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

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "bcachefs.h"
   3 #include "alloc_background.h"
   4 #include "alloc_foreground.h"
   5 #include "backpointers.h"
   6 #include "btree_cache.h"
   7 #include "btree_io.h"
   8 #include "btree_key_cache.h"
   9 #include "btree_update.h"
  10 #include "btree_update_interior.h"
  11 #include "btree_gc.h"
  12 #include "btree_write_buffer.h"
  13 #include "buckets.h"
  14 #include "buckets_waiting_for_journal.h"
  15 #include "clock.h"
  16 #include "debug.h"
  17 #include "ec.h"
  18 #include "error.h"
  19 #include "lru.h"
  20 #include "recovery.h"
  21 #include "trace.h"
  22 #include "varint.h"
  23
  24 #include <linux/kthread.h>
  25 #include <linux/math64.h>
  26 #include <linux/random.h>
  27 #include <linux/rculist.h>
  28 #include <linux/rcupdate.h>
  29 #include <linux/sched/task.h>
  30 #include <linux/sort.h>
  31
  32 /* Persistent alloc info: */
  33
  34 static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
  35 #define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8,
  36         BCH_ALLOC_FIELDS_V1()
  37 #undef x
  38 };
  39
  40 struct bkey_alloc_unpacked {
  41         u64             journal_seq;
  42         u8              gen;
  43         u8              oldest_gen;
  44         u8              data_type;
  45         bool            need_discard:1;
  46         bool            need_inc_gen:1;
  47 #define x(_name, _bits) u##_bits _name;
  48         BCH_ALLOC_FIELDS_V2()
  49 #undef  x
  50 };
  51
  52 static inline u64 alloc_field_v1_get(const struct bch_alloc *a,
  53                                      const void **p, unsigned field)
  54 {
  55         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  56         u64 v;
  57
  58         if (!(a->fields & (1 << field)))
  59                 return 0;
  60
  61         switch (bytes) {
  62         case 1:
  63                 v = *((const u8 *) *p);
  64                 break;
  65         case 2:
  66                 v = le16_to_cpup(*p);
  67                 break;
  68         case 4:
  69                 v = le32_to_cpup(*p);
  70                 break;
  71         case 8:
  72                 v = le64_to_cpup(*p);
  73                 break;
  74         default:
  75                 BUG();
  76         }
  77
  78         *p += bytes;
  79         return v;
  80 }
  81
  82 static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out,
  83                                  struct bkey_s_c k)
  84 {
  85         const struct bch_alloc *in = bkey_s_c_to_alloc(k).v;
  86         const void *d = in->data;
  87         unsigned idx = 0;
  88
  89         out->gen = in->gen;
  90
  91 #define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++);
  92         BCH_ALLOC_FIELDS_V1()
  93 #undef  x
  94 }
  95
  96 static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out,
  97                                 struct bkey_s_c k)
  98 {
  99         struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k);
 100         const u8 *in = a.v->data;
 101         const u8 *end = bkey_val_end(a);
 102         unsigned fieldnr = 0;
 103         int ret;
 104         u64 v;
 105
 106         out->gen        = a.v->gen;
 107         out->oldest_gen = a.v->oldest_gen;
 108         out->data_type  = a.v->data_type;
 109
 110 #define x(_name, _bits)                                                 \
 111         if (fieldnr < a.v->nr_fields) {                                 \
 112                 ret = bch2_varint_decode_fast(in, end, &v);             \
 113                 if (ret < 0)                                            \
 114                         return ret;                                     \
 115                 in += ret;                                              \
 116         } else {                                                        \
 117                 v = 0;                                                  \
 118         }                                                               \
 119         out->_name = v;                                                 \
 120         if (v != out->_name)                                            \
 121                 return -1;                                              \
 122         fieldnr++;
 123
 124         BCH_ALLOC_FIELDS_V2()
 125 #undef  x
 126         return 0;
 127 }
 128
 129 static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out,
 130                                 struct bkey_s_c k)
 131 {
 132         struct bkey_s_c_alloc_v3 a = bkey_s_c_to_alloc_v3(k);
 133         const u8 *in = a.v->data;
 134         const u8 *end = bkey_val_end(a);
 135         unsigned fieldnr = 0;
 136         int ret;
 137         u64 v;
 138
 139         out->gen        = a.v->gen;
 140         out->oldest_gen = a.v->oldest_gen;
 141         out->data_type  = a.v->data_type;
 142         out->need_discard = BCH_ALLOC_V3_NEED_DISCARD(a.v);
 143         out->need_inc_gen = BCH_ALLOC_V3_NEED_INC_GEN(a.v);
 144         out->journal_seq = le64_to_cpu(a.v->journal_seq);
 145
 146 #define x(_name, _bits)                                                 \
 147         if (fieldnr < a.v->nr_fields) {                                 \
 148                 ret = bch2_varint_decode_fast(in, end, &v);             \
 149                 if (ret < 0)                                            \
 150                         return ret;                                     \
 151                 in += ret;                                              \
 152         } else {                                                        \
 153                 v = 0;                                                  \
 154         }                                                               \
 155         out->_name = v;                                                 \
 156         if (v != out->_name)                                            \
 157                 return -1;                                              \
 158         fieldnr++;
 159
 160         BCH_ALLOC_FIELDS_V2()
 161 #undef  x
 162         return 0;
 163 }
 164
 165 static struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 166 {
 167         struct bkey_alloc_unpacked ret = { .gen = 0 };
 168
 169         switch (k.k->type) {
 170         case KEY_TYPE_alloc:
 171                 bch2_alloc_unpack_v1(&ret, k);
 172                 break;
 173         case KEY_TYPE_alloc_v2:
 174                 bch2_alloc_unpack_v2(&ret, k);
 175                 break;
 176         case KEY_TYPE_alloc_v3:
 177                 bch2_alloc_unpack_v3(&ret, k);
 178                 break;
 179         }
 180
 181         return ret;
 182 }
 183
 184 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
 185 {
 186         unsigned i, bytes = offsetof(struct bch_alloc, data);
 187
 188         for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++)
 189                 if (a->fields & (1 << i))
 190                         bytes += BCH_ALLOC_V1_FIELD_BYTES[i];
 191
 192         return DIV_ROUND_UP(bytes, sizeof(u64));
 193 }
 194
 195 int bch2_alloc_v1_invalid(struct bch_fs *c, struct bkey_s_c k,
 196                           enum bkey_invalid_flags flags,
 197                           struct printbuf *err)
 198 {
 199         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 200         int ret = 0;
 201
 202         /* allow for unknown fields */
 203         bkey_fsck_err_on(bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v), c, err,
 204                          alloc_v1_val_size_bad,
 205                          "incorrect value size (%zu < %u)",
 206                          bkey_val_u64s(a.k), bch_alloc_v1_val_u64s(a.v));
 207 fsck_err:
 208         return ret;
 209 }
 210
 211 int bch2_alloc_v2_invalid(struct bch_fs *c, struct bkey_s_c k,
 212                           enum bkey_invalid_flags flags,
 213                           struct printbuf *err)
 214 {
 215         struct bkey_alloc_unpacked u;
 216         int ret = 0;
 217
 218         bkey_fsck_err_on(bch2_alloc_unpack_v2(&u, k), c, err,
 219                          alloc_v2_unpack_error,
 220                          "unpack error");
 221 fsck_err:
 222         return ret;
 223 }
 224
 225 int bch2_alloc_v3_invalid(struct bch_fs *c, struct bkey_s_c k,
 226                           enum bkey_invalid_flags flags,
 227                           struct printbuf *err)
 228 {
 229         struct bkey_alloc_unpacked u;
 230         int ret = 0;
 231
 232         bkey_fsck_err_on(bch2_alloc_unpack_v3(&u, k), c, err,
 233                          alloc_v2_unpack_error,
 234                          "unpack error");
 235 fsck_err:
 236         return ret;
 237 }
 238
 239 int bch2_alloc_v4_invalid(struct bch_fs *c, struct bkey_s_c k,
 240                           enum bkey_invalid_flags flags, struct printbuf *err)
 241 {
 242         struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 243         int ret = 0;
 244
 245         bkey_fsck_err_on(alloc_v4_u64s(a.v) > bkey_val_u64s(k.k), c, err,
 246                          alloc_v4_val_size_bad,
 247                          "bad val size (%u > %zu)",
 248                          alloc_v4_u64s(a.v), bkey_val_u64s(k.k));
 249
 250         bkey_fsck_err_on(!BCH_ALLOC_V4_BACKPOINTERS_START(a.v) &&
 251                          BCH_ALLOC_V4_NR_BACKPOINTERS(a.v), c, err,
 252                          alloc_v4_backpointers_start_bad,
 253                          "invalid backpointers_start");
 254
 255         bkey_fsck_err_on(alloc_data_type(*a.v, a.v->data_type) != a.v->data_type, c, err,
 256                          alloc_key_data_type_bad,
 257                          "invalid data type (got %u should be %u)",
 258                          a.v->data_type, alloc_data_type(*a.v, a.v->data_type));
 259
 260         switch (a.v->data_type) {
 261         case BCH_DATA_free:
 262         case BCH_DATA_need_gc_gens:
 263         case BCH_DATA_need_discard:
 264                 bkey_fsck_err_on(bch2_bucket_sectors(*a.v) || a.v->stripe,
 265                                  c, err, alloc_key_empty_but_have_data,
 266                                  "empty data type free but have data");
