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 "buckets.h"
  13 #include "buckets_waiting_for_journal.h"
  14 #include "clock.h"
  15 #include "debug.h"
  16 #include "ec.h"
  17 #include "error.h"
  18 #include "lru.h"
  19 #include "recovery.h"
  20 #include "varint.h"
  21
  22 #include <linux/kthread.h>
  23 #include <linux/math64.h>
  24 #include <linux/random.h>
  25 #include <linux/rculist.h>
  26 #include <linux/rcupdate.h>
  27 #include <linux/sched/task.h>
  28 #include <linux/sort.h>
  29 #include <trace/events/bcachefs.h>
  30
  31 /* Persistent alloc info: */
  32
  33 static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
  34 #define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8,
  35         BCH_ALLOC_FIELDS_V1()
  36 #undef x
  37 };
  38
  39 struct bkey_alloc_unpacked {
  40         u64             journal_seq;
  41         u8              gen;
  42         u8              oldest_gen;
  43         u8              data_type;
  44         bool            need_discard:1;
  45         bool            need_inc_gen:1;
  46 #define x(_name, _bits) u##_bits _name;
  47         BCH_ALLOC_FIELDS_V2()
  48 #undef  x
  49 };
  50
  51 static inline u64 alloc_field_v1_get(const struct bch_alloc *a,
  52                                      const void **p, unsigned field)
  53 {
  54         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  55         u64 v;
  56
  57         if (!(a->fields & (1 << field)))
  58                 return 0;
  59
  60         switch (bytes) {
  61         case 1:
  62                 v = *((const u8 *) *p);
  63                 break;
  64         case 2:
  65                 v = le16_to_cpup(*p);
  66                 break;
  67         case 4:
  68                 v = le32_to_cpup(*p);
  69                 break;
  70         case 8:
  71                 v = le64_to_cpup(*p);
  72                 break;
  73         default:
  74                 BUG();
  75         }
  76
  77         *p += bytes;
  78         return v;
  79 }
  80
  81 static inline void alloc_field_v1_put(struct bkey_i_alloc *a, void **p,
  82                                       unsigned field, u64 v)
  83 {
  84         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  85
  86         if (!v)
  87                 return;
  88
  89         a->v.fields |= 1 << field;
  90
  91         switch (bytes) {
  92         case 1:
  93                 *((u8 *) *p) = v;
  94                 break;
  95         case 2:
  96                 *((__le16 *) *p) = cpu_to_le16(v);
  97                 break;
  98         case 4:
  99                 *((__le32 *) *p) = cpu_to_le32(v);
 100                 break;
 101         case 8:
 102                 *((__le64 *) *p) = cpu_to_le64(v);
 103                 break;
 104         default:
 105                 BUG();
 106         }
 107
 108         *p += bytes;
 109 }
 110
 111 static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out,
 112                                  struct bkey_s_c k)
 113 {
 114         const struct bch_alloc *in = bkey_s_c_to_alloc(k).v;
 115         const void *d = in->data;
 116         unsigned idx = 0;
 117
 118         out->gen = in->gen;
 119
 120 #define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++);
 121         BCH_ALLOC_FIELDS_V1()
 122 #undef  x
 123 }
 124
 125 static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out,
 126                                 struct bkey_s_c k)
 127 {
 128         struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k);
 129         const u8 *in = a.v->data;
 130         const u8 *end = bkey_val_end(a);
 131         unsigned fieldnr = 0;
 132         int ret;
 133         u64 v;
 134
 135         out->gen        = a.v->gen;
 136         out->oldest_gen = a.v->oldest_gen;
 137         out->data_type  = a.v->data_type;
 138
 139 #define x(_name, _bits)                                                 \
 140         if (fieldnr < a.v->nr_fields) {                                 \
 141                 ret = bch2_varint_decode_fast(in, end, &v);             \
 142                 if (ret < 0)                                            \
 143                         return ret;                                     \
 144                 in += ret;                                              \
 145         } else {                                                        \
 146                 v = 0;                                                  \
 147         }                                                               \
 148         out->_name = v;                                                 \
 149         if (v != out->_name)                                            \
 150                 return -1;                                              \
 151         fieldnr++;
 152
 153         BCH_ALLOC_FIELDS_V2()
 154 #undef  x
 155         return 0;
 156 }
 157
 158 static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out,
 159                                 struct bkey_s_c k)
 160 {
 161         struct bkey_s_c_alloc_v3 a = bkey_s_c_to_alloc_v3(k);
 162         const u8 *in = a.v->data;
 163         const u8 *end = bkey_val_end(a);
 164         unsigned fieldnr = 0;
 165         int ret;
 166         u64 v;
 167
 168         out->gen        = a.v->gen;
 169         out->oldest_gen = a.v->oldest_gen;
 170         out->data_type  = a.v->data_type;
 171         out->need_discard = BCH_ALLOC_V3_NEED_DISCARD(a.v);
 172         out->need_inc_gen = BCH_ALLOC_V3_NEED_INC_GEN(a.v);
 173         out->journal_seq = le64_to_cpu(a.v->journal_seq);
 174
 175 #define x(_name, _bits)                                                 \
 176         if (fieldnr < a.v->nr_fields) {                                 \
 177                 ret = bch2_varint_decode_fast(in, end, &v);             \
 178                 if (ret < 0)                                            \
 179                         return ret;                                     \
 180                 in += ret;                                              \
 181         } else {                                                        \
 182                 v = 0;                                                  \
 183         }                                                               \
 184         out->_name = v;                                                 \
 185         if (v != out->_name)                                            \
 186                 return -1;                                              \
 187         fieldnr++;
 188
 189         BCH_ALLOC_FIELDS_V2()
 190 #undef  x
 191         return 0;
 192 }
 193
 194 static struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 195 {
 196         struct bkey_alloc_unpacked ret = { .gen = 0 };
 197
 198         switch (k.k->type) {
 199         case KEY_TYPE_alloc:
 200                 bch2_alloc_unpack_v1(&ret, k);
 201                 break;
 202         case KEY_TYPE_alloc_v2:
 203                 bch2_alloc_unpack_v2(&ret, k);
 204                 break;
 205         case KEY_TYPE_alloc_v3:
 206                 bch2_alloc_unpack_v3(&ret, k);
 207                 break;
 208         }
 209
 210         return ret;
 211 }
 212
 213 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
 214 {
 215         unsigned i, bytes = offsetof(struct bch_alloc, data);
 216
 217         for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++)
 218                 if (a->fields & (1 << i))
 219                         bytes += BCH_ALLOC_V1_FIELD_BYTES[i];
 220
 221         return DIV_ROUND_UP(bytes, sizeof(u64));
 222 }
 223
 224 int bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k,
 225                           unsigned flags, struct printbuf *err)
 226 {
 227         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 228
 229         /* allow for unknown fields */
 230         if (bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v)) {
 231                 prt_printf(err, "incorrect value size (%zu < %u)",
 232                        bkey_val_u64s(a.k), bch_alloc_v1_val_u64s(a.v));
 233                 return -BCH_ERR_invalid_bkey;
 234         }
 235
 236         return 0;
 237 }
 238
 239 int bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k,
 240                           unsigned flags, struct printbuf *err)
 241 {
 242         struct bkey_alloc_unpacked u;
 243
 244         if (bch2_alloc_unpack_v2(&u, k)) {
 245                 prt_printf(err, "unpack error");
 246                 return -BCH_ERR_invalid_bkey;
 247         }
 248
 249         return 0;
 250 }
 251
 252 int bch2_alloc_v3_invalid(const struct bch_fs *c, struct bkey_s_c k,
 253                           unsigned flags, struct printbuf *err)
 254 {
 255         struct bkey_alloc_unpacked u;
 256
 257         if (bch2_alloc_unpack_v3(&u, k)) {
 258                 prt_printf(err, "unpack error");
 259                 return -BCH_ERR_invalid_bkey;
 260         }
 261
 262         return 0;
 263 }
 264
 265 int bch2_alloc_v4_invalid(const struct bch_fs *c, struct bkey_s_c k,
 266                           unsigned flags, struct printbuf *err)
 267 {
 268         struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 269         int rw = flags & WRITE;
 270
 271         if (alloc_v4_u64s(a.v) != bkey_val_u64s(k.k)) {
 272                 prt_printf(err, "bad val size (%lu != %u)",
 273                        bkey_val_u64s(k.k), alloc_v4_u64s(a.v));
 274                 return -BCH_ERR_invalid_bkey;
 275         }
 276
 277         if (!BCH_ALLOC_V4_BACKPOINTERS_START(a.v) &&
 278             BCH_ALLOC_V4_NR_BACKPOINTERS(a.v)) {
 279                 prt_printf(err, "invalid backpointers_start");
 280                 return -BCH_ERR_invalid_bkey;
 281         }
 282
 283         if (rw == WRITE &&
 284             !(flags & BKEY_INVALID_FROM_JOURNAL) &&
 285             test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags)) {
 286                 unsigned i, bp_len = 0;
 287
 288                 for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a.v); i++)
 289                         bp_len += alloc_v4_backpointers_c(a.v)[i].bucket_len;
 290
 291                 if (bp_len > a.v->dirty_sectors) {
 292                         prt_printf(err, "too many backpointers");
 293                         return -BCH_ERR_invalid_bkey;
 294                 }
 295         }
 296
 297         if (rw == WRITE) {
