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                           int rw, 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                           int rw, 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                           int rw, 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                           int rw, struct printbuf *err)
 267 {
 268         struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 269
 270         if (alloc_v4_u64s(a.v) != bkey_val_u64s(k.k)) {
 271                 prt_printf(err, "bad val size (%lu != %u)",
 272                        bkey_val_u64s(k.k), alloc_v4_u64s(a.v));
 273                 return -BCH_ERR_invalid_bkey;
 274         }
 275
 276         if (!BCH_ALLOC_V4_BACKPOINTERS_START(a.v) &&
 277             BCH_ALLOC_V4_NR_BACKPOINTERS(a.v)) {
 278                 prt_printf(err, "invalid backpointers_start");
 279                 return -BCH_ERR_invalid_bkey;
 280         }
 281
 282         /*
 283          * XXX this is wrong, we'll be checking updates that happened from
 284          * before BCH_FS_CHECK_BACKPOINTERS_DONE
 285          */
 286         if (rw == WRITE && test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags)) {
 287                 unsigned i, bp_len = 0;
 288
 289                 for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a.v); i++)
 290                         bp_len += alloc_v4_backpointers_c(a.v)[i].bucket_len;
 291
 292                 if (bp_len > a.v->dirty_sectors) {
 293                         prt_printf(err, "too many backpointers");
 294                         return -BCH_ERR_invalid_bkey;
 295                 }
 296         }
 297
 298         if (rw == WRITE) {
 299                 if (alloc_data_type(*a.v, a.v->data_type) != a.v->data_type) {
 300                         prt_printf(err, "invalid data type (got %u should be %u)",
 301                                a.v->data_type, alloc_data_type(*a.v, a.v->data_type));
 302                         return -BCH_ERR_invalid_bkey;
 303                 }
 304
 305                 switch (a.v->data_type) {
 306                 case BCH_DATA_free:
 307                 case BCH_DATA_need_gc_gens:
 308                 case BCH_DATA_need_discard:
 309                         if (a.v->dirty_sectors ||
 310                             a.v->cached_sectors ||
 311                             a.v->stripe) {
 312                                 prt_printf(err, "empty data type free but have data");
 313                                 return -BCH_ERR_invalid_bkey;
 314                         }
 315                         break;
 316                 case BCH_DATA_sb:
 317                 case BCH_DATA_journal:
 318                 case BCH_DATA_btree:
 319                 case BCH_DATA_user:
 320                 case BCH_DATA_parity:
 321                         if (!a.v->dirty_sectors) {
 322                                 prt_printf(err, "data_type %s but dirty_sectors==0",
 323                                        bch2_data_types[a.v->data_type]);
 324                                 return -BCH_ERR_invalid_bkey;
 325                         }
 326                         break;
 327                 case BCH_DATA_cached:
 328                         if (!a.v->cached_sectors ||
 329                             a.v->dirty_sectors ||
 330                             a.v->stripe) {
 331                                 prt_printf(err, "data type inconsistency");
 332                                 return -BCH_ERR_invalid_bkey;
 333                         }
 334
 335                         if (!a.v->io_time[READ] &&
 336                             test_bit(BCH_FS_CHECK_ALLOC_TO_LRU_REFS_DONE, &c->flags)) {
 337                                 prt_printf(err, "cached bucket with read_time == 0");
 338                                 return -BCH_ERR_invalid_bkey;
 339                         }
 340                         break;
 341                 case BCH_DATA_stripe:
 342                         if (!a.v->stripe) {
 343                                 prt_printf(err, "data_type %s but stripe==0",
 344                                        bch2_data_types[a.v->data_type]);
 345                                 return -BCH_ERR_invalid_bkey;
 346                         }
 347                         break;
 348                 }
 349         }
 350
 351         return 0;
 352 }
 353
 354 static inline u64 swab40(u64 x)
 355 {
 356         return (((x & 0x00000000ffULL) << 32)|
 357                 ((x & 0x000000ff00ULL) << 16)|
 358                 ((x & 0x0000ff0000ULL) >>  0)|
 359                 ((x & 0x00ff000000ULL) >> 16)|
 360                 ((x & 0xff00000000ULL) >> 32));
 361 }
 362
 363 void bch2_alloc_v4_swab(struct bkey_s k)
 364 {
 365         struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v;
 366         struct bch_backpointer *bp, *bps;
 367
 368         a->journal_seq          = swab64(a->journal_seq);
 369         a->flags                = swab32(a->flags);
 370         a->dirty_sectors        = swab32(a->dirty_sectors);
 371         a->cached_sectors       = swab32(a->cached_sectors);
 372         a->io_time[0]           = swab64(a->io_time[0]);
 373         a->io_time[1]           = swab64(a->io_time[1]);
 374         a->stripe               = swab32(a->stripe);
 375         a->nr_external_backpointers = swab32(a->nr_external_backpointers);
 376
 377         bps = alloc_v4_backpointers(a);
 378         for (bp = bps; bp < bps + BCH_ALLOC_V4_NR_BACKPOINTERS(a); bp++) {
 379                 bp->bucket_offset       = swab40(bp->bucket_offset);
 380                 bp->bucket_len          = swab32(bp->bucket_len);
 381                 bch2_bpos_swab(&bp->pos);
 382         }
 383 }
 384
 385 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 386 {
 387         struct bch_alloc_v4 _a;
 388         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &_a);
 389         const struct bch_backpointer *bps;
 390         unsigned i;
 391
 392         prt_newline(out);
 393         printbuf_indent_add(out, 2);
 394
 395         prt_printf(out, "gen %u oldest_gen %u data_type %s",
 396                a->gen, a->oldest_gen, bch2_data_types[a->data_type]);
 397         prt_newline(out);
 398         prt_printf(out, "journal_seq       %llu",       a->journal_seq);
 399         prt_newline(out);
 400         prt_printf(out, "need_discard      %llu",       BCH_ALLOC_V4_NEED_DISCARD(a));
 401         prt_newline(out);
 402         prt_printf(out, "need_inc_gen      %llu",       BCH_ALLOC_V4_NEED_INC_GEN(a));
 403         prt_newline(out);
 404         prt_printf(out, "dirty_sectors     %u", a->dirty_sectors);
 405         prt_newline(out);
 406         prt_printf(out, "cached_sectors    %u", a->cached_sectors);
 407         prt_newline(out);
 408         prt_printf(out, "stripe            %u", a->stripe);
 409         prt_newline(out);
 410         prt_printf(out, "stripe_redundancy %u", a->stripe_redundancy);
 411         prt_newline(out);
 412         prt_printf(out, "io_time[READ]     %llu",       a->io_time[READ]);
 413         prt_newline(out);
 414         prt_printf(out, "io_time[WRITE]    %llu",       a->io_time[WRITE]);
 415         prt_newline(out);
 416         prt_printf(out, "backpointers:     %llu",       BCH_ALLOC_V4_NR_BACKPOINTERS(a));
 417         printbuf_indent_add(out, 2);
 418
 419         bps = alloc_v4_backpointers_c(a);
 420         for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a); i++) {
 421                 prt_newline(out);
 422                 bch2_backpointer_to_text(out, &bps[i]);
 423         }
 424
 425         printbuf_indent_sub(out, 4);
 426 }
 427
 428 void __bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out)
 429 {
 430         if (k.k->type == KEY_TYPE_alloc_v4) {
 431                 int d;
 432
 433                 *out = *bkey_s_c_to_alloc_v4(k).v;
 434
 435                 d = (int) BCH_ALLOC_V4_U64s -
 436                         (int) (BCH_ALLOC_V4_BACKPOINTERS_START(out) ?: BCH_ALLOC_V4_U64s_V0);
 437                 if (unlikely(d > 0)) {
 438                         memset((u64 *) out + BCH_ALLOC_V4_BACKPOINTERS_START(out),
 439                                0,
 440                                d * sizeof(u64));
 441                         SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 442                 }
 443         } else {
 444                 struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 445
 446                 *out = (struct bch_alloc_v4) {
 447                         .journal_seq            = u.journal_seq,
 448                         .flags                  = u.need_discard,
 449                         .gen                    = u.gen,
 450                         .oldest_gen             = u.oldest_gen,
 451                         .data_type              = u.data_type,
 452                         .stripe_redundancy      = u.stripe_redundancy,
 453                         .dirty_sectors          = u.dirty_sectors,
 454                         .cached_sectors         = u.cached_sectors,
 455                         .io_time[READ]          = u.read_time,
 456                         .io_time[WRITE]         = u.write_time,
 457                         .stripe                 = u.stripe,
 458                 };
 459
 460                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 461         }
 462 }
 463
 464 static noinline struct bkey_i_alloc_v4 *
 465 __bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 466 {
 467         struct bkey_i_alloc_v4 *ret;
 468         unsigned bytes = k.k->type == KEY_TYPE_alloc_v4
 469                 ? bkey_bytes(k.k)
 470                 : sizeof(struct bkey_i_alloc_v4);
 471
 472         /*
 473          * Reserve space for one more backpointer here:
 474          * Not sketchy at doing it this way, nope...