 267                 break;
 268         case BCH_DATA_sb:
 269         case BCH_DATA_journal:
 270         case BCH_DATA_btree:
 271         case BCH_DATA_user:
 272         case BCH_DATA_parity:
 273                 bkey_fsck_err_on(!bch2_bucket_sectors_dirty(*a.v),
 274                                  c, err, alloc_key_dirty_sectors_0,
 275                                  "data_type %s but dirty_sectors==0",
 276                                  bch2_data_types[a.v->data_type]);
 277                 break;
 278         case BCH_DATA_cached:
 279                 bkey_fsck_err_on(!a.v->cached_sectors ||
 280                                  bch2_bucket_sectors_dirty(*a.v) ||
 281                                  a.v->stripe,
 282                                  c, err, alloc_key_cached_inconsistency,
 283                                  "data type inconsistency");
 284
 285                 bkey_fsck_err_on(!a.v->io_time[READ] &&
 286                                  c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_to_lru_refs,
 287                                  c, err, alloc_key_cached_but_read_time_zero,
 288                                  "cached bucket with read_time == 0");
 289                 break;
 290         case BCH_DATA_stripe:
 291                 break;
 292         }
 293 fsck_err:
 294         return ret;
 295 }
 296
 297 void bch2_alloc_v4_swab(struct bkey_s k)
 298 {
 299         struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v;
 300         struct bch_backpointer *bp, *bps;
 301
 302         a->journal_seq          = swab64(a->journal_seq);
 303         a->flags                = swab32(a->flags);
 304         a->dirty_sectors        = swab32(a->dirty_sectors);
 305         a->cached_sectors       = swab32(a->cached_sectors);
 306         a->io_time[0]           = swab64(a->io_time[0]);
 307         a->io_time[1]           = swab64(a->io_time[1]);
 308         a->stripe               = swab32(a->stripe);
 309         a->nr_external_backpointers = swab32(a->nr_external_backpointers);
 310         a->fragmentation_lru    = swab64(a->fragmentation_lru);
 311
 312         bps = alloc_v4_backpointers(a);
 313         for (bp = bps; bp < bps + BCH_ALLOC_V4_NR_BACKPOINTERS(a); bp++) {
 314                 bp->bucket_offset       = swab40(bp->bucket_offset);
 315                 bp->bucket_len          = swab32(bp->bucket_len);
 316                 bch2_bpos_swab(&bp->pos);
 317         }
 318 }
 319
 320 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 321 {
 322         struct bch_alloc_v4 _a;
 323         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &_a);
 324         unsigned i;
 325
 326         prt_newline(out);
 327         printbuf_indent_add(out, 2);
 328
 329         prt_printf(out, "gen %u oldest_gen %u data_type %s",
 330                a->gen, a->oldest_gen,
 331                a->data_type < BCH_DATA_NR
 332                ? bch2_data_types[a->data_type]
 333                : "(invalid data type)");
 334         prt_newline(out);
 335         prt_printf(out, "journal_seq       %llu",       a->journal_seq);
 336         prt_newline(out);
 337         prt_printf(out, "need_discard      %llu",       BCH_ALLOC_V4_NEED_DISCARD(a));
 338         prt_newline(out);
 339         prt_printf(out, "need_inc_gen      %llu",       BCH_ALLOC_V4_NEED_INC_GEN(a));
 340         prt_newline(out);
 341         prt_printf(out, "dirty_sectors     %u", a->dirty_sectors);
 342         prt_newline(out);
 343         prt_printf(out, "cached_sectors    %u", a->cached_sectors);
 344         prt_newline(out);
 345         prt_printf(out, "stripe            %u", a->stripe);
 346         prt_newline(out);
 347         prt_printf(out, "stripe_redundancy %u", a->stripe_redundancy);
 348         prt_newline(out);
 349         prt_printf(out, "io_time[READ]     %llu",       a->io_time[READ]);
 350         prt_newline(out);
 351         prt_printf(out, "io_time[WRITE]    %llu",       a->io_time[WRITE]);
 352         prt_newline(out);
 353         prt_printf(out, "fragmentation     %llu",       a->fragmentation_lru);
 354         prt_newline(out);
 355         prt_printf(out, "bp_start          %llu", BCH_ALLOC_V4_BACKPOINTERS_START(a));
 356         prt_newline(out);
 357
 358         if (BCH_ALLOC_V4_NR_BACKPOINTERS(a)) {
 359                 struct bkey_s_c_alloc_v4 a_raw = bkey_s_c_to_alloc_v4(k);
 360                 const struct bch_backpointer *bps = alloc_v4_backpointers_c(a_raw.v);
 361
 362                 prt_printf(out, "backpointers:     %llu", BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v));
 363                 printbuf_indent_add(out, 2);
 364
 365                 for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v); i++) {
 366                         prt_newline(out);
 367                         bch2_backpointer_to_text(out, &bps[i]);
 368                 }
 369
 370                 printbuf_indent_sub(out, 2);
 371         }
 372
 373         printbuf_indent_sub(out, 2);
 374 }
 375
 376 void __bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out)
 377 {
 378         if (k.k->type == KEY_TYPE_alloc_v4) {
 379                 void *src, *dst;
 380
 381                 *out = *bkey_s_c_to_alloc_v4(k).v;
 382
 383                 src = alloc_v4_backpointers(out);
 384                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 385                 dst = alloc_v4_backpointers(out);
 386
 387                 if (src < dst)
 388                         memset(src, 0, dst - src);
 389
 390                 SET_BCH_ALLOC_V4_NR_BACKPOINTERS(out, 0);
 391         } else {
 392                 struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 393
 394                 *out = (struct bch_alloc_v4) {
 395                         .journal_seq            = u.journal_seq,
 396                         .flags                  = u.need_discard,
 397                         .gen                    = u.gen,
 398                         .oldest_gen             = u.oldest_gen,
 399                         .data_type              = u.data_type,
 400                         .stripe_redundancy      = u.stripe_redundancy,
 401                         .dirty_sectors          = u.dirty_sectors,
 402                         .cached_sectors         = u.cached_sectors,
 403                         .io_time[READ]          = u.read_time,
 404                         .io_time[WRITE]         = u.write_time,
 405                         .stripe                 = u.stripe,
 406                 };
 407
 408                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 409         }
 410 }
 411
 412 static noinline struct bkey_i_alloc_v4 *
 413 __bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 414 {
 415         struct bkey_i_alloc_v4 *ret;
 416
 417         ret = bch2_trans_kmalloc(trans, max(bkey_bytes(k.k), sizeof(struct bkey_i_alloc_v4)));
 418         if (IS_ERR(ret))
 419                 return ret;
 420
 421         if (k.k->type == KEY_TYPE_alloc_v4) {
 422                 void *src, *dst;
 423
 424                 bkey_reassemble(&ret->k_i, k);
 425
 426                 src = alloc_v4_backpointers(&ret->v);
 427                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(&ret->v, BCH_ALLOC_V4_U64s);
 428                 dst = alloc_v4_backpointers(&ret->v);
 429
 430                 if (src < dst)
 431                         memset(src, 0, dst - src);
 432
 433                 SET_BCH_ALLOC_V4_NR_BACKPOINTERS(&ret->v, 0);
 434                 set_alloc_v4_u64s(ret);
 435         } else {
 436                 bkey_alloc_v4_init(&ret->k_i);
 437                 ret->k.p = k.k->p;
 438                 bch2_alloc_to_v4(k, &ret->v);
 439         }
 440         return ret;
 441 }
 442
 443 static inline struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut_inlined(struct btree_trans *trans, struct bkey_s_c k)
 444 {
 445         struct bkey_s_c_alloc_v4 a;
 446
 447         if (likely(k.k->type == KEY_TYPE_alloc_v4) &&
 448             ((a = bkey_s_c_to_alloc_v4(k), true) &&
 449              BCH_ALLOC_V4_NR_BACKPOINTERS(a.v) == 0))
 450                 return bch2_bkey_make_mut_noupdate_typed(trans, k, alloc_v4);
 451
 452         return __bch2_alloc_to_v4_mut(trans, k);
 453 }
 454
 455 struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 456 {
 457         return bch2_alloc_to_v4_mut_inlined(trans, k);
 458 }
 459
 460 struct bkey_i_alloc_v4 *
 461 bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
 462                               struct bpos pos)
 463 {
 464         struct bkey_s_c k;
 465         struct bkey_i_alloc_v4 *a;
 466         int ret;
 467
 468         k = bch2_bkey_get_iter(trans, iter, BTREE_ID_alloc, pos,
 469                              BTREE_ITER_WITH_UPDATES|
 470                              BTREE_ITER_CACHED|
 471                              BTREE_ITER_INTENT);
 472         ret = bkey_err(k);
 473         if (unlikely(ret))
 474                 return ERR_PTR(ret);
 475
 476         a = bch2_alloc_to_v4_mut_inlined(trans, k);
 477         ret = PTR_ERR_OR_ZERO(a);
 478         if (unlikely(ret))
 479                 goto err;
 480         return a;
 481 err:
 482         bch2_trans_iter_exit(trans, iter);
 483         return ERR_PTR(ret);
 484 }
 485
 486 static struct bpos alloc_gens_pos(struct bpos pos, unsigned *offset)
 487 {
 488         *offset = pos.offset & KEY_TYPE_BUCKET_GENS_MASK;
 489
 490         pos.offset >>= KEY_TYPE_BUCKET_GENS_BITS;
 491         return pos;
 492 }
 493
 494 static struct bpos bucket_gens_pos_to_alloc(struct bpos pos, unsigned offset)
 495 {
 496         pos.offset <<= KEY_TYPE_BUCKET_GENS_BITS;
 497         pos.offset += offset;
 498         return pos;
 499 }
 500
 501 static unsigned alloc_gen(struct bkey_s_c k, unsigned offset)
 502 {
 503         return k.k->type == KEY_TYPE_bucket_gens
 504                 ? bkey_s_c_to_bucket_gens(k).v->gens[offset]
 505                 : 0;
 506 }
 507
 508 int bch2_bucket_gens_invalid(struct bch_fs *c, struct bkey_s_c k,
 509                              enum bkey_invalid_flags flags,
 510                              struct printbuf *err)
 511 {
 512         int ret = 0;
 513