 298                 if (alloc_data_type(*a.v, a.v->data_type) != a.v->data_type) {
 299                         prt_printf(err, "invalid data type (got %u should be %u)",
 300                                a.v->data_type, alloc_data_type(*a.v, a.v->data_type));
 301                         return -BCH_ERR_invalid_bkey;
 302                 }
 303
 304                 switch (a.v->data_type) {
 305                 case BCH_DATA_free:
 306                 case BCH_DATA_need_gc_gens:
 307                 case BCH_DATA_need_discard:
 308                         if (a.v->dirty_sectors ||
 309                             a.v->cached_sectors ||
 310                             a.v->stripe) {
 311                                 prt_printf(err, "empty data type free but have data");
 312                                 return -BCH_ERR_invalid_bkey;
 313                         }
 314                         break;
 315                 case BCH_DATA_sb:
 316                 case BCH_DATA_journal:
 317                 case BCH_DATA_btree:
 318                 case BCH_DATA_user:
 319                 case BCH_DATA_parity:
 320                         if (!a.v->dirty_sectors) {
 321                                 prt_printf(err, "data_type %s but dirty_sectors==0",
 322                                        bch2_data_types[a.v->data_type]);
 323                                 return -BCH_ERR_invalid_bkey;
 324                         }
 325                         break;
 326                 case BCH_DATA_cached:
 327                         if (!a.v->cached_sectors ||
 328                             a.v->dirty_sectors ||
 329                             a.v->stripe) {
 330                                 prt_printf(err, "data type inconsistency");
 331                                 return -BCH_ERR_invalid_bkey;
 332                         }
 333
 334                         if (!a.v->io_time[READ] &&
 335                             test_bit(BCH_FS_CHECK_ALLOC_TO_LRU_REFS_DONE, &c->flags)) {
 336                                 prt_printf(err, "cached bucket with read_time == 0");
 337                                 return -BCH_ERR_invalid_bkey;
 338                         }
 339                         break;
 340                 case BCH_DATA_stripe:
 341                         if (!a.v->stripe) {
 342                                 prt_printf(err, "data_type %s but stripe==0",
 343                                        bch2_data_types[a.v->data_type]);
 344                                 return -BCH_ERR_invalid_bkey;
 345                         }
 346                         break;
 347                 }
 348         }
 349
 350         return 0;
 351 }
 352
 353 static inline u64 swab40(u64 x)
 354 {
 355         return (((x & 0x00000000ffULL) << 32)|
 356                 ((x & 0x000000ff00ULL) << 16)|
 357                 ((x & 0x0000ff0000ULL) >>  0)|
 358                 ((x & 0x00ff000000ULL) >> 16)|
 359                 ((x & 0xff00000000ULL) >> 32));
 360 }
 361
 362 void bch2_alloc_v4_swab(struct bkey_s k)
 363 {
 364         struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v;
 365         struct bch_backpointer *bp, *bps;
 366
 367         a->journal_seq          = swab64(a->journal_seq);
 368         a->flags                = swab32(a->flags);
 369         a->dirty_sectors        = swab32(a->dirty_sectors);
 370         a->cached_sectors       = swab32(a->cached_sectors);
 371         a->io_time[0]           = swab64(a->io_time[0]);
 372         a->io_time[1]           = swab64(a->io_time[1]);
 373         a->stripe               = swab32(a->stripe);
 374         a->nr_external_backpointers = swab32(a->nr_external_backpointers);
 375
 376         bps = alloc_v4_backpointers(a);
 377         for (bp = bps; bp < bps + BCH_ALLOC_V4_NR_BACKPOINTERS(a); bp++) {
 378                 bp->bucket_offset       = swab40(bp->bucket_offset);
 379                 bp->bucket_len          = swab32(bp->bucket_len);
 380                 bch2_bpos_swab(&bp->pos);
 381         }
 382 }
 383
 384 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 385 {
 386         struct bch_alloc_v4 _a;
 387         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &_a);
 388         unsigned i;
 389
 390         prt_newline(out);
 391         printbuf_indent_add(out, 2);
 392
 393         prt_printf(out, "gen %u oldest_gen %u data_type %s",
 394                a->gen, a->oldest_gen,
 395                a->data_type < BCH_DATA_NR
 396                ? bch2_data_types[a->data_type]
 397                : "(invalid data type)");
 398         prt_newline(out);
 399         prt_printf(out, "journal_seq       %llu",       a->journal_seq);
 400         prt_newline(out);
 401         prt_printf(out, "need_discard      %llu",       BCH_ALLOC_V4_NEED_DISCARD(a));
 402         prt_newline(out);
 403         prt_printf(out, "need_inc_gen      %llu",       BCH_ALLOC_V4_NEED_INC_GEN(a));
 404         prt_newline(out);
 405         prt_printf(out, "dirty_sectors     %u", a->dirty_sectors);
 406         prt_newline(out);
 407         prt_printf(out, "cached_sectors    %u", a->cached_sectors);
 408         prt_newline(out);
 409         prt_printf(out, "stripe            %u", a->stripe);
 410         prt_newline(out);
 411         prt_printf(out, "stripe_redundancy %u", a->stripe_redundancy);
 412         prt_newline(out);
 413         prt_printf(out, "io_time[READ]     %llu",       a->io_time[READ]);
 414         prt_newline(out);
 415         prt_printf(out, "io_time[WRITE]    %llu",       a->io_time[WRITE]);
 416         prt_newline(out);
 417         prt_printf(out, "bp_start          %llu", BCH_ALLOC_V4_BACKPOINTERS_START(a));
 418         prt_newline(out);
 419
 420         if (BCH_ALLOC_V4_NR_BACKPOINTERS(a)) {
 421                 struct bkey_s_c_alloc_v4 a_raw = bkey_s_c_to_alloc_v4(k);
 422                 const struct bch_backpointer *bps = alloc_v4_backpointers_c(a_raw.v);
 423
 424                 prt_printf(out, "backpointers:     %llu", BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v));
 425                 printbuf_indent_add(out, 2);
 426
 427                 for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v); i++) {
 428                         prt_newline(out);
 429                         bch2_backpointer_to_text(out, &bps[i]);
 430                 }
 431
 432                 printbuf_indent_sub(out, 2);
 433         }
 434
 435         printbuf_indent_sub(out, 2);
 436 }
 437
 438 void __bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out)
 439 {
 440         if (k.k->type == KEY_TYPE_alloc_v4) {
 441                 void *src, *dst;
 442
 443                 *out = *bkey_s_c_to_alloc_v4(k).v;
 444
 445                 src = alloc_v4_backpointers(out);
 446                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 447                 dst = alloc_v4_backpointers(out);
 448
 449                 if (src < dst)
 450                         memset(src, 0, dst - src);
 451         } else {
 452                 struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 453
 454                 *out = (struct bch_alloc_v4) {
 455                         .journal_seq            = u.journal_seq,
 456                         .flags                  = u.need_discard,
 457                         .gen                    = u.gen,
 458                         .oldest_gen             = u.oldest_gen,
 459                         .data_type              = u.data_type,
 460                         .stripe_redundancy      = u.stripe_redundancy,
 461                         .dirty_sectors          = u.dirty_sectors,
 462                         .cached_sectors         = u.cached_sectors,
 463                         .io_time[READ]          = u.read_time,
 464                         .io_time[WRITE]         = u.write_time,
 465                         .stripe                 = u.stripe,
 466                 };
 467
 468                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 469         }
 470 }
 471
 472 static noinline struct bkey_i_alloc_v4 *
 473 __bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 474 {
 475         struct bkey_i_alloc_v4 *ret;
 476         if (k.k->type == KEY_TYPE_alloc_v4) {
 477                 struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 478                 unsigned bytes = sizeof(struct bkey_i_alloc_v4) +
 479                         BCH_ALLOC_V4_NR_BACKPOINTERS(a.v) *
 480                         sizeof(struct bch_backpointer);
 481                 void *src, *dst;
 482
 483                 /*
 484                  * Reserve space for one more backpointer here:
 485                  * Not sketchy at doing it this way, nope...
 486                  */
 487                 ret = bch2_trans_kmalloc(trans, bytes + sizeof(struct bch_backpointer));
 488                 if (IS_ERR(ret))
 489                         return ret;
 490
 491                 bkey_reassemble(&ret->k_i, k);
 492
 493                 src = alloc_v4_backpointers(&ret->v);
 494                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(&ret->v, BCH_ALLOC_V4_U64s);
 495                 dst = alloc_v4_backpointers(&ret->v);
 496
 497                 memmove(dst, src, BCH_ALLOC_V4_NR_BACKPOINTERS(&ret->v) *
 498                         sizeof(struct bch_backpointer));
 499                 if (src < dst)
 500                         memset(src, 0, dst - src);
 501                 set_alloc_v4_u64s(ret);
 502         } else {
 503                 ret = bch2_trans_kmalloc(trans, sizeof(struct bkey_i_alloc_v4) +
 504                                          sizeof(struct bch_backpointer));
 505                 if (IS_ERR(ret))
 506                         return ret;
 507
 508                 bkey_alloc_v4_init(&ret->k_i);
 509                 ret->k.p = k.k->p;
 510                 bch2_alloc_to_v4(k, &ret->v);
 511         }
 512         return ret;
 513 }
 514
 515 static inline struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut_inlined(struct btree_trans *trans, struct bkey_s_c k)
 516 {
 517         if (likely(k.k->type == KEY_TYPE_alloc_v4) &&
 518             BCH_ALLOC_V4_BACKPOINTERS_START(bkey_s_c_to_alloc_v4(k).v) == BCH_ALLOC_V4_U64s) {
 519                 /*
 520                  * Reserve space for one more backpointer here:
 521                  * Not sketchy at doing it this way, nope...
