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
 418         if (k.k->type == KEY_TYPE_alloc_v4) {
 419                 struct bkey_s_c_alloc_v4 a_raw = bkey_s_c_to_alloc_v4(k);
 420                 const struct bch_backpointer *bps = alloc_v4_backpointers_c(a_raw.v);
 421
 422                 prt_printf(out, "bp_start          %llu", BCH_ALLOC_V4_BACKPOINTERS_START(a_raw.v));
 423                 prt_newline(out);
 424
 425                 prt_printf(out, "backpointers:     %llu", BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v));
 426                 printbuf_indent_add(out, 2);
 427
 428                 for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v); i++) {
 429                         prt_newline(out);
 430                         bch2_backpointer_to_text(out, &bps[i]);
 431                 }
 432
 433                 printbuf_indent_sub(out, 2);
 434         }
 435
 436         printbuf_indent_sub(out, 2);
 437 }
 438
 439 void __bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out)
 440 {
 441         if (k.k->type == KEY_TYPE_alloc_v4) {
 442                 void *src, *dst;
 443
 444                 *out = *bkey_s_c_to_alloc_v4(k).v;
 445
 446                 src = alloc_v4_backpointers(out);
 447                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 448                 dst = alloc_v4_backpointers(out);
 449
 450                 if (src < dst)
 451                         memset(src, 0, dst - src);
 452         } else {
 453                 struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 454
 455                 *out = (struct bch_alloc_v4) {
 456                         .journal_seq            = u.journal_seq,
 457                         .flags                  = u.need_discard,
 458                         .gen                    = u.gen,
 459                         .oldest_gen             = u.oldest_gen,
 460                         .data_type              = u.data_type,
 461                         .stripe_redundancy      = u.stripe_redundancy,
 462                         .dirty_sectors          = u.dirty_sectors,
 463                         .cached_sectors         = u.cached_sectors,
 464                         .io_time[READ]          = u.read_time,
 465                         .io_time[WRITE]         = u.write_time,
 466                         .stripe                 = u.stripe,
 467                 };
 468
 469                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 470         }
 471 }
 472
 473 static noinline struct bkey_i_alloc_v4 *
 474 __bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 475 {
 476         struct bkey_i_alloc_v4 *ret;
 477         if (k.k->type == KEY_TYPE_alloc_v4) {
 478                 struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 479                 unsigned bytes = sizeof(struct bkey_i_alloc_v4) +
 480                         BCH_ALLOC_V4_NR_BACKPOINTERS(a.v) *
 481                         sizeof(struct bch_backpointer);
 482                 void *src, *dst;
 483
 484                 /*
 485                  * Reserve space for one more backpointer here:
 486                  * Not sketchy at doing it this way, nope...
 487                  */
 488                 ret = bch2_trans_kmalloc(trans, bytes + sizeof(struct bch_backpointer));
 489                 if (IS_ERR(ret))
 490                         return ret;
 491
 492                 bkey_reassemble(&ret->k_i, k);
 493
 494                 src = alloc_v4_backpointers(&ret->v);
 495                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(&ret->v, BCH_ALLOC_V4_U64s);
 496                 dst = alloc_v4_backpointers(&ret->v);
 497
 498                 memmove(dst, src, BCH_ALLOC_V4_NR_BACKPOINTERS(&ret->v) *
 499                         sizeof(struct bch_backpointer));
 500                 if (src < dst)
 501                         memset(src, 0, dst - src);
 502                 set_alloc_v4_u64s(ret);
 503         } else {
 504                 ret = bch2_trans_kmalloc(trans, sizeof(struct bkey_i_alloc_v4) +
 505                                          sizeof(struct bch_backpointer));
 506                 if (IS_ERR(ret))
 507                         return ret;
 508
 509                 bkey_alloc_v4_init(&ret->k_i);
 510                 ret->k.p = k.k->p;
 511                 bch2_alloc_to_v4(k, &ret->v);
 512         }
 513         return ret;
 514 }
 515
 516 static inline struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut_inlined(struct btree_trans *trans, struct bkey_s_c k)
 517 {
 518         if (likely(k.k->type == KEY_TYPE_alloc_v4) &&
 519             BCH_ALLOC_V4_BACKPOINTERS_START(bkey_s_c_to_alloc_v4(k).v) == BCH_ALLOC_V4_U64s) {
 520                 /*
 521                  * Reserve space for one more backpointer here:
 522                  * Not sketchy at doing it this way, nope...
 523                  */
 524                 struct bkey_i_alloc_v4 *ret =
 525                         bch2_trans_kmalloc_nomemzero(trans, bkey_bytes(k.k) + sizeof(struct bch_backpointer));
 526                 if (!IS_ERR(ret))
 527                         bkey_reassemble(&ret->k_i, k);
 528                 return ret;
 529         }
 530
 531         return __bch2_alloc_to_v4_mut(trans, k);
 532 }
 533
 534 struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 535 {
 536         return bch2_alloc_to_v4_mut_inlined(trans, k);
 537 }
 538
 539 struct bkey_i_alloc_v4 *
 540 bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
 541                               struct bpos pos)
 542 {
 543         struct bkey_s_c k;
 544         struct bkey_i_alloc_v4 *a;
 545         int ret;
 546
 547         bch2_trans_iter_init(trans, iter, BTREE_ID_alloc, pos,
 548                              BTREE_ITER_WITH_UPDATES|
 549                              BTREE_ITER_CACHED|
 550                              BTREE_ITER_INTENT);
 551         k = bch2_btree_iter_peek_slot(iter);
 552         ret = bkey_err(k);
 553         if (unlikely(ret))
 554                 goto err;
 555
 556         a = bch2_alloc_to_v4_mut_inlined(trans, k);
 557         ret = PTR_ERR_OR_ZERO(a);
 558         if (unlikely(ret))
 559                 goto err;
 560         return a;
 561 err:
 562         bch2_trans_iter_exit(trans, iter);
 563         return ERR_PTR(ret);
 564 }
 565
 566 int bch2_alloc_read(struct bch_fs *c)
 567 {
 568         struct btree_trans trans;
 569         struct btree_iter iter;
 570         struct bkey_s_c k;
 571         struct bch_alloc_v4 a;
 572         struct bch_dev *ca;
 573         int ret;
 574
 575         bch2_trans_init(&trans, c, 0, 0);
 576
 577         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 578                            BTREE_ITER_PREFETCH, k, ret) {
 579                 /*
 580                  * Not a fsck error because this is checked/repaired by
 581                  * bch2_check_alloc_key() which runs later:
 582                  */
 583                 if (!bch2_dev_bucket_exists(c, k.k->p))
 584                         continue;
 585
 586                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 587
 588                 *bucket_gen(ca, k.k->p.offset) = bch2_alloc_to_v4(k, &a)->gen;
 589         }
 590         bch2_trans_iter_exit(&trans, &iter);
 591
 592         bch2_trans_exit(&trans);
 593
 594         if (ret)
 595                 bch_err(c, "error reading alloc info: %s", bch2_err_str(ret));
 596
 597         return ret;
 598 }
 599
 600 static struct bpos alloc_gens_pos(struct bpos pos, unsigned *offset)
 601 {
 602         *offset = pos.offset & KEY_TYPE_BUCKET_GENS_MASK;
 603
 604         pos.offset >>= KEY_TYPE_BUCKET_GENS_BITS;
 605         return pos;
 606 }
 607
 608 static struct bpos bucket_gens_pos_to_alloc(struct bpos pos, unsigned offset)
 609 {
 610         pos.offset <<= KEY_TYPE_BUCKET_GENS_BITS;
 611         pos.offset += offset;
 612         return pos;
 613 }
 614
 615 static unsigned alloc_gen(struct bkey_s_c k, unsigned offset)
 616 {
 617         return k.k->type == KEY_TYPE_bucket_gens
 618                 ? bkey_s_c_to_bucket_gens(k).v->gens[offset]
 619                 : 0;
 620 }
 621
 622 int bch2_bucket_gens_invalid(const struct bch_fs *c, struct bkey_s_c k,
 623                              unsigned flags, struct printbuf *err)
 624 {
 625         if (bkey_val_bytes(k.k) != sizeof(struct bch_bucket_gens)) {
 626                 prt_printf(err, "bad val size (%lu != %zu)",
 627                        bkey_val_bytes(k.k), sizeof(struct bch_bucket_gens));
 628                 return -BCH_ERR_invalid_bkey;
 629         }
 630
 631         return 0;
 632 }
 633
 634 void bch2_bucket_gens_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 635 {
 636         struct bkey_s_c_bucket_gens g = bkey_s_c_to_bucket_gens(k);
 637         unsigned i;
 638
 639         for (i = 0; i < ARRAY_SIZE(g.v->gens); i++) {
 640                 if (i)
 641                         prt_char(out, ' ');
 642                 prt_printf(out, "%u", g.v->gens[i]);
 643         }
 644 }
 645
 646 int bch2_bucket_gens_init(struct bch_fs *c)