 514         bkey_fsck_err_on(bkey_val_bytes(k.k) != sizeof(struct bch_bucket_gens), c, err,
 515                          bucket_gens_val_size_bad,
 516                          "bad val size (%zu != %zu)",
 517                          bkey_val_bytes(k.k), sizeof(struct bch_bucket_gens));
 518 fsck_err:
 519         return ret;
 520 }
 521
 522 void bch2_bucket_gens_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 523 {
 524         struct bkey_s_c_bucket_gens g = bkey_s_c_to_bucket_gens(k);
 525         unsigned i;
 526
 527         for (i = 0; i < ARRAY_SIZE(g.v->gens); i++) {
 528                 if (i)
 529                         prt_char(out, ' ');
 530                 prt_printf(out, "%u", g.v->gens[i]);
 531         }
 532 }
 533
 534 int bch2_bucket_gens_init(struct bch_fs *c)
 535 {
 536         struct btree_trans *trans = bch2_trans_get(c);
 537         struct bkey_i_bucket_gens g;
 538         bool have_bucket_gens_key = false;
 539         int ret;
 540
 541         ret = for_each_btree_key(trans, iter, BTREE_ID_alloc, POS_MIN,
 542                                  BTREE_ITER_PREFETCH, k, ({
 543                 /*
 544                  * Not a fsck error because this is checked/repaired by
 545                  * bch2_check_alloc_key() which runs later:
 546                  */
 547                 if (!bch2_dev_bucket_exists(c, k.k->p))
 548                         continue;
 549
 550                 struct bch_alloc_v4 a;
 551                 u8 gen = bch2_alloc_to_v4(k, &a)->gen;
 552                 unsigned offset;
 553                 struct bpos pos = alloc_gens_pos(iter.pos, &offset);
 554
 555                 if (have_bucket_gens_key && bkey_cmp(iter.pos, pos)) {
 556                         ret = commit_do(trans, NULL, NULL,
 557                                         BCH_TRANS_COMMIT_no_enospc,
 558                                 bch2_btree_insert_trans(trans, BTREE_ID_bucket_gens, &g.k_i, 0));
 559                         if (ret)
 560                                 break;
 561                         have_bucket_gens_key = false;
 562                 }
 563
 564                 if (!have_bucket_gens_key) {
 565                         bkey_bucket_gens_init(&g.k_i);
 566                         g.k.p = pos;
 567                         have_bucket_gens_key = true;
 568                 }
 569
 570                 g.v.gens[offset] = gen;
 571                 0;
 572         }));
 573
 574         if (have_bucket_gens_key && !ret)
 575                 ret = commit_do(trans, NULL, NULL,
 576                                 BCH_TRANS_COMMIT_no_enospc,
 577                         bch2_btree_insert_trans(trans, BTREE_ID_bucket_gens, &g.k_i, 0));
 578
 579         bch2_trans_put(trans);
 580
 581         bch_err_fn(c, ret);
 582         return ret;
 583 }
 584
 585 int bch2_alloc_read(struct bch_fs *c)
 586 {
 587         struct btree_trans *trans = bch2_trans_get(c);
 588         int ret;
 589
 590         down_read(&c->gc_lock);
 591
 592         if (c->sb.version_upgrade_complete >= bcachefs_metadata_version_bucket_gens) {
 593                 ret = for_each_btree_key(trans, iter, BTREE_ID_bucket_gens, POS_MIN,
 594                                          BTREE_ITER_PREFETCH, k, ({
 595                         u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
 596                         u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
 597
 598                         if (k.k->type != KEY_TYPE_bucket_gens)
 599                                 continue;
 600
 601                         const struct bch_bucket_gens *g = bkey_s_c_to_bucket_gens(k).v;
 602
 603                         /*
 604                          * Not a fsck error because this is checked/repaired by
 605                          * bch2_check_alloc_key() which runs later:
 606                          */
 607                         if (!bch2_dev_exists2(c, k.k->p.inode))
 608                                 continue;
 609
 610                         struct bch_dev *ca = bch_dev_bkey_exists(c, k.k->p.inode);
 611
 612                         for (u64 b = max_t(u64, ca->mi.first_bucket, start);
 613                              b < min_t(u64, ca->mi.nbuckets, end);
 614                              b++)
 615                                 *bucket_gen(ca, b) = g->gens[b & KEY_TYPE_BUCKET_GENS_MASK];
 616                         0;
 617                 }));
 618         } else {
 619                 ret = for_each_btree_key(trans, iter, BTREE_ID_alloc, POS_MIN,
 620                                          BTREE_ITER_PREFETCH, k, ({
 621                         /*
 622                          * Not a fsck error because this is checked/repaired by
 623                          * bch2_check_alloc_key() which runs later:
 624                          */
 625                         if (!bch2_dev_bucket_exists(c, k.k->p))
 626                                 continue;
 627
 628                         struct bch_dev *ca = bch_dev_bkey_exists(c, k.k->p.inode);
 629
 630                         struct bch_alloc_v4 a;
 631                         *bucket_gen(ca, k.k->p.offset) = bch2_alloc_to_v4(k, &a)->gen;
 632                         0;
 633                 }));
 634         }
 635
 636         bch2_trans_put(trans);
 637         up_read(&c->gc_lock);
 638
 639         bch_err_fn(c, ret);
 640         return ret;
 641 }
 642
 643 /* Free space/discard btree: */
 644
 645 static int bch2_bucket_do_index(struct btree_trans *trans,
 646                                 struct bkey_s_c alloc_k,
 647                                 const struct bch_alloc_v4 *a,
 648                                 bool set)
 649 {
 650         struct bch_fs *c = trans->c;
 651         struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 652         struct btree_iter iter;
 653         struct bkey_s_c old;
 654         struct bkey_i *k;
 655         enum btree_id btree;
 656         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 657         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 658         struct printbuf buf = PRINTBUF;
 659         int ret;
 660
 661         if (a->data_type != BCH_DATA_free &&
 662             a->data_type != BCH_DATA_need_discard)
 663                 return 0;
 664
 665         k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
 666         if (IS_ERR(k))
 667                 return PTR_ERR(k);
 668
 669         bkey_init(&k->k);
 670         k->k.type = new_type;
 671
 672         switch (a->data_type) {
 673         case BCH_DATA_free:
 674                 btree = BTREE_ID_freespace;
 675                 k->k.p = alloc_freespace_pos(alloc_k.k->p, *a);
 676                 bch2_key_resize(&k->k, 1);
 677                 break;
 678         case BCH_DATA_need_discard:
 679                 btree = BTREE_ID_need_discard;
 680                 k->k.p = alloc_k.k->p;
 681                 break;
 682         default:
 683                 return 0;
 684         }
 685
 686         old = bch2_bkey_get_iter(trans, &iter, btree,
 687                              bkey_start_pos(&k->k),
 688                              BTREE_ITER_INTENT);
 689         ret = bkey_err(old);
 690         if (ret)
 691                 return ret;
 692
 693         if (ca->mi.freespace_initialized &&
 694             c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info &&
 695             bch2_trans_inconsistent_on(old.k->type != old_type, trans,
 696                         "incorrect key when %s %s:%llu:%llu:0 (got %s should be %s)\n"
 697                         "  for %s",
 698                         set ? "setting" : "clearing",
 699                         bch2_btree_id_str(btree),
 700                         iter.pos.inode,
 701                         iter.pos.offset,
 702                         bch2_bkey_types[old.k->type],
 703                         bch2_bkey_types[old_type],
 704                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 705                 ret = -EIO;
 706                 goto err;
 707         }
 708
 709         ret = bch2_trans_update(trans, &iter, k, 0);
 710 err:
 711         bch2_trans_iter_exit(trans, &iter);
 712         printbuf_exit(&buf);
 713         return ret;
 714 }
 715
 716 static noinline int bch2_bucket_gen_update(struct btree_trans *trans,
 717                                            struct bpos bucket, u8 gen)
 718 {
 719         struct btree_iter iter;
 720         unsigned offset;
 721         struct bpos pos = alloc_gens_pos(bucket, &offset);
 722         struct bkey_i_bucket_gens *g;
 723         struct bkey_s_c k;
 724         int ret;
 725
 726         g = bch2_trans_kmalloc(trans, sizeof(*g));
 727         ret = PTR_ERR_OR_ZERO(g);
 728         if (ret)
 729                 return ret;
 730
 731         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_bucket_gens, pos,
 732                                BTREE_ITER_INTENT|
 733                                BTREE_ITER_WITH_UPDATES);
 734         ret = bkey_err(k);
 735         if (ret)
 736                 return ret;
 737
 738         if (k.k->type != KEY_TYPE_bucket_gens) {
 739                 bkey_bucket_gens_init(&g->k_i);
 740                 g->k.p = iter.pos;
 741         } else {
 742                 bkey_reassemble(&g->k_i, k);
 743         }
 744
 745         g->v.gens[offset] = gen;
 746
 747         ret = bch2_trans_update(trans, &iter, &g->k_i, 0);
 748         bch2_trans_iter_exit(trans, &iter);
 749         return ret;
 750 }
 751
 752 int bch2_trans_mark_alloc(struct btree_trans *trans,
 753                           enum btree_id btree_id, unsigned level,
 754                           struct bkey_s_c old, struct bkey_i *new,
 755                           unsigned flags)
 756 {
 757         struct bch_fs *c = trans->c;
 758         struct bch_alloc_v4 old_a_convert, *new_a;
 759         const struct bch_alloc_v4 *old_a;
 760         u64 old_lru, new_lru;
 761         int ret = 0;
 762
 763         /*
 764          * Deletion only happens in the device removal path, with
 765          * BTREE_TRIGGER_NORUN:
 766          */
 767         BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
 768
 769         old_a = bch2_alloc_to_v4(old, &old_a_convert);