 522                  */
 523                 struct bkey_i_alloc_v4 *ret =
 524                         bch2_trans_kmalloc_nomemzero(trans, bkey_bytes(k.k) + sizeof(struct bch_backpointer));
 525                 if (!IS_ERR(ret))
 526                         bkey_reassemble(&ret->k_i, k);
 527                 return ret;
 528         }
 529
 530         return __bch2_alloc_to_v4_mut(trans, k);
 531 }
 532
 533 struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 534 {
 535         return bch2_alloc_to_v4_mut_inlined(trans, k);
 536 }
 537
 538 struct bkey_i_alloc_v4 *
 539 bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
 540                               struct bpos pos)
 541 {
 542         struct bkey_s_c k;
 543         struct bkey_i_alloc_v4 *a;
 544         int ret;
 545
 546         bch2_trans_iter_init(trans, iter, BTREE_ID_alloc, pos,
 547                              BTREE_ITER_WITH_UPDATES|
 548                              BTREE_ITER_CACHED|
 549                              BTREE_ITER_INTENT);
 550         k = bch2_btree_iter_peek_slot(iter);
 551         ret = bkey_err(k);
 552         if (unlikely(ret))
 553                 goto err;
 554
 555         a = bch2_alloc_to_v4_mut_inlined(trans, k);
 556         ret = PTR_ERR_OR_ZERO(a);
 557         if (unlikely(ret))
 558                 goto err;
 559         return a;
 560 err:
 561         bch2_trans_iter_exit(trans, iter);
 562         return ERR_PTR(ret);
 563 }
 564
 565 int bch2_alloc_read(struct bch_fs *c)
 566 {
 567         struct btree_trans trans;
 568         struct btree_iter iter;
 569         struct bkey_s_c k;
 570         struct bch_alloc_v4 a;
 571         struct bch_dev *ca;
 572         int ret;
 573
 574         bch2_trans_init(&trans, c, 0, 0);
 575
 576         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 577                            BTREE_ITER_PREFETCH, k, ret) {
 578                 /*
 579                  * Not a fsck error because this is checked/repaired by
 580                  * bch2_check_alloc_key() which runs later:
 581                  */
 582                 if (!bch2_dev_bucket_exists(c, k.k->p))
 583                         continue;
 584
 585                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 586
 587                 *bucket_gen(ca, k.k->p.offset) = bch2_alloc_to_v4(k, &a)->gen;
 588         }
 589         bch2_trans_iter_exit(&trans, &iter);
 590
 591         bch2_trans_exit(&trans);
 592
 593         if (ret)
 594                 bch_err(c, "error reading alloc info: %s", bch2_err_str(ret));
 595
 596         return ret;
 597 }
 598
 599 static struct bpos alloc_gens_pos(struct bpos pos, unsigned *offset)
 600 {
 601         *offset = pos.offset & KEY_TYPE_BUCKET_GENS_MASK;
 602
 603         pos.offset >>= KEY_TYPE_BUCKET_GENS_BITS;
 604         return pos;
 605 }
 606
 607 static struct bpos bucket_gens_pos_to_alloc(struct bpos pos, unsigned offset)
 608 {
 609         pos.offset <<= KEY_TYPE_BUCKET_GENS_BITS;
 610         pos.offset += offset;
 611         return pos;
 612 }
 613
 614 static unsigned alloc_gen(struct bkey_s_c k, unsigned offset)
 615 {
 616         return k.k->type == KEY_TYPE_bucket_gens
 617                 ? bkey_s_c_to_bucket_gens(k).v->gens[offset]
 618                 : 0;
 619 }
 620
 621 int bch2_bucket_gens_invalid(const struct bch_fs *c, struct bkey_s_c k,
 622                              unsigned flags, struct printbuf *err)
 623 {
 624         if (bkey_val_bytes(k.k) != sizeof(struct bch_bucket_gens)) {
 625                 prt_printf(err, "bad val size (%lu != %zu)",
 626                        bkey_val_bytes(k.k), sizeof(struct bch_bucket_gens));
 627                 return -BCH_ERR_invalid_bkey;
 628         }
 629
 630         return 0;
 631 }
 632
 633 void bch2_bucket_gens_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 634 {
 635         struct bkey_s_c_bucket_gens g = bkey_s_c_to_bucket_gens(k);
 636         unsigned i;
 637
 638         for (i = 0; i < ARRAY_SIZE(g.v->gens); i++) {
 639                 if (i)
 640                         prt_char(out, ' ');
 641                 prt_printf(out, "%u", g.v->gens[i]);
 642         }
 643 }
 644
 645 int bch2_bucket_gens_init(struct bch_fs *c)
 646 {
 647         struct btree_trans trans;
 648         struct btree_iter iter;
 649         struct bkey_s_c k;
 650         struct bch_alloc_v4 a;
 651         struct bkey_i_bucket_gens g;
 652         bool have_bucket_gens_key = false;
 653         unsigned offset;
 654         struct bpos pos;
 655         u8 gen;
 656         int ret;
 657
 658         bch2_trans_init(&trans, c, 0, 0);
 659
 660         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 661                            BTREE_ITER_PREFETCH, k, ret) {
 662                 /*
 663                  * Not a fsck error because this is checked/repaired by
 664                  * bch2_check_alloc_key() which runs later:
 665                  */
 666                 if (!bch2_dev_bucket_exists(c, k.k->p))
 667                         continue;
 668
 669                 gen = bch2_alloc_to_v4(k, &a)->gen;
 670                 pos = alloc_gens_pos(iter.pos, &offset);
 671
 672                 if (have_bucket_gens_key && bkey_cmp(iter.pos, pos)) {
 673                         ret = commit_do(&trans, NULL, NULL,
 674                                         BTREE_INSERT_NOFAIL|
 675                                         BTREE_INSERT_LAZY_RW,
 676                                 __bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i, 0));
 677                         if (ret)
 678                                 break;
 679                         have_bucket_gens_key = false;
 680                 }
 681
 682                 if (!have_bucket_gens_key) {
 683                         bkey_bucket_gens_init(&g.k_i);
 684                         g.k.p = pos;
 685                         have_bucket_gens_key = true;
 686                 }
 687
 688                 g.v.gens[offset] = gen;
 689         }
 690         bch2_trans_iter_exit(&trans, &iter);
 691
 692         if (have_bucket_gens_key && !ret)
 693                 ret = commit_do(&trans, NULL, NULL,
 694                                 BTREE_INSERT_NOFAIL|
 695                                 BTREE_INSERT_LAZY_RW,
 696                         __bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i, 0));
 697
 698         bch2_trans_exit(&trans);
 699
 700         if (ret)
 701                 bch_err(c, "%s: error %s", __func__, bch2_err_str(ret));
 702
 703         return ret;
 704 }
 705
 706 int bch2_bucket_gens_read(struct bch_fs *c)
 707 {
 708         struct btree_trans trans;
 709         struct btree_iter iter;
 710         struct bkey_s_c k;
 711         const struct bch_bucket_gens *g;
 712         struct bch_dev *ca;
 713         u64 b;
 714         int ret;
 715
 716         bch2_trans_init(&trans, c, 0, 0);
 717
 718         for_each_btree_key(&trans, iter, BTREE_ID_bucket_gens, POS_MIN,
 719                            BTREE_ITER_PREFETCH, k, ret) {
 720                 u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
 721                 u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
 722
 723                 if (k.k->type != KEY_TYPE_bucket_gens)
 724                         continue;
 725
 726                 g = bkey_s_c_to_bucket_gens(k).v;
 727
 728                 /*
 729                  * Not a fsck error because this is checked/repaired by
 730                  * bch2_check_alloc_key() which runs later:
 731                  */
 732                 if (!bch2_dev_exists2(c, k.k->p.inode))
 733                         continue;
 734
 735                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 736
 737                 for (b = max_t(u64, ca->mi.first_bucket, start);
 738                      b < min_t(u64, ca->mi.nbuckets, end);
 739                      b++)
 740                         *bucket_gen(ca, b) = g->gens[b & KEY_TYPE_BUCKET_GENS_MASK];
 741         }
 742         bch2_trans_iter_exit(&trans, &iter);
 743
 744         bch2_trans_exit(&trans);
 745
 746         if (ret)
 747                 bch_err(c, "error reading alloc info: %s", bch2_err_str(ret));
 748
 749         return ret;
 750 }
 751
 752 /* Free space/discard btree: */
 753
 754 static int bch2_bucket_do_index(struct btree_trans *trans,
 755                                 struct bkey_s_c alloc_k,
 756                                 const struct bch_alloc_v4 *a,
 757                                 bool set)
 758 {
 759         struct bch_fs *c = trans->c;
 760         struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 761         struct btree_iter iter;
 762         struct bkey_s_c old;
 763         struct bkey_i *k;
 764         enum btree_id btree;
 765         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 766         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 767         struct printbuf buf = PRINTBUF;
 768         int ret;
 769
 770         if (a->data_type != BCH_DATA_free &&
 771             a->data_type != BCH_DATA_need_discard)
 772                 return 0;
 773
 774         k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
 775         if (IS_ERR(k))
 776                 return PTR_ERR(k);
 777
 778         bkey_init(&k->k);
 779         k->k.type = new_type;
 780
 781         switch (a->data_type) {
 782         case BCH_DATA_free:
 783                 btree = BTREE_ID_freespace;
 784                 k->k.p = alloc_freespace_pos(alloc_k.k->p, *a);
 785                 bch2_key_resize(&k->k, 1);
 786                 break;
 787         case BCH_DATA_need_discard:
 788                 btree = BTREE_ID_need_discard;
 789                 k->k.p = alloc_k.k->p;
 790                 break;
 791         default:
 792                 return 0;
 793         }
 794
 795         bch2_trans_iter_init(trans, &iter, btree,
 796                              bkey_start_pos(&k->k),
 797                              BTREE_ITER_INTENT);
 798         old = bch2_btree_iter_peek_slot(&iter);
 799         ret = bkey_err(old);
 800         if (ret)
 801                 goto err;
 802
 803         if (ca->mi.freespace_initialized &&
 804             test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags) &&
 805             bch2_trans_inconsistent_on(old.k->type != old_type, trans,
 806                         "incorrect key when %s %s btree (got %s should be %s)\n"
 807                         "  for %s",
