git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/alloc_background.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "bcachefs.h"
   3 #include "alloc_background.h"
   4 #include "alloc_foreground.h"
   5 #include "backpointers.h"
   6 #include "btree_cache.h"
   7 #include "btree_io.h"
   8 #include "btree_key_cache.h"
   9 #include "btree_update.h"
  10 #include "btree_update_interior.h"
  11 #include "btree_gc.h"
  12 #include "btree_write_buffer.h"
  13 #include "buckets.h"
  14 #include "buckets_waiting_for_journal.h"
  15 #include "clock.h"
  16 #include "debug.h"
  17 #include "ec.h"
  18 #include "error.h"
  19 #include "lru.h"
  20 #include "recovery.h"
  21 #include "trace.h"
  22 #include "varint.h"
  23
  24 #include <linux/kthread.h>
  25 #include <linux/math64.h>
  26 #include <linux/random.h>
  27 #include <linux/rculist.h>
  28 #include <linux/rcupdate.h>
  29 #include <linux/sched/task.h>
  30 #include <linux/sort.h>
  31
  32 /* Persistent alloc info: */
  33
  34 static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
  35 #define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8,
  36         BCH_ALLOC_FIELDS_V1()
  37 #undef x
  38 };
  39
  40 struct bkey_alloc_unpacked {
  41         u64             journal_seq;
  42         u8              gen;
  43         u8              oldest_gen;
  44         u8              data_type;
  45         bool            need_discard:1;
  46         bool            need_inc_gen:1;
  47 #define x(_name, _bits) u##_bits _name;
  48         BCH_ALLOC_FIELDS_V2()
  49 #undef  x
  50 };
  51
  52 static inline u64 alloc_field_v1_get(const struct bch_alloc *a,
  53                                      const void **p, unsigned field)
  54 {
  55         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  56         u64 v;
  57
  58         if (!(a->fields & (1 << field)))
  59                 return 0;
  60
  61         switch (bytes) {
  62         case 1:
  63                 v = *((const u8 *) *p);
  64                 break;
  65         case 2:
  66                 v = le16_to_cpup(*p);
  67                 break;
  68         case 4:
  69                 v = le32_to_cpup(*p);
  70                 break;
  71         case 8:
  72                 v = le64_to_cpup(*p);
  73                 break;
  74         default:
  75                 BUG();
  76         }
  77
  78         *p += bytes;
  79         return v;
  80 }
  81
  82 static inline void alloc_field_v1_put(struct bkey_i_alloc *a, void **p,
  83                                       unsigned field, u64 v)
  84 {
  85         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  86
  87         if (!v)
  88                 return;
  89
  90         a->v.fields |= 1 << field;
  91
  92         switch (bytes) {
  93         case 1:
  94                 *((u8 *) *p) = v;
  95                 break;
  96         case 2:
  97                 *((__le16 *) *p) = cpu_to_le16(v);
  98                 break;
  99         case 4:
 100                 *((__le32 *) *p) = cpu_to_le32(v);
 101                 break;
 102         case 8:
 103                 *((__le64 *) *p) = cpu_to_le64(v);
 104                 break;
 105         default:
 106                 BUG();
 107         }
 108
 109         *p += bytes;
 110 }
 111
 112 static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out,
 113                                  struct bkey_s_c k)
 114 {
 115         const struct bch_alloc *in = bkey_s_c_to_alloc(k).v;
 116         const void *d = in->data;
 117         unsigned idx = 0;
 118
 119         out->gen = in->gen;
 120
 121 #define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++);
 122         BCH_ALLOC_FIELDS_V1()
 123 #undef  x
 124 }
 125
 126 static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out,
 127                                 struct bkey_s_c k)
 128 {
 129         struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k);
 130         const u8 *in = a.v->data;
 131         const u8 *end = bkey_val_end(a);
 132         unsigned fieldnr = 0;
 133         int ret;
 134         u64 v;
 135
 136         out->gen        = a.v->gen;
 137         out->oldest_gen = a.v->oldest_gen;
 138         out->data_type  = a.v->data_type;
 139
 140 #define x(_name, _bits)                                                 \
 141         if (fieldnr < a.v->nr_fields) {                                 \
 142                 ret = bch2_varint_decode_fast(in, end, &v);             \
 143                 if (ret < 0)                                            \
 144                         return ret;                                     \
 145                 in += ret;                                              \
 146         } else {                                                        \
 147                 v = 0;                                                  \
 148         }                                                               \
 149         out->_name = v;                                                 \
 150         if (v != out->_name)                                            \
 151                 return -1;                                              \
 152         fieldnr++;
 153
 154         BCH_ALLOC_FIELDS_V2()
 155 #undef  x
 156         return 0;
 157 }
 158
 159 static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out,
 160                                 struct bkey_s_c k)
 161 {
 162         struct bkey_s_c_alloc_v3 a = bkey_s_c_to_alloc_v3(k);
 163         const u8 *in = a.v->data;
 164         const u8 *end = bkey_val_end(a);
 165         unsigned fieldnr = 0;
 166         int ret;
 167         u64 v;
 168
 169         out->gen        = a.v->gen;
 170         out->oldest_gen = a.v->oldest_gen;
 171         out->data_type  = a.v->data_type;
 172         out->need_discard = BCH_ALLOC_V3_NEED_DISCARD(a.v);
 173         out->need_inc_gen = BCH_ALLOC_V3_NEED_INC_GEN(a.v);
 174         out->journal_seq = le64_to_cpu(a.v->journal_seq);
 175
 176 #define x(_name, _bits)                                                 \
 177         if (fieldnr < a.v->nr_fields) {                                 \
 178                 ret = bch2_varint_decode_fast(in, end, &v);             \
 179                 if (ret < 0)                                            \
 180                         return ret;                                     \
 181                 in += ret;                                              \
 182         } else {                                                        \
 183                 v = 0;                                                  \
 184         }                                                               \
 185         out->_name = v;                                                 \
 186         if (v != out->_name)                                            \
 187                 return -1;                                              \
 188         fieldnr++;
 189
 190         BCH_ALLOC_FIELDS_V2()
 191 #undef  x
 192         return 0;
 193 }
 194
 195 static struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 196 {
 197         struct bkey_alloc_unpacked ret = { .gen = 0 };
 198
 199         switch (k.k->type) {
 200         case KEY_TYPE_alloc:
 201                 bch2_alloc_unpack_v1(&ret, k);
 202                 break;
 203         case KEY_TYPE_alloc_v2:
 204                 bch2_alloc_unpack_v2(&ret, k);
 205                 break;
 206         case KEY_TYPE_alloc_v3:
 207                 bch2_alloc_unpack_v3(&ret, k);
 208                 break;
 209         }
 210
 211         return ret;
 212 }
 213
 214 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
 215 {
 216         unsigned i, bytes = offsetof(struct bch_alloc, data);
 217
 218         for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++)
 219                 if (a->fields & (1 << i))
 220                         bytes += BCH_ALLOC_V1_FIELD_BYTES[i];
 221
 222         return DIV_ROUND_UP(bytes, sizeof(u64));
 223 }
 224
 225 int bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k,
 226                           enum bkey_invalid_flags flags,
 227                           struct printbuf *err)
 228 {
 229         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 230
 231         /* allow for unknown fields */
 232         if (bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v)) {
 233                 prt_printf(err, "incorrect value size (%zu < %u)",
 234                        bkey_val_u64s(a.k), bch_alloc_v1_val_u64s(a.v));
 235                 return -BCH_ERR_invalid_bkey;
 236         }
 237
 238         return 0;
 239 }
 240
 241 int bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k,
 242                           enum bkey_invalid_flags flags,
 243                           struct printbuf *err)
 244 {
 245         struct bkey_alloc_unpacked u;
 246
 247         if (bch2_alloc_unpack_v2(&u, k)) {
 248                 prt_printf(err, "unpack error");
 249                 return -BCH_ERR_invalid_bkey;
 250         }
 251
 252         return 0;
 253 }
 254
 255 int bch2_alloc_v3_invalid(const struct bch_fs *c, struct bkey_s_c k,
 256                           enum bkey_invalid_flags flags,
 257                           struct printbuf *err)
 258 {
 259         struct bkey_alloc_unpacked u;
 260
 261         if (bch2_alloc_unpack_v3(&u, k)) {
 262                 prt_printf(err, "unpack error");
 263                 return -BCH_ERR_invalid_bkey;
 264         }
 265
 266         return 0;
 267 }
 268
 269 int bch2_alloc_v4_invalid(const struct bch_fs *c, struct bkey_s_c k,
 270                           unsigned flags, struct printbuf *err)
 271 {
 272         struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 273         int rw = flags & WRITE;
 274
 275         if (alloc_v4_u64s(a.v) > bkey_val_u64s(k.k)) {
 276                 prt_printf(err, "bad val size (%u > %lu)",
 277                        alloc_v4_u64s(a.v), bkey_val_u64s(k.k));
 278                 return -BCH_ERR_invalid_bkey;
 279         }
 280
 281         if (!BCH_ALLOC_V4_BACKPOINTERS_START(a.v) &&
 282             BCH_ALLOC_V4_NR_BACKPOINTERS(a.v)) {
 283                 prt_printf(err, "invalid backpointers_start");
 284                 return -BCH_ERR_invalid_bkey;
 285         }
 286
 287         if (rw == WRITE &&
 288             !(flags & BKEY_INVALID_JOURNAL) &&
 289             c->curr_recovery_pass > BCH_RECOVERY_PASS_check_btree_backpointers) {
 290                 unsigned i, bp_len = 0;
 291
 292                 for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a.v); i++)
 293                         bp_len += alloc_v4_backpointers_c(a.v)[i].bucket_len;
 294