 770         new_a = &bkey_i_to_alloc_v4(new)->v;
 771
 772         new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
 773
 774         if (bch2_bucket_sectors(*new_a) > bch2_bucket_sectors(*old_a)) {
 775                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 776                 new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 777                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
 778                 SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
 779         }
 780
 781         if (data_type_is_empty(new_a->data_type) &&
 782             BCH_ALLOC_V4_NEED_INC_GEN(new_a) &&
 783             !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
 784                 new_a->gen++;
 785                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
 786         }
 787
 788         if (old_a->data_type != new_a->data_type ||
 789             (new_a->data_type == BCH_DATA_free &&
 790              alloc_freespace_genbits(*old_a) != alloc_freespace_genbits(*new_a))) {
 791                 ret =   bch2_bucket_do_index(trans, old, old_a, false) ?:
 792                         bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_a, true);
 793                 if (ret)
 794                         return ret;
 795         }
 796
 797         if (new_a->data_type == BCH_DATA_cached &&
 798             !new_a->io_time[READ])
 799                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 800
 801         old_lru = alloc_lru_idx_read(*old_a);
 802         new_lru = alloc_lru_idx_read(*new_a);
 803
 804         if (old_lru != new_lru) {
 805                 ret = bch2_lru_change(trans, new->k.p.inode,
 806                                       bucket_to_u64(new->k.p),
 807                                       old_lru, new_lru);
 808                 if (ret)
 809                         return ret;
 810         }
 811
 812         new_a->fragmentation_lru = alloc_lru_idx_fragmentation(*new_a,
 813                                         bch_dev_bkey_exists(c, new->k.p.inode));
 814
 815         if (old_a->fragmentation_lru != new_a->fragmentation_lru) {
 816                 ret = bch2_lru_change(trans,
 817                                 BCH_LRU_FRAGMENTATION_START,
 818                                 bucket_to_u64(new->k.p),
 819                                 old_a->fragmentation_lru, new_a->fragmentation_lru);
 820                 if (ret)
 821                         return ret;
 822         }
 823
 824         if (old_a->gen != new_a->gen) {
 825                 ret = bch2_bucket_gen_update(trans, new->k.p, new_a->gen);
 826                 if (ret)
 827                         return ret;
 828         }
 829
 830         /*
 831          * need to know if we're getting called from the invalidate path or
 832          * not:
 833          */
 834
 835         if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 836             old_a->cached_sectors) {
 837                 ret = bch2_update_cached_sectors_list(trans, new->k.p.inode,
 838                                                       -((s64) old_a->cached_sectors));
 839                 if (ret)
 840                         return ret;
 841         }
 842
 843         return 0;
 844 }
 845
 846 /*
 847  * This synthesizes deleted extents for holes, similar to BTREE_ITER_SLOTS for
 848  * extents style btrees, but works on non-extents btrees:
 849  */
 850 static struct bkey_s_c bch2_get_key_or_hole(struct btree_iter *iter, struct bpos end, struct bkey *hole)
 851 {
 852         struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
 853
 854         if (bkey_err(k))
 855                 return k;
 856
 857         if (k.k->type) {
 858                 return k;
 859         } else {
 860                 struct btree_iter iter2;
 861                 struct bpos next;
 862
 863                 bch2_trans_copy_iter(&iter2, iter);
 864
 865                 struct btree_path *path = btree_iter_path(iter->trans, iter);
 866                 if (!bpos_eq(path->l[0].b->key.k.p, SPOS_MAX))
 867                         end = bkey_min(end, bpos_nosnap_successor(path->l[0].b->key.k.p));
 868
 869                 end = bkey_min(end, POS(iter->pos.inode, iter->pos.offset + U32_MAX - 1));
 870
 871                 /*
 872                  * btree node min/max is a closed interval, upto takes a half
 873                  * open interval:
 874                  */
 875                 k = bch2_btree_iter_peek_upto(&iter2, end);
 876                 next = iter2.pos;
 877                 bch2_trans_iter_exit(iter->trans, &iter2);
 878
 879                 BUG_ON(next.offset >= iter->pos.offset + U32_MAX);
 880
 881                 if (bkey_err(k))
 882                         return k;
 883
 884                 bkey_init(hole);
 885                 hole->p = iter->pos;
 886
 887                 bch2_key_resize(hole, next.offset - iter->pos.offset);
 888                 return (struct bkey_s_c) { hole, NULL };
 889         }
 890 }
 891
 892 static bool next_bucket(struct bch_fs *c, struct bpos *bucket)
 893 {
 894         struct bch_dev *ca;
 895
 896         if (bch2_dev_bucket_exists(c, *bucket))
 897                 return true;
 898
 899         if (bch2_dev_exists2(c, bucket->inode)) {
 900                 ca = bch_dev_bkey_exists(c, bucket->inode);
 901
 902                 if (bucket->offset < ca->mi.first_bucket) {
 903                         bucket->offset = ca->mi.first_bucket;
 904                         return true;
 905                 }
 906
 907                 bucket->inode++;
 908                 bucket->offset = 0;
 909         }
 910
 911         rcu_read_lock();
 912         ca = __bch2_next_dev_idx(c, bucket->inode, NULL);
 913         if (ca)
 914                 *bucket = POS(ca->dev_idx, ca->mi.first_bucket);
 915         rcu_read_unlock();
 916
 917         return ca != NULL;
 918 }
 919
 920 static struct bkey_s_c bch2_get_key_or_real_bucket_hole(struct btree_iter *iter, struct bkey *hole)
 921 {
 922         struct bch_fs *c = iter->trans->c;
 923         struct bkey_s_c k;
 924 again:
 925         k = bch2_get_key_or_hole(iter, POS_MAX, hole);
 926         if (bkey_err(k))
 927                 return k;
 928
 929         if (!k.k->type) {
 930                 struct bpos bucket = bkey_start_pos(k.k);
 931
 932                 if (!bch2_dev_bucket_exists(c, bucket)) {
 933                         if (!next_bucket(c, &bucket))
 934                                 return bkey_s_c_null;
 935
 936                         bch2_btree_iter_set_pos(iter, bucket);
 937                         goto again;
 938                 }
 939
 940                 if (!bch2_dev_bucket_exists(c, k.k->p)) {
 941                         struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
 942
 943                         bch2_key_resize(hole, ca->mi.nbuckets - bucket.offset);
 944                 }
 945         }
 946
 947         return k;
 948 }
 949
 950 static noinline_for_stack
 951 int bch2_check_alloc_key(struct btree_trans *trans,
 952                          struct bkey_s_c alloc_k,
 953                          struct btree_iter *alloc_iter,
 954                          struct btree_iter *discard_iter,
 955                          struct btree_iter *freespace_iter,
 956                          struct btree_iter *bucket_gens_iter)
 957 {
 958         struct bch_fs *c = trans->c;
 959         struct bch_dev *ca;
 960         struct bch_alloc_v4 a_convert;
 961         const struct bch_alloc_v4 *a;
 962         unsigned discard_key_type, freespace_key_type;
 963         unsigned gens_offset;
 964         struct bkey_s_c k;
 965         struct printbuf buf = PRINTBUF;
 966         int ret;
 967
 968         if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
 969                         alloc_key_to_missing_dev_bucket,
 970                         "alloc key for invalid device:bucket %llu:%llu",
 971                         alloc_k.k->p.inode, alloc_k.k->p.offset))
 972                 return bch2_btree_delete_at(trans, alloc_iter, 0);
 973
 974         ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 975         if (!ca->mi.freespace_initialized)
 976                 return 0;
 977
 978         a = bch2_alloc_to_v4(alloc_k, &a_convert);
 979
 980         discard_key_type = a->data_type == BCH_DATA_need_discard ? KEY_TYPE_set : 0;
 981         bch2_btree_iter_set_pos(discard_iter, alloc_k.k->p);
 982         k = bch2_btree_iter_peek_slot(discard_iter);
 983         ret = bkey_err(k);
 984         if (ret)
 985                 goto err;
 986
 987         if (k.k->type != discard_key_type &&
 988             (c->opts.reconstruct_alloc ||
 989              fsck_err(c, need_discard_key_wrong,
 990                       "incorrect key in need_discard btree (got %s should be %s)\n"
 991                       "  %s",
 992                       bch2_bkey_types[k.k->type],
 993                       bch2_bkey_types[discard_key_type],
 994                       (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
 995                 struct bkey_i *update =
 996                         bch2_trans_kmalloc(trans, sizeof(*update));
 997
 998                 ret = PTR_ERR_OR_ZERO(update);
 999                 if (ret)
1000                         goto err;
1001
1002                 bkey_init(&update->k);
1003                 update->k.type  = discard_key_type;
1004                 update->k.p     = discard_iter->pos;
1005
1006                 ret = bch2_trans_update(trans, discard_iter, update, 0);
1007                 if (ret)
1008                         goto err;
1009         }
1010
1011         freespace_key_type = a->data_type == BCH_DATA_free ? KEY_TYPE_set : 0;
1012         bch2_btree_iter_set_pos(freespace_iter, alloc_freespace_pos(alloc_k.k->p, *a));
1013         k = bch2_btree_iter_peek_slot(freespace_iter);
1014         ret = bkey_err(k);
1015         if (ret)
1016                 goto err;
1017
1018         if (k.k->type != freespace_key_type &&
1019             (c->opts.reconstruct_alloc ||
1020              fsck_err(c, freespace_key_wrong,
1021                       "incorrect key in freespace btree (got %s should be %s)\n"
1022                       "  %s",