 808                         set ? "setting" : "clearing",
 809                         bch2_btree_ids[btree],
 810                         bch2_bkey_types[old.k->type],
 811                         bch2_bkey_types[old_type],
 812                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 813                 ret = -EIO;
 814                 goto err;
 815         }
 816
 817         ret = bch2_trans_update(trans, &iter, k, 0);
 818 err:
 819         bch2_trans_iter_exit(trans, &iter);
 820         printbuf_exit(&buf);
 821         return ret;
 822 }
 823
 824 static noinline int bch2_bucket_gen_update(struct btree_trans *trans,
 825                                            struct bpos bucket, u8 gen)
 826 {
 827         struct btree_iter iter;
 828         unsigned offset;
 829         struct bpos pos = alloc_gens_pos(bucket, &offset);
 830         struct bkey_i_bucket_gens *g;
 831         struct bkey_s_c k;
 832         int ret;
 833
 834         g = bch2_trans_kmalloc(trans, sizeof(*g));
 835         ret = PTR_ERR_OR_ZERO(g);
 836         if (ret)
 837                 return ret;
 838
 839         bch2_trans_iter_init(trans, &iter, BTREE_ID_bucket_gens, pos,
 840                              BTREE_ITER_INTENT|
 841                              BTREE_ITER_WITH_UPDATES);
 842         k = bch2_btree_iter_peek_slot(&iter);
 843         ret = bkey_err(k);
 844         if (ret)
 845                 goto err;
 846
 847         if (k.k->type != KEY_TYPE_bucket_gens) {
 848                 bkey_bucket_gens_init(&g->k_i);
 849                 g->k.p = iter.pos;
 850         } else {
 851                 bkey_reassemble(&g->k_i, k);
 852         }
 853
 854         g->v.gens[offset] = gen;
 855
 856         ret = bch2_trans_update(trans, &iter, &g->k_i, 0);
 857 err:
 858         bch2_trans_iter_exit(trans, &iter);
 859         return ret;
 860 }
 861
 862 int bch2_trans_mark_alloc(struct btree_trans *trans,
 863                           enum btree_id btree_id, unsigned level,
 864                           struct bkey_s_c old, struct bkey_i *new,
 865                           unsigned flags)
 866 {
 867         struct bch_fs *c = trans->c;
 868         struct bch_alloc_v4 old_a_convert, *new_a;
 869         const struct bch_alloc_v4 *old_a;
 870         u64 old_lru, new_lru;
 871         int ret = 0;
 872
 873         /*
 874          * Deletion only happens in the device removal path, with
 875          * BTREE_TRIGGER_NORUN:
 876          */
 877         BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
 878
 879         old_a = bch2_alloc_to_v4(old, &old_a_convert);
 880         new_a = &bkey_i_to_alloc_v4(new)->v;
 881
 882         new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
 883
 884         if (new_a->dirty_sectors > old_a->dirty_sectors ||
 885             new_a->cached_sectors > old_a->cached_sectors) {
 886                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 887                 new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 888                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
 889                 SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
 890         }
 891
 892         if (data_type_is_empty(new_a->data_type) &&
 893             BCH_ALLOC_V4_NEED_INC_GEN(new_a) &&
 894             !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
 895                 new_a->gen++;
 896                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
 897         }
 898
 899         if (old_a->data_type != new_a->data_type ||
 900             (new_a->data_type == BCH_DATA_free &&
 901              alloc_freespace_genbits(*old_a) != alloc_freespace_genbits(*new_a))) {
 902                 ret =   bch2_bucket_do_index(trans, old, old_a, false) ?:
 903                         bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_a, true);
 904                 if (ret)
 905                         return ret;
 906         }
 907
 908         if (new_a->data_type == BCH_DATA_cached &&
 909             !new_a->io_time[READ])
 910                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 911
 912         old_lru = alloc_lru_idx(*old_a);
 913         new_lru = alloc_lru_idx(*new_a);
 914
 915         if (old_lru != new_lru) {
 916                 ret = bch2_lru_change(trans, new->k.p.inode,
 917                                       bucket_to_u64(new->k.p),
 918                                       old_lru, new_lru);
 919                 if (ret)
 920                         return ret;
 921         }
 922
 923         if (old_a->gen != new_a->gen) {
 924                 ret = bch2_bucket_gen_update(trans, new->k.p, new_a->gen);
 925                 if (ret)
 926                         return ret;
 927         }
 928
 929         return 0;
 930 }
 931
 932 /*
 933  * This synthesizes deleted extents for holes, similar to BTREE_ITER_SLOTS for
 934  * extents style btrees, but works on non-extents btrees:
 935  */
 936 struct bkey_s_c bch2_get_key_or_hole(struct btree_iter *iter, struct bpos end, struct bkey *hole)
 937 {
 938         struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
 939
 940         if (bkey_err(k))
 941                 return k;
 942
 943         if (k.k->type) {
 944                 return k;
 945         } else {
 946                 struct btree_iter iter2;
 947                 struct bpos next;
 948
 949                 bch2_trans_copy_iter(&iter2, iter);
 950                 k = bch2_btree_iter_peek_upto(&iter2,
 951                                 bkey_min(bkey_min(end,
 952                                                   iter->path->l[0].b->key.k.p),
 953                                                   POS(iter->pos.inode, iter->pos.offset + U32_MAX - 1)));
 954                 next = iter2.pos;
 955                 bch2_trans_iter_exit(iter->trans, &iter2);
 956
 957                 BUG_ON(next.offset >= iter->pos.offset + U32_MAX);
 958
 959                 if (bkey_err(k))
 960                         return k;
 961
 962                 bkey_init(hole);
 963                 hole->p = iter->pos;
 964
 965                 bch2_key_resize(hole, next.offset - iter->pos.offset);
 966                 return (struct bkey_s_c) { hole, NULL };
 967         }
 968 }
 969
 970 static bool next_bucket(struct bch_fs *c, struct bpos *bucket)
 971 {
 972         struct bch_dev *ca;
 973         unsigned iter;
 974
 975         if (bch2_dev_bucket_exists(c, *bucket))
 976                 return true;
 977
 978         if (bch2_dev_exists2(c, bucket->inode)) {
 979                 ca = bch_dev_bkey_exists(c, bucket->inode);
 980
 981                 if (bucket->offset < ca->mi.first_bucket) {
 982                         bucket->offset = ca->mi.first_bucket;
 983                         return true;
 984                 }
 985
 986                 bucket->inode++;
 987                 bucket->offset = 0;
 988         }
 989
 990         rcu_read_lock();
 991         iter = bucket->inode;
 992         ca = __bch2_next_dev(c, &iter, NULL);
 993         if (ca)
 994                 bucket->offset = ca->mi.first_bucket;
 995         rcu_read_unlock();
 996
 997         return ca != NULL;
 998 }
 999
1000 struct bkey_s_c bch2_get_key_or_real_bucket_hole(struct btree_iter *iter, struct bkey *hole)
1001 {
1002         struct bch_fs *c = iter->trans->c;
1003         struct bkey_s_c k;
1004 again:
1005         k = bch2_get_key_or_hole(iter, POS_MAX, hole);
1006         if (bkey_err(k))
1007                 return k;
1008
1009         if (!k.k->type) {
1010                 struct bpos bucket = bkey_start_pos(k.k);
1011
1012                 if (!bch2_dev_bucket_exists(c, bucket)) {
1013                         if (!next_bucket(c, &bucket))
1014                                 return bkey_s_c_null;
1015
1016                         bch2_btree_iter_set_pos(iter, bucket);
1017                         goto again;
1018                 }
1019
1020                 if (!bch2_dev_bucket_exists(c, k.k->p)) {
1021                         struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
1022
1023                         bch2_key_resize(hole, ca->mi.nbuckets - bucket.offset);
1024                 }
1025         }
1026
1027         return k;
1028 }
1029
1030 static int bch2_check_alloc_key(struct btree_trans *trans,
1031                                 struct bkey_s_c alloc_k,
1032                                 struct btree_iter *alloc_iter,
1033                                 struct btree_iter *discard_iter,
1034                                 struct btree_iter *freespace_iter,
1035                                 struct btree_iter *bucket_gens_iter)
1036 {
1037         struct bch_fs *c = trans->c;
1038         struct bch_dev *ca;
1039         struct bch_alloc_v4 a_convert;
1040         const struct bch_alloc_v4 *a;
1041         unsigned discard_key_type, freespace_key_type;
1042         unsigned gens_offset;
1043         struct bkey_s_c k;
1044         struct printbuf buf = PRINTBUF;
1045         int ret;
1046
1047         if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
1048                         "alloc key for invalid device:bucket %llu:%llu",
1049                         alloc_k.k->p.inode, alloc_k.k->p.offset))
1050                 return bch2_btree_delete_at(trans, alloc_iter, 0);
1051
1052         ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
1053         if (!ca->mi.freespace_initialized)
1054                 return 0;
1055
1056         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1057
1058         discard_key_type = a->data_type == BCH_DATA_need_discard ? KEY_TYPE_set : 0;
1059         bch2_btree_iter_set_pos(discard_iter, alloc_k.k->p);
1060         k = bch2_btree_iter_peek_slot(discard_iter);
1061         ret = bkey_err(k);
1062         if (ret)
1063                 goto err;
1064
1065         if (k.k->type != discard_key_type &&
1066             (c->opts.reconstruct_alloc ||
1067              fsck_err(c, "incorrect key in need_discard btree (got %s should be %s)\n"
1068                       "  %s",
1069                       bch2_bkey_types[k.k->type],
1070                       bch2_bkey_types[discard_key_type],
1071                       (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1072                 struct bkey_i *update =
1073                         bch2_trans_kmalloc(trans, sizeof(*update));
1074
1075                 ret = PTR_ERR_OR_ZERO(update);
1076                 if (ret)
1077                         goto err;
1078
