bpos end)
1285 {
1286         if (!btree_id_is_extents(iter->btree_id)) {
1287                 return __bch2_check_discard_freespace_key(trans, iter);
1288         } else {
1289                 int ret;
1290
1291                 while (!bkey_eq(iter->pos, end) &&
1292                        !(ret = btree_trans_too_many_iters(trans) ?:
1293                                __bch2_check_discard_freespace_key(trans, iter)))
1294                         bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
1295
1296                 return ret;
1297         }
1298 }
1299
1300 /*
1301  * We've already checked that generation numbers in the bucket_gens btree are
1302  * valid for buckets that exist; this just checks for keys for nonexistent
1303  * buckets.
1304  */
1305 static noinline_for_stack
1306 int bch2_check_bucket_gens_key(struct btree_trans *trans,
1307                                struct btree_iter *iter,
1308                                struct bkey_s_c k)
1309 {
1310         struct bch_fs *c = trans->c;
1311         struct bkey_i_bucket_gens g;
1312         struct bch_dev *ca;
1313         u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
1314         u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
1315         u64 b;
1316         bool need_update = false, dev_exists;
1317         struct printbuf buf = PRINTBUF;
1318         int ret = 0;
1319
1320         BUG_ON(k.k->type != KEY_TYPE_bucket_gens);
1321         bkey_reassemble(&g.k_i, k);
1322
1323         /* if no bch_dev, skip out whether we repair or not */
1324         dev_exists = bch2_dev_exists2(c, k.k->p.inode);
1325         if (!dev_exists) {
1326                 if (fsck_err_on(!dev_exists, c,
1327                                 "bucket_gens key for invalid device:\n  %s",
1328                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1329                         ret = bch2_btree_delete_at(trans, iter, 0);
1330                 }
1331                 goto out;
1332         }
1333
1334         ca = bch_dev_bkey_exists(c, k.k->p.inode);
1335         if (fsck_err_on(end <= ca->mi.first_bucket ||
1336                         start >= ca->mi.nbuckets, c,
1337                         "bucket_gens key for invalid buckets:\n  %s",
1338                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1339                 ret = bch2_btree_delete_at(trans, iter, 0);
1340                 goto out;
1341         }
1342
1343         for (b = start; b < ca->mi.first_bucket; b++)
1344                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1345                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1346                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1347                         need_update = true;
1348                 }
1349
1350         for (b = ca->mi.nbuckets; b < end; b++)
1351                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1352                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1353                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1354                         need_update = true;
1355                 }
1356
1357         if (need_update) {
1358                 struct bkey_i *k;
1359
1360                 k = bch2_trans_kmalloc(trans, sizeof(g));
1361                 ret = PTR_ERR_OR_ZERO(k);
1362                 if (ret)
1363                         goto out;
1364
1365                 memcpy(k, &g, sizeof(g));
1366                 ret = bch2_trans_update(trans, iter, k, 0);
1367         }
1368 out:
1369 fsck_err:
1370         printbuf_exit(&buf);
1371         return ret;
1372 }
1373
1374 int bch2_check_alloc_info(struct bch_fs *c)
1375 {
1376         struct btree_trans trans;
1377         struct btree_iter iter, discard_iter, freespace_iter, bucket_gens_iter;
1378         struct bkey hole;
1379         struct bkey_s_c k;
1380         int ret = 0;
1381
1382         bch2_trans_init(&trans, c, 0, 0);
1383
1384         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc, POS_MIN,
1385                              BTREE_ITER_PREFETCH);
1386         bch2_trans_iter_init(&trans, &discard_iter, BTREE_ID_need_discard, POS_MIN,
1387                              BTREE_ITER_PREFETCH);
1388         bch2_trans_iter_init(&trans, &freespace_iter, BTREE_ID_freespace, POS_MIN,
1389                              BTREE_ITER_PREFETCH);
1390         bch2_trans_iter_init(&trans, &bucket_gens_iter, BTREE_ID_bucket_gens, POS_MIN,
1391                              BTREE_ITER_PREFETCH);
1392
1393         while (1) {
1394                 struct bpos next;
1395
1396                 bch2_trans_begin(&trans);
1397
1398                 k = bch2_get_key_or_real_bucket_hole(&iter, &hole);
1399                 ret = bkey_err(k);
1400                 if (ret)
1401                         goto bkey_err;
1402
1403                 if (!k.k)
1404                         break;
1405
1406                 if (k.k->type) {
1407                         next = bpos_nosnap_successor(k.k->p);
1408
1409                         ret = bch2_check_alloc_key(&trans,