1023                       bch2_bkey_types[k.k->type],
1024                       bch2_bkey_types[freespace_key_type],
1025                       (printbuf_reset(&buf),
1026                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1027                 struct bkey_i *update =
1028                         bch2_trans_kmalloc(trans, sizeof(*update));
1029
1030                 ret = PTR_ERR_OR_ZERO(update);
1031                 if (ret)
1032                         goto err;
1033
1034                 bkey_init(&update->k);
1035                 update->k.type  = freespace_key_type;
1036                 update->k.p     = freespace_iter->pos;
1037                 bch2_key_resize(&update->k, 1);
1038
1039                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1040                 if (ret)
1041                         goto err;
1042         }
1043
1044         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(alloc_k.k->p, &gens_offset));
1045         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1046         ret = bkey_err(k);
1047         if (ret)
1048                 goto err;
1049
1050         if (a->gen != alloc_gen(k, gens_offset) &&
1051             (c->opts.reconstruct_alloc ||
1052              fsck_err(c, bucket_gens_key_wrong,
1053                       "incorrect gen in bucket_gens btree (got %u should be %u)\n"
1054                       "  %s",
1055                       alloc_gen(k, gens_offset), a->gen,
1056                       (printbuf_reset(&buf),
1057                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1058                 struct bkey_i_bucket_gens *g =
1059                         bch2_trans_kmalloc(trans, sizeof(*g));
1060
1061                 ret = PTR_ERR_OR_ZERO(g);
1062                 if (ret)
1063                         goto err;
1064
1065                 if (k.k->type == KEY_TYPE_bucket_gens) {
1066                         bkey_reassemble(&g->k_i, k);
1067                 } else {
1068                         bkey_bucket_gens_init(&g->k_i);
1069                         g->k.p = alloc_gens_pos(alloc_k.k->p, &gens_offset);
1070                 }
1071
1072                 g->v.gens[gens_offset] = a->gen;
1073
1074                 ret = bch2_trans_update(trans, bucket_gens_iter, &g->k_i, 0);
1075                 if (ret)
1076                         goto err;
1077         }
1078 err:
1079 fsck_err:
1080         printbuf_exit(&buf);
1081         return ret;
1082 }
1083
1084 static noinline_for_stack
1085 int bch2_check_alloc_hole_freespace(struct btree_trans *trans,
1086                                     struct bpos start,
1087                                     struct bpos *end,
1088                                     struct btree_iter *freespace_iter)
1089 {
1090         struct bch_fs *c = trans->c;
1091         struct bch_dev *ca;
1092         struct bkey_s_c k;
1093         struct printbuf buf = PRINTBUF;
1094         int ret;
1095
1096         ca = bch_dev_bkey_exists(c, start.inode);
1097         if (!ca->mi.freespace_initialized)
1098                 return 0;
1099
1100         bch2_btree_iter_set_pos(freespace_iter, start);
1101
1102         k = bch2_btree_iter_peek_slot(freespace_iter);
1103         ret = bkey_err(k);
1104         if (ret)
1105                 goto err;
1106
1107         *end = bkey_min(k.k->p, *end);
1108
1109         if (k.k->type != KEY_TYPE_set &&
1110             (c->opts.reconstruct_alloc ||
1111              fsck_err(c, freespace_hole_missing,
1112                       "hole in alloc btree missing in freespace btree\n"
1113                       "  device %llu buckets %llu-%llu",
1114                       freespace_iter->pos.inode,
1115                       freespace_iter->pos.offset,
1116                       end->offset))) {
1117                 struct bkey_i *update =
1118                         bch2_trans_kmalloc(trans, sizeof(*update));
1119
1120                 ret = PTR_ERR_OR_ZERO(update);
1121                 if (ret)
1122                         goto err;
1123
1124                 bkey_init(&update->k);
1125                 update->k.type  = KEY_TYPE_set;
1126                 update->k.p     = freespace_iter->pos;
1127                 bch2_key_resize(&update->k,
1128                                 min_t(u64, U32_MAX, end->offset -
1129                                       freespace_iter->pos.offset));
1130
1131                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1132                 if (ret)
1133                         goto err;
1134         }
1135 err:
1136 fsck_err:
1137         printbuf_exit(&buf);
1138         return ret;
1139 }
1140
1141 static noinline_for_stack
1142 int bch2_check_alloc_hole_bucket_gens(struct btree_trans *trans,
1143                                       struct bpos start,
1144                                       struct bpos *end,
1145                                       struct btree_iter *bucket_gens_iter)
1146 {
1147         struct bch_fs *c = trans->c;
1148         struct bkey_s_c k;
1149         struct printbuf buf = PRINTBUF;
1150         unsigned i, gens_offset, gens_end_offset;
1151         int ret;
1152
1153         if (c->sb.version < bcachefs_metadata_version_bucket_gens)
1154                 return 0;
1155
1156         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(start, &gens_offset));
1157
1158         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1159         ret = bkey_err(k);
1160         if (ret)
1161                 goto err;
1162
1163         if (bkey_cmp(alloc_gens_pos(start, &gens_offset),
1164                      alloc_gens_pos(*end,  &gens_end_offset)))
1165                 gens_end_offset = KEY_TYPE_BUCKET_GENS_NR;
1166
1167         if (k.k->type == KEY_TYPE_bucket_gens) {
1168                 struct bkey_i_bucket_gens g;
1169                 bool need_update = false;
1170
1171                 bkey_reassemble(&g.k_i, k);
1172
1173                 for (i = gens_offset; i < gens_end_offset; i++) {
1174                         if (fsck_err_on(g.v.gens[i], c,
1175                                         bucket_gens_hole_wrong,
1176                                         "hole in alloc btree at %llu:%llu with nonzero gen in bucket_gens btree (%u)",
1177                                         bucket_gens_pos_to_alloc(k.k->p, i).inode,
1178                                         bucket_gens_pos_to_alloc(k.k->p, i).offset,
1179                                         g.v.gens[i])) {
1180                                 g.v.gens[i] = 0;
1181                                 need_update = true;
1182                         }
1183                 }
1184
1185                 if (need_update) {
1186                         struct bkey_i *u = bch2_trans_kmalloc(trans, sizeof(g));
1187
1188                         ret = PTR_ERR_OR_ZERO(u);
1189                         if (ret)
1190                                 goto err;
1191
1192                         memcpy(u, &g, sizeof(g));
1193
1194                         ret = bch2_trans_update(trans, bucket_gens_iter, u, 0);
1195                         if (ret)
1196                                 goto err;
1197                 }
1198         }
1199
1200         *end = bkey_min(*end, bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0));
1201 err:
1202 fsck_err:
1203         printbuf_exit(&buf);
1204         return ret;
1205 }
1206
1207 static noinline_for_stack int bch2_check_discard_freespace_key(struct btree_trans *trans,
1208                                               struct btree_iter *iter)
1209 {
1210         struct bch_fs *c = trans->c;
1211         struct btree_iter alloc_iter;
1212         struct bkey_s_c alloc_k;
1213         struct bch_alloc_v4 a_convert;
1214         const struct bch_alloc_v4 *a;
1215         u64 genbits;
1216         struct bpos pos;
1217         enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
1218                 ? BCH_DATA_need_discard
1219                 : BCH_DATA_free;
1220         struct printbuf buf = PRINTBUF;
1221         int ret;
1222
1223         pos = iter->pos;
1224         pos.offset &= ~(~0ULL << 56);
1225         genbits = iter->pos.offset & (~0ULL << 56);
1226
1227         alloc_k = bch2_bkey_get_iter(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
1228         ret = bkey_err(alloc_k);
1229         if (ret)
1230                 return ret;
1231
1232         if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
1233                         need_discard_freespace_key_to_invalid_dev_bucket,
1234                         "entry in %s btree for nonexistant dev:bucket %llu:%llu",
1235                         bch2_btree_id_str(iter->btree_id), pos.inode, pos.offset))
1236                 goto delete;
1237
1238         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1239
1240         if (fsck_err_on(a->data_type != state ||
1241                         (state == BCH_DATA_free &&
1242                          genbits != alloc_freespace_genbits(*a)), c,
1243                         need_discard_freespace_key_bad,
1244                         "%s\n  incorrectly set at %s:%llu:%llu:0 (free %u, genbits %llu should be %llu)",
1245                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
1246                         bch2_btree_id_str(iter->btree_id),
1247                         iter->pos.inode,
1248                         iter->pos.offset,
1249                         a->data_type == state,
1250                         genbits >> 56, alloc_freespace_genbits(*a) >> 56))
1251                 goto delete;
1252 out:
1253 fsck_err:
1254         set_btree_iter_dontneed(&alloc_iter);
1255         bch2_trans_iter_exit(trans, &alloc_iter);
1256         printbuf_exit(&buf);
1257         return ret;
1258 delete:
1259         ret =   bch2_btree_delete_extent_at(trans, iter,
1260                         iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0) ?:
1261                 bch2_trans_commit(trans, NULL, NULL,
1262                         BCH_TRANS_COMMIT_no_enospc);
1263         goto out;
1264 }
1265
1266 /*
1267  * We've already checked that generation numbers in the bucket_gens btree are
1268  * valid for buckets that exist; this just checks for keys for nonexistent
1269  * buckets.