 295                 if (bp_len > a.v->dirty_sectors) {
 296                         prt_printf(err, "too many backpointers");
 297                         return -BCH_ERR_invalid_bkey;
 298                 }
 299         }
 300
 301         if (rw == WRITE) {
 302                 if (alloc_data_type(*a.v, a.v->data_type) != a.v->data_type) {
 303                         prt_printf(err, "invalid data type (got %u should be %u)",
 304                                a.v->data_type, alloc_data_type(*a.v, a.v->data_type));
 305                         return -BCH_ERR_invalid_bkey;
 306                 }
 307
 308                 switch (a.v->data_type) {
 309                 case BCH_DATA_free:
 310                 case BCH_DATA_need_gc_gens:
 311                 case BCH_DATA_need_discard:
 312                         if (a.v->dirty_sectors ||
 313                             a.v->cached_sectors ||
 314                             a.v->stripe) {
 315                                 prt_printf(err, "empty data type free but have data");
 316                                 return -BCH_ERR_invalid_bkey;
 317                         }
 318                         break;
 319                 case BCH_DATA_sb:
 320                 case BCH_DATA_journal:
 321                 case BCH_DATA_btree:
 322                 case BCH_DATA_user:
 323                 case BCH_DATA_parity:
 324                         if (!a.v->dirty_sectors) {
 325                                 prt_printf(err, "data_type %s but dirty_sectors==0",
 326                                        bch2_data_types[a.v->data_type]);
 327                                 return -BCH_ERR_invalid_bkey;
 328                         }
 329                         break;
 330                 case BCH_DATA_cached:
 331                         if (!a.v->cached_sectors ||
 332                             a.v->dirty_sectors ||
 333                             a.v->stripe) {
 334                                 prt_printf(err, "data type inconsistency");
 335                                 return -BCH_ERR_invalid_bkey;
 336                         }
 337
 338                         if (!a.v->io_time[READ] &&
 339                             c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_to_lru_refs) {
 340                                 prt_printf(err, "cached bucket with read_time == 0");
 341                                 return -BCH_ERR_invalid_bkey;
 342                         }
 343                         break;
 344                 case BCH_DATA_stripe:
 345                         if (!a.v->stripe) {
 346                                 prt_printf(err, "data_type %s but stripe==0",
 347                                        bch2_data_types[a.v->data_type]);
 348                                 return -BCH_ERR_invalid_bkey;
 349                         }
 350                         break;
 351                 }
 352         }
 353
 354         return 0;
 355 }
 356
 357 static inline u64 swab40(u64 x)
 358 {
 359         return (((x & 0x00000000ffULL) << 32)|
 360                 ((x & 0x000000ff00ULL) << 16)|
 361                 ((x & 0x0000ff0000ULL) >>  0)|
 362                 ((x & 0x00ff000000ULL) >> 16)|
 363                 ((x & 0xff00000000ULL) >> 32));
 364 }
 365
 366 void bch2_alloc_v4_swab(struct bkey_s k)
 367 {
 368         struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v;
 369         struct bch_backpointer *bp, *bps;
 370
 371         a->journal_seq          = swab64(a->journal_seq);
 372         a->flags                = swab32(a->flags);
 373         a->dirty_sectors        = swab32(a->dirty_sectors);
 374         a->cached_sectors       = swab32(a->cached_sectors);
 375         a->io_time[0]           = swab64(a->io_time[0]);
 376         a->io_time[1]           = swab64(a->io_time[1]);
 377         a->stripe               = swab32(a->stripe);
 378         a->nr_external_backpointers = swab32(a->nr_external_backpointers);
 379
 380         bps = alloc_v4_backpointers(a);
 381         for (bp = bps; bp < bps + BCH_ALLOC_V4_NR_BACKPOINTERS(a); bp++) {
 382                 bp->bucket_offset       = swab40(bp->bucket_offset);
 383                 bp->bucket_len          = swab32(bp->bucket_len);
 384                 bch2_bpos_swab(&bp->pos);
 385         }
 386 }
 387
 388 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 389 {
 390         struct bch_alloc_v4 _a;
 391         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &_a);
 392         unsigned i;
 393
 394         prt_newline(out);
 395         printbuf_indent_add(out, 2);
 396
 397         prt_printf(out, "gen %u oldest_gen %u data_type %s",
 398                a->gen, a->oldest_gen,
 399                a->data_type < BCH_DATA_NR
 400                ? bch2_data_types[a->data_type]
 401                : "(invalid data type)");
 402         prt_newline(out);
 403         prt_printf(out, "journal_seq       %llu",       a->journal_seq);
 404         prt_newline(out);
 405         prt_printf(out, "need_discard      %llu",       BCH_ALLOC_V4_NEED_DISCARD(a));
 406         prt_newline(out);
 407         prt_printf(out, "need_inc_gen      %llu",       BCH_ALLOC_V4_NEED_INC_GEN(a));
 408         prt_newline(out);
 409         prt_printf(out, "dirty_sectors     %u", a->dirty_sectors);
 410         prt_newline(out);
 411         prt_printf(out, "cached_sectors    %u", a->cached_sectors);
 412         prt_newline(out);
 413         prt_printf(out, "stripe            %u", a->stripe);
 414         prt_newline(out);
 415         prt_printf(out, "stripe_redundancy %u", a->stripe_redundancy);
 416         prt_newline(out);
 417         prt_printf(out, "io_time[READ]     %llu",       a->io_time[READ]);
 418         prt_newline(out);
 419         prt_printf(out, "io_time[WRITE]    %llu",       a->io_time[WRITE]);
 420         prt_newline(out);
 421         prt_printf(out, "fragmentation     %llu",       a->fragmentation_lru);
 422         prt_newline(out);
 423         prt_printf(out, "bp_start          %llu", BCH_ALLOC_V4_BACKPOINTERS_START(a));
 424         prt_newline(out);
 425
 426         if (BCH_ALLOC_V4_NR_BACKPOINTERS(a)) {
 427                 struct bkey_s_c_alloc_v4 a_raw = bkey_s_c_to_alloc_v4(k);
 428                 const struct bch_backpointer *bps = alloc_v4_backpointers_c(a_raw.v);
 429
 430                 prt_printf(out, "backpointers:     %llu", BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v));
 431                 printbuf_indent_add(out, 2);
 432
 433                 for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v); i++) {
 434                         prt_newline(out);
 435                         bch2_backpointer_to_text(out, &bps[i]);
 436                 }
 437
 438                 printbuf_indent_sub(out, 2);
 439         }
 440
 441         printbuf_indent_sub(out, 2);
 442 }
 443
 444 void __bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out)
 445 {
 446         if (k.k->type == KEY_TYPE_alloc_v4) {
 447                 void *src, *dst;
 448
 449                 *out = *bkey_s_c_to_alloc_v4(k).v;
 450
 451                 src = alloc_v4_backpointers(out);
 452                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 453                 dst = alloc_v4_backpointers(out);
 454
 455                 if (src < dst)
 456                         memset(src, 0, dst - src);
 457
 458                 SET_BCH_ALLOC_V4_NR_BACKPOINTERS(out, 0);
 459         } else {
 460                 struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 461
 462                 *out = (struct bch_alloc_v4) {
 463                         .journal_seq            = u.journal_seq,
 464                         .flags                  = u.need_discard,
 465                         .gen                    = u.gen,
 466                         .oldest_gen             = u.oldest_gen,
 467                         .data_type              = u.data_type,
 468                         .stripe_redundancy      = u.stripe_redundancy,
 469                         .dirty_sectors          = u.dirty_sectors,
 470                         .cached_sectors         = u.cached_sectors,
 471                         .io_time[READ]          = u.read_time,
 472                         .io_time[WRITE]         = u.write_time,
 473                         .stripe                 = u.stripe,
 474                 };
 475
 476                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 477         }
 478 }
 479
 480 static noinline struct bkey_i_alloc_v4 *
 481 __bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 482 {
 483         struct bkey_i_alloc_v4 *ret;
 484
 485         ret = bch2_trans_kmalloc(trans, max(bkey_bytes(k.k), sizeof(struct bkey_i_alloc_v4)));
 486         if (IS_ERR(ret))
 487                 return ret;
 488
 489         if (k.k->type == KEY_TYPE_alloc_v4) {
 490                 void *src, *dst;
 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                 if (src < dst)
 499                         memset(src, 0, dst - src);
 500
 501                 SET_BCH_ALLOC_V4_NR_BACKPOINTERS(&ret->v, 0);
 502                 set_alloc_v4_u64s(ret);
 503         } else {
 504                 bkey_alloc_v4_init(&ret->k_i);
 505                 ret->k.p = k.k->p;
 506                 bch2_alloc_to_v4(k, &ret->v);
 507         }
 508         return ret;
 509 }
 510
 511 static inline struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut_inlined(struct btree_trans *trans, struct bkey_s_c k)
 512 {
 513         struct bkey_s_c_alloc_v4 a;
 514
 515         if (likely(k.k->type == KEY_TYPE_alloc_v4) &&
 516             ((a = bkey_s_c_to_alloc_v4(k), true) &&
 517              BCH_ALLOC_V4_NR_BACKPOINTERS(a.v) == 0))
 518                 return bch2_bkey_make_mut_noupdate_typed(trans, k, alloc_v4);
 519
 520         return __bch2_alloc_to_v4_mut(trans, k);
 521 }
 522
 523 struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 524 {
 525         return bch2_alloc_to_v4_mut_inlined(trans, k);
 526 }
 527
 528 struct bkey_i_alloc_v4 *
 529 bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
 530                               struct bpos pos)
 531 {
 532         struct bkey_s_c k;
 533         struct bkey_i_alloc_v4 *a;
 534         int ret;
 535
 536         k = bch2_bkey_get_iter(trans, iter, BTREE_ID_alloc, pos,
 537                              BTREE_ITER_WITH_UPDATES|
 538                              BTREE_ITER_CACHED|
 539                              BTREE_ITER_INTENT);
 540         ret = bkey_err(k);
 541         if (unlikely(ret))
 542                 return ERR_PTR(ret);
 543
 544         a = bch2_alloc_to_v4_mut_inlined(trans, k);
 545         ret = PTR_ERR_OR_ZERO(a);