1410                                                    k, &iter,
1411                                                    &discard_iter,
1412                                                    &freespace_iter,
1413                                                    &bucket_gens_iter);
1414                         if (ret)
1415                                 goto bkey_err;
1416                 } else {
1417                         next = k.k->p;
1418
1419                         ret = bch2_check_alloc_hole_freespace(&trans,
1420                                                     bkey_start_pos(k.k),
1421                                                     &next,
1422                                                     &freespace_iter) ?:
1423                                 bch2_check_alloc_hole_bucket_gens(&trans,
1424                                                     bkey_start_pos(k.k),
1425                                                     &next,
1426                                                     &bucket_gens_iter);
1427                         if (ret)
1428                                 goto bkey_err;
1429                 }
1430
1431                 ret = bch2_trans_commit(&trans, NULL, NULL,
1432                                         BTREE_INSERT_NOFAIL|
1433                                         BTREE_INSERT_LAZY_RW);
1434                 if (ret)
1435                         goto bkey_err;
1436
1437                 bch2_btree_iter_set_pos(&iter, next);
1438 bkey_err:
1439                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1440                         continue;
1441                 if (ret)
1442                         break;
1443         }
1444         bch2_trans_iter_exit(&trans, &bucket_gens_iter);
1445         bch2_trans_iter_exit(&trans, &freespace_iter);
1446         bch2_trans_iter_exit(&trans, &discard_iter);
1447         bch2_trans_iter_exit(&trans, &iter);
1448
1449         if (ret < 0)
1450                 goto err;
1451
1452         ret = for_each_btree_key2(&trans, iter,
1453                         BTREE_ID_need_discard, POS_MIN,
1454                         BTREE_ITER_PREFETCH, k,
1455                 bch2_check_discard_freespace_key(&trans, &iter, k.k->p)) ?:
1456               for_each_btree_key2(&trans, iter,
1457                         BTREE_ID_freespace, POS_MIN,
1458                         BTREE_ITER_PREFETCH, k,
1459                 bch2_check_discard_freespace_key(&trans, &iter, k.k->p)) ?:
1460               for_each_btree_key_commit(&trans, iter,
1461                         BTREE_ID_bucket_gens, POS_MIN,
1462                         BTREE_ITER_PREFETCH, k,
1463                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1464                 bch2_check_bucket_gens_key(&trans, &iter, k));
1465 err:
1466         bch2_trans_exit(&trans);
1467         if (ret)
1468                 bch_err_fn(c, ret);
1469         return ret;
1470 }
1471
1472 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
1473                                        struct btree_iter *alloc_iter)
1474 {
1475         struct bch_fs *c = trans->c;
1476         struct btree_iter lru_iter;
1477         struct bch_alloc_v4 a_convert;
1478         const struct bch_alloc_v4 *a;
1479         struct bkey_s_c alloc_k, lru_k;
1480         struct printbuf buf = PRINTBUF;
1481         int ret;
1482
1483         alloc_k = bch2_btree_iter_peek(alloc_iter);
1484         if (!alloc_k.k)
1485                 return 0;
1486
1487         ret = bkey_err(alloc_k);
1488         if (ret)
1489                 return ret;
1490
1491         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1492
1493         if (a->data_type != BCH_DATA_cached)
1494                 return 0;
1495
1496         lru_k = bch2_bkey_get_iter(trans, &lru_iter, BTREE_ID_lru,
1497                              lru_pos(alloc_k.k->p.inode,
1498                                      bucket_to_u64(alloc_k.k->p),
1499                                      a->io_time[READ]), 0);
1500         ret = bkey_err(lru_k);
1501         if (ret)
1502                 return ret;
1503
1504         if (fsck_err_on(!a->io_time[READ], c,
1505                         "cached bucket with read_time 0\n"
1506                         "  %s",
1507                 (printbuf_reset(&buf),
1508                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) ||
1509             fsck_err_on(lru_k.k->type != KEY_TYPE_set, c,
1510                         "missing lru entry\n"
1511                         "  %s",
1512                         (printbuf_reset(&buf),
1513                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
1514                 u64 read_time = a->io_time[READ] ?:
1515                         atomic64_read(&c->io_clock[READ].now);
1516
1517                 ret = bch2_lru_set(trans,
1518                                    alloc_k.k->p.inode,
1519                                    bucket_to_u64(alloc_k.k->p),
1520                                    read_time);
1521                 if (ret)
1522                         goto err;
1523