1270  */
1271 static noinline_for_stack
1272 int bch2_check_bucket_gens_key(struct btree_trans *trans,
1273                                struct btree_iter *iter,
1274                                struct bkey_s_c k)
1275 {
1276         struct bch_fs *c = trans->c;
1277         struct bkey_i_bucket_gens g;
1278         struct bch_dev *ca;
1279         u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
1280         u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
1281         u64 b;
1282         bool need_update = false, dev_exists;
1283         struct printbuf buf = PRINTBUF;
1284         int ret = 0;
1285
1286         BUG_ON(k.k->type != KEY_TYPE_bucket_gens);
1287         bkey_reassemble(&g.k_i, k);
1288
1289         /* if no bch_dev, skip out whether we repair or not */
1290         dev_exists = bch2_dev_exists2(c, k.k->p.inode);
1291         if (!dev_exists) {
1292                 if (fsck_err_on(!dev_exists, c,
1293                                 bucket_gens_to_invalid_dev,
1294                                 "bucket_gens key for invalid device:\n  %s",
1295                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1296                         ret = bch2_btree_delete_at(trans, iter, 0);
1297                 }
1298                 goto out;
1299         }
1300
1301         ca = bch_dev_bkey_exists(c, k.k->p.inode);
1302         if (fsck_err_on(end <= ca->mi.first_bucket ||
1303                         start >= ca->mi.nbuckets, c,
1304                         bucket_gens_to_invalid_buckets,
1305                         "bucket_gens key for invalid buckets:\n  %s",
1306                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1307                 ret = bch2_btree_delete_at(trans, iter, 0);
1308                 goto out;
1309         }
1310
1311         for (b = start; b < ca->mi.first_bucket; b++)
1312                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1313                                 bucket_gens_nonzero_for_invalid_buckets,
1314                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1315                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1316                         need_update = true;
1317                 }
1318
1319         for (b = ca->mi.nbuckets; b < end; b++)
1320                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1321                                 bucket_gens_nonzero_for_invalid_buckets,
1322                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1323                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1324                         need_update = true;
1325                 }
1326
1327         if (need_update) {
1328                 struct bkey_i *u = bch2_trans_kmalloc(trans, sizeof(g));
1329
1330                 ret = PTR_ERR_OR_ZERO(u);
1331                 if (ret)
1332                         goto out;
1333
1334                 memcpy(u, &g, sizeof(g));
1335                 ret = bch2_trans_update(trans, iter, u, 0);
1336         }
1337 out:
1338 fsck_err:
1339         printbuf_exit(&buf);
1340         return ret;
1341 }
1342
1343 int bch2_check_alloc_info(struct bch_fs *c)
1344 {
1345         struct btree_trans *trans = bch2_trans_get(c);
1346         struct btree_iter iter, discard_iter, freespace_iter, bucket_gens_iter;
1347         struct bkey hole;
1348         struct bkey_s_c k;
1349         int ret = 0;
1350
1351         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS_MIN,
1352                              BTREE_ITER_PREFETCH);
1353         bch2_trans_iter_init(trans, &discard_iter, BTREE_ID_need_discard, POS_MIN,
1354                              BTREE_ITER_PREFETCH);
1355         bch2_trans_iter_init(trans, &freespace_iter, BTREE_ID_freespace, POS_MIN,
1356                              BTREE_ITER_PREFETCH);
1357         bch2_trans_iter_init(trans, &bucket_gens_iter, BTREE_ID_bucket_gens, POS_MIN,
1358                              BTREE_ITER_PREFETCH);
1359
1360         while (1) {
1361                 struct bpos next;
1362
1363                 bch2_trans_begin(trans);
1364
1365                 k = bch2_get_key_or_real_bucket_hole(&iter, &hole);
1366                 ret = bkey_err(k);
1367                 if (ret)
1368                         goto bkey_err;
1369
1370                 if (!k.k)
1371                         break;
1372
1373                 if (k.k->type) {
1374                         next = bpos_nosnap_successor(k.k->p);
1375
1376                         ret = bch2_check_alloc_key(trans,
1377                                                    k, &iter,
1378                                                    &discard_iter,
1379                                                    &freespace_iter,
1380                                                    &bucket_gens_iter);
1381                         if (ret)
1382                                 goto bkey_err;
1383                 } else {
1384                         next = k.k->p;
1385
1386                         ret = bch2_check_alloc_hole_freespace(trans,
1387                                                     bkey_start_pos(k.k),
1388                                                     &next,
1389                                                     &freespace_iter) ?:
1390                                 bch2_check_alloc_hole_bucket_gens(trans,
1391                                                     bkey_start_pos(k.k),
1392                                                     &next,
1393                                                     &bucket_gens_iter);
1394                         if (ret)
1395                                 goto bkey_err;
1396                 }
1397
1398                 ret = bch2_trans_commit(trans, NULL, NULL,
1399                                         BCH_TRANS_COMMIT_no_enospc);
1400                 if (ret)
1401                         goto bkey_err;
1402
1403                 bch2_btree_iter_set_pos(&iter, next);
1404 bkey_err:
1405                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1406                         continue;
1407                 if (ret)
1408                         break;
1409         }
1410         bch2_trans_iter_exit(trans, &bucket_gens_iter);
1411         bch2_trans_iter_exit(trans, &freespace_iter);
1412         bch2_trans_iter_exit(trans, &discard_iter);
1413         bch2_trans_iter_exit(trans, &iter);
1414
1415         if (ret < 0)
1416                 goto err;
1417
1418         ret = for_each_btree_key(trans, iter,
1419                         BTREE_ID_need_discard, POS_MIN,
1420                         BTREE_ITER_PREFETCH, k,
1421                 bch2_check_discard_freespace_key(trans, &iter));
1422         if (ret)
1423                 goto err;
1424
1425         bch2_trans_iter_init(trans, &iter, BTREE_ID_freespace, POS_MIN,
1426                              BTREE_ITER_PREFETCH);
1427         while (1) {
1428                 bch2_trans_begin(trans);
1429                 k = bch2_btree_iter_peek(&iter);
1430                 if (!k.k)
1431                         break;
1432
1433                 ret = bkey_err(k) ?:
1434                         bch2_check_discard_freespace_key(trans, &iter);
1435                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
1436                         ret = 0;
1437                         continue;
1438                 }
1439                 if (ret) {
1440                         struct printbuf buf = PRINTBUF;
1441                         bch2_bkey_val_to_text(&buf, c, k);
1442
1443                         bch_err(c, "while checking %s", buf.buf);
1444                         printbuf_exit(&buf);
1445                         break;
1446                 }
1447
1448                 bch2_btree_iter_set_pos(&iter, bpos_nosnap_successor(iter.pos));
1449         }
1450         bch2_trans_iter_exit(trans, &iter);
1451         if (ret)
1452                 goto err;
1453
1454         ret = for_each_btree_key_commit(trans, iter,
1455                         BTREE_ID_bucket_gens, POS_MIN,
1456                         BTREE_ITER_PREFETCH, k,
1457                         NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1458                 bch2_check_bucket_gens_key(trans, &iter, k));
1459 err:
1460         bch2_trans_put(trans);
1461         bch_err_fn(c, ret);
1462         return ret;
1463 }
1464
1465 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
1466                                        struct btree_iter *alloc_iter)
1467 {
1468         struct bch_fs *c = trans->c;
1469         struct btree_iter lru_iter;
1470         struct bch_alloc_v4 a_convert;
1471         const struct bch_alloc_v4 *a;
1472         struct bkey_s_c alloc_k, lru_k;
1473         struct printbuf buf = PRINTBUF;
1474         int ret;
1475
1476         alloc_k = bch2_btree_iter_peek(alloc_iter);
1477         if (!alloc_k.k)
1478                 return 0;
1479
1480         ret = bkey_err(alloc_k);
1481         if (ret)
1482                 return ret;
1483
1484         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1485
1486         if (a->data_type != BCH_DATA_cached)
1487                 return 0;
1488
1489         if (fsck_err_on(!a->io_time[READ], c,
1490                         alloc_key_cached_but_read_time_zero,
1491                         "cached bucket with read_time 0\n"
1492                         "  %s",
1493                 (printbuf_reset(&buf),
1494                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
1495                 struct bkey_i_alloc_v4 *a_mut =
1496                         bch2_alloc_to_v4_mut(trans, alloc_k);
1497                 ret = PTR_ERR_OR_ZERO(a_mut);
1498                 if (ret)
1499                         goto err;
1500
1501                 a_mut->v.io_time[READ] = atomic64_read(&c->io_clock[READ].now);
1502                 ret = bch2_trans_update(trans, alloc_iter,
1503                                         &a_mut->k_i, BTREE_TRIGGER_NORUN);
1504                 if (ret)
1505                         goto err;
1506
1507                 a = &a_mut->v;
1508         }
1509
1510         lru_k = bch2_bkey_get_iter(trans, &lru_iter, BTREE_ID_lru,
1511                              lru_pos(alloc_k.k->p.inode,
1512                                      bucket_to_u64(alloc_k.k->p),
1513                                      a->io_time[READ]), 0);
1514         ret = bkey_err(lru_k);
1515         if (ret)
1516                 return ret;
1517
1518         if (fsck_err_on(lru_k.k->type != KEY_TYPE_set, c,
1519                         alloc_key_to_missing_lru_entry,
1520                         "missing lru entry\n"
1521                         "  %s",
1522                         (printbuf_reset(&buf),
1523                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
1524                 ret = bch2_lru_set(trans,
1525                                    alloc_k.k->p.inode,
1526                                    bucket_to_u64(alloc_k.k->p),
1527                                    a->io_time[READ]);
1528                 if (ret)
1529                         goto err;
1530         }
1531 err:
1532 fsck_err:
1533         bch2_trans_iter_exit(trans, &lru_iter);
1534         printbuf_exit(&buf);
1535         return ret;
1536 }
1537
1538 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
1539 {
1540         int ret = bch2_trans_run(c,
1541                 for_each_btree_key_commit(trans, iter, BTREE_ID_alloc,
1542                                 POS_MIN, BTREE_ITER_PREFETCH, k,