 546         if (unlikely(ret))
 547                 goto err;
 548         return a;
 549 err:
 550         bch2_trans_iter_exit(trans, iter);
 551         return ERR_PTR(ret);
 552 }
 553
 554 static struct bpos alloc_gens_pos(struct bpos pos, unsigned *offset)
 555 {
 556         *offset = pos.offset & KEY_TYPE_BUCKET_GENS_MASK;
 557
 558         pos.offset >>= KEY_TYPE_BUCKET_GENS_BITS;
 559         return pos;
 560 }
 561
 562 static struct bpos bucket_gens_pos_to_alloc(struct bpos pos, unsigned offset)
 563 {
 564         pos.offset <<= KEY_TYPE_BUCKET_GENS_BITS;
 565         pos.offset += offset;
 566         return pos;
 567 }
 568
 569 static unsigned alloc_gen(struct bkey_s_c k, unsigned offset)
 570 {
 571         return k.k->type == KEY_TYPE_bucket_gens
 572                 ? bkey_s_c_to_bucket_gens(k).v->gens[offset]
 573                 : 0;
 574 }
 575
 576 int bch2_bucket_gens_invalid(const struct bch_fs *c, struct bkey_s_c k,
 577                              enum bkey_invalid_flags flags,
 578                              struct printbuf *err)
 579 {
 580         if (bkey_val_bytes(k.k) != sizeof(struct bch_bucket_gens)) {
 581                 prt_printf(err, "bad val size (%lu != %zu)",
 582                        bkey_val_bytes(k.k), sizeof(struct bch_bucket_gens));
 583                 return -BCH_ERR_invalid_bkey;
 584         }
 585
 586         return 0;
 587 }
 588
 589 void bch2_bucket_gens_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 590 {
 591         struct bkey_s_c_bucket_gens g = bkey_s_c_to_bucket_gens(k);
 592         unsigned i;
 593
 594         for (i = 0; i < ARRAY_SIZE(g.v->gens); i++) {
 595                 if (i)
 596                         prt_char(out, ' ');
 597                 prt_printf(out, "%u", g.v->gens[i]);
 598         }
 599 }
 600
 601 int bch2_bucket_gens_init(struct bch_fs *c)
 602 {
 603         struct btree_trans trans;
 604         struct btree_iter iter;
 605         struct bkey_s_c k;
 606         struct bch_alloc_v4 a;
 607         struct bkey_i_bucket_gens g;
 608         bool have_bucket_gens_key = false;
 609         unsigned offset;
 610         struct bpos pos;
 611         u8 gen;
 612         int ret;
 613
 614         bch2_trans_init(&trans, c, 0, 0);
 615
 616         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 617                            BTREE_ITER_PREFETCH, k, ret) {
 618                 /*
 619                  * Not a fsck error because this is checked/repaired by
 620                  * bch2_check_alloc_key() which runs later:
 621                  */
 622                 if (!bch2_dev_bucket_exists(c, k.k->p))
 623                         continue;
 624
 625                 gen = bch2_alloc_to_v4(k, &a)->gen;
 626                 pos = alloc_gens_pos(iter.pos, &offset);
 627
 628                 if (have_bucket_gens_key && bkey_cmp(iter.pos, pos)) {
 629                         ret = commit_do(&trans, NULL, NULL,
 630                                         BTREE_INSERT_NOFAIL|
 631                                         BTREE_INSERT_LAZY_RW,
 632                                 __bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i, 0));
 633                         if (ret)
 634                                 break;
 635                         have_bucket_gens_key = false;
 636                 }
 637
 638                 if (!have_bucket_gens_key) {
 639                         bkey_bucket_gens_init(&g.k_i);
 640                         g.k.p = pos;
 641                         have_bucket_gens_key = true;
 642                 }
 643
 644                 g.v.gens[offset] = gen;
 645         }
 646         bch2_trans_iter_exit(&trans, &iter);
 647
 648         if (have_bucket_gens_key && !ret)
 649                 ret = commit_do(&trans, NULL, NULL,
 650                                 BTREE_INSERT_NOFAIL|
 651                                 BTREE_INSERT_LAZY_RW,
 652                         __bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i, 0));
 653
 654         bch2_trans_exit(&trans);
 655
 656         if (ret)
 657                 bch_err_fn(c, ret);
 658         return ret;
 659 }
 660
 661 int bch2_alloc_read(struct bch_fs *c)
 662 {
 663         struct btree_trans trans;
 664         struct btree_iter iter;
 665         struct bkey_s_c k;
 666         struct bch_dev *ca;
 667         int ret;
 668
 669         down_read(&c->gc_lock);
 670         bch2_trans_init(&trans, c, 0, 0);
 671
 672         if (c->sb.version_upgrade_complete >= bcachefs_metadata_version_bucket_gens) {
 673                 const struct bch_bucket_gens *g;
 674                 u64 b;
 675
 676                 for_each_btree_key(&trans, iter, BTREE_ID_bucket_gens, POS_MIN,
 677                                    BTREE_ITER_PREFETCH, k, ret) {
 678                         u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
 679                         u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
 680
 681                         if (k.k->type != KEY_TYPE_bucket_gens)
 682                                 continue;
 683
 684                         g = bkey_s_c_to_bucket_gens(k).v;
 685
 686                         /*
 687                          * Not a fsck error because this is checked/repaired by
 688                          * bch2_check_alloc_key() which runs later:
 689                          */
 690                         if (!bch2_dev_exists2(c, k.k->p.inode))
 691                                 continue;
 692
 693                         ca = bch_dev_bkey_exists(c, k.k->p.inode);
 694
 695                         for (b = max_t(u64, ca->mi.first_bucket, start);
 696                              b < min_t(u64, ca->mi.nbuckets, end);
 697                              b++)
 698                                 *bucket_gen(ca, b) = g->gens[b & KEY_TYPE_BUCKET_GENS_MASK];
 699                 }
 700                 bch2_trans_iter_exit(&trans, &iter);
 701         } else {
 702                 struct bch_alloc_v4 a;
 703
 704                 for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 705                                    BTREE_ITER_PREFETCH, k, ret) {
 706                         /*
 707                          * Not a fsck error because this is checked/repaired by
 708                          * bch2_check_alloc_key() which runs later:
 709                          */
 710                         if (!bch2_dev_bucket_exists(c, k.k->p))
 711                                 continue;
 712
 713                         ca = bch_dev_bkey_exists(c, k.k->p.inode);
 714
 715                         *bucket_gen(ca, k.k->p.offset) = bch2_alloc_to_v4(k, &a)->gen;
 716                 }
 717                 bch2_trans_iter_exit(&trans, &iter);
 718         }
 719
 720         bch2_trans_exit(&trans);
 721         up_read(&c->gc_lock);
 722
 723         if (ret)
 724                 bch_err_fn(c, ret);
 725
 726         return ret;
 727 }
 728
 729 /* Free space/discard btree: */
 730
 731 static int bch2_bucket_do_index(struct btree_trans *trans,
 732                                 struct bkey_s_c alloc_k,
 733                                 const struct bch_alloc_v4 *a,
 734                                 bool set)
 735 {
 736         struct bch_fs *c = trans->c;
 737         struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 738         struct btree_iter iter;
 739         struct bkey_s_c old;
 740         struct bkey_i *k;
 741         enum btree_id btree;
 742         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 743         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 744         struct printbuf buf = PRINTBUF;
 745         int ret;
 746
 747         if (a->data_type != BCH_DATA_free &&
 748             a->data_type != BCH_DATA_need_discard)
 749                 return 0;
 750
 751         k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
 752         if (IS_ERR(k))
 753                 return PTR_ERR(k);
 754
 755         bkey_init(&k->k);
 756         k->k.type = new_type;
 757
 758         switch (a->data_type) {
 759         case BCH_DATA_free:
 760                 btree = BTREE_ID_freespace;
 761                 k->k.p = alloc_freespace_pos(alloc_k.k->p, *a);
 762                 bch2_key_resize(&k->k, 1);
 763                 break;
 764         case BCH_DATA_need_discard:
 765                 btree = BTREE_ID_need_discard;
 766                 k->k.p = alloc_k.k->p;
 767                 break;
 768         default:
 769                 return 0;
 770         }
 771
 772         old = bch2_bkey_get_iter(trans, &iter, btree,
 773                              bkey_start_pos(&k->k),
 774                              BTREE_ITER_INTENT);
 775         ret = bkey_err(old);
 776         if (ret)
 777                 return ret;
 778
 779         if (ca->mi.freespace_initialized &&
 780             c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info &&
 781             bch2_trans_inconsistent_on(old.k->type != old_type, trans,
 782                         "incorrect key when %s %s:%llu:%llu:0 (got %s should be %s)\n"
 783                         "  for %s",
 784                         set ? "setting" : "clearing",
 785                         bch2_btree_ids[btree],
 786                         iter.pos.inode,
 787                         iter.pos.offset,
 788                         bch2_bkey_types[old.k->type],
 789                         bch2_bkey_types[old_type],
 790                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 791                 ret = -EIO;
 792                 goto err;
 793         }
 794
 795         ret = bch2_trans_update(trans, &iter, k, 0);
 796 err:
 797         bch2_trans_iter_exit(trans, &iter);
 798         printbuf_exit(&buf);
 799         return ret;
 800 }
 801
 802 static noinline int bch2_bucket_gen_update(struct btree_trans *trans,
 803                                            struct bpos bucket, u8 gen)
 804 {
 805         struct btree_iter iter;
 806         unsigned offset;
 807         struct bpos pos = alloc_gens_pos(bucket, &offset);
 808         struct bkey_i_bucket_gens *g;
 809         struct bkey_s_c k;
 810         int ret;
 811
 812         g = bch2_trans_kmalloc(trans, sizeof(*g));
 813         ret = PTR_ERR_OR_ZERO(g);
 814         if (ret)
 815                 return ret;
 816
 817         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_bucket_gens, pos,
 818                                BTREE_ITER_INTENT|
 819                                BTREE_ITER_WITH_UPDATES);
 820         ret = bkey_err(k);
 821         if (ret)
 822                 return ret;
 823
 824         if (k.k->type != KEY_TYPE_bucket_gens) {
 825                 bkey_bucket_gens_init(&g->k_i);
