git.sesse.net Git - bcachefs-tools-debian/blob - c_src/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_trigger_alloc(struct btree_trans *trans,
 753                        enum btree_id btree, unsigned level,
 754                        struct bkey_s_c old, struct bkey_s new,
 755                        unsigned flags)
 756 {
 757         struct bch_fs *c = trans->c;
 758         int ret = 0;
 759
 760         if (bch2_trans_inconsistent_on(!bch2_dev_bucket_exists(c, new.k->p), trans,
 761                                        "alloc key for invalid device or bucket"))
 762                 return -EIO;
 763
 764         struct bch_dev *ca = bch_dev_bkey_exists(c, new.k->p.inode);
 765
 766         struct bch_alloc_v4 old_a_convert;
 767         const struct bch_alloc_v4 *old_a = bch2_alloc_to_v4(old, &old_a_convert);
 768
 769         if (flags & BTREE_TRIGGER_TRANSACTIONAL) {
 770                 struct bch_alloc_v4 *new_a = bkey_s_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, new.s_c, 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                 u64 old_lru = alloc_lru_idx_read(*old_a);
 802                 u64 new_lru = alloc_lru_idx_read(*new_a);
 803                 if (old_lru != new_lru) {
 804                         ret = bch2_lru_change(trans, new.k->p.inode,
 805                                               bucket_to_u64(new.k->p),
 806                                               old_lru, new_lru);
 807                         if (ret)
 808                                 return ret;
 809                 }
 810
 811                 new_a->fragmentation_lru = alloc_lru_idx_fragmentation(*new_a,
 812                                                 bch_dev_bkey_exists(c, new.k->p.inode));
 813                 if (old_a->fragmentation_lru != new_a->fragmentation_lru) {
 814                         ret = bch2_lru_change(trans,
 815                                         BCH_LRU_FRAGMENTATION_START,
 816                                         bucket_to_u64(new.k->p),
 817                                         old_a->fragmentation_lru, new_a->fragmentation_lru);
 818                         if (ret)
 819                                 return ret;
 820                 }
 821
 822                 if (old_a->gen != new_a->gen) {
 823                         ret = bch2_bucket_gen_update(trans, new.k->p, new_a->gen);
 824                         if (ret)
 825                                 return ret;
 826                 }
 827
 828                 /*
 829                  * need to know if we're getting called from the invalidate path or
 830                  * not:
 831                  */
 832
 833                 if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 834                     old_a->cached_sectors) {
 835                         ret = bch2_update_cached_sectors_list(trans, new.k->p.inode,
 836                                                               -((s64) old_a->cached_sectors));
 837                         if (ret)
 838                                 return ret;
 839                 }
 840         }
 841
 842         if (!(flags & BTREE_TRIGGER_TRANSACTIONAL) && (flags & BTREE_TRIGGER_INSERT)) {
 843                 struct bch_alloc_v4 *new_a = bkey_s_to_alloc_v4(new).v;
 844                 u64 journal_seq = trans->journal_res.seq;
 845                 u64 bucket_journal_seq = new_a->journal_seq;
 846
 847                 if ((flags & BTREE_TRIGGER_INSERT) &&
 848                     data_type_is_empty(old_a->data_type) !=
 849                     data_type_is_empty(new_a->data_type) &&
 850                     new.k->type == KEY_TYPE_alloc_v4) {
 851                         struct bch_alloc_v4 *v = bkey_s_to_alloc_v4(new).v;
 852
 853                         /*
 854                          * If the btree updates referring to a bucket weren't flushed
 855                          * before the bucket became empty again, then the we don't have
 856                          * to wait on a journal flush before we can reuse the bucket:
 857                          */
 858                         v->journal_seq = bucket_journal_seq =
 859                                 data_type_is_empty(new_a->data_type) &&
 860                                 (journal_seq == v->journal_seq ||
 861                                  bch2_journal_noflush_seq(&c->journal, v->journal_seq))
 862                                 ? 0 : journal_seq;
 863                 }
 864
 865                 if (!data_type_is_empty(old_a->data_type) &&
 866                     data_type_is_empty(new_a->data_type) &&
 867                     bucket_journal_seq) {
 868                         ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 869                                         c->journal.flushed_seq_ondisk,
 870                                         new.k->p.inode, new.k->p.offset,
 871                                         bucket_journal_seq);
 872                         if (ret) {
 873                                 bch2_fs_fatal_error(c,
 874                                         "error setting bucket_needs_journal_commit: %i", ret);
 875                                 return ret;
 876                         }
 877                 }
 878
 879                 percpu_down_read(&c->mark_lock);
 880                 if (new_a->gen != old_a->gen)
 881                         *bucket_gen(ca, new.k->p.offset) = new_a->gen;
 882
 883                 bch2_dev_usage_update(c, ca, old_a, new_a, journal_seq, false);
 884
 885                 if (new_a->data_type == BCH_DATA_free &&
 886                     (!new_a->journal_seq || new_a->journal_seq < c->journal.flushed_seq_ondisk))
 887                         closure_wake_up(&c->freelist_wait);
 888
 889                 if (new_a->data_type == BCH_DATA_need_discard &&
 890                     (!bucket_journal_seq || bucket_journal_seq < c->journal.flushed_seq_ondisk))
 891                         bch2_do_discards(c);
 892
 893                 if (old_a->data_type != BCH_DATA_cached &&