 647 {
 648         struct btree_trans trans;
 649         struct btree_iter iter;
 650         struct bkey_s_c k;
 651         struct bch_alloc_v4 a;
 652         struct bkey_i_bucket_gens g;
 653         bool have_bucket_gens_key = false;
 654         unsigned offset;
 655         struct bpos pos;
 656         u8 gen;
 657         int ret;
 658
 659         bch2_trans_init(&trans, c, 0, 0);
 660
 661         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 662                            BTREE_ITER_PREFETCH, k, ret) {
 663                 /*
 664                  * Not a fsck error because this is checked/repaired by
 665                  * bch2_check_alloc_key() which runs later:
 666                  */
 667                 if (!bch2_dev_bucket_exists(c, k.k->p))
 668                         continue;
 669
 670                 gen = bch2_alloc_to_v4(k, &a)->gen;
 671                 pos = alloc_gens_pos(iter.pos, &offset);
 672
 673                 if (have_bucket_gens_key && bkey_cmp(iter.pos, pos)) {
 674                         ret = commit_do(&trans, NULL, NULL,
 675                                         BTREE_INSERT_NOFAIL|
 676                                         BTREE_INSERT_LAZY_RW,
 677                                 __bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i));
 678                         if (ret)
 679                                 break;
 680                         have_bucket_gens_key = false;
 681                 }
 682
 683                 if (!have_bucket_gens_key) {
 684                         bkey_bucket_gens_init(&g.k_i);
 685                         g.k.p = pos;
 686                         have_bucket_gens_key = true;
 687                 }
 688
 689                 g.v.gens[offset] = gen;
 690         }
 691         bch2_trans_iter_exit(&trans, &iter);
 692
 693         if (have_bucket_gens_key && !ret)
 694                 ret = commit_do(&trans, NULL, NULL,
 695                                 BTREE_INSERT_NOFAIL|
 696                                 BTREE_INSERT_LAZY_RW,
 697                         __bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i));
 698
 699         bch2_trans_exit(&trans);
 700
 701         if (ret)
 702                 bch_err(c, "%s: error %s", __func__, bch2_err_str(ret));
 703
 704         return ret;
 705 }
 706
 707 int bch2_bucket_gens_read(struct bch_fs *c)
 708 {
 709         struct btree_trans trans;
 710         struct btree_iter iter;
 711         struct bkey_s_c k;
 712         const struct bch_bucket_gens *g;
 713         struct bch_dev *ca;
 714         u64 b;
 715         int ret;
 716
 717         bch2_trans_init(&trans, c, 0, 0);
 718
 719         for_each_btree_key(&trans, iter, BTREE_ID_bucket_gens, POS_MIN,
 720                            BTREE_ITER_PREFETCH, k, ret) {
 721                 u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
 722                 u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
 723
 724                 if (k.k->type != KEY_TYPE_bucket_gens)
 725                         continue;
 726
 727                 g = bkey_s_c_to_bucket_gens(k).v;
 728
 729                 /*
 730                  * Not a fsck error because this is checked/repaired by
 731                  * bch2_check_alloc_key() which runs later:
 732                  */
 733                 if (!bch2_dev_exists2(c, k.k->p.inode))
 734                         continue;
 735
 736                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 737
 738                 for (b = max_t(u64, ca->mi.first_bucket, start);
 739                      b < min_t(u64, ca->mi.nbuckets, end);
 740                      b++)
 741                         *bucket_gen(ca, b) = g->gens[b & KEY_TYPE_BUCKET_GENS_MASK];
 742         }
 743         bch2_trans_iter_exit(&trans, &iter);
 744
 745         bch2_trans_exit(&trans);
 746
 747         if (ret)
 748                 bch_err(c, "error reading alloc info: %s", bch2_err_str(ret));
 749
 750         return ret;
 751 }
 752
 753 /* Free space/discard btree: */
 754
 755 static int bch2_bucket_do_index(struct btree_trans *trans,
 756                                 struct bkey_s_c alloc_k,
 757                                 const struct bch_alloc_v4 *a,
 758                                 bool set)
 759 {
 760         struct bch_fs *c = trans->c;
 761         struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 762         struct btree_iter iter;
 763         struct bkey_s_c old;
 764         struct bkey_i *k;
 765         enum btree_id btree;
 766         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 767         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 768         struct printbuf buf = PRINTBUF;
 769         int ret;
 770
 771         if (a->data_type != BCH_DATA_free &&
 772             a->data_type != BCH_DATA_need_discard)
 773                 return 0;
 774
 775         k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
 776         if (IS_ERR(k))
 777                 return PTR_ERR(k);
 778
 779         bkey_init(&k->k);
 780         k->k.type = new_type;
 781
 782         switch (a->data_type) {
 783         case BCH_DATA_free:
 784                 btree = BTREE_ID_freespace;
 785                 k->k.p = alloc_freespace_pos(alloc_k.k->p, *a);
 786                 bch2_key_resize(&k->k, 1);
 787                 break;
 788         case BCH_DATA_need_discard:
 789                 btree = BTREE_ID_need_discard;
 790                 k->k.p = alloc_k.k->p;
 791                 break;
 792         default:
 793                 return 0;
 794         }
 795
 796         bch2_trans_iter_init(trans, &iter, btree,
 797                              bkey_start_pos(&k->k),
 798                              BTREE_ITER_INTENT);
 799         old = bch2_btree_iter_peek_slot(&iter);
 800         ret = bkey_err(old);
 801         if (ret)
 802                 goto err;
 803
 804         if (ca->mi.freespace_initialized &&
 805             test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags) &&
 806             bch2_trans_inconsistent_on(old.k->type != old_type, trans,
 807                         "incorrect key when %s %s btree (got %s should be %s)\n"
 808                         "  for %s",
 809                         set ? "setting" : "clearing",
 810                         bch2_btree_ids[btree],
 811                         bch2_bkey_types[old.k->type],
 812                         bch2_bkey_types[old_type],
 813                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 814                 ret = -EIO;
 815                 goto err;
 816         }
 817
 818         ret = bch2_trans_update(trans, &iter, k, 0);
 819 err:
 820         bch2_trans_iter_exit(trans, &iter);
 821         printbuf_exit(&buf);
 822         return ret;
 823 }
 824
 825 static noinline int bch2_bucket_gen_update(struct btree_trans *trans,
 826                                            struct bpos bucket, u8 gen)
 827 {
 828         struct btree_iter iter;
 829         unsigned offset;
 830         struct bpos pos = alloc_gens_pos(bucket, &offset);
 831         struct bkey_i_bucket_gens *g;
 832         struct bkey_s_c k;
 833         int ret;
 834
 835         g = bch2_trans_kmalloc(trans, sizeof(*g));
 836         ret = PTR_ERR_OR_ZERO(g);
 837         if (ret)
 838                 return ret;
 839
 840         bch2_trans_iter_init(trans, &iter, BTREE_ID_bucket_gens, pos,
 841                              BTREE_ITER_INTENT|
 842                              BTREE_ITER_WITH_UPDATES);
 843         k = bch2_btree_iter_peek_slot(&iter);
 844         ret = bkey_err(k);
 845         if (ret)
 846                 goto err;
 847
 848         if (k.k->type != KEY_TYPE_bucket_gens) {
 849                 bkey_bucket_gens_init(&g->k_i);
 850                 g->k.p = iter.pos;
 851         } else {
 852                 bkey_reassemble(&g->k_i, k);
 853         }
 854
 855         g->v.gens[offset] = gen;
 856
 857         ret = bch2_trans_update(trans, &iter, &g->k_i, 0);
 858 err:
 859         bch2_trans_iter_exit(trans, &iter);
 860         return ret;
 861 }
 862
 863 int bch2_trans_mark_alloc(struct btree_trans *trans,
 864                           enum btree_id btree_id, unsigned level,
 865                           struct bkey_s_c old, struct bkey_i *new,
 866                           unsigned flags)
 867 {
 868         struct bch_fs *c = trans->c;
 869         struct bch_alloc_v4 old_a_convert, *new_a;
 870         const struct bch_alloc_v4 *old_a;
 871         u64 old_lru, new_lru;
 872         int ret = 0;
 873
 874         /*
 875          * Deletion only happens in the device removal path, with
 876          * BTREE_TRIGGER_NORUN:
 877          */
 878         BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
 879
 880         old_a = bch2_alloc_to_v4(old, &old_a_convert);
 881         new_a = &bkey_i_to_alloc_v4(new)->v;
 882
 883         new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
 884
 885         if (new_a->dirty_sectors > old_a->dirty_sectors ||
 886             new_a->cached_sectors > old_a->cached_sectors) {