 826                 g->k.p = iter.pos;
 827         } else {
 828                 bkey_reassemble(&g->k_i, k);
 829         }
 830
 831         g->v.gens[offset] = gen;
 832
 833         ret = bch2_trans_update(trans, &iter, &g->k_i, 0);
 834         bch2_trans_iter_exit(trans, &iter);
 835         return ret;
 836 }
 837
 838 int bch2_trans_mark_alloc(struct btree_trans *trans,
 839                           enum btree_id btree_id, unsigned level,
 840                           struct bkey_s_c old, struct bkey_i *new,
 841                           unsigned flags)
 842 {
 843         struct bch_fs *c = trans->c;
 844         struct bch_alloc_v4 old_a_convert, *new_a;
 845         const struct bch_alloc_v4 *old_a;
 846         u64 old_lru, new_lru;
 847         int ret = 0;
 848
 849         /*
 850          * Deletion only happens in the device removal path, with
 851          * BTREE_TRIGGER_NORUN:
 852          */
 853         BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
 854
 855         old_a = bch2_alloc_to_v4(old, &old_a_convert);
 856         new_a = &bkey_i_to_alloc_v4(new)->v;
 857
 858         new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
 859
 860         if (new_a->dirty_sectors > old_a->dirty_sectors ||
 861             new_a->cached_sectors > old_a->cached_sectors) {
 862                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 863                 new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 864                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
 865                 SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
 866         }
 867
 868         if (data_type_is_empty(new_a->data_type) &&
 869             BCH_ALLOC_V4_NEED_INC_GEN(new_a) &&
 870             !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
 871                 new_a->gen++;
 872                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
 873         }
 874
 875         if (old_a->data_type != new_a->data_type ||
 876             (new_a->data_type == BCH_DATA_free &&
 877              alloc_freespace_genbits(*old_a) != alloc_freespace_genbits(*new_a))) {
 878                 ret =   bch2_bucket_do_index(trans, old, old_a, false) ?:
 879                         bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_a, true);
 880                 if (ret)
 881                         return ret;
 882         }
 883
 884         if (new_a->data_type == BCH_DATA_cached &&
 885             !new_a->io_time[READ])
 886                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 887
 888         old_lru = alloc_lru_idx_read(*old_a);
 889         new_lru = alloc_lru_idx_read(*new_a);
 890
 891         if (old_lru != new_lru) {
 892                 ret = bch2_lru_change(trans, new->k.p.inode,
 893                                       bucket_to_u64(new->k.p),
 894                                       old_lru, new_lru);
 895                 if (ret)
 896                         return ret;
 897         }
 898
 899         new_a->fragmentation_lru = alloc_lru_idx_fragmentation(*new_a,
 900                                         bch_dev_bkey_exists(c, new->k.p.inode));
 901
 902         if (old_a->fragmentation_lru != new_a->fragmentation_lru) {
 903                 ret = bch2_lru_change(trans,
 904                                 BCH_LRU_FRAGMENTATION_START,
 905                                 bucket_to_u64(new->k.p),
 906                                 old_a->fragmentation_lru, new_a->fragmentation_lru);
 907                 if (ret)
 908                         return ret;
 909         }
 910
 911         if (old_a->gen != new_a->gen) {
 912                 ret = bch2_bucket_gen_update(trans, new->k.p, new_a->gen);
 913                 if (ret)
 914                         return ret;
 915         }
 916
 917         return 0;
 918 }
 919
 920 /*
 921  * This synthesizes deleted extents for holes, similar to BTREE_ITER_SLOTS for
 922  * extents style btrees, but works on non-extents btrees:
 923  */
 924 static struct bkey_s_c bch2_get_key_or_hole(struct btree_iter *iter, struct bpos end, struct bkey *hole)
 925 {
 926         struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
 927
 928         if (bkey_err(k))
 929                 return k;
 930
 931         if (k.k->type) {
 932                 return k;
 933         } else {
 934                 struct btree_iter iter2;
 935                 struct bpos next;
 936
 937                 bch2_trans_copy_iter(&iter2, iter);
 938
 939                 if (!bpos_eq(iter->path->l[0].b->key.k.p, SPOS_MAX))
 940                         end = bkey_min(end, bpos_nosnap_successor(iter->path->l[0].b->key.k.p));
 941
 942                 end = bkey_min(end, POS(iter->pos.inode, iter->pos.offset + U32_MAX - 1));
 943
 944                 /*
 945                  * btree node min/max is a closed interval, upto takes a half
 946                  * open interval:
 947                  */
 948                 k = bch2_btree_iter_peek_upto(&iter2, end);
 949                 next = iter2.pos;
 950                 bch2_trans_iter_exit(iter->trans, &iter2);
 951
 952                 BUG_ON(next.offset >= iter->pos.offset + U32_MAX);
 953
 954                 if (bkey_err(k))
 955                         return k;
 956
 957                 bkey_init(hole);
 958                 hole->p = iter->pos;
 959
 960                 bch2_key_resize(hole, next.offset - iter->pos.offset);
 961                 return (struct bkey_s_c) { hole, NULL };
 962         }
 963 }
 964
 965 static bool next_bucket(struct bch_fs *c, struct bpos *bucket)
 966 {
 967         struct bch_dev *ca;
 968         unsigned iter;
 969
 970         if (bch2_dev_bucket_exists(c, *bucket))
 971                 return true;
 972
 973         if (bch2_dev_exists2(c, bucket->inode)) {
 974                 ca = bch_dev_bkey_exists(c, bucket->inode);
 975
 976                 if (bucket->offset < ca->mi.first_bucket) {
 977                         bucket->offset = ca->mi.first_bucket;
 978                         return true;
 979                 }
 980
 981                 bucket->inode++;
 982                 bucket->offset = 0;
 983         }
 984
 985         rcu_read_lock();
 986         iter = bucket->inode;
 987         ca = __bch2_next_dev(c, &iter, NULL);
 988         if (ca)
 989                 *bucket = POS(ca->dev_idx, ca->mi.first_bucket);
 990         rcu_read_unlock();
 991
 992         return ca != NULL;
 993 }
 994
 995 static struct bkey_s_c bch2_get_key_or_real_bucket_hole(struct btree_iter *iter, struct bkey *hole)
 996 {
 997         struct bch_fs *c = iter->trans->c;
 998         struct bkey_s_c k;
 999 again:
1000         k = bch2_get_key_or_hole(iter, POS_MAX, hole);
1001         if (bkey_err(k))
1002                 return k;
1003
1004         if (!k.k->type) {
1005                 struct bpos bucket = bkey_start_pos(k.k);
1006
1007                 if (!bch2_dev_bucket_exists(c, bucket)) {
1008                         if (!next_bucket(c, &bucket))
1009                                 return bkey_s_c_null;
1010
1011                         bch2_btree_iter_set_pos(iter, bucket);
1012                         goto again;
1013                 }
1014
1015                 if (!bch2_dev_bucket_exists(c, k.k->p)) {
1016                         struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
1017
1018                         bch2_key_resize(hole, ca->mi.nbuckets - bucket.offset);
1019                 }
1020         }
1021
1022         return k;
1023 }
1024
1025 static noinline_for_stack
1026 int bch2_check_alloc_key(struct btree_trans *trans,
1027                          struct bkey_s_c alloc_k,
1028                          struct btree_iter *alloc_iter,
1029                          struct btree_iter *discard_iter,
1030                          struct btree_iter *freespace_iter,
1031                          struct btree_iter *bucket_gens_iter)
1032 {
1033         struct bch_fs *c = trans->c;
1034         struct bch_dev *ca;
1035         struct bch_alloc_v4 a_convert;
1036         const struct bch_alloc_v4 *a;
1037         unsigned discard_key_type, freespace_key_type;
1038         unsigned gens_offset;
1039         struct bkey_s_c k;
1040         struct printbuf buf = PRINTBUF;
1041         int ret;
1042
1043         if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
1044                         "alloc key for invalid device:bucket %llu:%llu",
1045                         alloc_k.k->p.inode, alloc_k.k->p.offset))
1046                 return bch2_btree_delete_at(trans, alloc_iter, 0);
1047
1048         ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
1049         if (!ca->mi.freespace_initialized)
1050                 return 0;
1051
1052         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1053
1054         discard_key_type = a->data_type == BCH_DATA_need_discard ? KEY_TYPE_set : 0;
1055         bch2_btree_iter_set_pos(discard_iter, alloc_k.k->p);
1056         k = bch2_btree_iter_peek_slot(discard_iter);
1057         ret = bkey_err(k);
1058         if (ret)
1059                 goto err;
1060
1061         if (k.k->type != discard_key_type &&
1062             (c->opts.reconstruct_alloc ||
1063              fsck_err(c, "incorrect key in need_discard btree (got %s should be %s)\n"
1064                       "  %s",
1065                       bch2_bkey_types[k.k->type],
1066                       bch2_bkey_types[discard_key_type],
1067                       (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1068                 struct bkey_i *update =
1069                         bch2_trans_kmalloc(trans, sizeof(*update));
1070
1071                 ret = PTR_ERR_OR_ZERO(update);
1072                 if (ret)
1073                         goto err;
1074
1075                 bkey_init(&update->k);
1076                 update->k.type  = discard_key_type;
1077                 update->k.p     = discard_iter->pos;
1078
1079                 ret = bch2_trans_update(trans, discard_iter, update, 0);
1080                 if (ret)
1081                         goto err;
1082         }
1083
1084         freespace_key_type = a->data_type == BCH_DATA_free ? KEY_TYPE_set : 0;
1085         bch2_btree_iter_set_pos(freespace_iter, alloc_freespace_pos(alloc_k.k->p, *a));
1086         k = bch2_btree_iter_peek_slot(freespace_iter);
1087         ret = bkey_err(k);
1088         if (ret)
1089                 goto err;
1090
1091         if (k.k->type != freespace_key_type &&
1092             (c->opts.reconstruct_alloc ||
1093              fsck_err(c, "incorrect key in freespace btree (got %s should be %s)\n"
1094                       "  %s",