 475          */
 476         ret = bch2_trans_kmalloc(trans, bytes + sizeof(struct bch_backpointer));
 477         if (IS_ERR(ret))
 478                 return ret;
 479
 480         if (k.k->type == KEY_TYPE_alloc_v4) {
 481                 struct bch_backpointer *src, *dst;
 482
 483                 bkey_reassemble(&ret->k_i, k);
 484
 485                 src = alloc_v4_backpointers(&ret->v);
 486                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(&ret->v, BCH_ALLOC_V4_U64s);
 487                 dst = alloc_v4_backpointers(&ret->v);
 488
 489                 memmove(dst, src, BCH_ALLOC_V4_NR_BACKPOINTERS(&ret->v) *
 490                         sizeof(struct bch_backpointer));
 491                 memset(src, 0, dst - src);
 492                 set_alloc_v4_u64s(ret);
 493         } else {
 494                 bkey_alloc_v4_init(&ret->k_i);
 495                 ret->k.p = k.k->p;
 496                 bch2_alloc_to_v4(k, &ret->v);
 497         }
 498         return ret;
 499 }
 500
 501 static inline struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut_inlined(struct btree_trans *trans, struct bkey_s_c k)
 502 {
 503         if (likely(k.k->type == KEY_TYPE_alloc_v4) &&
 504             BCH_ALLOC_V4_BACKPOINTERS_START(bkey_s_c_to_alloc_v4(k).v) == BCH_ALLOC_V4_U64s) {
 505                 /*
 506                  * Reserve space for one more backpointer here:
 507                  * Not sketchy at doing it this way, nope...
 508                  */
 509                 struct bkey_i_alloc_v4 *ret =
 510                         bch2_trans_kmalloc_nomemzero(trans, bkey_bytes(k.k) + sizeof(struct bch_backpointer));
 511                 if (!IS_ERR(ret)) {
 512                         bkey_reassemble(&ret->k_i, k);
 513                         memset((void *) ret + bkey_bytes(k.k), 0, sizeof(struct bch_backpointer));
 514                 }
 515                 return ret;
 516         }
 517
 518         return __bch2_alloc_to_v4_mut(trans, k);
 519 }
 520
 521 struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 522 {
 523         return bch2_alloc_to_v4_mut_inlined(trans, k);
 524 }
 525
 526 struct bkey_i_alloc_v4 *
 527 bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
 528                               struct bpos pos)
 529 {
 530         struct bkey_s_c k;
 531         struct bkey_i_alloc_v4 *a;
 532         int ret;
 533
 534         bch2_trans_iter_init(trans, iter, BTREE_ID_alloc, pos,
 535                              BTREE_ITER_WITH_UPDATES|
 536                              BTREE_ITER_CACHED|
 537                              BTREE_ITER_INTENT);
 538         k = bch2_btree_iter_peek_slot(iter);
 539         ret = bkey_err(k);
 540         if (unlikely(ret))
 541                 goto err;
 542
 543         a = bch2_alloc_to_v4_mut_inlined(trans, k);
 544         ret = PTR_ERR_OR_ZERO(a);
 545         if (unlikely(ret))
 546                 goto err;
 547         return a;
 548 err:
 549         bch2_trans_iter_exit(trans, iter);
 550         return ERR_PTR(ret);
 551 }
 552
 553 int bch2_alloc_read(struct bch_fs *c)
 554 {
 555         struct btree_trans trans;
 556         struct btree_iter iter;
 557         struct bkey_s_c k;
 558         struct bch_alloc_v4 a;
 559         struct bch_dev *ca;
 560         int ret;
 561
 562         bch2_trans_init(&trans, c, 0, 0);
 563
 564         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 565                            BTREE_ITER_PREFETCH, k, ret) {
 566                 /*
 567                  * Not a fsck error because this is checked/repaired by
 568                  * bch2_check_alloc_key() which runs later:
 569                  */
 570                 if (!bch2_dev_bucket_exists(c, k.k->p))
 571                         continue;
 572
 573                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 574
 575                 *bucket_gen(ca, k.k->p.offset) = bch2_alloc_to_v4(k, &a)->gen;
 576         }
 577         bch2_trans_iter_exit(&trans, &iter);
 578
 579         bch2_trans_exit(&trans);
 580
 581         if (ret)
 582                 bch_err(c, "error reading alloc info: %s", bch2_err_str(ret));
 583
 584         return ret;
 585 }
 586
 587 static struct bpos alloc_gens_pos(struct bpos pos, unsigned *offset)
 588 {
 589         *offset = pos.offset & KEY_TYPE_BUCKET_GENS_MASK;
 590
 591         pos.offset >>= KEY_TYPE_BUCKET_GENS_BITS;
 592         return pos;
 593 }
 594
 595 static struct bpos bucket_gens_pos_to_alloc(struct bpos pos, unsigned offset)
 596 {
 597         pos.offset <<= KEY_TYPE_BUCKET_GENS_BITS;
 598         pos.offset += offset;
 599         return pos;
 600 }
 601
 602 static unsigned alloc_gen(struct bkey_s_c k, unsigned offset)
 603 {
 604         return k.k->type == KEY_TYPE_bucket_gens
 605                 ? bkey_s_c_to_bucket_gens(k).v->gens[offset]
 606                 : 0;
 607 }
 608
 609 int bch2_bucket_gens_invalid(const struct bch_fs *c, struct bkey_s_c k,
 610                              int rw, struct printbuf *err)
 611 {
 612         if (bkey_val_bytes(k.k) != sizeof(struct bch_bucket_gens)) {
 613                 prt_printf(err, "bad val size (%lu != %zu)",
 614                        bkey_val_bytes(k.k), sizeof(struct bch_bucket_gens));
 615                 return -BCH_ERR_invalid_bkey;
 616         }
 617
 618         return 0;
 619 }
 620
 621 void bch2_bucket_gens_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 622 {
 623         struct bkey_s_c_bucket_gens g = bkey_s_c_to_bucket_gens(k);
 624         unsigned i;
 625
 626         for (i = 0; i < ARRAY_SIZE(g.v->gens); i++) {
 627                 if (i)
 628                         prt_char(out, ' ');
 629                 prt_printf(out, "%u", g.v->gens[i]);
 630         }
 631 }
 632
 633 int bch2_bucket_gens_init(struct bch_fs *c)
 634 {
 635         struct btree_trans trans;
 636         struct btree_iter iter;
 637         struct bkey_s_c k;
 638         struct bch_alloc_v4 a;
 639         struct bkey_i_bucket_gens g;
 640         bool have_bucket_gens_key = false;
 641         unsigned offset;
 642         struct bpos pos;
 643         u8 gen;
 644         int ret;
 645
 646         bch2_trans_init(&trans, c, 0, 0);
 647
 648         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 649                            BTREE_ITER_PREFETCH, k, ret) {
 650                 /*
 651                  * Not a fsck error because this is checked/repaired by
 652                  * bch2_check_alloc_key() which runs later:
 653                  */
 654                 if (!bch2_dev_bucket_exists(c, k.k->p))
 655                         continue;
 656
 657                 gen = bch2_alloc_to_v4(k, &a)->gen;
 658                 pos = alloc_gens_pos(iter.pos, &offset);
 659
 660                 if (have_bucket_gens_key && bkey_cmp(iter.pos, pos)) {
 661                         ret = commit_do(&trans, NULL, NULL,
 662                                         BTREE_INSERT_NOFAIL|
 663                                         BTREE_INSERT_LAZY_RW,
 664                                 __bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i));
 665                         if (ret)
 666                                 break;
 667                         have_bucket_gens_key = false;
 668                 }
 669
 670                 if (!have_bucket_gens_key) {
 671                         bkey_bucket_gens_init(&g.k_i);
 672                         g.k.p = pos;
 673                         have_bucket_gens_key = true;
 674                 }
 675
 676                 g.v.gens[offset] = gen;
 677         }
 678         bch2_trans_iter_exit(&trans, &iter);
 679
 680         if (have_bucket_gens_key && !ret)
 681                 ret = commit_do(&trans, NULL, NULL,
 682                                 BTREE_INSERT_NOFAIL|
 683                                 BTREE_INSERT_LAZY_RW,
 684                         __bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i));
 685
 686         bch2_trans_exit(&trans);
 687
 688         if (ret)
 689                 bch_err(c, "%s: error %s", __func__, bch2_err_str(ret));
 690
 691         return ret;
 692 }
 693
 694 int bch2_bucket_gens_read(struct bch_fs *c)
 695 {
 696         struct btree_trans trans;
 697         struct btree_iter iter;
 698         struct bkey_s_c k;
 699         const struct bch_bucket_gens *g;
 700         struct bch_dev *ca;
 701         u64 b;
 702         int ret;
 703
 704         bch2_trans_init(&trans, c, 0, 0);
 705
 706         for_each_btree_key(&trans, iter, BTREE_ID_bucket_gens, POS_MIN,
 707                            BTREE_ITER_PREFETCH, k, ret) {
 708                 u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
 709                 u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
 710
 711                 if (k.k->type != KEY_TYPE_bucket_gens)
 712                         continue;
 713
 714                 g = bkey_s_c_to_bucket_gens(k).v;
 715
 716                 /*
 717                  * Not a fsck error because this is checked/repaired by
 718                  * bch2_check_alloc_key() which runs later:
 719                  */
 720                 if (!bch2_dev_exists2(c, k.k->p.inode))
 721                         continue;
 722
 723                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 724
 725                 for (b = max_t(u64, ca->mi.first_bucket, start);
 726                      b < min_t(u64, ca->mi.nbuckets, end);
 727                      b++)
 728                         *bucket_gen(ca, b) = g->gens[b & KEY_TYPE_BUCKET_GENS_MASK];
 729         }