1524                 if (a->io_time[READ] != read_time) {
1525                         struct bkey_i_alloc_v4 *a_mut =
1526                                 bch2_alloc_to_v4_mut(trans, alloc_k);
1527                         ret = PTR_ERR_OR_ZERO(a_mut);
1528                         if (ret)
1529                                 goto err;
1530
1531                         a_mut->v.io_time[READ] = read_time;
1532                         ret = bch2_trans_update(trans, alloc_iter,
1533                                                 &a_mut->k_i, BTREE_TRIGGER_NORUN);
1534                         if (ret)
1535                                 goto err;
1536                 }
1537         }
1538 err:
1539 fsck_err:
1540         bch2_trans_iter_exit(trans, &lru_iter);
1541         printbuf_exit(&buf);
1542         return ret;
1543 }
1544
1545 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
1546 {
1547         struct btree_iter iter;
1548         struct bkey_s_c k;
1549         int ret = 0;
1550
1551         ret = bch2_trans_run(c,
1552                 for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1553                                 POS_MIN, BTREE_ITER_PREFETCH, k,
1554                                 NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1555                         bch2_check_alloc_to_lru_ref(&trans, &iter)));
1556         if (ret)
1557                 bch_err_fn(c, ret);
1558         return ret;
1559 }
1560
1561 static int bch2_discard_one_bucket(struct btree_trans *trans,
1562                                    struct btree_iter *need_discard_iter,
1563                                    struct bpos *discard_pos_done,
1564                                    u64 *seen,
1565                                    u64 *open,
1566                                    u64 *need_journal_commit,
1567                                    u64 *discarded)
1568 {
1569         struct bch_fs *c = trans->c;
1570         struct bpos pos = need_discard_iter->pos;
1571         struct btree_iter iter = { NULL };
1572         struct bkey_s_c k;
1573         struct bch_dev *ca;
1574         struct bkey_i_alloc_v4 *a;
1575         struct printbuf buf = PRINTBUF;
1576         int ret = 0;
1577
1578         ca = bch_dev_bkey_exists(c, pos.inode);
1579         if (!percpu_ref_tryget(&ca->io_ref)) {
1580                 bch2_btree_iter_set_pos(need_discard_iter, POS(pos.inode + 1, 0));
1581                 return 0;
1582         }
1583
1584         if (bch2_bucket_is_open_safe(c, pos.inode, pos.offset)) {
1585                 (*open)++;
1586                 goto out;
1587         }
1588
1589         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
1590                         c->journal.flushed_seq_ondisk,
1591                         pos.inode, pos.offset)) {
1592                 (*need_journal_commit)++;
1593                 goto out;
1594         }
1595
1596         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_alloc,
1597                                need_discard_iter->pos,
1598                                BTREE_ITER_CACHED);
1599         ret = bkey_err(k);
1600         if (ret)
1601                 goto out;
1602
1603         a = bch2_alloc_to_v4_mut(trans, k);
1604         ret = PTR_ERR_OR_ZERO(a);
1605         if (ret)
1606                 goto out;
1607
1608         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
1609                 a->v.gen++;
1610                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1611                 goto write;
1612         }
1613
1614         if (a->v.journal_seq > c->journal.flushed_seq_ondisk) {
1615                 if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
1616                         bch2_trans_inconsistent(trans,
1617                                 "clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
1618                                 "%s",
1619                                 a->v.journal_seq,
1620                                 c->journal.flushed_seq_ondisk,
1621                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1622                         ret = -EIO;
1623                 }
1624                 goto out;
1625         }
1626
1627         if (a->v.data_type != BCH_DATA_need_discard) {
1628                 if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
1629                         bch2_trans_inconsistent(trans,
1630                                 "bucket incorrectly set in need_discard btree\n"
1631                                 "%s",
1632                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1633                         ret = -EIO;
1634                 }
1635
1636                 goto out;
1637         }
1638
1639         if (!bkey_eq(*discard_pos_done, iter.pos) &&
1640             ca->mi.discard && !c->opts.nochanges) {
1641                 /*
1642                  * This works without any other locks because this is the only
1643                  * thread that removes items from the need_discard tree
1644                  */
1645                 bch2_trans_unlock(trans);