1095                       bch2_bkey_types[k.k->type],
1096                       bch2_bkey_types[freespace_key_type],
1097                       (printbuf_reset(&buf),
1098                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1099                 struct bkey_i *update =
1100                         bch2_trans_kmalloc(trans, sizeof(*update));
1101
1102                 ret = PTR_ERR_OR_ZERO(update);
1103                 if (ret)
1104                         goto err;
1105
1106                 bkey_init(&update->k);
1107                 update->k.type  = freespace_key_type;
1108                 update->k.p     = freespace_iter->pos;
1109                 bch2_key_resize(&update->k, 1);
1110
1111                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1112                 if (ret)
1113                         goto err;
1114         }
1115
1116         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(alloc_k.k->p, &gens_offset));
1117         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1118         ret = bkey_err(k);
1119         if (ret)
1120                 goto err;
1121
1122         if (a->gen != alloc_gen(k, gens_offset) &&
1123             (c->opts.reconstruct_alloc ||
1124              fsck_err(c, "incorrect gen in bucket_gens btree (got %u should be %u)\n"
1125                       "  %s",
1126                       alloc_gen(k, gens_offset), a->gen,
1127                       (printbuf_reset(&buf),
1128                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1129                 struct bkey_i_bucket_gens *g =
1130                         bch2_trans_kmalloc(trans, sizeof(*g));
1131
1132                 ret = PTR_ERR_OR_ZERO(g);
1133                 if (ret)
1134                         goto err;
1135
1136                 if (k.k->type == KEY_TYPE_bucket_gens) {
1137                         bkey_reassemble(&g->k_i, k);
1138                 } else {
1139                         bkey_bucket_gens_init(&g->k_i);
1140                         g->k.p = alloc_gens_pos(alloc_k.k->p, &gens_offset);
1141                 }
1142
1143                 g->v.gens[gens_offset] = a->gen;
1144
1145                 ret = bch2_trans_update(trans, bucket_gens_iter, &g->k_i, 0);
1146                 if (ret)
1147                         goto err;
1148         }
1149 err:
1150 fsck_err:
1151         printbuf_exit(&buf);
1152         return ret;
1153 }
1154
1155 static noinline_for_stack
1156 int bch2_check_alloc_hole_freespace(struct btree_trans *trans,
1157                                     struct bpos start,
1158                                     struct bpos *end,
1159                                     struct btree_iter *freespace_iter)
1160 {
1161         struct bch_fs *c = trans->c;
1162         struct bch_dev *ca;
1163         struct bkey_s_c k;
1164         struct printbuf buf = PRINTBUF;
1165         int ret;
1166
1167         ca = bch_dev_bkey_exists(c, start.inode);
1168         if (!ca->mi.freespace_initialized)
1169                 return 0;
1170
1171         bch2_btree_iter_set_pos(freespace_iter, start);
1172
1173         k = bch2_btree_iter_peek_slot(freespace_iter);
1174         ret = bkey_err(k);
1175         if (ret)
1176                 goto err;
1177
1178         *end = bkey_min(k.k->p, *end);
1179
1180         if (k.k->type != KEY_TYPE_set &&
1181             (c->opts.reconstruct_alloc ||
1182              fsck_err(c, "hole in alloc btree missing in freespace btree\n"
1183                       "  device %llu buckets %llu-%llu",
1184                       freespace_iter->pos.inode,
1185                       freespace_iter->pos.offset,
1186                       end->offset))) {
1187                 struct bkey_i *update =
1188                         bch2_trans_kmalloc(trans, sizeof(*update));
1189
1190                 ret = PTR_ERR_OR_ZERO(update);
1191                 if (ret)
1192                         goto err;
1193
1194                 bkey_init(&update->k);
1195                 update->k.type  = KEY_TYPE_set;
1196                 update->k.p     = freespace_iter->pos;
1197                 bch2_key_resize(&update->k,
1198                                 min_t(u64, U32_MAX, end->offset -
1199                                       freespace_iter->pos.offset));
1200
1201                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1202                 if (ret)
1203                         goto err;
1204         }
1205 err:
1206 fsck_err:
1207         printbuf_exit(&buf);
1208         return ret;
1209 }
1210
1211 static noinline_for_stack
1212 int bch2_check_alloc_hole_bucket_gens(struct btree_trans *trans,
1213                                       struct bpos start,
1214                                       struct bpos *end,
1215                                       struct btree_iter *bucket_gens_iter)
1216 {
1217         struct bch_fs *c = trans->c;
1218         struct bkey_s_c k;
1219         struct printbuf buf = PRINTBUF;
1220         unsigned i, gens_offset, gens_end_offset;
1221         int ret;
1222
1223         if (c->sb.version < bcachefs_metadata_version_bucket_gens)
1224                 return 0;
1225
1226         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(start, &gens_offset));
1227
1228         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1229         ret = bkey_err(k);
1230         if (ret)
1231                 goto err;
1232
1233         if (bkey_cmp(alloc_gens_pos(start, &gens_offset),
1234                      alloc_gens_pos(*end,  &gens_end_offset)))
1235                 gens_end_offset = KEY_TYPE_BUCKET_GENS_NR;
1236
1237         if (k.k->type == KEY_TYPE_bucket_gens) {
1238                 struct bkey_i_bucket_gens g;
1239                 bool need_update = false;
1240
1241                 bkey_reassemble(&g.k_i, k);
1242
1243                 for (i = gens_offset; i < gens_end_offset; i++) {
1244                         if (fsck_err_on(g.v.gens[i], c,
1245                                         "hole in alloc btree at %llu:%llu with nonzero gen in bucket_gens btree (%u)",
1246                                         bucket_gens_pos_to_alloc(k.k->p, i).inode,
1247                                         bucket_gens_pos_to_alloc(k.k->p, i).offset,
1248                                         g.v.gens[i])) {
1249                                 g.v.gens[i] = 0;
1250                                 need_update = true;
1251                         }
1252                 }
1253
1254                 if (need_update) {
1255                         struct bkey_i *k = bch2_trans_kmalloc(trans, sizeof(g));
1256
1257                         ret = PTR_ERR_OR_ZERO(k);
1258                         if (ret)
1259                                 goto err;
1260
1261                         memcpy(k, &g, sizeof(g));
1262
1263                         ret = bch2_trans_update(trans, bucket_gens_iter, k, 0);
1264                         if (ret)
1265                                 goto err;
1266                 }
1267         }
1268
1269         *end = bkey_min(*end, bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0));
1270 err:
1271 fsck_err:
1272         printbuf_exit(&buf);
1273         return ret;
1274 }
1275
1276 static noinline_for_stack int __bch2_check_discard_freespace_key(struct btree_trans *trans,
1277                                               struct btree_iter *iter)
1278 {
1279         struct bch_fs *c = trans->c;
1280         struct btree_iter alloc_iter;
1281         struct bkey_s_c alloc_k;
1282         struct bch_alloc_v4 a_convert;
1283         const struct bch_alloc_v4 *a;
1284         u64 genbits;
1285         struct bpos pos;
1286         enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
1287                 ? BCH_DATA_need_discard
1288                 : BCH_DATA_free;
1289         struct printbuf buf = PRINTBUF;
1290         int ret;
1291
1292         pos = iter->pos;
1293         pos.offset &= ~(~0ULL << 56);
1294         genbits = iter->pos.offset & (~0ULL << 56);
1295
1296         alloc_k = bch2_bkey_get_iter(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
1297         ret = bkey_err(alloc_k);
1298         if (ret)
1299                 return ret;
1300
1301         if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
1302                         "entry in %s btree for nonexistant dev:bucket %llu:%llu",
1303                         bch2_btree_ids[iter->btree_id], pos.inode, pos.offset))
1304                 goto delete;
1305
1306         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1307
1308         if (fsck_err_on(a->data_type != state ||
1309                         (state == BCH_DATA_free &&
1310                          genbits != alloc_freespace_genbits(*a)), c,
1311                         "%s\n  incorrectly set at %s:%llu:%llu:0 (free %u, genbits %llu should be %llu)",
1312                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
1313                         bch2_btree_ids[iter->btree_id],
1314                         iter->pos.inode,
1315                         iter->pos.offset,
1316                         a->data_type == state,
1317                         genbits >> 56, alloc_freespace_genbits(*a) >> 56))
1318                 goto delete;
1319 out:
1320 fsck_err:
1321         set_btree_iter_dontneed(&alloc_iter);
1322         bch2_trans_iter_exit(trans, &alloc_iter);
1323         printbuf_exit(&buf);
1324         return ret;
1325 delete:
1326         ret =   bch2_btree_delete_extent_at(trans, iter,
1327                         iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0) ?:
1328                 bch2_trans_commit(trans, NULL, NULL,
1329                         BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW);
1330         goto out;
1331 }
1332
1333 static int bch2_check_discard_freespace_key(struct btree_trans *trans,
1334                                             struct btree_iter *iter,
1335                                             struct bpos end)
1336 {
1337         if (!btree_node_type_is_extents(iter->btree_id)) {
1338                 return __bch2_check_discard_freespace_key(trans, iter);
1339         } else {
1340                 int ret;
1341
1342                 while (!bkey_eq(iter->pos, end) &&
1343                        !(ret = btree_trans_too_many_iters(trans) ?:
1344                                __bch2_check_discard_freespace_key(trans, iter)))
1345                         bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
1346
1347                 return ret;
1348         }
1349 }
1350
1351 /*
1352  * We've already checked that generation numbers in the bucket_gens btree are
1353  * valid for buckets that exist; this just checks for keys for nonexistent
1354  * buckets.