 730         bch2_trans_iter_exit(&trans, &iter);
 731
 732         bch2_trans_exit(&trans);
 733
 734         if (ret)
 735                 bch_err(c, "error reading alloc info: %s", bch2_err_str(ret));
 736
 737         return ret;
 738 }
 739
 740 /* Free space/discard btree: */
 741
 742 static int bch2_bucket_do_index(struct btree_trans *trans,
 743                                 struct bkey_s_c alloc_k,
 744                                 const struct bch_alloc_v4 *a,
 745                                 bool set)
 746 {
 747         struct bch_fs *c = trans->c;
 748         struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 749         struct btree_iter iter;
 750         struct bkey_s_c old;
 751         struct bkey_i *k;
 752         enum btree_id btree;
 753         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 754         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 755         struct printbuf buf = PRINTBUF;
 756         int ret;
 757
 758         if (a->data_type != BCH_DATA_free &&
 759             a->data_type != BCH_DATA_need_discard)
 760                 return 0;
 761
 762         k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
 763         if (IS_ERR(k))
 764                 return PTR_ERR(k);
 765
 766         bkey_init(&k->k);
 767         k->k.type = new_type;
 768
 769         switch (a->data_type) {
 770         case BCH_DATA_free:
 771                 btree = BTREE_ID_freespace;
 772                 k->k.p = alloc_freespace_pos(alloc_k.k->p, *a);
 773                 bch2_key_resize(&k->k, 1);
 774                 break;
 775         case BCH_DATA_need_discard:
 776                 btree = BTREE_ID_need_discard;
 777                 k->k.p = alloc_k.k->p;
 778                 break;
 779         default:
 780                 return 0;
 781         }
 782
 783         bch2_trans_iter_init(trans, &iter, btree,
 784                              bkey_start_pos(&k->k),
 785                              BTREE_ITER_INTENT);
 786         old = bch2_btree_iter_peek_slot(&iter);
 787         ret = bkey_err(old);
 788         if (ret)
 789                 goto err;
 790
 791         if (ca->mi.freespace_initialized &&
 792             bch2_trans_inconsistent_on(old.k->type != old_type, trans,
 793                         "incorrect key when %s %s btree (got %s should be %s)\n"
 794                         "  for %s",
 795                         set ? "setting" : "clearing",
 796                         bch2_btree_ids[btree],
 797                         bch2_bkey_types[old.k->type],
 798                         bch2_bkey_types[old_type],
 799                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 800                 ret = -EIO;
 801                 goto err;
 802         }
 803
 804         ret = bch2_trans_update(trans, &iter, k, 0);
 805 err:
 806         bch2_trans_iter_exit(trans, &iter);
 807         printbuf_exit(&buf);
 808         return ret;
 809 }
 810
 811 static noinline int bch2_bucket_gen_update(struct btree_trans *trans,
 812                                            struct bpos bucket, u8 gen)
 813 {
 814         struct btree_iter iter;
 815         unsigned offset;
 816         struct bpos pos = alloc_gens_pos(bucket, &offset);
 817         struct bkey_i_bucket_gens *g;
 818         struct bkey_s_c k;
 819         int ret;
 820
 821         g = bch2_trans_kmalloc(trans, sizeof(*g));
 822         ret = PTR_ERR_OR_ZERO(g);
 823         if (ret)
 824                 return ret;
 825
 826         bch2_trans_iter_init(trans, &iter, BTREE_ID_bucket_gens, pos,
 827                              BTREE_ITER_INTENT|
 828                              BTREE_ITER_WITH_UPDATES);
 829         k = bch2_btree_iter_peek_slot(&iter);
 830         ret = bkey_err(k);
 831         if (ret)
 832                 goto err;
 833
 834         if (k.k->type != KEY_TYPE_bucket_gens) {
 835                 bkey_bucket_gens_init(&g->k_i);
 836                 g->k.p = iter.pos;
 837         } else {
 838                 bkey_reassemble(&g->k_i, k);
 839         }
 840
 841         g->v.gens[offset] = gen;
 842
 843         ret = bch2_trans_update(trans, &iter, &g->k_i, 0);
 844 err:
 845         bch2_trans_iter_exit(trans, &iter);
 846         return ret;
 847 }
 848
 849 int bch2_trans_mark_alloc(struct btree_trans *trans,
 850                           enum btree_id btree_id, unsigned level,
 851                           struct bkey_s_c old, struct bkey_i *new,
 852                           unsigned flags)
 853 {
 854         struct bch_fs *c = trans->c;
 855         struct bch_alloc_v4 old_a_convert, *new_a;
 856         const struct bch_alloc_v4 *old_a;
 857         u64 old_lru, new_lru;
 858         int ret = 0;
 859
 860         /*
 861          * Deletion only happens in the device removal path, with
 862          * BTREE_TRIGGER_NORUN:
 863          */
 864         BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
 865
 866         old_a = bch2_alloc_to_v4(old, &old_a_convert);
 867         new_a = &bkey_i_to_alloc_v4(new)->v;
 868
 869         new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
 870
 871         if (new_a->dirty_sectors > old_a->dirty_sectors ||
 872             new_a->cached_sectors > old_a->cached_sectors) {
 873                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 874                 new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 875                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
 876                 SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
 877         }
 878
 879         if (data_type_is_empty(new_a->data_type) &&
 880             BCH_ALLOC_V4_NEED_INC_GEN(new_a) &&
 881             !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
 882                 new_a->gen++;
 883                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
 884         }
 885
 886         if (old_a->data_type != new_a->data_type ||
 887             (new_a->data_type == BCH_DATA_free &&
 888              alloc_freespace_genbits(*old_a) != alloc_freespace_genbits(*new_a))) {
 889                 ret =   bch2_bucket_do_index(trans, old, old_a, false) ?:
 890                         bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_a, true);
 891                 if (ret)
 892                         return ret;
 893         }
 894
 895         if (new_a->data_type == BCH_DATA_cached &&
 896             !new_a->io_time[READ])
 897                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 898
 899         old_lru = alloc_lru_idx(*old_a);
 900         new_lru = alloc_lru_idx(*new_a);
 901
 902         if (old_lru != new_lru) {
 903                 ret = bch2_lru_change(trans, new->k.p.inode, new->k.p.offset,
 904                                       old_lru, &new_lru, old);
 905                 if (ret)
 906                         return ret;
 907
 908                 if (new_a->data_type == BCH_DATA_cached)
 909                         new_a->io_time[READ] = new_lru;
 910         }
 911
 912         if (old_a->gen != new_a->gen) {
 913                 ret = bch2_bucket_gen_update(trans, new->k.p, new_a->gen);
 914                 if (ret)
 915                         return ret;
 916         }
 917
 918         return 0;
 919 }
 920
 921 /*
 922  * This synthesizes deleted extents for holes, similar to BTREE_ITER_SLOTS for
 923  * extents style btrees, but works on non-extents btrees:
 924  */
 925 struct bkey_s_c bch2_get_key_or_hole(struct btree_iter *iter, struct bpos end, struct bkey *hole)
 926 {
 927         struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
 928
 929         if (bkey_err(k))
 930                 return k;
 931
 932         if (k.k->type) {
 933                 return k;
 934         } else {
 935                 struct btree_iter iter2;
 936                 struct bpos next;
 937
 938                 bch2_trans_copy_iter(&iter2, iter);
 939                 k = bch2_btree_iter_peek_upto(&iter2,
 940                                 bkey_min(bkey_min(end,
 941                                                   iter->path->l[0].b->key.k.p),
 942                                                   POS(iter->pos.inode, iter->pos.offset + U32_MAX - 1)));
 943                 next = iter2.pos;
 944                 bch2_trans_iter_exit(iter->trans, &iter2);
 945
 946                 BUG_ON(next.offset >= iter->pos.offset + U32_MAX);
 947
 948                 if (bkey_err(k))
 949                         return k;
 950
 951                 bkey_init(hole);
 952                 hole->p = iter->pos;
 953
 954                 bch2_key_resize(hole, next.offset - iter->pos.offset);
 955                 return (struct bkey_s_c) { hole, NULL };
 956         }
 957 }
 958
 959 static bool next_bucket(struct bch_fs *c, struct bpos *bucket)
 960 {
 961         struct bch_dev *ca;
 962         unsigned iter;
 963
 964         if (bch2_dev_bucket_exists(c, *bucket))
 965                 return true;
 966
 967         if (bch2_dev_exists2(c, bucket->inode)) {
 968                 ca = bch_dev_bkey_exists(c, bucket->inode);
 969
 970                 if (bucket->offset < ca->mi.first_bucket) {
 971                         bucket->offset = ca->mi.first_bucket;
 972                         return true;
 973                 }
 974
 975                 bucket->inode++;
 976                 bucket->offset = 0;
 977         }
 978
 979         rcu_read_lock();
 980         iter = bucket->inode;
 981         ca = __bch2_next_dev(c, &iter, NULL);
 982         if (ca)
 983                 bucket->offset = ca->mi.first_bucket;
 984         rcu_read_unlock();