 887                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 888                 new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 889                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
 890                 SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
 891         }
 892
 893         if (data_type_is_empty(new_a->data_type) &&
 894             BCH_ALLOC_V4_NEED_INC_GEN(new_a) &&
 895             !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
 896                 new_a->gen++;
 897                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
 898         }
 899
 900         if (old_a->data_type != new_a->data_type ||
 901             (new_a->data_type == BCH_DATA_free &&
 902              alloc_freespace_genbits(*old_a) != alloc_freespace_genbits(*new_a))) {
 903                 ret =   bch2_bucket_do_index(trans, old, old_a, false) ?:
 904                         bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_a, true);
 905                 if (ret)
 906                         return ret;
 907         }
 908
 909         if (new_a->data_type == BCH_DATA_cached &&
 910             !new_a->io_time[READ])
 911                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 912
 913         old_lru = alloc_lru_idx(*old_a);
 914         new_lru = alloc_lru_idx(*new_a);
 915
 916         if (old_lru != new_lru) {
 917                 ret = bch2_lru_change(trans, new->k.p.inode,
 918                                       bucket_to_u64(new->k.p),
 919                                       old_lru, new_lru);
 920                 if (ret)
 921                         return ret;
 922         }
 923
 924         if (old_a->gen != new_a->gen) {
 925                 ret = bch2_bucket_gen_update(trans, new->k.p, new_a->gen);
 926                 if (ret)
 927                         return ret;
 928         }
 929
 930         return 0;
 931 }
 932
 933 /*
 934  * This synthesizes deleted extents for holes, similar to BTREE_ITER_SLOTS for
 935  * extents style btrees, but works on non-extents btrees:
 936  */
 937 struct bkey_s_c bch2_get_key_or_hole(struct btree_iter *iter, struct bpos end, struct bkey *hole)
 938 {
 939         struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
 940
 941         if (bkey_err(k))
 942                 return k;
 943
 944         if (k.k->type) {
 945                 return k;
 946         } else {
 947                 struct btree_iter iter2;
 948                 struct bpos next;
 949
 950                 bch2_trans_copy_iter(&iter2, iter);
 951                 k = bch2_btree_iter_peek_upto(&iter2,
 952                                 bkey_min(bkey_min(end,
 953                                                   iter->path->l[0].b->key.k.p),
 954                                                   POS(iter->pos.inode, iter->pos.offset + U32_MAX - 1)));
 955                 next = iter2.pos;
 956                 bch2_trans_iter_exit(iter->trans, &iter2);
 957
 958                 BUG_ON(next.offset >= iter->pos.offset + U32_MAX);
 959
 960                 if (bkey_err(k))
 961                         return k;
 962
 963                 bkey_init(hole);
 964                 hole->p = iter->pos;
 965
 966                 bch2_key_resize(hole, next.offset - iter->pos.offset);
 967                 return (struct bkey_s_c) { hole, NULL };
 968         }
 969 }
 970
 971 static bool next_bucket(struct bch_fs *c, struct bpos *bucket)
 972 {
 973         struct bch_dev *ca;
 974         unsigned iter;
 975
 976         if (bch2_dev_bucket_exists(c, *bucket))
 977                 return true;
 978
 979         if (bch2_dev_exists2(c, bucket->inode)) {
 980                 ca = bch_dev_bkey_exists(c, bucket->inode);
 981
 982                 if (bucket->offset < ca->mi.first_bucket) {
 983                         bucket->offset = ca->mi.first_bucket;
 984                         return true;
 985                 }
 986
 987                 bucket->inode++;
 988                 bucket->offset = 0;
 989         }
 990
 991         rcu_read_lock();
 992         iter = bucket->inode;
 993         ca = __bch2_next_dev(c, &iter, NULL);
 994         if (ca)
 995                 bucket->offset = ca->mi.first_bucket;
 996         rcu_read_unlock();
 997
 998         return ca != NULL;
 999 }
1000
1001 struct bkey_s_c bch2_get_key_or_real_bucket_hole(struct btree_iter *iter, struct bkey *hole)
1002 {
1003         struct bch_fs *c = iter->trans->c;
1004         struct bkey_s_c k;
1005 again:
1006         k = bch2_get_key_or_hole(iter, POS_MAX, hole);
1007         if (bkey_err(k))
1008                 return k;
1009
1010         if (!k.k->type) {
1011                 struct bpos bucket = bkey_start_pos(k.k);
1012
1013                 if (!bch2_dev_bucket_exists(c, bucket)) {
1014                         if (!next_bucket(c, &bucket))
1015                                 return bkey_s_c_null;
1016
1017                         bch2_btree_iter_set_pos(iter, bucket);
1018                         goto again;
1019                 }
1020
1021                 if (!bch2_dev_bucket_exists(c, k.k->p)) {
1022                         struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
1023
1024                         bch2_key_resize(hole, ca->mi.nbuckets - bucket.offset);
1025                 }
1026         }
1027
1028         return k;
1029 }
1030
1031 static int bch2_check_alloc_key(struct btree_trans *trans,
1032                                 struct bkey_s_c alloc_k,
1033                                 struct btree_iter *alloc_iter,
1034                                 struct btree_iter *discard_iter,
1035                                 struct btree_iter *freespace_iter,
1036                                 struct btree_iter *bucket_gens_iter)
1037 {
1038         struct bch_fs *c = trans->c;
1039         struct bch_dev *ca;
1040         struct bch_alloc_v4 a_convert;
1041         const struct bch_alloc_v4 *a;
1042         unsigned discard_key_type, freespace_key_type;
1043         unsigned gens_offset;
1044         struct bkey_s_c k;
1045         struct printbuf buf = PRINTBUF;
1046         int ret;
1047
1048         if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
1049                         "alloc key for invalid device:bucket %llu:%llu",
1050                         alloc_k.k->p.inode, alloc_k.k->p.offset))
1051                 return bch2_btree_delete_at(trans, alloc_iter, 0);
1052
1053         ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
1054         if (!ca->mi.freespace_initialized)
1055                 return 0;
1056
1057         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1058
1059         discard_key_type = a->data_type == BCH_DATA_need_discard ? KEY_TYPE_set : 0;
1060         bch2_btree_iter_set_pos(discard_iter, alloc_k.k->p);
1061         k = bch2_btree_iter_peek_slot(discard_iter);
1062         ret = bkey_err(k);
1063         if (ret)
1064                 goto err;
1065
1066         if (k.k->type != discard_key_type &&
1067             (c->opts.reconstruct_alloc ||
1068              fsck_err(c, "incorrect key in need_discard btree (got %s should be %s)\n"
1069                       "  %s",
1070                       bch2_bkey_types[k.k->type],
1071                       bch2_bkey_types[discard_key_type],
1072                       (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1073                 struct bkey_i *update =
1074                         bch2_trans_kmalloc(trans, sizeof(*update));
1075
1076                 ret = PTR_ERR_OR_ZERO(update);
1077                 if (ret)
1078                         goto err;
1079
1080                 bkey_init(&update->k);
1081                 update->k.type  = discard_key_type;
1082                 update->k.p     = discard_iter->pos;
1083
1084                 ret = bch2_trans_update(trans, discard_iter, update, 0);
1085                 if (ret)
1086                         goto err;
1087         }
1088
1089         freespace_key_type = a->data_type == BCH_DATA_free ? KEY_TYPE_set : 0;
1090         bch2_btree_iter_set_pos(freespace_iter, alloc_freespace_pos(alloc_k.k->p, *a));
1091         k = bch2_btree_iter_peek_slot(freespace_iter);
1092         ret = bkey_err(k);
1093         if (ret)
1094                 goto err;
1095
1096         if (k.k->type != freespace_key_type &&
1097             (c->opts.reconstruct_alloc ||
1098              fsck_err(c, "incorrect key in freespace btree (got %s should be %s)\n"
1099                       "  %s",
1100                       bch2_bkey_types[k.k->type],
1101                       bch2_bkey_types[freespace_key_type],
1102                       (printbuf_reset(&buf),
1103                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1104                 struct bkey_i *update =
1105                         bch2_trans_kmalloc(trans, sizeof(*update));
1106
1107                 ret = PTR_ERR_OR_ZERO(update);
1108                 if (ret)
1109                         goto err;
1110
1111                 bkey_init(&update->k);