1355  */
1356 static noinline_for_stack
1357 int bch2_check_bucket_gens_key(struct btree_trans *trans,
1358                                struct btree_iter *iter,
1359                                struct bkey_s_c k)
1360 {
1361         struct bch_fs *c = trans->c;
1362         struct bkey_i_bucket_gens g;
1363         struct bch_dev *ca;
1364         u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
1365         u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
1366         u64 b;
1367         bool need_update = false, dev_exists;
1368         struct printbuf buf = PRINTBUF;
1369         int ret = 0;
1370
1371         BUG_ON(k.k->type != KEY_TYPE_bucket_gens);
1372         bkey_reassemble(&g.k_i, k);
1373
1374         /* if no bch_dev, skip out whether we repair or not */
1375         dev_exists = bch2_dev_exists2(c, k.k->p.inode);
1376         if (!dev_exists) {
1377                 if (fsck_err_on(!dev_exists, c,
1378                                 "bucket_gens key for invalid device:\n  %s",
1379                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1380                         ret = bch2_btree_delete_at(trans, iter, 0);
1381                 }
1382                 goto out;
1383         }
1384
1385         ca = bch_dev_bkey_exists(c, k.k->p.inode);
1386         if (fsck_err_on(end <= ca->mi.first_bucket ||
1387                         start >= ca->mi.nbuckets, c,
1388                         "bucket_gens key for invalid buckets:\n  %s",
1389                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1390                 ret = bch2_btree_delete_at(trans, iter, 0);
1391                 goto out;
1392         }
1393
1394         for (b = start; b < ca->mi.first_bucket; b++)
1395                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1396                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1397                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1398                         need_update = true;
1399                 }
1400
1401         for (b = ca->mi.nbuckets; b < end; b++)
1402                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1403                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1404                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1405                         need_update = true;
1406                 }
1407
1408         if (need_update) {
1409                 struct bkey_i *k;
1410
1411                 k = bch2_trans_kmalloc(trans, sizeof(g));
1412                 ret = PTR_ERR_OR_ZERO(k);
1413                 if (ret)
1414                         goto out;
1415
1416                 memcpy(k, &g, sizeof(g));
1417                 ret = bch2_trans_update(trans, iter, k, 0);
1418         }
1419 out:
1420 fsck_err:
1421         printbuf_exit(&buf);
1422         return ret;
1423 }
1424
1425 int bch2_check_alloc_info(struct bch_fs *c)
1426 {
1427         struct btree_trans trans;
1428         struct btree_iter iter, discard_iter, freespace_iter, bucket_gens_iter;
1429         struct bkey hole;
1430         struct bkey_s_c k;
1431         int ret = 0;
1432
1433         bch2_trans_init(&trans, c, 0, 0);
1434
1435         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc, POS_MIN,
1436                              BTREE_ITER_PREFETCH);
1437         bch2_trans_iter_init(&trans, &discard_iter, BTREE_ID_need_discard, POS_MIN,
1438                              BTREE_ITER_PREFETCH);
1439         bch2_trans_iter_init(&trans, &freespace_iter, BTREE_ID_freespace, POS_MIN,
1440                              BTREE_ITER_PREFETCH);
1441         bch2_trans_iter_init(&trans, &bucket_gens_iter, BTREE_ID_bucket_gens, POS_MIN,
1442                              BTREE_ITER_PREFETCH);
1443
1444         while (1) {
1445                 struct bpos next;
1446
1447                 bch2_trans_begin(&trans);
1448
1449                 k = bch2_get_key_or_real_bucket_hole(&iter, &hole);
1450                 ret = bkey_err(k);
1451                 if (ret)
1452                         goto bkey_err;
1453
1454                 if (!k.k)
1455                         break;
1456
1457                 if (k.k->type) {
1458                         next = bpos_nosnap_successor(k.k->p);
1459
1460                         ret = bch2_check_alloc_key(&trans,
1461                                                    k, &iter,
1462                                                    &discard_iter,
1463                                                    &freespace_iter,
1464                                                    &bucket_gens_iter);
1465                         if (ret)
1466                                 goto bkey_err;
1467                 } else {
1468                         next = k.k->p;
1469
1470                         ret = bch2_check_alloc_hole_freespace(&trans,
1471                                                     bkey_start_pos(k.k),
1472                                                     &next,
1473                                                     &freespace_iter) ?:
1474                                 bch2_check_alloc_hole_bucket_gens(&trans,
1475                                                     bkey_start_pos(k.k),
1476                                                     &next,
1477                                                     &bucket_gens_iter);
1478                         if (ret)
1479                                 goto bkey_err;
1480                 }
1481
1482                 ret = bch2_trans_commit(&trans, NULL, NULL,
1483                                         BTREE_INSERT_NOFAIL|
1484                                         BTREE_INSERT_LAZY_RW);
1485                 if (ret)
1486                         goto bkey_err;
1487
1488                 bch2_btree_iter_set_pos(&iter, next);
1489 bkey_err:
1490                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1491                         continue;
1492                 if (ret)
1493                         break;
1494         }
1495         bch2_trans_iter_exit(&trans, &bucket_gens_iter);
1496         bch2_trans_iter_exit(&trans, &freespace_iter);
1497         bch2_trans_iter_exit(&trans, &discard_iter);
1498         bch2_trans_iter_exit(&trans, &iter);
1499
1500         if (ret < 0)
1501                 goto err;
1502
1503         ret = for_each_btree_key2(&trans, iter,
1504                         BTREE_ID_need_discard, POS_MIN,
1505                         BTREE_ITER_PREFETCH, k,
1506                 bch2_check_discard_freespace_key(&trans, &iter, k.k->p)) ?:
1507               for_each_btree_key2(&trans, iter,
1508                         BTREE_ID_freespace, POS_MIN,
1509                         BTREE_ITER_PREFETCH, k,
1510                 bch2_check_discard_freespace_key(&trans, &iter, k.k->p)) ?:
1511               for_each_btree_key_commit(&trans, iter,
1512                         BTREE_ID_bucket_gens, POS_MIN,
1513                         BTREE_ITER_PREFETCH, k,
1514                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1515                 bch2_check_bucket_gens_key(&trans, &iter, k));
1516 err:
1517         bch2_trans_exit(&trans);
1518         if (ret)
1519                 bch_err_fn(c, ret);
1520         return ret;
1521 }
1522
1523 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
1524                                        struct btree_iter *alloc_iter)
1525 {
1526         struct bch_fs *c = trans->c;
1527         struct btree_iter lru_iter;
1528         struct bch_alloc_v4 a_convert;
1529         const struct bch_alloc_v4 *a;
1530         struct bkey_s_c alloc_k, lru_k;
1531         struct printbuf buf = PRINTBUF;
1532         int ret;
1533
1534         alloc_k = bch2_btree_iter_peek(alloc_iter);
1535         if (!alloc_k.k)
1536                 return 0;
1537
1538         ret = bkey_err(alloc_k);
1539         if (ret)
1540                 return ret;
1541
1542         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1543
1544         if (a->data_type != BCH_DATA_cached)
1545                 return 0;
1546
1547         lru_k = bch2_bkey_get_iter(trans, &lru_iter, BTREE_ID_lru,
1548                              lru_pos(alloc_k.k->p.inode,
1549                                      bucket_to_u64(alloc_k.k->p),
1550                                      a->io_time[READ]), 0);
1551         ret = bkey_err(lru_k);
1552         if (ret)
1553                 return ret;
1554
1555         if (fsck_err_on(!a->io_time[READ], c,
1556                         "cached bucket with read_time 0\n"
1557                         "  %s",
1558                 (printbuf_reset(&buf),
1559                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) ||
1560             fsck_err_on(lru_k.k->type != KEY_TYPE_set, c,
1561                         "missing lru entry\n"
1562                         "  %s",
1563                         (printbuf_reset(&buf),
1564                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
1565                 u64 read_time = a->io_time[READ] ?:
1566                         atomic64_read(&c->io_clock[READ].now);
1567
1568                 ret = bch2_lru_set(trans,
1569                                    alloc_k.k->p.inode,
1570                                    bucket_to_u64(alloc_k.k->p),
1571                                    read_time);
1572                 if (ret)
1573                         goto err;
1574
1575                 if (a->io_time[READ] != read_time) {
1576                         struct bkey_i_alloc_v4 *a_mut =
1577                                 bch2_alloc_to_v4_mut(trans, alloc_k);
1578                         ret = PTR_ERR_OR_ZERO(a_mut);
1579                         if (ret)
1580                                 goto err;
1581
1582                         a_mut->v.io_time[READ] = read_time;
1583                         ret = bch2_trans_update(trans, alloc_iter,
1584                                                 &a_mut->k_i, BTREE_TRIGGER_NORUN);
1585                         if (ret)
1586                                 goto err;
1587                 }
1588         }
1589 err:
1590 fsck_err:
1591         bch2_trans_iter_exit(trans, &lru_iter);
1592         printbuf_exit(&buf);
1593         return ret;
1594 }
1595
1596 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
1597 {
1598         struct btree_iter iter;
1599         struct bkey_s_c k;
1600         int ret = 0;
1601
1602         ret = bch2_trans_run(c,
1603                 for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1604                                 POS_MIN, BTREE_ITER_PREFETCH, k,
1605                                 NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1606                         bch2_check_alloc_to_lru_ref(&trans, &iter)));
1607         if (ret)
1608                 bch_err_fn(c, ret);
1609         return ret;
1610 }
1611
1612 static int bch2_discard_one_bucket(struct btree_trans *trans,
1613                                    struct btree_iter *need_discard_iter,
1614                                    struct bpos *discard_pos_done,
1615                                    u64 *seen,
1616                                    u64 *open,
1617                                    u64 *need_journal_commit,
1618                                    u64 *discarded)
1619 {
1620         struct bch_fs *c = trans->c;