1079                 bkey_init(&update->k);
1080                 update->k.type  = discard_key_type;
1081                 update->k.p     = discard_iter->pos;
1082
1083                 ret = bch2_trans_update(trans, discard_iter, update, 0);
1084                 if (ret)
1085                         goto err;
1086         }
1087
1088         freespace_key_type = a->data_type == BCH_DATA_free ? KEY_TYPE_set : 0;
1089         bch2_btree_iter_set_pos(freespace_iter, alloc_freespace_pos(alloc_k.k->p, *a));
1090         k = bch2_btree_iter_peek_slot(freespace_iter);
1091         ret = bkey_err(k);
1092         if (ret)
1093                 goto err;
1094
1095         if (k.k->type != freespace_key_type &&
1096             (c->opts.reconstruct_alloc ||
1097              fsck_err(c, "incorrect key in freespace btree (got %s should be %s)\n"
1098                       "  %s",
1099                       bch2_bkey_types[k.k->type],
1100                       bch2_bkey_types[freespace_key_type],
1101                       (printbuf_reset(&buf),
1102                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1103                 struct bkey_i *update =
1104                         bch2_trans_kmalloc(trans, sizeof(*update));
1105
1106                 ret = PTR_ERR_OR_ZERO(update);
1107                 if (ret)
1108                         goto err;
1109
1110                 bkey_init(&update->k);
1111                 update->k.type  = freespace_key_type;
1112                 update->k.p     = freespace_iter->pos;
1113                 bch2_key_resize(&update->k, 1);
1114
1115                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1116                 if (ret)
1117                         goto err;
1118         }
1119
1120         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(alloc_k.k->p, &gens_offset));
1121         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1122         ret = bkey_err(k);
1123         if (ret)
1124                 goto err;
1125
1126         if (a->gen != alloc_gen(k, gens_offset) &&
1127             (c->opts.reconstruct_alloc ||
1128              fsck_err(c, "incorrect gen in bucket_gens btree (got %u should be %u)\n"
1129                       "  %s",
1130                       alloc_gen(k, gens_offset), a->gen,
1131                       (printbuf_reset(&buf),
1132                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1133                 struct bkey_i_bucket_gens *g =
1134                         bch2_trans_kmalloc(trans, sizeof(*g));
1135
1136                 ret = PTR_ERR_OR_ZERO(g);
1137                 if (ret)
1138                         goto err;
1139
1140                 if (k.k->type == KEY_TYPE_bucket_gens) {
1141                         bkey_reassemble(&g->k_i, k);
1142                 } else {
1143                         bkey_bucket_gens_init(&g->k_i);
1144                         g->k.p = alloc_gens_pos(alloc_k.k->p, &gens_offset);
1145                 }
1146
1147                 g->v.gens[gens_offset] = a->gen;
1148
1149                 ret = bch2_trans_update(trans, bucket_gens_iter, &g->k_i, 0);
1150                 if (ret)
1151                         goto err;
1152         }
1153 err:
1154 fsck_err:
1155         printbuf_exit(&buf);
1156         return ret;
1157 }
1158
1159 static int bch2_check_alloc_hole_freespace(struct btree_trans *trans,
1160                                  struct bpos start,
1161                                  struct bpos *end,
1162                                  struct btree_iter *freespace_iter)
1163 {
1164         struct bch_fs *c = trans->c;
1165         struct bch_dev *ca;
1166         struct bkey_s_c k;
1167         struct printbuf buf = PRINTBUF;
1168         int ret;
1169
1170         ca = bch_dev_bkey_exists(c, start.inode);
1171         if (!ca->mi.freespace_initialized)
1172                 return 0;
1173
1174         bch2_btree_iter_set_pos(freespace_iter, start);
1175
1176         k = bch2_btree_iter_peek_slot(freespace_iter);
1177         ret = bkey_err(k);
1178         if (ret)
1179                 goto err;
1180
1181         *end = bkey_min(k.k->p, *end);
1182
1183         if (k.k->type != KEY_TYPE_set &&
1184             (c->opts.reconstruct_alloc ||
1185              fsck_err(c, "hole in alloc btree missing in freespace btree\n"
1186                       "  device %llu buckets %llu-%llu",
1187                       freespace_iter->pos.inode,
1188                       freespace_iter->pos.offset,
1189                       end->offset))) {
1190                 struct bkey_i *update =
1191                         bch2_trans_kmalloc(trans, sizeof(*update));
1192
1193                 ret = PTR_ERR_OR_ZERO(update);
1194                 if (ret)
1195                         goto err;
1196
1197                 bkey_init(&update->k);
1198                 update->k.type  = KEY_TYPE_set;
1199                 update->k.p     = freespace_iter->pos;
1200                 bch2_key_resize(&update->k,
1201                                 min_t(u64, U32_MAX, end->offset -
1202                                       freespace_iter->pos.offset));
1203
1204                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1205                 if (ret)
1206                         goto err;
1207         }
1208 err:
1209 fsck_err:
1210         printbuf_exit(&buf);
1211         return ret;
1212 }
1213
1214 static int bch2_check_alloc_hole_bucket_gens(struct btree_trans *trans,
1215                                  struct bpos start,
1216                                  struct bpos *end,
1217                                  struct btree_iter *bucket_gens_iter)
1218 {
1219         struct bch_fs *c = trans->c;
1220         struct bkey_s_c k;
1221         struct printbuf buf = PRINTBUF;
1222         unsigned i, gens_offset, gens_end_offset;
1223         int ret;
1224
1225         if (c->sb.version < bcachefs_metadata_version_bucket_gens &&
1226             !c->opts.version_upgrade)
1227                 return 0;
1228
1229         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(start, &gens_offset));
1230
1231         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1232         ret = bkey_err(k);
1233         if (ret)
1234                 goto err;
1235
1236         if (bkey_cmp(alloc_gens_pos(start, &gens_offset),
1237                      alloc_gens_pos(*end,  &gens_end_offset)))
1238                 gens_end_offset = KEY_TYPE_BUCKET_GENS_NR;
1239
1240         if (k.k->type == KEY_TYPE_bucket_gens) {
1241                 struct bkey_i_bucket_gens g;
1242                 bool need_update = false;
1243
1244                 bkey_reassemble(&g.k_i, k);
1245
1246                 for (i = gens_offset; i < gens_end_offset; i++) {
1247                         if (fsck_err_on(g.v.gens[i], c,
1248                                         "hole in alloc btree at %llu:%llu with nonzero gen in bucket_gens btree (%u)",
1249                                         bucket_gens_pos_to_alloc(k.k->p, i).inode,
1250                                         bucket_gens_pos_to_alloc(k.k->p, i).offset,
1251                                         g.v.gens[i])) {
1252                                 g.v.gens[i] = 0;
1253                                 need_update = true;
1254                         }
1255                 }
1256
1257                 if (need_update) {
1258                         struct bkey_i *k = bch2_trans_kmalloc(trans, sizeof(g));
1259
1260                         ret = PTR_ERR_OR_ZERO(k);
1261                         if (ret)
1262                                 goto err;
1263
1264                         memcpy(k, &g, sizeof(g));
1265
1266                         ret = bch2_trans_update(trans, bucket_gens_iter, k, 0);
1267                         if (ret)
1268                                 goto err;
1269                 }
1270         }
1271
1272         *end = bkey_min(*end, bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0));
1273 err:
1274 fsck_err:
1275         printbuf_exit(&buf);
1276         return ret;
1277 }
1278
1279 static int bch2_check_discard_freespace_key(struct btree_trans *trans,
1280                                             struct btree_iter *iter)
1281 {
1282         struct bch_fs *c = trans->c;
1283         struct btree_iter alloc_iter;
1284         struct bkey_s_c alloc_k;
1285         struct bch_alloc_v4 a_convert;
1286         const struct bch_alloc_v4 *a;
1287         u64 genbits;
1288         struct bpos pos;
1289         enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
1290                 ? BCH_DATA_need_discard
1291                 : BCH_DATA_free;
1292         struct printbuf buf = PRINTBUF;
1293         int ret;
1294
1295         pos = iter->pos;
1296         pos.offset &= ~(~0ULL << 56);
1297         genbits = iter->pos.offset & (~0ULL << 56);
1298
1299         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
1300
1301         if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
1302                         "entry in %s btree for nonexistant dev:bucket %llu:%llu",
1303                         bch2_btree_ids[iter->btree_id], pos.inode, pos.offset))
1304                 goto delete;
1305
1306         alloc_k = bch2_btree_iter_peek_slot(&alloc_iter);
1307         ret = bkey_err(alloc_k);
1308         if (ret)
1309                 goto err;
1310
1311         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1312
1313         if (fsck_err_on(a->data_type != state ||
1314                         (state == BCH_DATA_free &&
1315                          genbits != alloc_freespace_genbits(*a)), c,
1316                         "%s\n  incorrectly set in %s index (free %u, genbits %llu should be %llu)",
1317                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
1318                         bch2_btree_ids[iter->btree_id],
1319                         a->data_type == state,
1320                         genbits >> 56, alloc_freespace_genbits(*a) >> 56))
1321                 goto delete;
1322 out:
1323 err:
1324 fsck_err:
1325         bch2_trans_iter_exit(trans, &alloc_iter);
1326         printbuf_exit(&buf);
1327         return ret;
1328 delete:
1329         ret = bch2_btree_delete_extent_at(trans, iter,
1330                         iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0);
1331         goto out;
1332 }
1333
1334 /*
1335  * We've already checked that generation numbers in the bucket_gens btree are
1336  * valid for buckets that exist; this just checks for keys for nonexistent
1337  * buckets.