1112                 update->k.type  = freespace_key_type;
1113                 update->k.p     = freespace_iter->pos;
1114                 bch2_key_resize(&update->k, 1);
1115
1116                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1117                 if (ret)
1118                         goto err;
1119         }
1120
1121         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(alloc_k.k->p, &gens_offset));
1122         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1123         ret = bkey_err(k);
1124         if (ret)
1125                 goto err;
1126
1127         if (a->gen != alloc_gen(k, gens_offset) &&
1128             (c->opts.reconstruct_alloc ||
1129              fsck_err(c, "incorrect gen in bucket_gens btree (got %u should be %u)\n"
1130                       "  %s",
1131                       alloc_gen(k, gens_offset), a->gen,
1132                       (printbuf_reset(&buf),
1133                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1134                 struct bkey_i_bucket_gens *g =
1135                         bch2_trans_kmalloc(trans, sizeof(*g));
1136
1137                 ret = PTR_ERR_OR_ZERO(g);
1138                 if (ret)
1139                         goto err;
1140
1141                 if (k.k->type == KEY_TYPE_bucket_gens) {
1142                         bkey_reassemble(&g->k_i, k);
1143                 } else {
1144                         bkey_bucket_gens_init(&g->k_i);
1145                         g->k.p = alloc_gens_pos(alloc_k.k->p, &gens_offset);
1146                 }
1147
1148                 g->v.gens[gens_offset] = a->gen;
1149
1150                 ret = bch2_trans_update(trans, bucket_gens_iter, &g->k_i, 0);
1151                 if (ret)
1152                         goto err;
1153         }
1154 err:
1155 fsck_err:
1156         printbuf_exit(&buf);
1157         return ret;
1158 }
1159
1160 static int bch2_check_alloc_hole_freespace(struct btree_trans *trans,
1161                                  struct bpos start,
1162                                  struct bpos *end,
1163                                  struct btree_iter *freespace_iter)
1164 {
1165         struct bch_fs *c = trans->c;
1166         struct bch_dev *ca;
1167         struct bkey_s_c k;
1168         struct printbuf buf = PRINTBUF;
1169         int ret;
1170
1171         ca = bch_dev_bkey_exists(c, start.inode);
1172         if (!ca->mi.freespace_initialized)
1173                 return 0;
1174
1175         bch2_btree_iter_set_pos(freespace_iter, start);
1176
1177         k = bch2_btree_iter_peek_slot(freespace_iter);
1178         ret = bkey_err(k);
1179         if (ret)
1180                 goto err;
1181
1182         *end = bkey_min(k.k->p, *end);
1183
1184         if (k.k->type != KEY_TYPE_set &&
1185             (c->opts.reconstruct_alloc ||
1186              fsck_err(c, "hole in alloc btree missing in freespace btree\n"
1187                       "  device %llu buckets %llu-%llu",
1188                       freespace_iter->pos.inode,
1189                       freespace_iter->pos.offset,
1190                       end->offset))) {
1191                 struct bkey_i *update =
1192                         bch2_trans_kmalloc(trans, sizeof(*update));
1193
1194                 ret = PTR_ERR_OR_ZERO(update);
1195                 if (ret)
1196                         goto err;
1197
1198                 bkey_init(&update->k);
1199                 update->k.type  = KEY_TYPE_set;
1200                 update->k.p     = freespace_iter->pos;
1201                 bch2_key_resize(&update->k,
1202                                 min_t(u64, U32_MAX, end->offset -
1203                                       freespace_iter->pos.offset));
1204
1205                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1206                 if (ret)
1207                         goto err;
1208         }
1209 err:
1210 fsck_err:
1211         printbuf_exit(&buf);
1212         return ret;
1213 }
1214
1215 static int bch2_check_alloc_hole_bucket_gens(struct btree_trans *trans,
1216                                  struct bpos start,
1217                                  struct bpos *end,
1218                                  struct btree_iter *bucket_gens_iter)
1219 {
1220         struct bch_fs *c = trans->c;
1221         struct bkey_s_c k;
1222         struct printbuf buf = PRINTBUF;
1223         unsigned i, gens_offset, gens_end_offset;
1224         int ret;
1225
1226         if (c->sb.version < bcachefs_metadata_version_bucket_gens &&
1227             !c->opts.version_upgrade)
1228                 return 0;
1229
1230         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(start, &gens_offset));
1231
1232         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1233         ret = bkey_err(k);
1234         if (ret)
1235                 goto err;
1236
1237         if (bkey_cmp(alloc_gens_pos(start, &gens_offset),
1238                      alloc_gens_pos(*end,  &gens_end_offset)))
1239                 gens_end_offset = KEY_TYPE_BUCKET_GENS_NR;
1240
1241         if (k.k->type == KEY_TYPE_bucket_gens) {
1242                 struct bkey_i_bucket_gens g;
1243                 bool need_update = false;
1244
1245                 bkey_reassemble(&g.k_i, k);
1246
1247                 for (i = gens_offset; i < gens_end_offset; i++) {
1248                         if (fsck_err_on(g.v.gens[i], c,
1249                                         "hole in alloc btree at %llu:%llu with nonzero gen in bucket_gens btree (%u)",
1250                                         bucket_gens_pos_to_alloc(k.k->p, i).inode,
1251                                         bucket_gens_pos_to_alloc(k.k->p, i).offset,
1252                                         g.v.gens[i])) {
1253                                 g.v.gens[i] = 0;
1254                                 need_update = true;
1255                         }
1256                 }
1257
1258                 if (need_update) {
1259                         struct bkey_i *k = bch2_trans_kmalloc(trans, sizeof(g));
1260
1261                         ret = PTR_ERR_OR_ZERO(k);
1262                         if (ret)
1263                                 goto err;
1264
1265                         memcpy(k, &g, sizeof(g));
1266
1267                         ret = bch2_trans_update(trans, bucket_gens_iter, k, 0);
1268                         if (ret)
1269                                 goto err;
1270                 }
1271         }
1272
1273         *end = bkey_min(*end, bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0));
1274 err:
1275 fsck_err:
1276         printbuf_exit(&buf);
1277         return ret;
1278 }
1279
1280 static int bch2_check_discard_freespace_key(struct btree_trans *trans,
1281                                             struct btree_iter *iter)
1282 {
1283         struct bch_fs *c = trans->c;
1284         struct btree_iter alloc_iter;
1285         struct bkey_s_c alloc_k;
1286         struct bch_alloc_v4 a_convert;
1287         const struct bch_alloc_v4 *a;
1288         u64 genbits;
1289         struct bpos pos;
1290         enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
1291                 ? BCH_DATA_need_discard
1292                 : BCH_DATA_free;
1293         struct printbuf buf = PRINTBUF;
1294         int ret;
1295
1296         pos = iter->pos;
1297         pos.offset &= ~(~0ULL << 56);
1298         genbits = iter->pos.offset & (~0ULL << 56);
1299
1300         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
1301
1302         if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
1303                         "entry in %s btree for nonexistant dev:bucket %llu:%llu",
1304                         bch2_btree_ids[iter->btree_id], pos.inode, pos.offset))
1305                 goto delete;
1306
1307         alloc_k = bch2_btree_iter_peek_slot(&alloc_iter);
1308         ret = bkey_err(alloc_k);
1309         if (ret)
1310                 goto err;
1311
1312         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1313
1314         if (fsck_err_on(a->data_type != state ||
1315                         (state == BCH_DATA_free &&
1316                          genbits != alloc_freespace_genbits(*a)), c,
1317                         "%s\n  incorrectly set in %s index (free %u, genbits %llu should be %llu)",
1318                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
1319                         bch2_btree_ids[iter->btree_id],
1320                         a->data_type == state,
1321                         genbits >> 56, alloc_freespace_genbits(*a) >> 56))
1322                 goto delete;
1323 out:
1324 err:
1325 fsck_err:
1326         bch2_trans_iter_exit(trans, &alloc_iter);
1327         printbuf_exit(&buf);
1328         return ret;
1329 delete:
1330         ret = bch2_btree_delete_extent_at(trans, iter,
1331                         iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0);
1332         goto out;
1333 }
1334
1335 /*
1336  * We've already checked that generation numbers in the bucket_gens btree are
1337  * valid for buckets that exist; this just checks for keys for nonexistent
1338  * buckets.