 985
 986         return ca != NULL;
 987 }
 988
 989 struct bkey_s_c bch2_get_key_or_real_bucket_hole(struct btree_iter *iter, struct bkey *hole)
 990 {
 991         struct bch_fs *c = iter->trans->c;
 992         struct bkey_s_c k;
 993 again:
 994         k = bch2_get_key_or_hole(iter, POS_MAX, hole);
 995         if (bkey_err(k))
 996                 return k;
 997
 998         if (!k.k->type) {
 999                 struct bpos bucket = bkey_start_pos(k.k);
1000
1001                 if (!bch2_dev_bucket_exists(c, bucket)) {
1002                         if (!next_bucket(c, &bucket))
1003                                 return bkey_s_c_null;
1004
1005                         bch2_btree_iter_set_pos(iter, bucket);
1006                         goto again;
1007                 }
1008
1009                 if (!bch2_dev_bucket_exists(c, k.k->p)) {
1010                         struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
1011
1012                         bch2_key_resize(hole, ca->mi.nbuckets - bucket.offset);
1013                 }
1014         }
1015
1016         return k;
1017 }
1018
1019 static int bch2_check_alloc_key(struct btree_trans *trans,
1020                                 struct bkey_s_c alloc_k,
1021                                 struct btree_iter *alloc_iter,
1022                                 struct btree_iter *discard_iter,
1023                                 struct btree_iter *freespace_iter,
1024                                 struct btree_iter *bucket_gens_iter)
1025 {
1026         struct bch_fs *c = trans->c;
1027         struct bch_dev *ca;
1028         struct bch_alloc_v4 a_convert;
1029         const struct bch_alloc_v4 *a;
1030         unsigned discard_key_type, freespace_key_type;
1031         unsigned gens_offset;
1032         struct bkey_s_c k;
1033         struct printbuf buf = PRINTBUF;
1034         int ret;
1035
1036         if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
1037                         "alloc key for invalid device:bucket %llu:%llu",
1038                         alloc_k.k->p.inode, alloc_k.k->p.offset))
1039                 return bch2_btree_delete_at(trans, alloc_iter, 0);
1040
1041         ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
1042         if (!ca->mi.freespace_initialized)
1043                 return 0;
1044
1045         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1046
1047         discard_key_type = a->data_type == BCH_DATA_need_discard ? KEY_TYPE_set : 0;
1048         bch2_btree_iter_set_pos(discard_iter, alloc_k.k->p);
1049         k = bch2_btree_iter_peek_slot(discard_iter);
1050         ret = bkey_err(k);
1051         if (ret)
1052                 goto err;
1053
1054         if (k.k->type != discard_key_type &&
1055             (c->opts.reconstruct_alloc ||
1056              fsck_err(c, "incorrect key in need_discard btree (got %s should be %s)\n"
1057                       "  %s",
1058                       bch2_bkey_types[k.k->type],
1059                       bch2_bkey_types[discard_key_type],
1060                       (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1061                 struct bkey_i *update =
1062                         bch2_trans_kmalloc(trans, sizeof(*update));
1063
1064                 ret = PTR_ERR_OR_ZERO(update);
1065                 if (ret)
1066                         goto err;
1067
1068                 bkey_init(&update->k);
1069                 update->k.type  = discard_key_type;
1070                 update->k.p     = discard_iter->pos;
1071
1072                 ret = bch2_trans_update(trans, discard_iter, update, 0);
1073                 if (ret)
1074                         goto err;
1075         }
1076
1077         freespace_key_type = a->data_type == BCH_DATA_free ? KEY_TYPE_set : 0;
1078         bch2_btree_iter_set_pos(freespace_iter, alloc_freespace_pos(alloc_k.k->p, *a));
1079         k = bch2_btree_iter_peek_slot(freespace_iter);
1080         ret = bkey_err(k);
1081         if (ret)
1082                 goto err;
1083
1084         if (k.k->type != freespace_key_type &&
1085             (c->opts.reconstruct_alloc ||
1086              fsck_err(c, "incorrect key in freespace btree (got %s should be %s)\n"
1087                       "  %s",
1088                       bch2_bkey_types[k.k->type],
1089                       bch2_bkey_types[freespace_key_type],
1090                       (printbuf_reset(&buf),
1091                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1092                 struct bkey_i *update =
1093                         bch2_trans_kmalloc(trans, sizeof(*update));
1094
1095                 ret = PTR_ERR_OR_ZERO(update);
1096                 if (ret)
1097                         goto err;
1098
1099                 bkey_init(&update->k);
1100                 update->k.type  = freespace_key_type;
1101                 update->k.p     = freespace_iter->pos;
1102                 bch2_key_resize(&update->k, 1);
1103
1104                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1105                 if (ret)
1106                         goto err;
1107         }
1108
1109         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(alloc_k.k->p, &gens_offset));
1110         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1111         ret = bkey_err(k);
1112         if (ret)
1113                 goto err;
1114
1115         if (a->gen != alloc_gen(k, gens_offset) &&
1116             (c->opts.reconstruct_alloc ||
1117              fsck_err(c, "incorrect gen in bucket_gens btree (got %u should be %u)\n"
1118                       "  %s",
1119                       alloc_gen(k, gens_offset), a->gen,
1120                       (printbuf_reset(&buf),
1121                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1122                 struct bkey_i_bucket_gens *g =
1123                         bch2_trans_kmalloc(trans, sizeof(*g));
1124
1125                 ret = PTR_ERR_OR_ZERO(g);
1126                 if (ret)
1127                         goto err;
1128
1129                 if (k.k->type == KEY_TYPE_bucket_gens) {
1130                         bkey_reassemble(&g->k_i, k);
1131                 } else {
1132                         bkey_bucket_gens_init(&g->k_i);
1133                         g->k.p = alloc_gens_pos(alloc_k.k->p, &gens_offset);
1134                 }
1135
1136                 g->v.gens[gens_offset] = a->gen;
1137
1138                 ret = bch2_trans_update(trans, bucket_gens_iter, &g->k_i, 0);
1139                 if (ret)
1140                         goto err;
1141         }
1142 err:
1143 fsck_err:
1144         printbuf_exit(&buf);
1145         return ret;
1146 }
1147
1148 static int bch2_check_alloc_hole_freespace(struct btree_trans *trans,
1149                                  struct bpos start,
1150                                  struct bpos *end,
1151                                  struct btree_iter *freespace_iter)
1152 {
1153         struct bch_fs *c = trans->c;
1154         struct bch_dev *ca;
1155         struct bkey_s_c k;
1156         struct printbuf buf = PRINTBUF;
1157         int ret;
1158
1159         ca = bch_dev_bkey_exists(c, start.inode);
1160         if (!ca->mi.freespace_initialized)
1161                 return 0;
1162
1163         bch2_btree_iter_set_pos(freespace_iter, start);
1164
1165         k = bch2_btree_iter_peek_slot(freespace_iter);
1166         ret = bkey_err(k);
1167         if (ret)
1168                 goto err;
1169
1170         *end = bkey_min(k.k->p, *end);
1171
1172         if (k.k->type != KEY_TYPE_set &&
1173             (c->opts.reconstruct_alloc ||
1174              fsck_err(c, "hole in alloc btree missing in freespace btree\n"
1175                       "  device %llu buckets %llu-%llu",
1176                       freespace_iter->pos.inode,
1177                       freespace_iter->pos.offset,
1178                       end->offset))) {
1179                 struct bkey_i *update =
1180                         bch2_trans_kmalloc(trans, sizeof(*update));
1181
1182                 ret = PTR_ERR_OR_ZERO(update);
1183                 if (ret)
1184                         goto err;
1185
1186                 bkey_init(&update->k);
1187                 update->k.type  = KEY_TYPE_set;
1188                 update->k.p     = freespace_iter->pos;
1189                 bch2_key_resize(&update->k,
1190                                 min_t(u64, U32_MAX, end->offset -
1191                                       freespace_iter->pos.offset));
1192
1193                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1194                 if (ret)
1195                         goto err;
1196         }
1197 err:
1198 fsck_err:
1199         printbuf_exit(&buf);
1200         return ret;
1201 }
1202
1203 static int bch2_check_alloc_hole_bucket_gens(struct btree_trans *trans,
1204                                  struct bpos start,
1205                                  struct bpos *end,
1206                                  struct btree_iter *bucket_gens_iter)
1207 {
1208         struct bch_fs *c = trans->c;
1209         struct bkey_s_c k;
1210         struct printbuf buf = PRINTBUF;
1211         unsigned i, gens_offset, gens_end_offset;
1212         int ret;
1213
1214         if (c->sb.version < bcachefs_metadata_version_bucket_gens &&
1215             !c->opts.version_upgrade)
1216                 return 0;
1217
1218         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(start, &gens_offset));
1219
1220         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1221         ret = bkey_err(k);
1222         if (ret)