1338  */
1339 static int bch2_check_bucket_gens_key(struct btree_trans *trans,
1340                                       struct btree_iter *iter,
1341                                       struct bkey_s_c k)
1342 {
1343         struct bch_fs *c = trans->c;
1344         struct bkey_i_bucket_gens g;
1345         struct bch_dev *ca;
1346         u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
1347         u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
1348         u64 b;
1349         bool need_update = false;
1350         struct printbuf buf = PRINTBUF;
1351         int ret = 0;
1352
1353         BUG_ON(k.k->type != KEY_TYPE_bucket_gens);
1354         bkey_reassemble(&g.k_i, k);
1355
1356         if (fsck_err_on(!bch2_dev_exists2(c, k.k->p.inode), c,
1357                         "bucket_gens key for invalid device:\n  %s",
1358                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1359                 ret = bch2_btree_delete_at(trans, iter, 0);
1360                 goto out;
1361         }
1362
1363         ca = bch_dev_bkey_exists(c, k.k->p.inode);
1364         if (fsck_err_on(end <= ca->mi.first_bucket ||
1365                         start >= ca->mi.nbuckets, c,
1366                         "bucket_gens key for invalid buckets:\n  %s",
1367                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1368                 ret = bch2_btree_delete_at(trans, iter, 0);
1369                 goto out;
1370         }
1371
1372         for (b = start; b < ca->mi.first_bucket; b++)
1373                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1374                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1375                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1376                         need_update = true;
1377                 }
1378
1379         for (b = ca->mi.nbuckets; b < end; b++)
1380                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1381                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1382                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1383                         need_update = true;
1384                 }
1385
1386         if (need_update) {
1387                 struct bkey_i *k;
1388
1389                 k = bch2_trans_kmalloc(trans, sizeof(g));
1390                 ret = PTR_ERR_OR_ZERO(k);
1391                 if (ret)
1392                         goto out;
1393
1394                 memcpy(k, &g, sizeof(g));
1395                 ret = bch2_trans_update(trans, iter, k, 0);
1396         }
1397 out:
1398 fsck_err:
1399         printbuf_exit(&buf);
1400         return ret;
1401 }
1402
1403 int bch2_check_alloc_info(struct bch_fs *c)
1404 {
1405         struct btree_trans trans;
1406         struct btree_iter iter, discard_iter, freespace_iter, bucket_gens_iter;
1407         struct bkey hole;
1408         struct bkey_s_c k;
1409         int ret = 0;
1410
1411         bch2_trans_init(&trans, c, 0, 0);
1412
1413         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc, POS_MIN,
1414                              BTREE_ITER_PREFETCH);
1415         bch2_trans_iter_init(&trans, &discard_iter, BTREE_ID_need_discard, POS_MIN,
1416                              BTREE_ITER_PREFETCH);
1417         bch2_trans_iter_init(&trans, &freespace_iter, BTREE_ID_freespace, POS_MIN,
1418                              BTREE_ITER_PREFETCH);
1419         bch2_trans_iter_init(&trans, &bucket_gens_iter, BTREE_ID_bucket_gens, POS_MIN,
1420                              BTREE_ITER_PREFETCH);
1421
1422         while (1) {
1423                 struct bpos next;
1424
1425                 bch2_trans_begin(&trans);
1426
1427                 k = bch2_get_key_or_real_bucket_hole(&iter, &hole);
1428                 ret = bkey_err(k);
1429                 if (ret)
1430                         goto bkey_err;
1431
1432                 if (!k.k)
1433                         break;
1434
1435                 if (k.k->type) {
1436                         next = bpos_nosnap_successor(k.k->p);
1437
1438                         ret = bch2_check_alloc_key(&trans,
1439                                                    k, &iter,
1440                                                    &discard_iter,
1441                                                    &freespace_iter,
1442                                                    &bucket_gens_iter);
1443                         if (ret)
1444                                 goto bkey_err;
1445                 } else {
1446                         next = k.k->p;
1447
1448                         ret = bch2_check_alloc_hole_freespace(&trans,
1449                                                     bkey_start_pos(k.k),
1450                                                     &next,
1451                                                     &freespace_iter) ?:
1452                                 bch2_check_alloc_hole_bucket_gens(&trans,
1453                                                     bkey_start_pos(k.k),
1454                                                     &next,
1455                                                     &bucket_gens_iter);
1456                         if (ret)
1457                                 goto bkey_err;
1458                 }
1459
1460                 ret = bch2_trans_commit(&trans, NULL, NULL,
1461                                         BTREE_INSERT_NOFAIL|
1462                                         BTREE_INSERT_LAZY_RW);
1463                 if (ret)
1464                         goto bkey_err;
1465
1466                 bch2_btree_iter_set_pos(&iter, next);
1467 bkey_err:
1468                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1469                         continue;
1470                 if (ret)
1471                         break;
1472         }
1473         bch2_trans_iter_exit(&trans, &bucket_gens_iter);
1474         bch2_trans_iter_exit(&trans, &freespace_iter);
1475         bch2_trans_iter_exit(&trans, &discard_iter);
1476         bch2_trans_iter_exit(&trans, &iter);
1477
1478         if (ret < 0)
1479                 goto err;
1480
1481         ret = for_each_btree_key_commit(&trans, iter,
1482                         BTREE_ID_need_discard, POS_MIN,
1483                         BTREE_ITER_PREFETCH, k,
1484                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1485                 bch2_check_discard_freespace_key(&trans, &iter)) ?:
1486               for_each_btree_key_commit(&trans, iter,
1487                         BTREE_ID_freespace, POS_MIN,
1488                         BTREE_ITER_PREFETCH, k,
1489                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1490                 bch2_check_discard_freespace_key(&trans, &iter)) ?:
1491               for_each_btree_key_commit(&trans, iter,
1492                         BTREE_ID_bucket_gens, POS_MIN,
1493                         BTREE_ITER_PREFETCH, k,
1494                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1495                 bch2_check_bucket_gens_key(&trans, &iter, k));
1496 err:
1497         bch2_trans_exit(&trans);
1498         return ret < 0 ? ret : 0;
1499 }
1500
1501 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
1502                                        struct btree_iter *alloc_iter)
1503 {
1504         struct bch_fs *c = trans->c;
1505         struct btree_iter lru_iter;
1506         struct bch_alloc_v4 a_convert;
1507         const struct bch_alloc_v4 *a;
1508         struct bkey_s_c alloc_k, k;
1509         struct printbuf buf = PRINTBUF;
1510         int ret;
1511
1512         alloc_k = bch2_btree_iter_peek(alloc_iter);
1513         if (!alloc_k.k)
1514                 return 0;
1515
1516         ret = bkey_err(alloc_k);
1517         if (ret)
1518                 return ret;
1519
1520         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1521
1522         if (a->data_type != BCH_DATA_cached)
1523                 return 0;
1524
1525         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
1526                              lru_pos(alloc_k.k->p.inode,
1527                                      bucket_to_u64(alloc_k.k->p),
1528                                      a->io_time[READ]), 0);
1529         k = bch2_btree_iter_peek_slot(&lru_iter);
1530         ret = bkey_err(k);
1531         if (ret)
1532                 goto err;
1533
1534         if (fsck_err_on(!a->io_time[READ], c,
1535                         "cached bucket with read_time 0\n"
1536                         "  %s",
1537                 (printbuf_reset(&buf),
1538                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) ||
1539             fsck_err_on(k.k->type != KEY_TYPE_set, c,
1540                         "missing lru entry\n"
1541                         "  %s",
1542                         (printbuf_reset(&buf),
1543                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
1544                 u64 read_time = a->io_time[READ] ?:
1545                         atomic64_read(&c->io_clock[READ].now);
1546
1547                 ret = bch2_lru_set(trans,
1548                                    alloc_k.k->p.inode,
1549                                    bucket_to_u64(alloc_k.k->p),
1550                                    read_time);
1551                 if (ret)
1552                         goto err;
1553
1554                 if (a->io_time[READ] != read_time) {
1555                         struct bkey_i_alloc_v4 *a_mut =
1556                                 bch2_alloc_to_v4_mut(trans, alloc_k);
1557                         ret = PTR_ERR_OR_ZERO(a_mut);
1558                         if (ret)
1559                                 goto err;
1560
1561                         a_mut->v.io_time[READ] = read_time;
1562                         ret = bch2_trans_update(trans, alloc_iter,
1563                                                 &a_mut->k_i, BTREE_TRIGGER_NORUN);
1564                         if (ret)
1565                                 goto err;
1566                 }
1567         }
1568 err:
1569 fsck_err:
1570         bch2_trans_iter_exit(trans, &lru_iter);
1571         printbuf_exit(&buf);
1572         return ret;
1573 }
1574
1575 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
1576 {
1577         struct btree_trans trans;
1578         struct btree_iter iter;
1579         struct bkey_s_c k;
1580         int ret = 0;
1581
1582         bch2_trans_init(&trans, c, 0, 0);
1583
1584         for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1585                         POS_MIN, BTREE_ITER_PREFETCH, k,
1586                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1587                 bch2_check_alloc_to_lru_ref(&trans, &iter));
1588
1589         bch2_trans_exit(&trans);
1590         return ret < 0 ? ret : 0;
1591 }
1592
1593 static int bch2_discard_one_bucket(struct btree_trans *trans,
1594                                    struct btree_iter *need_discard_iter,
1595                                    struct bpos *discard_pos_done,
1596                                    u64 *seen,
1597                                    u64 *open,
1598                                    u64 *need_journal_commit,
1599                                    u64 *discarded)
1600 {
1601         struct bch_fs *c = trans->c;
1602         struct bpos pos = need_discard_iter->pos;
1603         struct btree_iter iter = { NULL };
1604         struct bkey_s_c k;
1605         struct bch_dev *ca;
1606         struct bkey_i_alloc_v4 *a;