1339  */
1340 static int bch2_check_bucket_gens_key(struct btree_trans *trans,
1341                                       struct btree_iter *iter,
1342                                       struct bkey_s_c k)
1343 {
1344         struct bch_fs *c = trans->c;
1345         struct bkey_i_bucket_gens g;
1346         struct bch_dev *ca;
1347         u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
1348         u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
1349         u64 b;
1350         bool need_update = false;
1351         struct printbuf buf = PRINTBUF;
1352         int ret = 0;
1353
1354         BUG_ON(k.k->type != KEY_TYPE_bucket_gens);
1355         bkey_reassemble(&g.k_i, k);
1356
1357         if (fsck_err_on(!bch2_dev_exists2(c, k.k->p.inode), c,
1358                         "bucket_gens key for invalid device:\n  %s",
1359                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1360                 ret = bch2_btree_delete_at(trans, iter, 0);
1361                 goto out;
1362         }
1363
1364         ca = bch_dev_bkey_exists(c, k.k->p.inode);
1365         if (fsck_err_on(end <= ca->mi.first_bucket ||
1366                         start >= ca->mi.nbuckets, c,
1367                         "bucket_gens key for invalid buckets:\n  %s",
1368                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1369                 ret = bch2_btree_delete_at(trans, iter, 0);
1370                 goto out;
1371         }
1372
1373         for (b = start; b < ca->mi.first_bucket; b++)
1374                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1375                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1376                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1377                         need_update = true;
1378                 }
1379
1380         for (b = ca->mi.nbuckets; b < end; b++)
1381                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1382                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1383                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1384                         need_update = true;
1385                 }
1386
1387         if (need_update) {
1388                 struct bkey_i *k;
1389
1390                 k = bch2_trans_kmalloc(trans, sizeof(g));
1391                 ret = PTR_ERR_OR_ZERO(k);
1392                 if (ret)
1393                         goto out;
1394
1395                 memcpy(k, &g, sizeof(g));
1396                 ret = bch2_trans_update(trans, iter, k, 0);
1397         }
1398 out:
1399 fsck_err:
1400         printbuf_exit(&buf);
1401         return ret;
1402 }
1403
1404 int bch2_check_alloc_info(struct bch_fs *c)
1405 {
1406         struct btree_trans trans;
1407         struct btree_iter iter, discard_iter, freespace_iter, bucket_gens_iter;
1408         struct bkey hole;
1409         struct bkey_s_c k;
1410         int ret = 0;
1411
1412         bch2_trans_init(&trans, c, 0, 0);
1413
1414         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc, POS_MIN,
1415                              BTREE_ITER_PREFETCH);
1416         bch2_trans_iter_init(&trans, &discard_iter, BTREE_ID_need_discard, POS_MIN,
1417                              BTREE_ITER_PREFETCH);
1418         bch2_trans_iter_init(&trans, &freespace_iter, BTREE_ID_freespace, POS_MIN,
1419                              BTREE_ITER_PREFETCH);
1420         bch2_trans_iter_init(&trans, &bucket_gens_iter, BTREE_ID_bucket_gens, POS_MIN,
1421                              BTREE_ITER_PREFETCH);
1422
1423         while (1) {
1424                 struct bpos next;
1425
1426                 bch2_trans_begin(&trans);
1427
1428                 k = bch2_get_key_or_real_bucket_hole(&iter, &hole);
1429                 ret = bkey_err(k);
1430                 if (ret)
1431                         goto bkey_err;
1432
1433                 if (!k.k)
1434                         break;
1435
1436                 if (k.k->type) {
1437                         next = bpos_nosnap_successor(k.k->p);
1438
1439                         ret = bch2_check_alloc_key(&trans,
1440                                                    k, &iter,
1441                                                    &discard_iter,
1442                                                    &freespace_iter,
1443                                                    &bucket_gens_iter);
1444                         if (ret)
1445                                 goto bkey_err;
1446                 } else {
1447                         next = k.k->p;
1448
1449                         ret = bch2_check_alloc_hole_freespace(&trans,
1450                                                     bkey_start_pos(k.k),
1451                                                     &next,
1452                                                     &freespace_iter) ?:
1453                                 bch2_check_alloc_hole_bucket_gens(&trans,
1454                                                     bkey_start_pos(k.k),
1455                                                     &next,
1456                                                     &bucket_gens_iter);
1457                         if (ret)
1458                                 goto bkey_err;
1459                 }
1460
1461                 ret = bch2_trans_commit(&trans, NULL, NULL,
1462                                         BTREE_INSERT_NOFAIL|
1463                                         BTREE_INSERT_LAZY_RW);
1464                 if (ret)
1465                         goto bkey_err;
1466
1467                 bch2_btree_iter_set_pos(&iter, next);
1468 bkey_err:
1469                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1470                         continue;
1471                 if (ret)
1472                         break;
1473         }
1474         bch2_trans_iter_exit(&trans, &bucket_gens_iter);
1475         bch2_trans_iter_exit(&trans, &freespace_iter);
1476         bch2_trans_iter_exit(&trans, &discard_iter);
1477         bch2_trans_iter_exit(&trans, &iter);
1478
1479         if (ret < 0)
1480                 goto err;
1481
1482         ret = for_each_btree_key_commit(&trans, iter,
1483                         BTREE_ID_need_discard, POS_MIN,
1484                         BTREE_ITER_PREFETCH, k,
1485                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1486                 bch2_check_discard_freespace_key(&trans, &iter)) ?:
1487               for_each_btree_key_commit(&trans, iter,
1488                         BTREE_ID_freespace, POS_MIN,
1489                         BTREE_ITER_PREFETCH, k,
1490                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1491                 bch2_check_discard_freespace_key(&trans, &iter)) ?:
1492               for_each_btree_key_commit(&trans, iter,
1493                         BTREE_ID_bucket_gens, POS_MIN,
1494                         BTREE_ITER_PREFETCH, k,
1495                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1496                 bch2_check_bucket_gens_key(&trans, &iter, k));
1497 err:
1498         bch2_trans_exit(&trans);
1499         return ret < 0 ? ret : 0;
1500 }
1501
1502 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
1503                                        struct btree_iter *alloc_iter)
1504 {
1505         struct bch_fs *c = trans->c;
1506         struct btree_iter lru_iter;
1507         struct bch_alloc_v4 a_convert;
1508         const struct bch_alloc_v4 *a;
1509         struct bkey_s_c alloc_k, k;
1510         struct printbuf buf = PRINTBUF;
1511         int ret;
1512
1513         alloc_k = bch2_btree_iter_peek(alloc_iter);
1514         if (!alloc_k.k)
1515                 return 0;
1516
1517         ret = bkey_err(alloc_k);
1518         if (ret)
1519                 return ret;
1520
1521         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1522
1523         if (a->data_type != BCH_DATA_cached)
1524                 return 0;
1525
1526         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
1527                              lru_pos(alloc_k.k->p.inode,
1528                                      bucket_to_u64(alloc_k.k->p),
1529                                      a->io_time[READ]), 0);
1530         k = bch2_btree_iter_peek_slot(&lru_iter);
1531         ret = bkey_err(k);
1532         if (ret)
1533                 goto err;
1534
1535         if (fsck_err_on(!a->io_time[READ], c,
1536                         "cached bucket with read_time 0\n"
1537                         "  %s",
1538                 (printbuf_reset(&buf),
1539                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) ||
1540             fsck_err_on(k.k->type != KEY_TYPE_set, c,