1543                                 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1544                         bch2_check_alloc_to_lru_ref(trans, &iter)));
1545         bch_err_fn(c, ret);
1546         return ret;
1547 }
1548
1549 static int bch2_discard_one_bucket(struct btree_trans *trans,
1550                                    struct btree_iter *need_discard_iter,
1551                                    struct bpos *discard_pos_done,
1552                                    u64 *seen,
1553                                    u64 *open,
1554                                    u64 *need_journal_commit,
1555                                    u64 *discarded)
1556 {
1557         struct bch_fs *c = trans->c;
1558         struct bpos pos = need_discard_iter->pos;
1559         struct btree_iter iter = { NULL };
1560         struct bkey_s_c k;
1561         struct bch_dev *ca;
1562         struct bkey_i_alloc_v4 *a;
1563         struct printbuf buf = PRINTBUF;
1564         int ret = 0;
1565
1566         ca = bch_dev_bkey_exists(c, pos.inode);
1567         if (!percpu_ref_tryget(&ca->io_ref)) {
1568                 bch2_btree_iter_set_pos(need_discard_iter, POS(pos.inode + 1, 0));
1569                 return 0;
1570         }
1571
1572         if (bch2_bucket_is_open_safe(c, pos.inode, pos.offset)) {
1573                 (*open)++;
1574                 goto out;
1575         }
1576
1577         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
1578                         c->journal.flushed_seq_ondisk,
1579                         pos.inode, pos.offset)) {
1580                 (*need_journal_commit)++;
1581                 goto out;
1582         }
1583
1584         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_alloc,
1585                                need_discard_iter->pos,
1586                                BTREE_ITER_CACHED);
1587         ret = bkey_err(k);
1588         if (ret)
1589                 goto out;
1590
1591         a = bch2_alloc_to_v4_mut(trans, k);
1592         ret = PTR_ERR_OR_ZERO(a);
1593         if (ret)
1594                 goto out;
1595
1596         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
1597                 a->v.gen++;
1598                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1599                 goto write;
1600         }
1601
1602         if (a->v.journal_seq > c->journal.flushed_seq_ondisk) {
1603                 if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
1604                         bch2_trans_inconsistent(trans,
1605                                 "clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
1606                                 "%s",
1607                                 a->v.journal_seq,
1608                                 c->journal.flushed_seq_ondisk,
1609                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1610                         ret = -EIO;
1611                 }
1612                 goto out;
1613         }
1614
1615         if (a->v.data_type != BCH_DATA_need_discard) {
1616                 if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
1617                         bch2_trans_inconsistent(trans,
1618                                 "bucket incorrectly set in need_discard btree\n"
1619                                 "%s",
1620                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1621                         ret = -EIO;
1622                 }
1623
1624                 goto out;
1625         }
1626
1627         if (!bkey_eq(*discard_pos_done, iter.pos) &&
1628             ca->mi.discard && !c->opts.nochanges) {
1629                 /*
1630                  * This works without any other locks because this is the only
1631                  * thread that removes items from the need_discard tree
1632                  */
1633                 bch2_trans_unlock(trans);
1634                 blkdev_issue_discard(ca->disk_sb.bdev,
1635                                      k.k->p.offset * ca->mi.bucket_size,
1636                                      ca->mi.bucket_size,
1637                                      GFP_KERNEL);
1638                 *discard_pos_done = iter.pos;
1639
1640                 ret = bch2_trans_relock_notrace(trans);
1641                 if (ret)
1642                         goto out;
1643         }
1644
1645         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
1646         a->v.data_type = alloc_data_type(a->v, a->v.data_type);
1647 write:
1648         ret =   bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1649                 bch2_trans_commit(trans, NULL, NULL,
1650                                   BCH_WATERMARK_btree|
1651                                   BCH_TRANS_COMMIT_no_enospc);
1652         if (ret)
1653                 goto out;
1654
1655         count_event(c, bucket_discard);
1656         (*discarded)++;
1657 out:
1658         (*seen)++;
1659         bch2_trans_iter_exit(trans, &iter);
1660         percpu_ref_put(&ca->io_ref);
1661         printbuf_exit(&buf);
1662         return ret;
1663 }
1664
1665 static void bch2_do_discards_work(struct work_struct *work)
1666 {
1667         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
1668         u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
1669         struct bpos discard_pos_done = POS_MAX;
1670         int ret;
1671
1672         /*
1673          * We're doing the commit in bch2_discard_one_bucket instead of using
1674          * for_each_btree_key_commit() so that we can increment counters after
1675          * successful commit:
1676          */
1677         ret = bch2_trans_run(c,
1678                 for_each_btree_key(trans, iter,
1679                                    BTREE_ID_need_discard, POS_MIN, 0, k,
1680                         bch2_discard_one_bucket(trans, &iter, &discard_pos_done,
1681                                                 &seen,
1682                                                 &open,
1683                                                 &need_journal_commit,
1684                                                 &discarded)));
1685
1686         if (need_journal_commit * 2 > seen)
1687                 bch2_journal_flush_async(&c->journal, NULL);
1688
1689         bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1690
1691         trace_discard_buckets(c, seen, open, need_journal_commit, discarded,
1692                               bch2_err_str(ret));
1693 }
1694
1695 void bch2_do_discards(struct bch_fs *c)
1696 {
1697         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_discard) &&
1698             !queue_work(c->write_ref_wq, &c->discard_work))
1699                 bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1700 }
1701
1702 static int invalidate_one_bucket(struct btree_trans *trans,
1703                                  struct btree_iter *lru_iter,
1704                                  struct bkey_s_c lru_k,
1705                                  s64 *nr_to_invalidate)
1706 {
1707         struct bch_fs *c = trans->c;
1708         struct btree_iter alloc_iter = { NULL };
1709         struct bkey_i_alloc_v4 *a = NULL;
1710         struct printbuf buf = PRINTBUF;
1711         struct bpos bucket = u64_to_bucket(lru_k.k->p.offset);
1712         unsigned cached_sectors;
1713         int ret = 0;
1714
1715         if (*nr_to_invalidate <= 0)
1716                 return 1;
1717
1718         if (!bch2_dev_bucket_exists(c, bucket)) {
1719                 prt_str(&buf, "lru entry points to invalid bucket");
1720                 goto err;
1721         }
1722
1723         if (bch2_bucket_is_open_safe(c, bucket.inode, bucket.offset))
1724                 return 0;
1725
1726         a = bch2_trans_start_alloc_update(trans, &alloc_iter, bucket);
1727         ret = PTR_ERR_OR_ZERO(a);
1728         if (ret)
1729                 goto out;
1730
1731         /* We expect harmless races here due to the btree write buffer: */
1732         if (lru_pos_time(lru_iter->pos) != alloc_lru_idx_read(a->v))
1733                 goto out;
1734
1735         BUG_ON(a->v.data_type != BCH_DATA_cached);
1736
1737         if (!a->v.cached_sectors)
1738                 bch_err(c, "invalidating empty bucket, confused");
1739
1740         cached_sectors = a->v.cached_sectors;
1741
1742         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1743         a->v.gen++;
1744         a->v.data_type          = 0;
1745         a->v.dirty_sectors      = 0;
1746         a->v.cached_sectors     = 0;
1747         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1748         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1749
1750         ret =   bch2_trans_update(trans, &alloc_iter, &a->k_i,
1751                                 BTREE_TRIGGER_BUCKET_INVALIDATE) ?:
1752                 bch2_trans_commit(trans, NULL, NULL,
1753                                   BCH_WATERMARK_btree|
1754                                   BCH_TRANS_COMMIT_no_enospc);
1755         if (ret)
1756                 goto out;
1757
1758         trace_and_count(c, bucket_invalidate, c, bucket.inode, bucket.offset, cached_sectors);
1759         --*nr_to_invalidate;
1760 out:
1761         bch2_trans_iter_exit(trans, &alloc_iter);
1762         printbuf_exit(&buf);
1763         return ret;
1764 err:
1765         prt_str(&buf, "\n  lru key: ");
1766         bch2_bkey_val_to_text(&buf, c, lru_k);
1767
1768         prt_str(&buf, "\n  lru entry: ");
1769         bch2_lru_pos_to_text(&buf, lru_iter->pos);
1770
1771         prt_str(&buf, "\n  alloc key: ");
1772         if (!a)
1773                 bch2_bpos_to_text(&buf, bucket);
1774         else
1775                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
1776
1777         bch_err(c, "%s", buf.buf);
1778         if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_lrus) {
1779                 bch2_inconsistent_error(c);
1780                 ret = -EINVAL;
1781         }
1782
1783         goto out;
1784 }
1785
1786 static void bch2_do_invalidates_work(struct work_struct *work)
1787 {
1788         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1789         struct btree_trans *trans = bch2_trans_get(c);
1790         int ret = 0;
1791
1792         ret = bch2_btree_write_buffer_tryflush(trans);
1793         if (ret)
1794                 goto err;
1795
1796         for_each_member_device(c, ca) {
1797                 s64 nr_to_invalidate =
1798                         should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
1799
1800                 ret = for_each_btree_key_upto(trans, iter, BTREE_ID_lru,
1801                                 lru_pos(ca->dev_idx, 0, 0),
1802                                 lru_pos(ca->dev_idx, U64_MAX, LRU_TIME_MAX),
1803                                 BTREE_ITER_INTENT, k,
1804                         invalidate_one_bucket(trans, &iter, k, &nr_to_invalidate));
1805
1806                 if (ret < 0) {
1807                         percpu_ref_put(&ca->ref);
1808                         break;
1809                 }
1810         }
1811 err:
1812         bch2_trans_put(trans);
1813         bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1814 }
1815
1816 void bch2_do_invalidates(struct bch_fs *c)
1817 {
1818         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_invalidate) &&
1819             !queue_work(c->write_ref_wq, &c->invalidate_work))
1820                 bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1821 }
1822
1823 int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca,
1824                             u64 bucket_start, u64 bucket_end)
1825 {
1826         struct btree_trans *trans = bch2_trans_get(c);