1223                 goto err;
1224
1225         if (bkey_cmp(alloc_gens_pos(start, &gens_offset),
1226                      alloc_gens_pos(*end,  &gens_end_offset)))
1227                 gens_end_offset = KEY_TYPE_BUCKET_GENS_NR;
1228
1229         if (k.k->type == KEY_TYPE_bucket_gens) {
1230                 struct bkey_i_bucket_gens g;
1231                 bool need_update = false;
1232
1233                 bkey_reassemble(&g.k_i, k);
1234
1235                 for (i = gens_offset; i < gens_end_offset; i++) {
1236                         if (fsck_err_on(g.v.gens[i], c,
1237                                         "hole in alloc btree at %llu:%llu with nonzero gen in bucket_gens btree (%u)",
1238                                         bucket_gens_pos_to_alloc(k.k->p, i).inode,
1239                                         bucket_gens_pos_to_alloc(k.k->p, i).offset,
1240                                         g.v.gens[i])) {
1241                                 g.v.gens[i] = 0;
1242                                 need_update = true;
1243                         }
1244                 }
1245
1246                 if (need_update) {
1247                         ret = bch2_trans_update(trans, bucket_gens_iter, &g.k_i, 0);
1248                         if (ret)
1249                                 goto err;
1250                 }
1251         }
1252
1253         *end = bkey_min(*end, bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0));
1254 err:
1255 fsck_err:
1256         printbuf_exit(&buf);
1257         return ret;
1258 }
1259
1260 static int bch2_check_discard_freespace_key(struct btree_trans *trans,
1261                                             struct btree_iter *iter)
1262 {
1263         struct bch_fs *c = trans->c;
1264         struct btree_iter alloc_iter;
1265         struct bkey_s_c alloc_k;
1266         struct bch_alloc_v4 a_convert;
1267         const struct bch_alloc_v4 *a;
1268         u64 genbits;
1269         struct bpos pos;
1270         enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
1271                 ? BCH_DATA_need_discard
1272                 : BCH_DATA_free;
1273         struct printbuf buf = PRINTBUF;
1274         int ret;
1275
1276         pos = iter->pos;
1277         pos.offset &= ~(~0ULL << 56);
1278         genbits = iter->pos.offset & (~0ULL << 56);
1279
1280         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
1281
1282         if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
1283                         "entry in %s btree for nonexistant dev:bucket %llu:%llu",
1284                         bch2_btree_ids[iter->btree_id], pos.inode, pos.offset))
1285                 goto delete;
1286
1287         alloc_k = bch2_btree_iter_peek_slot(&alloc_iter);
1288         ret = bkey_err(alloc_k);
1289         if (ret)
1290                 goto err;
1291
1292         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1293
1294         if (fsck_err_on(a->data_type != state ||
1295                         (state == BCH_DATA_free &&
1296                          genbits != alloc_freespace_genbits(*a)), c,
1297                         "%s\n  incorrectly set in %s index (free %u, genbits %llu should be %llu)",
1298                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
1299                         bch2_btree_ids[iter->btree_id],
1300                         a->data_type == state,
1301                         genbits >> 56, alloc_freespace_genbits(*a) >> 56))
1302                 goto delete;
1303 out:
1304 err:
1305 fsck_err:
1306         bch2_trans_iter_exit(trans, &alloc_iter);
1307         printbuf_exit(&buf);
1308         return ret;
1309 delete:
1310         ret = bch2_btree_delete_extent_at(trans, iter,
1311                         iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0);
1312         goto out;
1313 }
1314
1315 /*
1316  * We've already checked that generation numbers in the bucket_gens btree are
1317  * valid for buckets that exist; this just checks for keys for nonexistent
1318  * buckets.
1319  */
1320 static int bch2_check_bucket_gens_key(struct btree_trans *trans,
1321                                       struct btree_iter *iter,
1322                                       struct bkey_s_c k)
1323 {
1324         struct bch_fs *c = trans->c;
1325         struct bkey_i_bucket_gens g;
1326         struct bch_dev *ca;
1327         u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
1328         u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
1329         u64 b;
1330         bool need_update = false;
1331         struct printbuf buf = PRINTBUF;
1332         int ret = 0;
1333
1334         BUG_ON(k.k->type != KEY_TYPE_bucket_gens);
1335         bkey_reassemble(&g.k_i, k);
1336
1337         if (fsck_err_on(!bch2_dev_exists2(c, k.k->p.inode), c,
1338                         "bucket_gens key for invalid device:\n  %s",
1339                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1340                 ret = bch2_btree_delete_at(trans, iter, 0);
1341                 goto out;
1342         }
1343
1344         ca = bch_dev_bkey_exists(c, k.k->p.inode);
1345         if (fsck_err_on(end <= ca->mi.first_bucket ||
1346                         start >= ca->mi.nbuckets, c,
1347                         "bucket_gens key for invalid buckets:\n  %s",
1348                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1349                 ret = bch2_btree_delete_at(trans, iter, 0);
1350                 goto out;
1351         }
1352
1353         for (b = start; b < ca->mi.first_bucket; b++)
1354                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1355                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1356                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1357                         need_update = true;
1358                 }
1359
1360         for (b = ca->mi.nbuckets; b < end; b++)
1361                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1362                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1363                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1364                         need_update = true;
1365                 }
1366
1367         if (need_update) {
1368                 struct bkey_i *k;
1369
1370                 k = bch2_trans_kmalloc(trans, sizeof(g));
1371                 ret = PTR_ERR_OR_ZERO(k);
1372                 if (ret)
1373                         return ret;
1374
1375                 memcpy(k, &g, sizeof(g));
1376                 ret = bch2_trans_update(trans, iter, k, 0);
1377         }
1378 out:
1379 fsck_err:
1380         printbuf_exit(&buf);
1381         return ret;
1382 }
1383
1384 int bch2_check_alloc_info(struct bch_fs *c)
1385 {
1386         struct btree_trans trans;
1387         struct btree_iter iter, discard_iter, freespace_iter, bucket_gens_iter;
1388         struct bkey hole;
1389         struct bkey_s_c k;
1390         int ret = 0;
1391
1392         bch2_trans_init(&trans, c, 0, 0);
1393
1394         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc, POS_MIN,
1395                              BTREE_ITER_PREFETCH);
1396         bch2_trans_iter_init(&trans, &discard_iter, BTREE_ID_need_discard, POS_MIN,
1397                              BTREE_ITER_PREFETCH);
1398         bch2_trans_iter_init(&trans, &freespace_iter, BTREE_ID_freespace, POS_MIN,
1399                              BTREE_ITER_PREFETCH);
1400         bch2_trans_iter_init(&trans, &bucket_gens_iter, BTREE_ID_bucket_gens, POS_MIN,
1401                              BTREE_ITER_PREFETCH);
1402
1403         while (1) {
1404                 struct bpos next;
1405
1406                 bch2_trans_begin(&trans);
1407
1408                 k = bch2_get_key_or_real_bucket_hole(&iter, &hole);
1409                 ret = bkey_err(k);
1410                 if (ret)
1411                         goto bkey_err;
1412
1413                 if (!k.k)
1414                         break;
1415
1416                 if (k.k->type) {
1417                         next = bpos_nosnap_successor(k.k->p);
1418
1419                         ret = bch2_check_alloc_key(&trans,
1420                                                    k, &iter,
1421                                                    &discard_iter,
1422                                                    &freespace_iter,
1423                                                    &bucket_gens_iter);
1424                         if (ret)
1425                                 break;
1426                 } else {
1427                         next = k.k->p;
1428
1429                         ret = bch2_check_alloc_hole_freespace(&trans,
1430                                                     bkey_start_pos(k.k),
1431                                                     &next,
1432                                                     &freespace_iter) ?:
1433                                 bch2_check_alloc_hole_bucket_gens(&trans,
1434                                                     bkey_start_pos(k.k),
1435                                                     &next,
1436                                                     &bucket_gens_iter);
1437                         if (ret)
1438                                 goto bkey_err;
1439                 }
1440
1441                 ret = bch2_trans_commit(&trans, NULL, NULL,
1442                                         BTREE_INSERT_NOFAIL|
1443                                         BTREE_INSERT_LAZY_RW);
1444                 if (ret)
1445                         goto bkey_err;
1446
1447                 bch2_btree_iter_set_pos(&iter, next);
1448 bkey_err:
1449                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1450                         continue;
1451                 if (ret)
1452                         break;
1453         }
1454         bch2_trans_iter_exit(&trans, &bucket_gens_iter);
1455         bch2_trans_iter_exit(&trans, &freespace_iter);
1456         bch2_trans_iter_exit(&trans, &discard_iter);
1457         bch2_trans_iter_exit(&trans, &iter);
1458
1459         if (ret < 0)
1460                 goto err;
1461
1462         ret = for_each_btree_key_commit(&trans, iter,
1463                         BTREE_ID_need_discard, POS_MIN,
1464                         BTREE_ITER_PREFETCH, k,
1465                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1466                 bch2_check_discard_freespace_key(&trans, &iter)) ?:
1467               for_each_btree_key_commit(&trans, iter,
1468                         BTREE_ID_freespace, POS_MIN,
1469                         BTREE_ITER_PREFETCH, k,
1470                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1471                 bch2_check_discard_freespace_key(&trans, &iter)) ?:
1472               for_each_btree_key_commit(&trans, iter,
1473                         BTREE_ID_bucket_gens, POS_MIN,
1474                         BTREE_ITER_PREFETCH, k,
1475                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1476                 bch2_check_bucket_gens_key(&trans, &iter, k));
1477 err:
1478         bch2_trans_exit(&trans);
1479         return ret < 0 ? ret : 0;
1480 }
1481
1482 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
1483                                        struct btree_iter *alloc_iter)
1484 {
1485         struct bch_fs *c = trans->c;
1486         struct btree_iter lru_iter;
1487         struct bch_alloc_v4 a_convert;
1488         const struct bch_alloc_v4 *a;
1489         struct bkey_s_c alloc_k, k;
1490         struct printbuf buf = PRINTBUF;
1491         struct printbuf buf2 = PRINTBUF;
1492         int ret;
1493
1494         alloc_k = bch2_btree_iter_peek(alloc_iter);
1495         if (!alloc_k.k)
1496                 return 0;
1497
1498         ret = bkey_err(alloc_k);
1499         if (ret)
1500                 return ret;
1501
1502         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1503
1504         if (a->data_type != BCH_DATA_cached)
1505                 return 0;
1506
1507         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
1508                              POS(alloc_k.k->p.inode, a->io_time[READ]), 0);
1509
1510         k = bch2_btree_iter_peek_slot(&lru_iter);
1511         ret = bkey_err(k);
1512         if (ret)
1513                 goto err;
1514
1515         if (fsck_err_on(!a->io_time[READ], c,
1516                         "cached bucket with read_time 0\n"
1517                         "  %s",
1518                 (printbuf_reset(&buf),
1519                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) ||
1520             fsck_err_on(k.k->type != KEY_TYPE_lru ||
1521                         le64_to_cpu(bkey_s_c_to_lru(k).v->idx) != alloc_k.k->p.offset, c,
1522                         "incorrect/missing lru entry\n"
1523                         "  %s\n"
1524                         "  %s",
1525                         (printbuf_reset(&buf),
1526                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
1527                         (bch2_bkey_val_to_text(&buf2, c, k), buf2.buf))) {
1528                 u64 read_time = a->io_time[READ] ?:
1529                         atomic64_read(&c->io_clock[READ].now);
1530
1531                 ret = bch2_lru_set(trans,
1532                                    alloc_k.k->p.inode,
1533                                    alloc_k.k->p.offset,
1534                                    &read_time);
1535                 if (ret)
1536                         goto err;
1537
1538                 if (a->io_time[READ] != read_time) {
1539                         struct bkey_i_alloc_v4 *a_mut =
1540                                 bch2_alloc_to_v4_mut(trans, alloc_k);
1541                         ret = PTR_ERR_OR_ZERO(a_mut);
1542                         if (ret)
1543                                 goto err;
1544
1545                         a_mut->v.io_time[READ] = read_time;
1546                         ret = bch2_trans_update(trans, alloc_iter,
1547                                                 &a_mut->k_i, BTREE_TRIGGER_NORUN);
1548                         if (ret)
1549                                 goto err;
1550                 }
1551         }
1552 err:
1553 fsck_err:
1554         bch2_trans_iter_exit(trans, &lru_iter);
1555         printbuf_exit(&buf2);
1556         printbuf_exit(&buf);
1557         return ret;
1558 }
1559
1560 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
1561 {
1562         struct btree_trans trans;
1563         struct btree_iter iter;
1564         struct bkey_s_c k;
1565         int ret = 0;
1566
1567         bch2_trans_init(&trans, c, 0, 0);
1568
1569         for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1570                         POS_MIN, BTREE_ITER_PREFETCH, k,
1571                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1572                 bch2_check_alloc_to_lru_ref(&trans, &iter));
1573
1574         bch2_trans_exit(&trans);
1575         return ret < 0 ? ret : 0;
1576 }
1577
1578 static int bch2_discard_one_bucket(struct btree_trans *trans,
1579                                    struct btree_iter *need_discard_iter,
1580                                    struct bpos *discard_pos_done,
1581                                    u64 *seen,
1582                                    u64 *open,
1583                                    u64 *need_journal_commit,
1584                                    u64 *discarded)
1585 {
1586         struct bch_fs *c = trans->c;
1587         struct bpos pos = need_discard_iter->pos;
1588         struct btree_iter iter = { NULL };
1589         struct bkey_s_c k;
1590         struct bch_dev *ca;
1591         struct bkey_i_alloc_v4 *a;
1592         struct printbuf buf = PRINTBUF;
1593         bool did_discard = false;
1594         int ret = 0;
1595
1596         ca = bch_dev_bkey_exists(c, pos.inode);
1597         if (!percpu_ref_tryget(&ca->io_ref)) {
1598                 bch2_btree_iter_set_pos(need_discard_iter, POS(pos.inode + 1, 0));
1599                 return 0;
1600         }
1601
1602         if (bch2_bucket_is_open_safe(c, pos.inode, pos.offset)) {
1603                 (*open)++;
1604                 goto out;
1605         }
1606
1607         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
1608                         c->journal.flushed_seq_ondisk,
1609                         pos.inode, pos.offset)) {
1610                 (*need_journal_commit)++;
1611                 goto out;
1612         }
1613
1614         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
1615                              need_discard_iter->pos,
1616                              BTREE_ITER_CACHED);
1617         k = bch2_btree_iter_peek_slot(&iter);
1618         ret = bkey_err(k);
1619         if (ret)
1620                 goto out;
1621
1622         a = bch2_alloc_to_v4_mut(trans, k);
1623         ret = PTR_ERR_OR_ZERO(a);
1624         if (ret)
1625                 goto out;
1626
1627         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
1628                 a->v.gen++;
1629                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1630                 goto write;
1631         }
1632
1633         if (bch2_trans_inconsistent_on(a->v.journal_seq > c->journal.flushed_seq_ondisk, trans,
1634                         "clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
1635                         "%s",
1636                         a->v.journal_seq,
1637                         c->journal.flushed_seq_ondisk,
1638                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1639                 ret = -EIO;
1640                 goto out;
1641         }
1642
1643         if (bch2_trans_inconsistent_on(a->v.data_type != BCH_DATA_need_discard, trans,
1644                         "bucket incorrectly set in need_discard btree\n"
1645                         "%s",
1646                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1647                 ret = -EIO;
1648                 goto out;
1649         }
1650
1651         if (!bkey_eq(*discard_pos_done, iter.pos) &&
1652             ca->mi.discard && !c->opts.nochanges) {
1653                 /*
1654                  * This works without any other locks because this is the only
1655                  * thread that removes items from the need_discard tree
1656                  */
1657                 bch2_trans_unlock(trans);
1658                 blkdev_issue_discard(ca->disk_sb.bdev,
1659                                      k.k->p.offset * ca->mi.bucket_size,
1660                                      ca->mi.bucket_size,
1661                                      GFP_KERNEL);
1662
1663                 ret = bch2_trans_relock(trans);
1664                 if (ret)
1665                         goto out;
1666         }
1667
1668         *discard_pos_done = iter.pos;
1669         did_discard = true;
1670
1671         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
1672         a->v.data_type = alloc_data_type(a->v, a->v.data_type);
1673 write:
1674         ret =   bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1675                 bch2_trans_commit(trans, NULL, NULL,
1676                                   BTREE_INSERT_USE_RESERVE|BTREE_INSERT_NOFAIL);
1677         if (ret)