1541                         "missing lru entry\n"
1542                         "  %s",
1543                         (printbuf_reset(&buf),
1544                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
1545                 u64 read_time = a->io_time[READ] ?:
1546                         atomic64_read(&c->io_clock[READ].now);
1547
1548                 ret = bch2_lru_set(trans,
1549                                    alloc_k.k->p.inode,
1550                                    bucket_to_u64(alloc_k.k->p),
1551                                    read_time);
1552                 if (ret)
1553                         goto err;
1554
1555                 if (a->io_time[READ] != read_time) {
1556                         struct bkey_i_alloc_v4 *a_mut =
1557                                 bch2_alloc_to_v4_mut(trans, alloc_k);
1558                         ret = PTR_ERR_OR_ZERO(a_mut);
1559                         if (ret)
1560                                 goto err;
1561
1562                         a_mut->v.io_time[READ] = read_time;
1563                         ret = bch2_trans_update(trans, alloc_iter,
1564                                                 &a_mut->k_i, BTREE_TRIGGER_NORUN);
1565                         if (ret)
1566                                 goto err;
1567                 }
1568         }
1569 err:
1570 fsck_err:
1571         bch2_trans_iter_exit(trans, &lru_iter);
1572         printbuf_exit(&buf);
1573         return ret;
1574 }
1575
1576 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
1577 {
1578         struct btree_trans trans;
1579         struct btree_iter iter;
1580         struct bkey_s_c k;
1581         int ret = 0;
1582
1583         bch2_trans_init(&trans, c, 0, 0);
1584
1585         for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1586                         POS_MIN, BTREE_ITER_PREFETCH, k,
1587                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1588                 bch2_check_alloc_to_lru_ref(&trans, &iter));
1589
1590         bch2_trans_exit(&trans);
1591         return ret < 0 ? ret : 0;
1592 }
1593
1594 static int bch2_discard_one_bucket(struct btree_trans *trans,
1595                                    struct btree_iter *need_discard_iter,
1596                                    struct bpos *discard_pos_done,
1597                                    u64 *seen,
1598                                    u64 *open,
1599                                    u64 *need_journal_commit,
1600                                    u64 *discarded)
1601 {
1602         struct bch_fs *c = trans->c;
1603         struct bpos pos = need_discard_iter->pos;
1604         struct btree_iter iter = { NULL };
1605         struct bkey_s_c k;
1606         struct bch_dev *ca;
1607         struct bkey_i_alloc_v4 *a;
1608         struct printbuf buf = PRINTBUF;
1609         int ret = 0;
1610
1611         ca = bch_dev_bkey_exists(c, pos.inode);
1612         if (!percpu_ref_tryget(&ca->io_ref)) {
1613                 bch2_btree_iter_set_pos(need_discard_iter, POS(pos.inode + 1, 0));
1614                 return 0;
1615         }
1616
1617         if (bch2_bucket_is_open_safe(c, pos.inode, pos.offset)) {
1618                 (*open)++;
1619                 goto out;
1620         }
1621
1622         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
1623                         c->journal.flushed_seq_ondisk,
1624                         pos.inode, pos.offset)) {
1625                 (*need_journal_commit)++;
1626                 goto out;
1627         }
1628
1629         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
1630                              need_discard_iter->pos,
1631                              BTREE_ITER_CACHED);
1632         k = bch2_btree_iter_peek_slot(&iter);
1633         ret = bkey_err(k);
1634         if (ret)
1635                 goto out;
1636
1637         a = bch2_alloc_to_v4_mut(trans, k);
1638         ret = PTR_ERR_OR_ZERO(a);
1639         if (ret)
1640                 goto out;
1641
1642         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
1643                 a->v.gen++;
1644                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1645                 goto write;
1646         }
1647
1648         if (a->v.journal_seq > c->journal.flushed_seq_ondisk) {
1649                 if (test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
1650                         bch2_trans_inconsistent(trans,
1651                                 "clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
1652                                 "%s",
1653                                 a->v.journal_seq,
1654                                 c->journal.flushed_seq_ondisk,
1655                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1656                         ret = -EIO;
1657                 }
1658                 goto out;
1659         }
1660
1661         if (a->v.data_type != BCH_DATA_need_discard) {
1662                 if (test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
1663                         bch2_trans_inconsistent(trans,
1664                                 "bucket incorrectly set in need_discard btree\n"
1665                                 "%s",
1666                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1667                         ret = -EIO;
1668                 }
1669
1670                 goto out;
1671         }
1672
1673         if (!bkey_eq(*discard_pos_done, iter.pos) &&
1674             ca->mi.discard && !c->opts.nochanges) {
1675                 /*
1676                  * This works without any other locks because this is the only
1677                  * thread that removes items from the need_discard tree
1678                  */
1679                 bch2_trans_unlock(trans);
1680                 blkdev_issue_discard(ca->disk_sb.bdev,
1681                                      k.k->p.offset * ca->mi.bucket_size,
1682                                      ca->mi.bucket_size,
1683                                      GFP_KERNEL);
1684                 *discard_pos_done = iter.pos;
1685
1686                 ret = bch2_trans_relock_notrace(trans);
1687                 if (ret)
1688                         goto out;
1689         }
1690
1691         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
1692         a->v.data_type = alloc_data_type(a->v, a->v.data_type);
1693 write:
1694         ret =   bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1695                 bch2_trans_commit(trans, NULL, NULL,
1696                                   BTREE_INSERT_USE_RESERVE|BTREE_INSERT_NOFAIL);
1697         if (ret)
1698                 goto out;
1699
1700         this_cpu_inc(c->counters[BCH_COUNTER_bucket_discard]);
1701         (*discarded)++;
1702 out:
1703         (*seen)++;
1704         bch2_trans_iter_exit(trans, &iter);
1705         percpu_ref_put(&ca->io_ref);
1706         printbuf_exit(&buf);
1707         return ret;
1708 }
1709
1710 static void bch2_do_discards_work(struct work_struct *work)
1711 {
1712         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
1713         struct btree_trans trans;
1714         struct btree_iter iter;
1715         struct bkey_s_c k;
1716         u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
1717         struct bpos discard_pos_done = POS_MAX;
1718         int ret;
1719
1720         bch2_trans_init(&trans, c, 0, 0);
1721
1722         /*
1723          * We're doing the commit in bch2_discard_one_bucket instead of using
1724          * for_each_btree_key_commit() so that we can increment counters after
1725          * successful commit:
1726          */
1727         ret = for_each_btree_key2(&trans, iter,
1728                         BTREE_ID_need_discard, POS_MIN, 0, k,
1729                 bch2_discard_one_bucket(&trans, &iter, &discard_pos_done,
1730                                         &seen,
1731                                         &open,
1732                                         &need_journal_commit,
1733                                         &discarded));
1734
1735         bch2_trans_exit(&trans);
1736
1737         if (need_journal_commit * 2 > seen)
1738                 bch2_journal_flush_async(&c->journal, NULL);
1739
1740         bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1741
1742         trace_discard_buckets(c, seen, open, need_journal_commit, discarded,
1743                               bch2_err_str(ret));
1744 }
1745
1746 void bch2_do_discards(struct bch_fs *c)
1747 {
1748         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_discard) &&
1749             !queue_work(system_long_wq, &c->discard_work))
1750                 bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1751 }