1827         struct btree_iter iter;
1828         struct bkey_s_c k;
1829         struct bkey hole;
1830         struct bpos end = POS(ca->dev_idx, bucket_end);
1831         struct bch_member *m;
1832         unsigned long last_updated = jiffies;
1833         int ret;
1834
1835         BUG_ON(bucket_start > bucket_end);
1836         BUG_ON(bucket_end > ca->mi.nbuckets);
1837
1838         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
1839                 POS(ca->dev_idx, max_t(u64, ca->mi.first_bucket, bucket_start)),
1840                 BTREE_ITER_PREFETCH);
1841         /*
1842          * Scan the alloc btree for every bucket on @ca, and add buckets to the
1843          * freespace/need_discard/need_gc_gens btrees as needed:
1844          */
1845         while (1) {
1846                 if (last_updated + HZ * 10 < jiffies) {
1847                         bch_info(ca, "%s: currently at %llu/%llu",
1848                                  __func__, iter.pos.offset, ca->mi.nbuckets);
1849                         last_updated = jiffies;
1850                 }
1851
1852                 bch2_trans_begin(trans);
1853
1854                 if (bkey_ge(iter.pos, end)) {
1855                         ret = 0;
1856                         break;
1857                 }
1858
1859                 k = bch2_get_key_or_hole(&iter, end, &hole);
1860                 ret = bkey_err(k);
1861                 if (ret)
1862                         goto bkey_err;
1863
1864                 if (k.k->type) {
1865                         /*
1866                          * We process live keys in the alloc btree one at a
1867                          * time:
1868                          */
1869                         struct bch_alloc_v4 a_convert;
1870                         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1871
1872                         ret =   bch2_bucket_do_index(trans, k, a, true) ?:
1873                                 bch2_trans_commit(trans, NULL, NULL,
1874                                                   BCH_TRANS_COMMIT_no_enospc);
1875                         if (ret)
1876                                 goto bkey_err;
1877
1878                         bch2_btree_iter_advance(&iter);
1879                 } else {
1880                         struct bkey_i *freespace;
1881
1882                         freespace = bch2_trans_kmalloc(trans, sizeof(*freespace));
1883                         ret = PTR_ERR_OR_ZERO(freespace);
1884                         if (ret)
1885                                 goto bkey_err;
1886
1887                         bkey_init(&freespace->k);
1888                         freespace->k.type       = KEY_TYPE_set;
1889                         freespace->k.p          = k.k->p;
1890                         freespace->k.size       = k.k->size;
1891
1892                         ret = bch2_btree_insert_trans(trans, BTREE_ID_freespace, freespace, 0) ?:
1893                                 bch2_trans_commit(trans, NULL, NULL,
1894                                                   BCH_TRANS_COMMIT_no_enospc);
1895                         if (ret)
1896                                 goto bkey_err;
1897
1898                         bch2_btree_iter_set_pos(&iter, k.k->p);
1899                 }
1900 bkey_err:
1901                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1902                         continue;
1903                 if (ret)
1904                         break;
1905         }
1906
1907         bch2_trans_iter_exit(trans, &iter);
1908         bch2_trans_put(trans);
1909
1910         if (ret < 0) {
1911                 bch_err_msg(ca, ret, "initializing free space");
1912                 return ret;
1913         }
1914
1915         mutex_lock(&c->sb_lock);
1916         m = bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx);
1917         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1918         mutex_unlock(&c->sb_lock);
1919
1920         return 0;
1921 }
1922
1923 int bch2_fs_freespace_init(struct bch_fs *c)
1924 {
1925         int ret = 0;
1926         bool doing_init = false;
1927
1928         /*
1929          * We can crash during the device add path, so we need to check this on
1930          * every mount:
1931          */
1932
1933         for_each_member_device(c, ca) {
1934                 if (ca->mi.freespace_initialized)
1935                         continue;
1936
1937                 if (!doing_init) {
1938                         bch_info(c, "initializing freespace");
1939                         doing_init = true;
1940                 }
1941
1942                 ret = bch2_dev_freespace_init(c, ca, 0, ca->mi.nbuckets);
1943                 if (ret) {
1944                         percpu_ref_put(&ca->ref);
1945                         bch_err_fn(c, ret);
1946                         return ret;
1947                 }
1948         }
1949
1950         if (doing_init) {
1951                 mutex_lock(&c->sb_lock);
1952                 bch2_write_super(c);
1953                 mutex_unlock(&c->sb_lock);
1954                 bch_verbose(c, "done initializing freespace");
1955         }
1956
1957         return 0;
1958 }
1959
1960 /* Bucket IO clocks: */
1961
1962 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
1963                               size_t bucket_nr, int rw)
1964 {
1965         struct bch_fs *c = trans->c;
1966         struct btree_iter iter;
1967         struct bkey_i_alloc_v4 *a;
1968         u64 now;
1969         int ret = 0;
1970
1971         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
1972         ret = PTR_ERR_OR_ZERO(a);
1973         if (ret)
1974                 return ret;
1975
1976         now = atomic64_read(&c->io_clock[rw].now);
1977         if (a->v.io_time[rw] == now)
1978                 goto out;
1979
1980         a->v.io_time[rw] = now;
1981
1982         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1983                 bch2_trans_commit(trans, NULL, NULL, 0);
1984 out:
1985         bch2_trans_iter_exit(trans, &iter);
1986         return ret;
1987 }
1988
1989 /* Startup/shutdown (ro/rw): */
1990
1991 void bch2_recalc_capacity(struct bch_fs *c)
1992 {
1993         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
1994         unsigned bucket_size_max = 0;
1995         unsigned long ra_pages = 0;
1996
1997         lockdep_assert_held(&c->state_lock);
1998
1999         for_each_online_member(c, ca) {
2000                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
2001
2002                 ra_pages += bdi->ra_pages;
2003         }
2004
2005         bch2_set_ra_pages(c, ra_pages);
2006
2007         for_each_rw_member(c, ca) {
2008                 u64 dev_reserve = 0;
2009
2010                 /*
2011                  * We need to reserve buckets (from the number
2012                  * of currently available buckets) against
2013                  * foreground writes so that mainly copygc can
2014                  * make forward progress.
2015                  *
2016                  * We need enough to refill the various reserves
2017                  * from scratch - copygc will use its entire
2018                  * reserve all at once, then run against when
2019                  * its reserve is refilled (from the formerly
2020                  * available buckets).
2021                  *
2022                  * This reserve is just used when considering if
2023                  * allocations for foreground writes must wait -
2024                  * not -ENOSPC calculations.
2025                  */
2026
2027                 dev_reserve += ca->nr_btree_reserve * 2;
2028                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
2029
2030                 dev_reserve += 1;       /* btree write point */
2031                 dev_reserve += 1;       /* copygc write point */
2032                 dev_reserve += 1;       /* rebalance write point */
2033
2034                 dev_reserve *= ca->mi.bucket_size;
2035
2036                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
2037                                              ca->mi.first_bucket);
2038
2039                 reserved_sectors += dev_reserve * 2;
2040
2041                 bucket_size_max = max_t(unsigned, bucket_size_max,
2042                                         ca->mi.bucket_size);
2043         }
2044
2045         gc_reserve = c->opts.gc_reserve_bytes
2046                 ? c->opts.gc_reserve_bytes >> 9
2047                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
2048
2049         reserved_sectors = max(gc_reserve, reserved_sectors);
2050
2051         reserved_sectors = min(reserved_sectors, capacity);
2052
2053         c->capacity = capacity - reserved_sectors;
2054
2055         c->bucket_size_max = bucket_size_max;
2056
2057         /* Wake up case someone was waiting for buckets */
2058         closure_wake_up(&c->freelist_wait);
2059 }
2060
2061 u64 bch2_min_rw_member_capacity(struct bch_fs *c)
2062 {
2063         u64 ret = U64_MAX;
2064
2065         for_each_rw_member(c, ca)
2066                 ret = min(ret, ca->mi.nbuckets * ca->mi.bucket_size);
2067         return ret;
2068 }
2069
2070 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
2071 {
2072         struct open_bucket *ob;
2073         bool ret = false;
2074
2075         for (ob = c->open_buckets;
2076              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
2077              ob++) {
2078                 spin_lock(&ob->lock);
2079                 if (ob->valid && !ob->on_partial_list &&
2080                     ob->dev == ca->dev_idx)
2081                         ret = true;
2082                 spin_unlock(&ob->lock);
2083         }
2084
2085         return ret;
2086 }
2087
2088 /* device goes ro: */
2089 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
2090 {
2091         unsigned i;
2092
2093         /* First, remove device from allocation groups: */
2094
2095         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2096                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
2097
2098         /*
2099          * Capacity is calculated based off of devices in allocation groups:
2100          */
2101         bch2_recalc_capacity(c);
2102
2103         bch2_open_buckets_stop(c, ca, false);
2104
2105         /*
2106          * Wake up threads that were blocked on allocation, so they can notice
2107          * the device can no longer be removed and the capacity has changed:
2108          */
2109         closure_wake_up(&c->freelist_wait);
2110
2111         /*
2112          * journal_res_get() can block waiting for free space in the journal -
2113          * it needs to notice there may not be devices to allocate from anymore:
2114          */
2115         wake_up(&c->journal.wait);
2116
2117         /* Now wait for any in flight writes: */
2118
2119         closure_wait_event(&c->open_buckets_wait,
2120                            !bch2_dev_has_open_write_point(c, ca));
2121 }
2122
2123 /* device goes rw: */
2124 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
2125 {
2126         unsigned i;
2127
2128         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2129                 if (ca->mi.data_allowed & (1 << i))
2130                         set_bit(ca->dev_idx, c->rw_devs[i].d);
2131 }
2132
2133 void bch2_fs_allocator_background_init(struct bch_fs *c)
2134 {
2135         spin_lock_init(&c->freelist_lock);
2136         INIT_WORK(&c->discard_work, bch2_do_discards_work);
2137         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
2138 }