1646                 blkdev_issue_discard(ca->disk_sb.bdev,
1647                                      k.k->p.offset * ca->mi.bucket_size,
1648                                      ca->mi.bucket_size,
1649                                      GFP_KERNEL);
1650                 *discard_pos_done = iter.pos;
1651
1652                 ret = bch2_trans_relock_notrace(trans);
1653                 if (ret)
1654                         goto out;
1655         }
1656
1657         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
1658         a->v.data_type = alloc_data_type(a->v, a->v.data_type);
1659 write:
1660         ret =   bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1661                 bch2_trans_commit(trans, NULL, NULL,
1662                                   BCH_WATERMARK_btree|
1663                                   BTREE_INSERT_NOFAIL);
1664         if (ret)
1665                 goto out;
1666
1667         this_cpu_inc(c->counters[BCH_COUNTER_bucket_discard]);
1668         (*discarded)++;
1669 out:
1670         (*seen)++;
1671         bch2_trans_iter_exit(trans, &iter);
1672         percpu_ref_put(&ca->io_ref);
1673         printbuf_exit(&buf);
1674         return ret;
1675 }
1676
1677 static void bch2_do_discards_work(struct work_struct *work)
1678 {
1679         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
1680         struct btree_trans trans;
1681         struct btree_iter iter;
1682         struct bkey_s_c k;
1683         u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
1684         struct bpos discard_pos_done = POS_MAX;
1685         int ret;
1686
1687         bch2_trans_init(&trans, c, 0, 0);
1688
1689         /*
1690          * We're doing the commit in bch2_discard_one_bucket instead of using
1691          * for_each_btree_key_commit() so that we can increment counters after
1692          * successful commit:
1693          */
1694         ret = for_each_btree_key2(&trans, iter,
1695                         BTREE_ID_need_discard, POS_MIN, 0, k,
1696                 bch2_discard_one_bucket(&trans, &iter, &discard_pos_done,
1697                                         &seen,
1698                                         &open,
1699                                         &need_journal_commit,
1700                                         &discarded));
1701
1702         bch2_trans_exit(&trans);
1703
1704         if (need_journal_commit * 2 > seen)
1705                 bch2_journal_flush_async(&c->journal, NULL);
1706
1707         bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1708
1709         trace_discard_buckets(c, seen, open, need_journal_commit, discarded,
1710                               bch2_err_str(ret));
1711 }
1712
1713 void bch2_do_discards(struct bch_fs *c)
1714 {
1715         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_discard) &&
1716             !queue_work(c->write_ref_wq, &c->discard_work))
1717                 bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1718 }
1719
1720 static int invalidate_one_bucket(struct btree_trans *trans,
1721                                  struct btree_iter *lru_iter,
1722                                  struct bkey_s_c lru_k,
1723                                  s64 *nr_to_invalidate)
1724 {
1725         struct bch_fs *c = trans->c;
1726         struct btree_iter alloc_iter = { NULL };
1727         struct bkey_i_alloc_v4 *a = NULL;
1728         struct printbuf buf = PRINTBUF;
1729         struct bpos bucket = u64_to_bucket(lru_k.k->p.offset);
1730         unsigned cached_sectors;
1731         int ret = 0;
1732
1733         if (*nr_to_invalidate <= 0)
1734                 return 1;
1735
1736         if (!bch2_dev_bucket_exists(c, bucket)) {
1737                 prt_str(&buf, "lru entry points to invalid bucket");
1738                 goto err;
1739         }
1740
1741         if (bch2_bucket_is_open_safe(c, bucket.inode, bucket.offset))
1742                 return 0;
1743
1744         a = bch2_trans_start_alloc_update(trans, &alloc_iter, bucket);
1745         ret = PTR_ERR_OR_ZERO(a);
1746         if (ret)
1747                 goto out;
1748
1749         /* We expect harmless races here due to the btree write buffer: */
1750         if (lru_pos_time(lru_iter->pos) != alloc_lru_idx_read(a->v))
1751                 goto out;
1752
1753         BUG_ON(a->v.data_type != BCH_DATA_cached);
1754
1755         if (!a->v.cached_sectors)
1756                 bch_err(c, "invalidating empty bucket, confused");
1757
1758         cached_sectors = a->v.cached_sectors;
1759
1760         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1761         a->v.gen++;
1762         a->v.data_type          = 0;
1763         a->v.dirty_sectors      = 0;
1764         a->v.cached_sectors     = 0;
1765         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1766         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1767
1768         ret =   bch2_trans_update(trans, &alloc_iter, &a->k_i,
1769                                 BTREE_TRIGGER_BUCKET_INVALIDATE) ?:
1770                 bch2_trans_commit(trans, NULL, NULL,