1607         struct printbuf buf = PRINTBUF;
1608         int ret = 0;
1609
1610         ca = bch_dev_bkey_exists(c, pos.inode);
1611         if (!percpu_ref_tryget(&ca->io_ref)) {
1612                 bch2_btree_iter_set_pos(need_discard_iter, POS(pos.inode + 1, 0));
1613                 return 0;
1614         }
1615
1616         if (bch2_bucket_is_open_safe(c, pos.inode, pos.offset)) {
1617                 (*open)++;
1618                 goto out;
1619         }
1620
1621         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
1622                         c->journal.flushed_seq_ondisk,
1623                         pos.inode, pos.offset)) {
1624                 (*need_journal_commit)++;
1625                 goto out;
1626         }
1627
1628         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
1629                              need_discard_iter->pos,
1630                              BTREE_ITER_CACHED);
1631         k = bch2_btree_iter_peek_slot(&iter);
1632         ret = bkey_err(k);
1633         if (ret)
1634                 goto out;
1635
1636         a = bch2_alloc_to_v4_mut(trans, k);
1637         ret = PTR_ERR_OR_ZERO(a);
1638         if (ret)
1639                 goto out;
1640
1641         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
1642                 a->v.gen++;
1643                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1644                 goto write;
1645         }
1646
1647         if (a->v.journal_seq > c->journal.flushed_seq_ondisk) {
1648                 if (test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
1649                         bch2_trans_inconsistent(trans,
1650                                 "clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
1651                                 "%s",
1652                                 a->v.journal_seq,
1653                                 c->journal.flushed_seq_ondisk,
1654                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1655                         ret = -EIO;
1656                 }
1657                 goto out;
1658         }
1659
1660         if (a->v.data_type != BCH_DATA_need_discard) {
1661                 if (test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
1662                         bch2_trans_inconsistent(trans,
1663                                 "bucket incorrectly set in need_discard btree\n"
1664                                 "%s",
1665                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1666                         ret = -EIO;
1667                 }
1668
1669                 goto out;
1670         }
1671
1672         if (!bkey_eq(*discard_pos_done, iter.pos) &&
1673             ca->mi.discard && !c->opts.nochanges) {
1674                 /*
1675                  * This works without any other locks because this is the only
1676                  * thread that removes items from the need_discard tree
1677                  */
1678                 bch2_trans_unlock(trans);
1679                 blkdev_issue_discard(ca->disk_sb.bdev,
1680                                      k.k->p.offset * ca->mi.bucket_size,
1681                                      ca->mi.bucket_size,
1682                                      GFP_KERNEL);
1683                 *discard_pos_done = iter.pos;
1684
1685                 ret = bch2_trans_relock_notrace(trans);
1686                 if (ret)
1687                         goto out;
1688         }
1689
1690         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
1691         a->v.data_type = alloc_data_type(a->v, a->v.data_type);
1692 write:
1693         ret =   bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1694                 bch2_trans_commit(trans, NULL, NULL,
1695                                   BTREE_INSERT_USE_RESERVE|BTREE_INSERT_NOFAIL);
1696         if (ret)
1697                 goto out;
1698
1699         this_cpu_inc(c->counters[BCH_COUNTER_bucket_discard]);
1700         (*discarded)++;
1701 out:
1702         (*seen)++;
1703         bch2_trans_iter_exit(trans, &iter);
1704         percpu_ref_put(&ca->io_ref);
1705         printbuf_exit(&buf);
1706         return ret;
1707 }
1708
1709 static void bch2_do_discards_work(struct work_struct *work)
1710 {
1711         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
1712         struct btree_trans trans;
1713         struct btree_iter iter;
1714         struct bkey_s_c k;
1715         u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
1716         struct bpos discard_pos_done = POS_MAX;
1717         int ret;
1718
1719         bch2_trans_init(&trans, c, 0, 0);
1720
1721         /*
1722          * We're doing the commit in bch2_discard_one_bucket instead of using
1723          * for_each_btree_key_commit() so that we can increment counters after
1724          * successful commit:
1725          */
1726         ret = for_each_btree_key2(&trans, iter,
1727                         BTREE_ID_need_discard, POS_MIN, 0, k,
1728                 bch2_discard_one_bucket(&trans, &iter, &discard_pos_done,
1729                                         &seen,
1730                                         &open,
1731                                         &need_journal_commit,
1732                                         &discarded));
1733
1734         bch2_trans_exit(&trans);
1735
1736         if (need_journal_commit * 2 > seen)
1737                 bch2_journal_flush_async(&c->journal, NULL);
1738
1739         bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1740
1741         trace_discard_buckets(c, seen, open, need_journal_commit, discarded,
1742                               bch2_err_str(ret));
1743 }
1744
1745 void bch2_do_discards(struct bch_fs *c)
1746 {
1747         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_discard) &&
1748             !queue_work(system_long_wq, &c->discard_work))
1749                 bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1750 }
1751
1752 static int invalidate_one_bucket(struct btree_trans *trans,
1753                                  struct btree_iter *lru_iter,
1754                                  struct bkey_s_c lru_k,
1755                                  s64 *nr_to_invalidate)
1756 {
1757         struct bch_fs *c = trans->c;
1758         struct btree_iter alloc_iter = { NULL };
1759         struct bkey_i_alloc_v4 *a = NULL;
1760         struct printbuf buf = PRINTBUF;
1761         struct bpos bucket = u64_to_bucket(lru_k.k->p.offset);
1762         unsigned cached_sectors;
1763         int ret = 0;
1764
1765         if (*nr_to_invalidate <= 0)
1766                 return 1;
1767
1768         if (!bch2_dev_bucket_exists(c, bucket)) {
1769                 prt_str(&buf, "lru entry points to invalid bucket");
1770                 goto err;
1771         }
1772
1773         a = bch2_trans_start_alloc_update(trans, &alloc_iter, bucket);
1774         ret = PTR_ERR_OR_ZERO(a);
1775         if (ret)
1776                 goto out;
1777
1778         if (lru_pos_time(lru_iter->pos) != alloc_lru_idx(a->v)) {
1779                 prt_str(&buf, "alloc key does not point back to lru entry when invalidating bucket:");
1780                 goto err;
1781         }
1782
1783         if (a->v.data_type != BCH_DATA_cached) {
1784                 prt_str(&buf, "lru entry points to non cached bucket:");
1785                 goto err;
1786         }
1787
1788         if (!a->v.cached_sectors)
1789                 bch_err(c, "invalidating empty bucket, confused");
1790
1791         cached_sectors = a->v.cached_sectors;
1792
1793         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1794         a->v.gen++;
1795         a->v.data_type          = 0;
1796         a->v.dirty_sectors      = 0;
1797         a->v.cached_sectors     = 0;
1798         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1799         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1800
1801         ret =   bch2_trans_update(trans, &alloc_iter, &a->k_i,
1802                                 BTREE_TRIGGER_BUCKET_INVALIDATE) ?:
1803                 bch2_trans_commit(trans, NULL, NULL,
1804                                   BTREE_INSERT_USE_RESERVE|BTREE_INSERT_NOFAIL);
1805         if (ret)
1806                 goto out;
1807
1808         trace_and_count(c, bucket_invalidate, c, bucket.inode, bucket.offset, cached_sectors);
1809         --*nr_to_invalidate;
1810 out:
1811         bch2_trans_iter_exit(trans, &alloc_iter);
1812         printbuf_exit(&buf);
1813         return ret;
1814 err:
1815         prt_str(&buf, "\n  lru key: ");
1816         bch2_bkey_val_to_text(&buf, c, lru_k);
1817
1818         prt_str(&buf, "\n  lru entry: ");
1819         bch2_lru_pos_to_text(&buf, lru_iter->pos);
1820
1821         prt_str(&buf, "\n  alloc key: ");
1822         if (!a)
1823                 bch2_bpos_to_text(&buf, bucket);
1824         else
1825                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
1826
1827         bch_err(c, "%s", buf.buf);
1828         if (test_bit(BCH_FS_CHECK_LRUS_DONE, &c->flags)) {
1829                 bch2_inconsistent_error(c);
1830                 ret = -EINVAL;
1831         }
1832
1833         goto out;
1834 }
1835
1836 static void bch2_do_invalidates_work(struct work_struct *work)
1837 {
1838         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1839         struct bch_dev *ca;
1840         struct btree_trans trans;
1841         struct btree_iter iter;
1842         struct bkey_s_c k;
1843         unsigned i;
1844         int ret = 0;
1845
1846         bch2_trans_init(&trans, c, 0, 0);
1847
1848         for_each_member_device(ca, c, i) {
1849                 s64 nr_to_invalidate =
1850                         should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
1851
1852                 ret = for_each_btree_key2_upto(&trans, iter, BTREE_ID_lru,
1853                                 lru_pos(ca->dev_idx, 0, 0),
1854                                 lru_pos(ca->dev_idx, U64_MAX, LRU_TIME_MAX),
1855                                 BTREE_ITER_INTENT, k,
1856                         invalidate_one_bucket(&trans, &iter, k, &nr_to_invalidate));
1857
1858                 if (ret < 0) {
1859                         percpu_ref_put(&ca->ref);
1860                         break;
1861                 }
1862         }
1863
1864         bch2_trans_exit(&trans);
1865         bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1866 }
1867
1868 void bch2_do_invalidates(struct bch_fs *c)
1869 {
1870         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_invalidate) &&
1871             !queue_work(system_long_wq, &c->invalidate_work))
1872                 bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1873 }
1874
1875 static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca)
1876 {
1877         struct btree_trans trans;
1878         struct btree_iter iter;
1879         struct bkey_s_c k;
1880         struct bkey hole;
1881         struct bpos end = POS(ca->dev_idx, ca->mi.nbuckets);
1882         struct bch_member *m;
1883         int ret;
1884
1885         bch2_trans_init(&trans, c, 0, 0);
1886
1887         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc,
1888                              POS(ca->dev_idx, ca->mi.first_bucket),
1889                              BTREE_ITER_PREFETCH);
1890         /*
1891          * Scan the alloc btree for every bucket on @ca, and add buckets to the
1892          * freespace/need_discard/need_gc_gens btrees as needed:
1893          */
1894         while (1) {