1621         struct bpos pos = need_discard_iter->pos;
1622         struct btree_iter iter = { NULL };
1623         struct bkey_s_c k;
1624         struct bch_dev *ca;
1625         struct bkey_i_alloc_v4 *a;
1626         struct printbuf buf = PRINTBUF;
1627         int ret = 0;
1628
1629         ca = bch_dev_bkey_exists(c, pos.inode);
1630         if (!percpu_ref_tryget(&ca->io_ref)) {
1631                 bch2_btree_iter_set_pos(need_discard_iter, POS(pos.inode + 1, 0));
1632                 return 0;
1633         }
1634
1635         if (bch2_bucket_is_open_safe(c, pos.inode, pos.offset)) {
1636                 (*open)++;
1637                 goto out;
1638         }
1639
1640         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
1641                         c->journal.flushed_seq_ondisk,
1642                         pos.inode, pos.offset)) {
1643                 (*need_journal_commit)++;
1644                 goto out;
1645         }
1646
1647         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_alloc,
1648                                need_discard_iter->pos,
1649                                BTREE_ITER_CACHED);
1650         ret = bkey_err(k);
1651         if (ret)
1652                 goto out;
1653
1654         a = bch2_alloc_to_v4_mut(trans, k);
1655         ret = PTR_ERR_OR_ZERO(a);
1656         if (ret)
1657                 goto out;
1658
1659         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
1660                 a->v.gen++;
1661                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1662                 goto write;
1663         }
1664
1665         if (a->v.journal_seq > c->journal.flushed_seq_ondisk) {
1666                 if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
1667                         bch2_trans_inconsistent(trans,
1668                                 "clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
1669                                 "%s",
1670                                 a->v.journal_seq,
1671                                 c->journal.flushed_seq_ondisk,
1672                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1673                         ret = -EIO;
1674                 }
1675                 goto out;
1676         }
1677
1678         if (a->v.data_type != BCH_DATA_need_discard) {
1679                 if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
1680                         bch2_trans_inconsistent(trans,
1681                                 "bucket incorrectly set in need_discard btree\n"
1682                                 "%s",
1683                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1684                         ret = -EIO;
1685                 }
1686
1687                 goto out;
1688         }
1689
1690         if (!bkey_eq(*discard_pos_done, iter.pos) &&
1691             ca->mi.discard && !c->opts.nochanges) {
1692                 /*
1693                  * This works without any other locks because this is the only
1694                  * thread that removes items from the need_discard tree
1695                  */
1696                 bch2_trans_unlock(trans);
1697                 blkdev_issue_discard(ca->disk_sb.bdev,
1698                                      k.k->p.offset * ca->mi.bucket_size,
1699                                      ca->mi.bucket_size,
1700                                      GFP_KERNEL);
1701                 *discard_pos_done = iter.pos;
1702
1703                 ret = bch2_trans_relock_notrace(trans);
1704                 if (ret)
1705                         goto out;
1706         }
1707
1708         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
1709         a->v.data_type = alloc_data_type(a->v, a->v.data_type);
1710 write:
1711         ret =   bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1712                 bch2_trans_commit(trans, NULL, NULL,
1713                                   BCH_WATERMARK_btree|
1714                                   BTREE_INSERT_NOFAIL);
1715         if (ret)
1716                 goto out;
1717
1718         this_cpu_inc(c->counters[BCH_COUNTER_bucket_discard]);
1719         (*discarded)++;
1720 out:
1721         (*seen)++;
1722         bch2_trans_iter_exit(trans, &iter);
1723         percpu_ref_put(&ca->io_ref);
1724         printbuf_exit(&buf);
1725         return ret;
1726 }
1727
1728 static void bch2_do_discards_work(struct work_struct *work)
1729 {
1730         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
1731         struct btree_trans trans;
1732         struct btree_iter iter;
1733         struct bkey_s_c k;
1734         u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
1735         struct bpos discard_pos_done = POS_MAX;
1736         int ret;
1737
1738         bch2_trans_init(&trans, c, 0, 0);
1739
1740         /*
1741          * We're doing the commit in bch2_discard_one_bucket instead of using
1742          * for_each_btree_key_commit() so that we can increment counters after
1743          * successful commit:
1744          */
1745         ret = for_each_btree_key2(&trans, iter,
1746                         BTREE_ID_need_discard, POS_MIN, 0, k,
1747                 bch2_discard_one_bucket(&trans, &iter, &discard_pos_done,
1748                                         &seen,
1749                                         &open,
1750                                         &need_journal_commit,
1751                                         &discarded));
1752
1753         bch2_trans_exit(&trans);
1754
1755         if (need_journal_commit * 2 > seen)
1756                 bch2_journal_flush_async(&c->journal, NULL);
1757
1758         bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1759
1760         trace_discard_buckets(c, seen, open, need_journal_commit, discarded,
1761                               bch2_err_str(ret));
1762 }
1763
1764 void bch2_do_discards(struct bch_fs *c)
1765 {
1766         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_discard) &&
1767             !queue_work(c->write_ref_wq, &c->discard_work))
1768                 bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1769 }
1770
1771 static int invalidate_one_bucket(struct btree_trans *trans,
1772                                  struct btree_iter *lru_iter,
1773                                  struct bkey_s_c lru_k,
1774                                  s64 *nr_to_invalidate)
1775 {
1776         struct bch_fs *c = trans->c;
1777         struct btree_iter alloc_iter = { NULL };
1778         struct bkey_i_alloc_v4 *a = NULL;
1779         struct printbuf buf = PRINTBUF;
1780         struct bpos bucket = u64_to_bucket(lru_k.k->p.offset);
1781         unsigned cached_sectors;
1782         int ret = 0;
1783
1784         if (*nr_to_invalidate <= 0)
1785                 return 1;
1786
1787         if (!bch2_dev_bucket_exists(c, bucket)) {
1788                 prt_str(&buf, "lru entry points to invalid bucket");
1789                 goto err;
1790         }
1791
1792         if (bch2_bucket_is_open_safe(c, bucket.inode, bucket.offset))
1793                 return 0;
1794
1795         a = bch2_trans_start_alloc_update(trans, &alloc_iter, bucket);
1796         ret = PTR_ERR_OR_ZERO(a);
1797         if (ret)
1798                 goto out;
1799
1800         /* We expect harmless races here due to the btree write buffer: */
1801         if (lru_pos_time(lru_iter->pos) != alloc_lru_idx_read(a->v))
1802                 goto out;
1803
1804         BUG_ON(a->v.data_type != BCH_DATA_cached);
1805
1806         if (!a->v.cached_sectors)
1807                 bch_err(c, "invalidating empty bucket, confused");
1808
1809         cached_sectors = a->v.cached_sectors;
1810
1811         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1812         a->v.gen++;
1813         a->v.data_type          = 0;
1814         a->v.dirty_sectors      = 0;
1815         a->v.cached_sectors     = 0;
1816         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1817         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1818
1819         ret =   bch2_trans_update(trans, &alloc_iter, &a->k_i,
1820                                 BTREE_TRIGGER_BUCKET_INVALIDATE) ?:
1821                 bch2_trans_commit(trans, NULL, NULL,
1822                                   BCH_WATERMARK_btree|
1823                                   BTREE_INSERT_NOFAIL);
1824         if (ret)
1825                 goto out;
1826
1827         trace_and_count(c, bucket_invalidate, c, bucket.inode, bucket.offset, cached_sectors);
1828         --*nr_to_invalidate;
1829 out:
1830         bch2_trans_iter_exit(trans, &alloc_iter);
1831         printbuf_exit(&buf);
1832         return ret;
1833 err:
1834         prt_str(&buf, "\n  lru key: ");
1835         bch2_bkey_val_to_text(&buf, c, lru_k);
1836
1837         prt_str(&buf, "\n  lru entry: ");
1838         bch2_lru_pos_to_text(&buf, lru_iter->pos);
1839
1840         prt_str(&buf, "\n  alloc key: ");
1841         if (!a)
1842                 bch2_bpos_to_text(&buf, bucket);
1843         else
1844                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
1845
1846         bch_err(c, "%s", buf.buf);
1847         if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_lrus) {
1848                 bch2_inconsistent_error(c);
1849                 ret = -EINVAL;
1850         }
1851
1852         goto out;
1853 }
1854
1855 static void bch2_do_invalidates_work(struct work_struct *work)
1856 {
1857         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1858         struct bch_dev *ca;
1859         struct btree_trans trans;
1860         struct btree_iter iter;
1861         struct bkey_s_c k;
1862         unsigned i;
1863         int ret = 0;
1864
1865         bch2_trans_init(&trans, c, 0, 0);
1866
1867         ret = bch2_btree_write_buffer_flush(&trans);
1868         if (ret)
1869                 goto err;
1870
1871         for_each_member_device(ca, c, i) {
1872                 s64 nr_to_invalidate =
1873                         should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
1874
1875                 ret = for_each_btree_key2_upto(&trans, iter, BTREE_ID_lru,
1876                                 lru_pos(ca->dev_idx, 0, 0),
1877                                 lru_pos(ca->dev_idx, U64_MAX, LRU_TIME_MAX),
1878                                 BTREE_ITER_INTENT, k,
1879                         invalidate_one_bucket(&trans, &iter, k, &nr_to_invalidate));
1880
1881                 if (ret < 0) {
1882                         percpu_ref_put(&ca->ref);
1883                         break;
1884                 }
1885         }
1886 err:
1887         bch2_trans_exit(&trans);
1888         bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1889 }
1890
1891 void bch2_do_invalidates(struct bch_fs *c)
1892 {
1893         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_invalidate) &&
1894             !queue_work(c->write_ref_wq, &c->invalidate_work))
1895                 bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1896 }
1897
1898 static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca,
1899                                    unsigned long *last_updated)
1900 {
1901         struct btree_trans trans;
1902         struct btree_iter iter;
1903         struct bkey_s_c k;
1904         struct bkey hole;
1905         struct bpos end = POS(ca->dev_idx, ca->mi.nbuckets);