1771                                   BCH_WATERMARK_btree|
1772                                   BTREE_INSERT_NOFAIL);
1773         if (ret)
1774                 goto out;
1775
1776         trace_and_count(c, bucket_invalidate, c, bucket.inode, bucket.offset, cached_sectors);
1777         --*nr_to_invalidate;
1778 out:
1779         bch2_trans_iter_exit(trans, &alloc_iter);
1780         printbuf_exit(&buf);
1781         return ret;
1782 err:
1783         prt_str(&buf, "\n  lru key: ");
1784         bch2_bkey_val_to_text(&buf, c, lru_k);
1785
1786         prt_str(&buf, "\n  lru entry: ");
1787         bch2_lru_pos_to_text(&buf, lru_iter->pos);
1788
1789         prt_str(&buf, "\n  alloc key: ");
1790         if (!a)
1791                 bch2_bpos_to_text(&buf, bucket);
1792         else
1793                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
1794
1795         bch_err(c, "%s", buf.buf);
1796         if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_lrus) {
1797                 bch2_inconsistent_error(c);
1798                 ret = -EINVAL;
1799         }
1800
1801         goto out;
1802 }
1803
1804 static void bch2_do_invalidates_work(struct work_struct *work)
1805 {
1806         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1807         struct bch_dev *ca;
1808         struct btree_trans trans;
1809         struct btree_iter iter;
1810         struct bkey_s_c k;
1811         unsigned i;
1812         int ret = 0;
1813
1814         bch2_trans_init(&trans, c, 0, 0);
1815
1816         ret = bch2_btree_write_buffer_flush(&trans);
1817         if (ret)
1818                 goto err;
1819
1820         for_each_member_device(ca, c, i) {
1821                 s64 nr_to_invalidate =
1822                         should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
1823
1824                 ret = for_each_btree_key2_upto(&trans, iter, BTREE_ID_lru,
1825                                 lru_pos(ca->dev_idx, 0, 0),
1826                                 lru_pos(ca->dev_idx, U64_MAX, LRU_TIME_MAX),
1827                                 BTREE_ITER_INTENT, k,
1828                         invalidate_one_bucket(&trans, &iter, k, &nr_to_invalidate));
1829
1830                 if (ret < 0) {
1831                         percpu_ref_put(&ca->ref);
1832                         break;
1833                 }
1834         }
1835 err:
1836         bch2_trans_exit(&trans);
1837         bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1838 }
1839
1840 void bch2_do_invalidates(struct bch_fs *c)
1841 {
1842         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_invalidate) &&
1843             !queue_work(c->write_ref_wq, &c->invalidate_work))
1844                 bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1845 }
1846
1847 static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca,
1848                                    unsigned long *last_updated)
1849 {
1850         struct btree_trans trans;
1851         struct btree_iter iter;
1852         struct bkey_s_c k;
1853         struct bkey hole;
1854         struct bpos end = POS(ca->dev_idx, ca->mi.nbuckets);
1855         struct bch_member *m;
1856         int ret;
1857
1858         bch2_trans_init(&trans, c, 0, 0);
1859
1860         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc,
1861                              POS(ca->dev_idx, ca->mi.first_bucket),
1862                              BTREE_ITER_PREFETCH);
1863         /*
1864          * Scan the alloc btree for every bucket on @ca, and add buckets to the
1865          * freespace/need_discard/need_gc_gens btrees as needed:
1866          */
1867         while (1) {
1868                 if (*last_updated + HZ * 10 < jiffies) {
1869                         bch_info(ca, "%s: currently at %llu/%llu",
1870                                  __func__, iter.pos.offset, ca->mi.nbuckets);
1871                         *last_updated = jiffies;
1872                 }
1873
1874                 bch2_trans_begin(&trans);
1875
1876                 if (bkey_ge(iter.pos, end)) {
1877                         ret = 0;
1878                         break;
1879                 }
1880
1881                 k = bch2_get_key_or_hole(&iter, end, &hole);
1882                 ret = bkey_err(k);
1883                 if (ret)
1884                         goto bkey_err;
1885
1886                 if (k.k->type) {
1887                         /*
1888                          * We process live keys in the alloc btree one at a
1889                          * time:
1890                          */
1891                         struct bch_alloc_v4 a_convert;
1892                         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1893
1894                         ret =   bch2_bucket_do_index(&trans, k, a, true) ?:
1895                                 bch2_trans_commit(&trans, NULL, NULL,
1896                                                   BTREE_INSERT_LAZY_RW|
1897                                                   BTREE_INSERT_NOFAIL);
1898                         if (ret)
1899                                 goto bkey_err;