1678                 goto out;
1679
1680         if (did_discard) {
1681                 this_cpu_inc(c->counters[BCH_COUNTER_bucket_discard]);
1682                 (*discarded)++;
1683         }
1684 out:
1685         bch2_trans_iter_exit(trans, &iter);
1686         percpu_ref_put(&ca->io_ref);
1687         printbuf_exit(&buf);
1688         return ret;
1689 }
1690
1691 static void bch2_do_discards_work(struct work_struct *work)
1692 {
1693         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
1694         struct btree_trans trans;
1695         struct btree_iter iter;
1696         struct bkey_s_c k;
1697         u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
1698         struct bpos discard_pos_done = POS_MAX;
1699         int ret;
1700
1701         bch2_trans_init(&trans, c, 0, 0);
1702
1703         /*
1704          * We're doing the commit in bch2_discard_one_bucket instead of using
1705          * for_each_btree_key_commit() so that we can increment counters after
1706          * successful commit:
1707          */
1708         ret = for_each_btree_key2(&trans, iter,
1709                         BTREE_ID_need_discard, POS_MIN, 0, k,
1710                 bch2_discard_one_bucket(&trans, &iter, &discard_pos_done,
1711                                         &seen,
1712                                         &open,
1713                                         &need_journal_commit,
1714                                         &discarded));
1715
1716         bch2_trans_exit(&trans);
1717
1718         if (need_journal_commit * 2 > seen)
1719                 bch2_journal_flush_async(&c->journal, NULL);
1720
1721         percpu_ref_put(&c->writes);
1722
1723         trace_discard_buckets(c, seen, open, need_journal_commit, discarded,
1724                               bch2_err_str(ret));
1725 }
1726
1727 void bch2_do_discards(struct bch_fs *c)
1728 {
1729         if (percpu_ref_tryget_live(&c->writes) &&
1730             !queue_work(system_long_wq, &c->discard_work))
1731                 percpu_ref_put(&c->writes);
1732 }
1733
1734 static int invalidate_one_bucket(struct btree_trans *trans,
1735                                  struct btree_iter *lru_iter, struct bkey_s_c k,
1736                                  unsigned dev_idx, s64 *nr_to_invalidate)
1737 {
1738         struct bch_fs *c = trans->c;
1739         struct btree_iter alloc_iter = { NULL };
1740         struct bkey_i_alloc_v4 *a;
1741         struct bpos bucket;
1742         struct printbuf buf = PRINTBUF;
1743         unsigned cached_sectors;
1744         int ret = 0;
1745
1746         if (*nr_to_invalidate <= 0 || k.k->p.inode != dev_idx)
1747                 return 1;
1748
1749         if (k.k->type != KEY_TYPE_lru) {
1750                 prt_printf(&buf, "non lru key in lru btree:\n  ");
1751                 bch2_bkey_val_to_text(&buf, c, k);
1752
1753                 if (!test_bit(BCH_FS_CHECK_LRUS_DONE, &c->flags)) {
1754                         bch_err(c, "%s", buf.buf);
1755                 } else {
1756                         bch2_trans_inconsistent(trans, "%s", buf.buf);
1757                         ret = -EINVAL;
1758                 }
1759
1760                 goto out;
1761         }
1762
1763         bucket = POS(dev_idx, le64_to_cpu(bkey_s_c_to_lru(k).v->idx));
1764
1765         a = bch2_trans_start_alloc_update(trans, &alloc_iter, bucket);
1766         ret = PTR_ERR_OR_ZERO(a);
1767         if (ret)
1768                 goto out;
1769
1770         if (k.k->p.offset != alloc_lru_idx(a->v)) {
1771                 prt_printf(&buf, "alloc key does not point back to lru entry when invalidating bucket:\n  ");
1772                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
1773                 prt_printf(&buf, "\n  ");
1774                 bch2_bkey_val_to_text(&buf, c, k);
1775
1776                 if (!test_bit(BCH_FS_CHECK_LRUS_DONE, &c->flags)) {
1777                         bch_err(c, "%s", buf.buf);
1778                 } else {
1779                         bch2_trans_inconsistent(trans, "%s", buf.buf);
1780                         ret = -EINVAL;
1781                 }
1782
1783                 goto out;
1784         }
1785
1786         if (!a->v.cached_sectors)
1787                 bch_err(c, "invalidating empty bucket, confused");
1788
1789         cached_sectors = a->v.cached_sectors;
1790
1791         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1792         a->v.gen++;
1793         a->v.data_type          = 0;
1794         a->v.dirty_sectors      = 0;
1795         a->v.cached_sectors     = 0;
1796         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1797         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1798
1799         ret =   bch2_trans_update(trans, &alloc_iter, &a->k_i,
1800                                 BTREE_TRIGGER_BUCKET_INVALIDATE) ?:
1801                 bch2_trans_commit(trans, NULL, NULL,
1802                                   BTREE_INSERT_USE_RESERVE|BTREE_INSERT_NOFAIL);
1803         if (ret)
1804                 goto out;
1805
1806         trace_and_count(c, bucket_invalidate, c, bucket.inode, bucket.offset, cached_sectors);
1807         --*nr_to_invalidate;
1808 out:
1809         bch2_trans_iter_exit(trans, &alloc_iter);
1810         printbuf_exit(&buf);
1811         return ret;
1812 }
1813
1814 static void bch2_do_invalidates_work(struct work_struct *work)
1815 {
1816         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1817         struct bch_dev *ca;
1818         struct btree_trans trans;
1819         struct btree_iter iter;
1820         struct bkey_s_c k;
1821         unsigned i;
1822         int ret = 0;
1823
1824         bch2_trans_init(&trans, c, 0, 0);
1825
1826         for_each_member_device(ca, c, i) {
1827                 s64 nr_to_invalidate =
1828                         should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
1829
1830                 ret = for_each_btree_key2(&trans, iter, BTREE_ID_lru,
1831                                 POS(ca->dev_idx, 0), BTREE_ITER_INTENT, k,
1832                         invalidate_one_bucket(&trans, &iter, k, ca->dev_idx, &nr_to_invalidate));
1833
1834                 if (ret < 0) {
1835                         percpu_ref_put(&ca->ref);
1836                         break;
1837                 }
1838         }
1839
1840         bch2_trans_exit(&trans);
1841         percpu_ref_put(&c->writes);
1842 }
1843
1844 void bch2_do_invalidates(struct bch_fs *c)
1845 {
1846         if (percpu_ref_tryget_live(&c->writes) &&
1847             !queue_work(system_long_wq, &c->invalidate_work))
1848                 percpu_ref_put(&c->writes);
1849 }
1850
1851 static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca)
1852 {
1853         struct btree_trans trans;
1854         struct btree_iter iter;
1855         struct bkey_s_c k;
1856         struct bkey hole;
1857         struct bpos end = POS(ca->dev_idx, ca->mi.nbuckets);
1858         struct bch_member *m;
1859         int ret;
1860
1861         bch2_trans_init(&trans, c, 0, 0);
1862
1863         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc,
1864                              POS(ca->dev_idx, ca->mi.first_bucket),
1865                              BTREE_ITER_PREFETCH);
1866         /*
1867          * Scan the alloc btree for every bucket on @ca, and add buckets to the
1868          * freespace/need_discard/need_gc_gens btrees as needed:
1869          */
1870         while (1) {
1871                 bch2_trans_begin(&trans);
1872
1873                 if (bkey_ge(iter.pos, end)) {
1874                         ret = 0;
1875                         break;
1876                 }
1877
1878                 k = bch2_get_key_or_hole(&iter, end, &hole);
1879                 ret = bkey_err(k);
1880                 if (ret)
1881                         goto bkey_err;
1882
1883                 if (k.k->type) {
1884                         /*
1885                          * We process live keys in the alloc btree one at a
1886                          * time:
1887                          */
1888                         struct bch_alloc_v4 a_convert;
1889                         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1890
1891                         ret =   bch2_bucket_do_index(&trans, k, a, true) ?:
1892                                 bch2_trans_commit(&trans, NULL, NULL,
1893                                                   BTREE_INSERT_LAZY_RW|
1894                                                   BTREE_INSERT_NOFAIL);
1895                         if (ret)
1896                                 goto bkey_err;
1897
1898                         bch2_btree_iter_advance(&iter);
1899                 } else {
1900                         struct bkey_i *freespace;
1901
1902                         freespace = bch2_trans_kmalloc(&trans, sizeof(*freespace));
1903                         ret = PTR_ERR_OR_ZERO(freespace);
1904                         if (ret)
1905                                 goto bkey_err;
1906
1907                         bkey_init(&freespace->k);
1908                         freespace->k.type       = KEY_TYPE_set;
1909                         freespace->k.p          = k.k->p;
1910                         freespace->k.size       = k.k->size;
1911
1912                         ret = __bch2_btree_insert(&trans, BTREE_ID_freespace, freespace) ?:
1913                                 bch2_trans_commit(&trans, NULL, NULL,
1914                                                   BTREE_INSERT_LAZY_RW|
1915                                                   BTREE_INSERT_NOFAIL);
1916                         if (ret)