1906         struct bch_member *m;
1907         int ret;
1908
1909         bch2_trans_init(&trans, c, 0, 0);
1910
1911         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc,
1912                              POS(ca->dev_idx, ca->mi.first_bucket),
1913                              BTREE_ITER_PREFETCH);
1914         /*
1915          * Scan the alloc btree for every bucket on @ca, and add buckets to the
1916          * freespace/need_discard/need_gc_gens btrees as needed:
1917          */
1918         while (1) {
1919                 if (*last_updated + HZ * 10 < jiffies) {
1920                         bch_info(ca, "%s: currently at %llu/%llu",
1921                                  __func__, iter.pos.offset, ca->mi.nbuckets);
1922                         *last_updated = jiffies;
1923                 }
1924
1925                 bch2_trans_begin(&trans);
1926
1927                 if (bkey_ge(iter.pos, end)) {
1928                         ret = 0;
1929                         break;
1930                 }
1931
1932                 k = bch2_get_key_or_hole(&iter, end, &hole);
1933                 ret = bkey_err(k);
1934                 if (ret)
1935                         goto bkey_err;
1936
1937                 if (k.k->type) {
1938                         /*
1939                          * We process live keys in the alloc btree one at a
1940                          * time:
1941                          */
1942                         struct bch_alloc_v4 a_convert;
1943                         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1944
1945                         ret =   bch2_bucket_do_index(&trans, k, a, true) ?:
1946                                 bch2_trans_commit(&trans, NULL, NULL,
1947                                                   BTREE_INSERT_LAZY_RW|
1948                                                   BTREE_INSERT_NOFAIL);
1949                         if (ret)
1950                                 goto bkey_err;
1951
1952                         bch2_btree_iter_advance(&iter);
1953                 } else {
1954                         struct bkey_i *freespace;
1955
1956                         freespace = bch2_trans_kmalloc(&trans, sizeof(*freespace));
1957                         ret = PTR_ERR_OR_ZERO(freespace);
1958                         if (ret)
1959                                 goto bkey_err;
1960
1961                         bkey_init(&freespace->k);
1962                         freespace->k.type       = KEY_TYPE_set;
1963                         freespace->k.p          = k.k->p;
1964                         freespace->k.size       = k.k->size;
1965
1966                         ret = __bch2_btree_insert(&trans, BTREE_ID_freespace, freespace, 0) ?:
1967                                 bch2_trans_commit(&trans, NULL, NULL,
1968                                                   BTREE_INSERT_LAZY_RW|
1969                                                   BTREE_INSERT_NOFAIL);
1970                         if (ret)
1971                                 goto bkey_err;
1972
1973                         bch2_btree_iter_set_pos(&iter, k.k->p);
1974                 }
1975 bkey_err:
1976                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1977                         continue;
1978                 if (ret)
1979                         break;
1980         }
1981
1982         bch2_trans_iter_exit(&trans, &iter);
1983         bch2_trans_exit(&trans);
1984
1985         if (ret < 0) {
1986                 bch_err(ca, "error initializing free space: %s", bch2_err_str(ret));
1987                 return ret;
1988         }
1989
1990         mutex_lock(&c->sb_lock);
1991         m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx;
1992         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1993         mutex_unlock(&c->sb_lock);
1994
1995         return 0;
1996 }
1997
1998 int bch2_fs_freespace_init(struct bch_fs *c)
1999 {
2000         struct bch_dev *ca;
2001         unsigned i;
2002         int ret = 0;
2003         bool doing_init = false;
2004         unsigned long last_updated = jiffies;
2005
2006         /*
2007          * We can crash during the device add path, so we need to check this on
2008          * every mount:
2009          */
2010
2011         for_each_member_device(ca, c, i) {
2012                 if (ca->mi.freespace_initialized)
2013                         continue;
2014
2015                 if (!doing_init) {
2016                         bch_info(c, "initializing freespace");
2017                         doing_init = true;
2018                 }
2019
2020                 ret = bch2_dev_freespace_init(c, ca, &last_updated);
2021                 if (ret) {
2022                         percpu_ref_put(&ca->ref);
2023                         bch_err_fn(c, ret);
2024                         return ret;
2025                 }
2026         }
2027
2028         if (doing_init) {
2029                 mutex_lock(&c->sb_lock);
2030                 bch2_write_super(c);
2031                 mutex_unlock(&c->sb_lock);
2032                 bch_verbose(c, "done initializing freespace");
2033         }
2034
2035         return 0;
2036 }
2037
2038 /* Bucket IO clocks: */
2039
2040 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
2041                               size_t bucket_nr, int rw)
2042 {
2043         struct bch_fs *c = trans->c;
2044         struct btree_iter iter;
2045         struct bkey_i_alloc_v4 *a;
2046         u64 now;
2047         int ret = 0;
2048
2049         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
2050         ret = PTR_ERR_OR_ZERO(a);
2051         if (ret)
2052                 return ret;
2053
2054         now = atomic64_read(&c->io_clock[rw].now);
2055         if (a->v.io_time[rw] == now)
2056                 goto out;
2057
2058         a->v.io_time[rw] = now;
2059
2060         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
2061                 bch2_trans_commit(trans, NULL, NULL, 0);
2062 out:
2063         bch2_trans_iter_exit(trans, &iter);
2064         return ret;
2065 }
2066
2067 /* Startup/shutdown (ro/rw): */
2068
2069 void bch2_recalc_capacity(struct bch_fs *c)
2070 {
2071         struct bch_dev *ca;
2072         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
2073         unsigned bucket_size_max = 0;
2074         unsigned long ra_pages = 0;
2075         unsigned i;
2076
2077         lockdep_assert_held(&c->state_lock);
2078
2079         for_each_online_member(ca, c, i) {
2080                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
2081
2082                 ra_pages += bdi->ra_pages;
2083         }
2084
2085         bch2_set_ra_pages(c, ra_pages);
2086
2087         for_each_rw_member(ca, c, i) {
2088                 u64 dev_reserve = 0;
2089
2090                 /*
2091                  * We need to reserve buckets (from the number
2092                  * of currently available buckets) against
2093                  * foreground writes so that mainly copygc can
2094                  * make forward progress.
2095                  *
2096                  * We need enough to refill the various reserves
2097                  * from scratch - copygc will use its entire
2098                  * reserve all at once, then run against when
2099                  * its reserve is refilled (from the formerly
2100                  * available buckets).
2101                  *
2102                  * This reserve is just used when considering if
2103                  * allocations for foreground writes must wait -
2104                  * not -ENOSPC calculations.
2105                  */
2106
2107                 dev_reserve += ca->nr_btree_reserve * 2;
2108                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
2109
2110                 dev_reserve += 1;       /* btree write point */
2111                 dev_reserve += 1;       /* copygc write point */
2112                 dev_reserve += 1;       /* rebalance write point */
2113
2114                 dev_reserve *= ca->mi.bucket_size;
2115
2116                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
2117                                              ca->mi.first_bucket);
2118
2119                 reserved_sectors += dev_reserve * 2;
2120
2121                 bucket_size_max = max_t(unsigned, bucket_size_max,
2122                                         ca->mi.bucket_size);
2123         }
2124
2125         gc_reserve = c->opts.gc_reserve_bytes
2126                 ? c->opts.gc_reserve_bytes >> 9
2127                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
2128
2129         reserved_sectors = max(gc_reserve, reserved_sectors);
2130
2131         reserved_sectors = min(reserved_sectors, capacity);
2132
2133         c->capacity = capacity - reserved_sectors;
2134
2135         c->bucket_size_max = bucket_size_max;
2136
2137         /* Wake up case someone was waiting for buckets */
2138         closure_wake_up(&c->freelist_wait);
2139 }
2140
2141 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
2142 {
2143         struct open_bucket *ob;
2144         bool ret = false;
2145
2146         for (ob = c->open_buckets;
2147              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
2148              ob++) {
2149                 spin_lock(&ob->lock);
2150                 if (ob->valid && !ob->on_partial_list &&
2151                     ob->dev == ca->dev_idx)
2152                         ret = true;
2153                 spin_unlock(&ob->lock);
2154         }
2155
2156         return ret;
2157 }
2158
2159 /* device goes ro: */
2160 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
2161 {
2162         unsigned i;
2163
2164         /* First, remove device from allocation groups: */
2165
2166         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2167                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
2168
2169         /*
2170          * Capacity is calculated based off of devices in allocation groups:
2171          */
2172         bch2_recalc_capacity(c);
2173
2174         bch2_open_buckets_stop(c, ca, false);
2175
2176         /*
2177          * Wake up threads that were blocked on allocation, so they can notice
2178          * the device can no longer be removed and the capacity has changed:
2179          */
2180         closure_wake_up(&c->freelist_wait);
2181
2182         /*
2183          * journal_res_get() can block waiting for free space in the journal -
2184          * it needs to notice there may not be devices to allocate from anymore:
2185          */
2186         wake_up(&c->journal.wait);
2187
2188         /* Now wait for any in flight writes: */
2189
2190         closure_wait_event(&c->open_buckets_wait,
2191                            !bch2_dev_has_open_write_point(c, ca));
2192 }
2193
2194 /* device goes rw: */
2195 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
2196 {
2197         unsigned i;
2198
2199         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2200                 if (ca->mi.data_allowed & (1 << i))
2201                         set_bit(ca->dev_idx, c->rw_devs[i].d);
2202 }
2203
2204 void bch2_fs_allocator_background_init(struct bch_fs *c)
2205 {
2206         spin_lock_init(&c->freelist_lock);
2207         INIT_WORK(&c->discard_work, bch2_do_discards_work);
2208         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
2209 }