1752
1753 static int invalidate_one_bucket(struct btree_trans *trans,
1754                                  struct btree_iter *lru_iter,
1755                                  struct bkey_s_c lru_k,
1756                                  s64 *nr_to_invalidate)
1757 {
1758         struct bch_fs *c = trans->c;
1759         struct btree_iter alloc_iter = { NULL };
1760         struct bkey_i_alloc_v4 *a = NULL;
1761         struct printbuf buf = PRINTBUF;
1762         struct bpos bucket = u64_to_bucket(lru_k.k->p.offset);
1763         unsigned cached_sectors;
1764         int ret = 0;
1765
1766         if (*nr_to_invalidate <= 0)
1767                 return 1;
1768
1769         if (!bch2_dev_bucket_exists(c, bucket)) {
1770                 prt_str(&buf, "lru entry points to invalid bucket");
1771                 goto err;
1772         }
1773
1774         a = bch2_trans_start_alloc_update(trans, &alloc_iter, bucket);
1775         ret = PTR_ERR_OR_ZERO(a);
1776         if (ret)
1777                 goto out;
1778
1779         if (lru_pos_time(lru_iter->pos) != alloc_lru_idx(a->v)) {
1780                 prt_str(&buf, "alloc key does not point back to lru entry when invalidating bucket:");
1781                 goto err;
1782         }
1783
1784         if (a->v.data_type != BCH_DATA_cached) {
1785                 prt_str(&buf, "lru entry points to non cached bucket:");
1786                 goto err;
1787         }
1788
1789         if (!a->v.cached_sectors)
1790                 bch_err(c, "invalidating empty bucket, confused");
1791
1792         cached_sectors = a->v.cached_sectors;
1793
1794         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1795         a->v.gen++;
1796         a->v.data_type          = 0;
1797         a->v.dirty_sectors      = 0;
1798         a->v.cached_sectors     = 0;
1799         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1800         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1801
1802         ret =   bch2_trans_update(trans, &alloc_iter, &a->k_i,
1803                                 BTREE_TRIGGER_BUCKET_INVALIDATE) ?:
1804                 bch2_trans_commit(trans, NULL, NULL,
1805                                   BTREE_INSERT_USE_RESERVE|BTREE_INSERT_NOFAIL);
1806         if (ret)
1807                 goto out;
1808
1809         trace_and_count(c, bucket_invalidate, c, bucket.inode, bucket.offset, cached_sectors);
1810         --*nr_to_invalidate;
1811 out:
1812         bch2_trans_iter_exit(trans, &alloc_iter);
1813         printbuf_exit(&buf);
1814         return ret;
1815 err:
1816         prt_str(&buf, "\n  lru key: ");
1817         bch2_bkey_val_to_text(&buf, c, lru_k);
1818
1819         prt_str(&buf, "\n  lru entry: ");
1820         bch2_lru_pos_to_text(&buf, lru_iter->pos);
1821
1822         prt_str(&buf, "\n  alloc key: ");
1823         if (!a)
1824                 bch2_bpos_to_text(&buf, bucket);
1825         else
1826                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
1827
1828         bch_err(c, "%s", buf.buf);
1829         if (test_bit(BCH_FS_CHECK_LRUS_DONE, &c->flags)) {
1830                 bch2_inconsistent_error(c);
1831                 ret = -EINVAL;
1832         }
1833
1834         goto out;
1835 }
1836
1837 static void bch2_do_invalidates_work(struct work_struct *work)
1838 {
1839         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1840         struct bch_dev *ca;
1841         struct btree_trans trans;
1842         struct btree_iter iter;
1843         struct bkey_s_c k;
1844         unsigned i;
1845         int ret = 0;
1846
1847         bch2_trans_init(&trans, c, 0, 0);
1848
1849         for_each_member_device(ca, c, i) {
1850                 s64 nr_to_invalidate =
1851                         should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
1852
1853                 ret = for_each_btree_key2_upto(&trans, iter, BTREE_ID_lru,
1854                                 lru_pos(ca->dev_idx, 0, 0),
1855                                 lru_pos(ca->dev_idx, U64_MAX, LRU_TIME_MAX),
1856                                 BTREE_ITER_INTENT, k,
1857                         invalidate_one_bucket(&trans, &iter, k, &nr_to_invalidate));
1858
1859                 if (ret < 0) {
1860                         percpu_ref_put(&ca->ref);
1861                         break;
1862                 }
1863         }
1864
1865         bch2_trans_exit(&trans);
1866         bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1867 }
1868
1869 void bch2_do_invalidates(struct bch_fs *c)
1870 {
1871         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_invalidate) &&
1872             !queue_work(system_long_wq, &c->invalidate_work))
1873                 bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1874 }
1875
1876 static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca)
1877 {
1878         struct btree_trans trans;
1879         struct btree_iter iter;
1880         struct bkey_s_c k;
1881         struct bkey hole;
1882         struct bpos end = POS(ca->dev_idx, ca->mi.nbuckets);
1883         struct bch_member *m;
1884         int ret;
1885
1886         bch2_trans_init(&trans, c, 0, 0);
1887
1888         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc,
1889                              POS(ca->dev_idx, ca->mi.first_bucket),
1890                              BTREE_ITER_PREFETCH);
1891         /*
1892          * Scan the alloc btree for every bucket on @ca, and add buckets to the
1893          * freespace/need_discard/need_gc_gens btrees as needed:
1894          */
1895         while (1) {
1896                 bch2_trans_begin(&trans);
1897
1898                 if (bkey_ge(iter.pos, end)) {
1899                         ret = 0;
1900                         break;
1901                 }
1902
1903                 k = bch2_get_key_or_hole(&iter, end, &hole);
1904                 ret = bkey_err(k);
1905                 if (ret)
1906                         goto bkey_err;
1907
1908                 if (k.k->type) {
1909                         /*
1910                          * We process live keys in the alloc btree one at a
1911                          * time:
1912                          */
1913                         struct bch_alloc_v4 a_convert;
1914                         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1915
1916                         ret =   bch2_bucket_do_index(&trans, k, a, true) ?:
1917                                 bch2_trans_commit(&trans, NULL, NULL,
1918                                                   BTREE_INSERT_LAZY_RW|
1919                                                   BTREE_INSERT_NOFAIL);
1920                         if (ret)
1921                                 goto bkey_err;
1922
1923                         bch2_btree_iter_advance(&iter);
1924                 } else {
1925                         struct bkey_i *freespace;
1926
1927                         freespace = bch2_trans_kmalloc(&trans, sizeof(*freespace));
1928                         ret = PTR_ERR_OR_ZERO(freespace);
1929                         if (ret)
1930                                 goto bkey_err;
1931
1932                         bkey_init(&freespace->k);
1933                         freespace->k.type       = KEY_TYPE_set;
1934                         freespace->k.p          = k.k->p;
1935                         freespace->k.size       = k.k->size;
1936
1937                         ret = __bch2_btree_insert(&trans, BTREE_ID_freespace, freespace) ?:
1938                                 bch2_trans_commit(&trans, NULL, NULL,
1939                                                   BTREE_INSERT_LAZY_RW|
1940                                                   BTREE_INSERT_NOFAIL);
1941                         if (ret)
1942                                 goto bkey_err;
1943
1944                         bch2_btree_iter_set_pos(&iter, k.k->p);
1945                 }
1946 bkey_err:
1947                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1948                         continue;
1949                 if (ret)
1950                         break;
1951         }
1952
1953         bch2_trans_iter_exit(&trans, &iter);
1954         bch2_trans_exit(&trans);
1955
1956         if (ret < 0) {
1957                 bch_err(ca, "error initializing free space: %s", bch2_err_str(ret));
1958                 return ret;
1959         }
1960
1961         mutex_lock(&c->sb_lock);
1962         m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx;
1963         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1964         mutex_unlock(&c->sb_lock);
1965
1966         return 0;
1967 }
1968
1969 int bch2_fs_freespace_init(struct bch_fs *c)