1900
1901                         bch2_btree_iter_advance(&iter);
1902                 } else {
1903                         struct bkey_i *freespace;
1904
1905                         freespace = bch2_trans_kmalloc(&trans, sizeof(*freespace));
1906                         ret = PTR_ERR_OR_ZERO(freespace);
1907                         if (ret)
1908                                 goto bkey_err;
1909
1910                         bkey_init(&freespace->k);
1911                         freespace->k.type       = KEY_TYPE_set;
1912                         freespace->k.p          = k.k->p;
1913                         freespace->k.size       = k.k->size;
1914
1915                         ret = __bch2_btree_insert(&trans, BTREE_ID_freespace, freespace, 0) ?:
1916                                 bch2_trans_commit(&trans, NULL, NULL,
1917                                                   BTREE_INSERT_LAZY_RW|
1918                                                   BTREE_INSERT_NOFAIL);
1919                         if (ret)
1920                                 goto bkey_err;
1921
1922                         bch2_btree_iter_set_pos(&iter, k.k->p);
1923                 }
1924 bkey_err:
1925                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1926                         continue;
1927                 if (ret)
1928                         break;
1929         }
1930
1931         bch2_trans_iter_exit(&trans, &iter);
1932         bch2_trans_exit(&trans);
1933
1934         if (ret < 0) {
1935                 bch_err(ca, "error initializing free space: %s", bch2_err_str(ret));
1936                 return ret;
1937         }
1938
1939         mutex_lock(&c->sb_lock);
1940         m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx;
1941         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1942         mutex_unlock(&c->sb_lock);
1943
1944         return 0;
1945 }
1946
1947 int bch2_fs_freespace_init(struct bch_fs *c)
1948 {
1949         struct bch_dev *ca;
1950         unsigned i;
1951         int ret = 0;
1952         bool doing_init = false;
1953         unsigned long last_updated = jiffies;
1954
1955         /*
1956          * We can crash during the device add path, so we need to check this on
1957          * every mount:
1958          */
1959
1960         for_each_member_device(ca, c, i) {
1961                 if (ca->mi.freespace_initialized)
1962                         continue;
1963
1964                 if (!doing_init) {
1965                         bch_info(c, "initializing freespace");
1966                         doing_init = true;
1967                 }
1968
1969                 ret = bch2_dev_freespace_init(c, ca, &last_updated);
1970                 if (ret) {
1971                         percpu_ref_put(&ca->ref);
1972                         bch_err_fn(c, ret);
1973                         return ret;
1974                 }
1975         }
1976
1977         if (doing_init) {
1978                 mutex_lock(&c->sb_lock);
1979                 bch2_write_super(c);
1980                 mutex_unlock(&c->sb_lock);
1981                 bch_verbose(c, "done initializing freespace");
1982         }
1983
1984         return 0;
1985 }
1986
1987 /* Bucket IO clocks: */
1988
1989 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
1990                               size_t bucket_nr, int rw)
1991 {
1992         struct bch_fs *c = trans->c;
1993         struct btree_iter iter;
1994         struct bkey_i_alloc_v4 *a;
1995         u64 now;
1996         int ret = 0;
1997
1998         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
1999         ret = PTR_ERR_OR_ZERO(a);
2000         if (ret)
2001                 return ret;
2002
2003         now = atomic64_read(&c->io_clock[rw].now);
2004         if (a->v.io_time[rw] == now)
2005                 goto out;
2006
2007         a->v.io_time[rw] = now;
2008
2009         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
2010                 bch2_trans_commit(trans, NULL, NULL, 0);
2011 out:
2012         bch2_trans_iter_exit(trans, &iter);
2013         return ret;
2014 }
2015
2016 /* Startup/shutdown (ro/rw): */
2017
2018 void bch2_recalc_capacity(struct bch_fs *c)
2019 {
2020         struct bch_dev *ca;
2021         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
2022         unsigned bucket_size_max = 0;
2023         unsigned long ra_pages = 0;
2024         unsigned i;
2025
2026         lockdep_assert_held(&c->state_lock);
2027
2028         for_each_online_member(ca, c, i) {
2029                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
2030
2031                 ra_pages += bdi->ra_pages;
2032         }
2033
2034         bch2_set_ra_pages(c, ra_pages);
2035
2036         for_each_rw_member(ca, c, i) {
2037                 u64 dev_reserve = 0;
2038
2039                 /*
2040                  * We need to reserve buckets (from the number
2041                  * of currently available buckets) against
2042                  * foreground writes so that mainly copygc can
2043                  * make forward progress.