1917                                 goto bkey_err;
1918
1919                         bch2_btree_iter_set_pos(&iter, k.k->p);
1920                 }
1921 bkey_err:
1922                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1923                         continue;
1924                 if (ret)
1925                         break;
1926         }
1927
1928         bch2_trans_iter_exit(&trans, &iter);
1929         bch2_trans_exit(&trans);
1930
1931         if (ret < 0) {
1932                 bch_err(ca, "error initializing free space: %s", bch2_err_str(ret));
1933                 return ret;
1934         }
1935
1936         mutex_lock(&c->sb_lock);
1937         m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx;
1938         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1939         mutex_unlock(&c->sb_lock);
1940
1941         return 0;
1942 }
1943
1944 int bch2_fs_freespace_init(struct bch_fs *c)
1945 {
1946         struct bch_dev *ca;
1947         unsigned i;
1948         int ret = 0;
1949         bool doing_init = false;
1950
1951         /*
1952          * We can crash during the device add path, so we need to check this on
1953          * every mount:
1954          */
1955
1956         for_each_member_device(ca, c, i) {
1957                 if (ca->mi.freespace_initialized)
1958                         continue;
1959
1960                 if (!doing_init) {
1961                         bch_info(c, "initializing freespace");
1962                         doing_init = true;
1963                 }
1964
1965                 ret = bch2_dev_freespace_init(c, ca);
1966                 if (ret) {
1967                         percpu_ref_put(&ca->ref);
1968                         return ret;
1969                 }
1970         }
1971
1972         if (doing_init) {
1973                 mutex_lock(&c->sb_lock);
1974                 bch2_write_super(c);
1975                 mutex_unlock(&c->sb_lock);
1976
1977                 bch_verbose(c, "done initializing freespace");
1978         }
1979
1980         return ret;
1981 }
1982
1983 /* Bucket IO clocks: */
1984
1985 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
1986                               size_t bucket_nr, int rw)
1987 {
1988         struct bch_fs *c = trans->c;
1989         struct btree_iter iter;
1990         struct bkey_i_alloc_v4 *a;
1991         u64 now;
1992         int ret = 0;
1993
1994         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
1995         ret = PTR_ERR_OR_ZERO(a);
1996         if (ret)
1997                 return ret;
1998
1999         now = atomic64_read(&c->io_clock[rw].now);
2000         if (a->v.io_time[rw] == now)
2001                 goto out;
2002
2003         a->v.io_time[rw] = now;
2004
2005         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
2006                 bch2_trans_commit(trans, NULL, NULL, 0);
2007 out:
2008         bch2_trans_iter_exit(trans, &iter);
2009         return ret;
2010 }
2011
2012 /* Startup/shutdown (ro/rw): */
2013
2014 void bch2_recalc_capacity(struct bch_fs *c)
2015 {
2016         struct bch_dev *ca;
2017         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
2018         unsigned bucket_size_max = 0;
2019         unsigned long ra_pages = 0;
2020         unsigned i;
2021
2022         lockdep_assert_held(&c->state_lock);
2023
2024         for_each_online_member(ca, c, i) {
2025                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
2026
2027                 ra_pages += bdi->ra_pages;
2028         }
2029
2030         bch2_set_ra_pages(c, ra_pages);
2031
2032         for_each_rw_member(ca, c, i) {
2033                 u64 dev_reserve = 0;
2034
2035                 /*
2036                  * We need to reserve buckets (from the number
2037                  * of currently available buckets) against
2038                  * foreground writes so that mainly copygc can
2039                  * make forward progress.
2040                  *
2041                  * We need enough to refill the various reserves
2042                  * from scratch - copygc will use its entire
2043                  * reserve all at once, then run against when
2044                  * its reserve is refilled (from the formerly
2045                  * available buckets).
2046                  *
2047                  * This reserve is just used when considering if
2048                  * allocations for foreground writes must wait -
2049                  * not -ENOSPC calculations.
2050                  */
2051
2052                 dev_reserve += ca->nr_btree_reserve * 2;
2053                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
2054
2055                 dev_reserve += 1;       /* btree write point */
2056                 dev_reserve += 1;       /* copygc write point */
2057                 dev_reserve += 1;       /* rebalance write point */
2058
2059                 dev_reserve *= ca->mi.bucket_size;
2060
2061                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
2062                                              ca->mi.first_bucket);
2063
2064                 reserved_sectors += dev_reserve * 2;
2065
2066                 bucket_size_max = max_t(unsigned, bucket_size_max,
2067                                         ca->mi.bucket_size);
2068         }
2069
2070         gc_reserve = c->opts.gc_reserve_bytes
2071                 ? c->opts.gc_reserve_bytes >> 9
2072                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
2073
2074         reserved_sectors = max(gc_reserve, reserved_sectors);
2075
2076         reserved_sectors = min(reserved_sectors, capacity);
2077
2078         c->capacity = capacity - reserved_sectors;
2079
2080         c->bucket_size_max = bucket_size_max;
2081
2082         /* Wake up case someone was waiting for buckets */
2083         closure_wake_up(&c->freelist_wait);
2084 }
2085
2086 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
2087 {
2088         struct open_bucket *ob;
2089         bool ret = false;
2090
2091         for (ob = c->open_buckets;
2092              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
2093              ob++) {
2094                 spin_lock(&ob->lock);
2095                 if (ob->valid && !ob->on_partial_list &&
2096                     ob->dev == ca->dev_idx)
2097                         ret = true;
2098                 spin_unlock(&ob->lock);
2099         }
2100
2101         return ret;
2102 }
2103
2104 /* device goes ro: */
2105 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
2106 {
2107         unsigned i;
2108
2109         /* First, remove device from allocation groups: */
2110
2111         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2112                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
2113
2114         /*
2115          * Capacity is calculated based off of devices in allocation groups:
2116          */
2117         bch2_recalc_capacity(c);
2118
2119         /* Next, close write points that point to this device... */
2120         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
2121                 bch2_writepoint_stop(c, ca, &c->write_points[i]);
2122
2123         bch2_writepoint_stop(c, ca, &c->copygc_write_point);
2124         bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
2125         bch2_writepoint_stop(c, ca, &c->btree_write_point);
2126
2127         mutex_lock(&c->btree_reserve_cache_lock);
2128         while (c->btree_reserve_cache_nr) {
2129                 struct btree_alloc *a =
2130                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
2131
2132                 bch2_open_buckets_put(c, &a->ob);
2133         }
2134         mutex_unlock(&c->btree_reserve_cache_lock);
2135
2136         while (1) {
2137                 struct open_bucket *ob;
2138
2139                 spin_lock(&c->freelist_lock);
2140                 if (!ca->open_buckets_partial_nr) {
2141                         spin_unlock(&c->freelist_lock);
2142                         break;
2143                 }
2144                 ob = c->open_buckets +
2145                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
2146                 ob->on_partial_list = false;
2147                 spin_unlock(&c->freelist_lock);
2148
2149                 bch2_open_bucket_put(c, ob);
2150         }
2151
2152         bch2_ec_stop_dev(c, ca);
2153
2154         /*
2155          * Wake up threads that were blocked on allocation, so they can notice
2156          * the device can no longer be removed and the capacity has changed:
2157          */
2158         closure_wake_up(&c->freelist_wait);
2159
2160         /*
2161          * journal_res_get() can block waiting for free space in the journal -
2162          * it needs to notice there may not be devices to allocate from anymore:
2163          */
2164         wake_up(&c->journal.wait);
2165
2166         /* Now wait for any in flight writes: */
2167
2168         closure_wait_event(&c->open_buckets_wait,
2169                            !bch2_dev_has_open_write_point(c, ca));
2170 }
2171
2172 /* device goes rw: */
2173 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
2174 {
2175         unsigned i;
2176
2177         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2178                 if (ca->mi.data_allowed & (1 << i))
2179                         set_bit(ca->dev_idx, c->rw_devs[i].d);
2180 }
2181
2182 void bch2_fs_allocator_background_init(struct bch_fs *c)
2183 {
2184         spin_lock_init(&c->freelist_lock);
2185         INIT_WORK(&c->discard_work, bch2_do_discards_work);
2186         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
2187 }