1970 {
1971         struct bch_dev *ca;
1972         unsigned i;
1973         int ret = 0;
1974         bool doing_init = false;
1975
1976         /*
1977          * We can crash during the device add path, so we need to check this on
1978          * every mount:
1979          */
1980
1981         for_each_member_device(ca, c, i) {
1982                 if (ca->mi.freespace_initialized)
1983                         continue;
1984
1985                 if (!doing_init) {
1986                         bch_info(c, "initializing freespace");
1987                         doing_init = true;
1988                 }
1989
1990                 ret = bch2_dev_freespace_init(c, ca);
1991                 if (ret) {
1992                         percpu_ref_put(&ca->ref);
1993                         return ret;
1994                 }
1995         }
1996
1997         if (doing_init) {
1998                 mutex_lock(&c->sb_lock);
1999                 bch2_write_super(c);
2000                 mutex_unlock(&c->sb_lock);
2001
2002                 bch_verbose(c, "done initializing freespace");
2003         }
2004
2005         return ret;
2006 }
2007
2008 /* Bucket IO clocks: */
2009
2010 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
2011                               size_t bucket_nr, int rw)
2012 {
2013         struct bch_fs *c = trans->c;
2014         struct btree_iter iter;
2015         struct bkey_i_alloc_v4 *a;
2016         u64 now;
2017         int ret = 0;
2018
2019         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
2020         ret = PTR_ERR_OR_ZERO(a);
2021         if (ret)
2022                 return ret;
2023
2024         now = atomic64_read(&c->io_clock[rw].now);
2025         if (a->v.io_time[rw] == now)
2026                 goto out;
2027
2028         a->v.io_time[rw] = now;
2029
2030         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
2031                 bch2_trans_commit(trans, NULL, NULL, 0);
2032 out:
2033         bch2_trans_iter_exit(trans, &iter);
2034         return ret;
2035 }
2036
2037 /* Startup/shutdown (ro/rw): */
2038
2039 void bch2_recalc_capacity(struct bch_fs *c)
2040 {
2041         struct bch_dev *ca;
2042         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
2043         unsigned bucket_size_max = 0;
2044         unsigned long ra_pages = 0;
2045         unsigned i;
2046
2047         lockdep_assert_held(&c->state_lock);
2048
2049         for_each_online_member(ca, c, i) {
2050                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
2051
2052                 ra_pages += bdi->ra_pages;
2053         }
2054
2055         bch2_set_ra_pages(c, ra_pages);
2056
2057         for_each_rw_member(ca, c, i) {
2058                 u64 dev_reserve = 0;
2059
2060                 /*
2061                  * We need to reserve buckets (from the number
2062                  * of currently available buckets) against
2063                  * foreground writes so that mainly copygc can
2064                  * make forward progress.
2065                  *
2066                  * We need enough to refill the various reserves
2067                  * from scratch - copygc will use its entire
2068                  * reserve all at once, then run against when
2069                  * its reserve is refilled (from the formerly
2070                  * available buckets).
2071                  *
2072                  * This reserve is just used when considering if
2073                  * allocations for foreground writes must wait -
2074                  * not -ENOSPC calculations.
2075                  */
2076
2077                 dev_reserve += ca->nr_btree_reserve * 2;
2078                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
2079
2080                 dev_reserve += 1;       /* btree write point */
2081                 dev_reserve += 1;       /* copygc write point */
2082                 dev_reserve += 1;       /* rebalance write point */
2083
2084                 dev_reserve *= ca->mi.bucket_size;
2085
2086                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
2087                                              ca->mi.first_bucket);
2088
2089                 reserved_sectors += dev_reserve * 2;
2090
2091                 bucket_size_max = max_t(unsigned, bucket_size_max,
2092                                         ca->mi.bucket_size);
2093         }
2094
2095         gc_reserve = c->opts.gc_reserve_bytes
2096                 ? c->opts.gc_reserve_bytes >> 9
2097                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
2098
2099         reserved_sectors = max(gc_reserve, reserved_sectors);
2100
2101         reserved_sectors = min(reserved_sectors, capacity);
2102
2103         c->capacity = capacity - reserved_sectors;
2104
2105         c->bucket_size_max = bucket_size_max;
2106
2107         /* Wake up case someone was waiting for buckets */
2108         closure_wake_up(&c->freelist_wait);
2109 }
2110
2111 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
2112 {
2113         struct open_bucket *ob;
2114         bool ret = false;
2115
2116         for (ob = c->open_buckets;
2117              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
2118              ob++) {
2119                 spin_lock(&ob->lock);
2120                 if (ob->valid && !ob->on_partial_list &&
2121                     ob->dev == ca->dev_idx)
2122                         ret = true;
2123                 spin_unlock(&ob->lock);
2124         }
2125
2126         return ret;
2127 }
2128
2129 /* device goes ro: */
2130 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
2131 {
2132         unsigned i;
2133
2134         /* First, remove device from allocation groups: */
2135
2136         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2137                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
2138
2139         /*
2140          * Capacity is calculated based off of devices in allocation groups:
2141          */
2142         bch2_recalc_capacity(c);
2143
2144         /* Next, close write points that point to this device... */
2145         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
2146                 bch2_writepoint_stop(c, ca, &c->write_points[i]);
2147
2148         bch2_writepoint_stop(c, ca, &c->copygc_write_point);
2149         bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
2150         bch2_writepoint_stop(c, ca, &c->btree_write_point);
2151
2152         mutex_lock(&c->btree_reserve_cache_lock);
2153         while (c->btree_reserve_cache_nr) {
2154                 struct btree_alloc *a =
2155                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
2156
2157                 bch2_open_buckets_put(c, &a->ob);
2158         }
2159         mutex_unlock(&c->btree_reserve_cache_lock);
2160
2161         while (1) {
2162                 struct open_bucket *ob;
2163
2164                 spin_lock(&c->freelist_lock);
2165                 if (!ca->open_buckets_partial_nr) {
2166                         spin_unlock(&c->freelist_lock);
2167                         break;
2168                 }
2169                 ob = c->open_buckets +
2170                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
2171                 ob->on_partial_list = false;
2172                 spin_unlock(&c->freelist_lock);
2173
2174                 bch2_open_bucket_put(c, ob);
2175         }
2176
2177         bch2_ec_stop_dev(c, ca);
2178
2179         /*
2180          * Wake up threads that were blocked on allocation, so they can notice
2181          * the device can no longer be removed and the capacity has changed:
2182          */
2183         closure_wake_up(&c->freelist_wait);
2184
2185         /*
2186          * journal_res_get() can block waiting for free space in the journal -
2187          * it needs to notice there may not be devices to allocate from anymore:
2188          */
2189         wake_up(&c->journal.wait);
2190
2191         /* Now wait for any in flight writes: */
2192
2193         closure_wait_event(&c->open_buckets_wait,
2194                            !bch2_dev_has_open_write_point(c, ca));
2195 }
2196
2197 /* device goes rw: */
2198 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
2199 {
2200         unsigned i;
2201
2202         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2203                 if (ca->mi.data_allowed & (1 << i))
2204                         set_bit(ca->dev_idx, c->rw_devs[i].d);
2205 }
2206
2207 void bch2_fs_allocator_background_init(struct bch_fs *c)
2208 {
2209         spin_lock_init(&c->freelist_lock);
2210         INIT_WORK(&c->discard_work, bch2_do_discards_work);
2211         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
2212 }