1895                 bch2_trans_begin(&trans);
1896
1897                 if (bkey_ge(iter.pos, end)) {
1898                         ret = 0;
1899                         break;
1900                 }
1901
1902                 k = bch2_get_key_or_hole(&iter, end, &hole);
1903                 ret = bkey_err(k);
1904                 if (ret)
1905                         goto bkey_err;
1906
1907                 if (k.k->type) {
1908                         /*
1909                          * We process live keys in the alloc btree one at a
1910                          * time:
1911                          */
1912                         struct bch_alloc_v4 a_convert;
1913                         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1914
1915                         ret =   bch2_bucket_do_index(&trans, k, a, true) ?:
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_advance(&iter);
1923                 } else {
1924                         struct bkey_i *freespace;
1925
1926                         freespace = bch2_trans_kmalloc(&trans, sizeof(*freespace));
1927                         ret = PTR_ERR_OR_ZERO(freespace);
1928                         if (ret)
1929                                 goto bkey_err;
1930
1931                         bkey_init(&freespace->k);
1932                         freespace->k.type       = KEY_TYPE_set;
1933                         freespace->k.p          = k.k->p;
1934                         freespace->k.size       = k.k->size;
1935
1936                         ret = __bch2_btree_insert(&trans, BTREE_ID_freespace, freespace, 0) ?:
1937                                 bch2_trans_commit(&trans, NULL, NULL,
1938                                                   BTREE_INSERT_LAZY_RW|
1939                                                   BTREE_INSERT_NOFAIL);
1940                         if (ret)
1941                                 goto bkey_err;
1942
1943                         bch2_btree_iter_set_pos(&iter, k.k->p);
1944                 }
1945 bkey_err:
1946                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1947                         continue;
1948                 if (ret)
1949                         break;
1950         }
1951
1952         bch2_trans_iter_exit(&trans, &iter);
1953         bch2_trans_exit(&trans);
1954
1955         if (ret < 0) {
1956                 bch_err(ca, "error initializing free space: %s", bch2_err_str(ret));
1957                 return ret;
1958         }
1959
1960         mutex_lock(&c->sb_lock);
1961         m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx;
1962         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1963         mutex_unlock(&c->sb_lock);
1964
1965         return 0;
1966 }
1967
1968 int bch2_fs_freespace_init(struct bch_fs *c)
1969 {
1970         struct bch_dev *ca;
1971         unsigned i;
1972         int ret = 0;
1973         bool doing_init = false;
1974
1975         /*
1976          * We can crash during the device add path, so we need to check this on
1977          * every mount:
1978          */
1979
1980         for_each_member_device(ca, c, i) {
1981                 if (ca->mi.freespace_initialized)
1982                         continue;
1983
1984                 if (!doing_init) {
1985                         bch_info(c, "initializing freespace");
1986                         doing_init = true;
1987                 }
1988
1989                 ret = bch2_dev_freespace_init(c, ca);
1990                 if (ret) {
1991                         percpu_ref_put(&ca->ref);
1992                         return ret;
1993                 }
1994         }
1995
1996         if (doing_init) {
1997                 mutex_lock(&c->sb_lock);
1998                 bch2_write_super(c);
1999                 mutex_unlock(&c->sb_lock);
2000
2001                 bch_verbose(c, "done initializing freespace");
2002         }
2003
2004         return ret;
2005 }
2006
2007 /* Bucket IO clocks: */
2008
2009 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
2010                               size_t bucket_nr, int rw)
2011 {
2012         struct bch_fs *c = trans->c;
2013         struct btree_iter iter;
2014         struct bkey_i_alloc_v4 *a;
2015         u64 now;
2016         int ret = 0;
2017
2018         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
2019         ret = PTR_ERR_OR_ZERO(a);
2020         if (ret)
2021                 return ret;
2022
2023         now = atomic64_read(&c->io_clock[rw].now);
2024         if (a->v.io_time[rw] == now)
2025                 goto out;
2026
2027         a->v.io_time[rw] = now;
2028
2029         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
2030                 bch2_trans_commit(trans, NULL, NULL, 0);
2031 out:
2032         bch2_trans_iter_exit(trans, &iter);
2033         return ret;
2034 }
2035
2036 /* Startup/shutdown (ro/rw): */
2037
2038 void bch2_recalc_capacity(struct bch_fs *c)
2039 {
2040         struct bch_dev *ca;
2041         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
2042         unsigned bucket_size_max = 0;
2043         unsigned long ra_pages = 0;
2044         unsigned i;
2045
2046         lockdep_assert_held(&c->state_lock);
2047
2048         for_each_online_member(ca, c, i) {
2049                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
2050
2051                 ra_pages += bdi->ra_pages;
2052         }
2053
2054         bch2_set_ra_pages(c, ra_pages);
2055
2056         for_each_rw_member(ca, c, i) {
2057                 u64 dev_reserve = 0;
2058
2059                 /*
2060                  * We need to reserve buckets (from the number
2061                  * of currently available buckets) against
2062                  * foreground writes so that mainly copygc can
2063                  * make forward progress.
2064                  *
2065                  * We need enough to refill the various reserves
2066                  * from scratch - copygc will use its entire
2067                  * reserve all at once, then run against when
2068                  * its reserve is refilled (from the formerly
2069                  * available buckets).
2070                  *
2071                  * This reserve is just used when considering if
2072                  * allocations for foreground writes must wait -
2073                  * not -ENOSPC calculations.
2074                  */
2075
2076                 dev_reserve += ca->nr_btree_reserve * 2;
2077                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
2078
2079                 dev_reserve += 1;       /* btree write point */
2080                 dev_reserve += 1;       /* copygc write point */
2081                 dev_reserve += 1;       /* rebalance write point */
2082
2083                 dev_reserve *= ca->mi.bucket_size;
2084
2085                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
2086                                              ca->mi.first_bucket);
2087
2088                 reserved_sectors += dev_reserve * 2;
2089
2090                 bucket_size_max = max_t(unsigned, bucket_size_max,
2091                                         ca->mi.bucket_size);
2092         }
2093
2094         gc_reserve = c->opts.gc_reserve_bytes
2095                 ? c->opts.gc_reserve_bytes >> 9
2096                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
2097
2098         reserved_sectors = max(gc_reserve, reserved_sectors);
2099
2100         reserved_sectors = min(reserved_sectors, capacity);
2101
2102         c->capacity = capacity - reserved_sectors;
2103
2104         c->bucket_size_max = bucket_size_max;
2105
2106         /* Wake up case someone was waiting for buckets */
2107         closure_wake_up(&c->freelist_wait);
2108 }
2109
2110 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
2111 {
2112         struct open_bucket *ob;
2113         bool ret = false;
2114
2115         for (ob = c->open_buckets;
2116              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
2117              ob++) {
2118                 spin_lock(&ob->lock);
2119                 if (ob->valid && !ob->on_partial_list &&
2120                     ob->dev == ca->dev_idx)
2121                         ret = true;
2122                 spin_unlock(&ob->lock);
2123         }
2124
2125         return ret;
2126 }
2127
2128 /* device goes ro: */
2129 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
2130 {
2131         unsigned i;
2132
2133         /* First, remove device from allocation groups: */
2134
2135         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2136                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
2137
2138         /*
2139          * Capacity is calculated based off of devices in allocation groups:
2140          */
2141         bch2_recalc_capacity(c);
2142
2143         /* Next, close write points that point to this device... */
2144         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
2145                 bch2_writepoint_stop(c, ca, &c->write_points[i]);
2146
2147         bch2_writepoint_stop(c, ca, &c->copygc_write_point);
2148         bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
2149         bch2_writepoint_stop(c, ca, &c->btree_write_point);
2150
2151         mutex_lock(&c->btree_reserve_cache_lock);
2152         while (c->btree_reserve_cache_nr) {
2153                 struct btree_alloc *a =
2154                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
2155
2156                 bch2_open_buckets_put(c, &a->ob);
2157         }
2158         mutex_unlock(&c->btree_reserve_cache_lock);
2159
2160         while (1) {
2161                 struct open_bucket *ob;
2162
2163                 spin_lock(&c->freelist_lock);
2164                 if (!ca->open_buckets_partial_nr) {
2165                         spin_unlock(&c->freelist_lock);
2166                         break;
2167                 }
2168                 ob = c->open_buckets +
2169                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
2170                 ob->on_partial_list = false;
2171                 spin_unlock(&c->freelist_lock);
2172
2173                 bch2_open_bucket_put(c, ob);
2174         }
2175
2176         bch2_ec_stop_dev(c, ca);
2177
2178         /*
2179          * Wake up threads that were blocked on allocation, so they can notice
2180          * the device can no longer be removed and the capacity has changed:
2181          */
2182         closure_wake_up(&c->freelist_wait);
2183
2184         /*
2185          * journal_res_get() can block waiting for free space in the journal -
2186          * it needs to notice there may not be devices to allocate from anymore:
2187          */
2188         wake_up(&c->journal.wait);
2189
2190         /* Now wait for any in flight writes: */
2191
2192         closure_wait_event(&c->open_buckets_wait,
2193                            !bch2_dev_has_open_write_point(c, ca));
2194 }
2195
2196 /* device goes rw: */
2197 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
2198 {
2199         unsigned i;
2200
2201         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2202                 if (ca->mi.data_allowed & (1 << i))
2203                         set_bit(ca->dev_idx, c->rw_devs[i].d);
2204 }
2205
2206 void bch2_fs_allocator_background_init(struct bch_fs *c)
2207 {
2208         spin_lock_init(&c->freelist_lock);
2209         INIT_WORK(&c->discard_work, bch2_do_discards_work);
2210         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
2211 }