2044                  *
2045                  * We need enough to refill the various reserves
2046                  * from scratch - copygc will use its entire
2047                  * reserve all at once, then run against when
2048                  * its reserve is refilled (from the formerly
2049                  * available buckets).
2050                  *
2051                  * This reserve is just used when considering if
2052                  * allocations for foreground writes must wait -
2053                  * not -ENOSPC calculations.
2054                  */
2055
2056                 dev_reserve += ca->nr_btree_reserve * 2;
2057                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
2058
2059                 dev_reserve += 1;       /* btree write point */
2060                 dev_reserve += 1;       /* copygc write point */
2061                 dev_reserve += 1;       /* rebalance write point */
2062
2063                 dev_reserve *= ca->mi.bucket_size;
2064
2065                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
2066                                              ca->mi.first_bucket);
2067
2068                 reserved_sectors += dev_reserve * 2;
2069
2070                 bucket_size_max = max_t(unsigned, bucket_size_max,
2071                                         ca->mi.bucket_size);
2072         }
2073
2074         gc_reserve = c->opts.gc_reserve_bytes
2075                 ? c->opts.gc_reserve_bytes >> 9
2076                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
2077
2078         reserved_sectors = max(gc_reserve, reserved_sectors);
2079
2080         reserved_sectors = min(reserved_sectors, capacity);
2081
2082         c->capacity = capacity - reserved_sectors;
2083
2084         c->bucket_size_max = bucket_size_max;
2085
2086         /* Wake up case someone was waiting for buckets */
2087         closure_wake_up(&c->freelist_wait);
2088 }
2089
2090 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
2091 {
2092         struct open_bucket *ob;
2093         bool ret = false;
2094
2095         for (ob = c->open_buckets;
2096              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
2097              ob++) {
2098                 spin_lock(&ob->lock);
2099                 if (ob->valid && !ob->on_partial_list &&
2100                     ob->dev == ca->dev_idx)
2101                         ret = true;
2102                 spin_unlock(&ob->lock);
2103         }
2104
2105         return ret;
2106 }
2107
2108 /* device goes ro: */
2109 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
2110 {
2111         unsigned i;
2112
2113         /* First, remove device from allocation groups: */
2114
2115         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2116                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
2117
2118         /*
2119          * Capacity is calculated based off of devices in allocation groups:
2120          */
2121         bch2_recalc_capacity(c);
2122
2123         bch2_open_buckets_stop(c, ca, false);
2124
2125         /*
2126          * Wake up threads that were blocked on allocation, so they can notice
2127          * the device can no longer be removed and the capacity has changed:
2128          */
2129         closure_wake_up(&c->freelist_wait);
2130
2131         /*
2132          * journal_res_get() can block waiting for free space in the journal -
2133          * it needs to notice there may not be devices to allocate from anymore:
2134          */
2135         wake_up(&c->journal.wait);
2136
2137         /* Now wait for any in flight writes: */
2138
2139         closure_wait_event(&c->open_buckets_wait,
2140                            !bch2_dev_has_open_write_point(c, ca));
2141 }
2142
2143 /* device goes rw: */
2144 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
2145 {
2146         unsigned i;
2147
2148         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2149                 if (ca->mi.data_allowed & (1 << i))
2150                         set_bit(ca->dev_idx, c->rw_devs[i].d);
2151 }
2152
2153 void bch2_fs_allocator_background_init(struct bch_fs *c)
2154 {
2155         spin_lock_init(&c->freelist_lock);
2156         INIT_WORK(&c->discard_work, bch2_do_discards_work);
2157         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
2158 }