 894                     new_a->data_type == BCH_DATA_cached &&
 895                     should_invalidate_buckets(ca, bch2_dev_usage_read(ca)))
 896                         bch2_do_invalidates(c);
 897
 898                 if (new_a->data_type == BCH_DATA_need_gc_gens)
 899                         bch2_do_gc_gens(c);
 900                 percpu_up_read(&c->mark_lock);
 901         }
 902
 903         if ((flags & BTREE_TRIGGER_GC) &&
 904             (flags & BTREE_TRIGGER_BUCKET_INVALIDATE)) {
 905                 struct bch_alloc_v4 new_a_convert;
 906                 const struct bch_alloc_v4 *new_a = bch2_alloc_to_v4(new.s_c, &new_a_convert);
 907
 908                 percpu_down_read(&c->mark_lock);
 909                 struct bucket *g = gc_bucket(ca, new.k->p.offset);
 910
 911                 bucket_lock(g);
 912
 913                 g->gen_valid            = 1;
 914                 g->gen                  = new_a->gen;
 915                 g->data_type            = new_a->data_type;
 916                 g->stripe               = new_a->stripe;
 917                 g->stripe_redundancy    = new_a->stripe_redundancy;
 918                 g->dirty_sectors        = new_a->dirty_sectors;
 919                 g->cached_sectors       = new_a->cached_sectors;
 920
 921                 bucket_unlock(g);
 922                 percpu_up_read(&c->mark_lock);
 923         }
 924
 925         return 0;
 926 }
 927
 928 /*
 929  * This synthesizes deleted extents for holes, similar to BTREE_ITER_SLOTS for
 930  * extents style btrees, but works on non-extents btrees:
 931  */
 932 static struct bkey_s_c bch2_get_key_or_hole(struct btree_iter *iter, struct bpos end, struct bkey *hole)
 933 {
 934         struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
 935
 936         if (bkey_err(k))
 937                 return k;
 938
 939         if (k.k->type) {
 940                 return k;
 941         } else {
 942                 struct btree_iter iter2;
 943                 struct bpos next;
 944
 945                 bch2_trans_copy_iter(&iter2, iter);
 946
 947                 struct btree_path *path = btree_iter_path(iter->trans, iter);
 948                 if (!bpos_eq(path->l[0].b->key.k.p, SPOS_MAX))
 949                         end = bkey_min(end, bpos_nosnap_successor(path->l[0].b->key.k.p));
 950
 951                 end = bkey_min(end, POS(iter->pos.inode, iter->pos.offset + U32_MAX - 1));
 952
 953                 /*
 954                  * btree node min/max is a closed interval, upto takes a half
 955                  * open interval:
 956                  */
 957                 k = bch2_btree_iter_peek_upto(&iter2, end);
 958                 next = iter2.pos;
 959                 bch2_trans_iter_exit(iter->trans, &iter2);
 960
 961                 BUG_ON(next.offset >= iter->pos.offset + U32_MAX);
 962
 963                 if (bkey_err(k))
 964                         return k;
 965
 966                 bkey_init(hole);
 967                 hole->p = iter->pos;
 968
 969                 bch2_key_resize(hole, next.offset - iter->pos.offset);
 970                 return (struct bkey_s_c) { hole, NULL };
 971         }
 972 }
 973
 974 static bool next_bucket(struct bch_fs *c, struct bpos *bucket)
 975 {
 976         struct bch_dev *ca;
 977
 978         if (bch2_dev_bucket_exists(c, *bucket))
 979                 return true;
 980
 981         if (bch2_dev_exists2(c, bucket->inode)) {
 982                 ca = bch_dev_bkey_exists(c, bucket->inode);
 983
 984                 if (bucket->offset < ca->mi.first_bucket) {
 985                         bucket->offset = ca->mi.first_bucket;
 986                         return true;
 987                 }
 988
 989                 bucket->inode++;
 990                 bucket->offset = 0;
 991         }
 992
 993         rcu_read_lock();
 994         ca = __bch2_next_dev_idx(c, bucket->inode, NULL);
 995         if (ca)
 996                 *bucket = POS(ca->dev_idx, ca->mi.first_bucket);
 997         rcu_read_unlock();
 998
 999         return ca != NULL;
1000 }
1001
1002 static struct bkey_s_c bch2_get_key_or_real_bucket_hole(struct btree_iter *iter, struct bkey *hole)
1003 {
1004         struct bch_fs *c = iter->trans->c;
1005         struct bkey_s_c k;
1006 again:
1007         k = bch2_get_key_or_hole(iter, POS_MAX, hole);
1008         if (bkey_err(k))
1009                 return k;
1010
1011         if (!k.k->type) {
1012                 struct bpos bucket = bkey_start_pos(k.k);
1013
1014                 if (!bch2_dev_bucket_exists(c, bucket)) {
1015                         if (!next_bucket(c, &bucket))
1016                                 return bkey_s_c_null;
1017
1018                         bch2_btree_iter_set_pos(iter, bucket);
1019                         goto again;
1020                 }
1021
1022                 if (!bch2_dev_bucket_exists(c, k.k->p)) {
1023                         struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
1024
1025                         bch2_key_resize(hole, ca->mi.nbuckets - bucket.offset);
1026                 }
1027         }
1028
1029         return k;
1030 }
1031
1032 static noinline_for_stack
1033 int bch2_check_alloc_key(struct btree_trans *trans,
1034                          struct bkey_s_c alloc_k,
1035                          struct btree_iter *alloc_iter,
1036                          struct btree_iter *discard_iter,
1037                          struct btree_iter *freespace_iter,
1038                          struct btree_iter *bucket_gens_iter)
1039 {
1040         struct bch_fs *c = trans->c;
1041         struct bch_dev *ca;
1042         struct bch_alloc_v4 a_convert;
1043         const struct bch_alloc_v4 *a;
1044         unsigned discard_key_type, freespace_key_type;
1045         unsigned gens_offset;
1046         struct bkey_s_c k;
1047         struct printbuf buf = PRINTBUF;
1048         int ret;
1049
1050         if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
1051                         alloc_key_to_missing_dev_bucket,
1052                         "alloc key for invalid device:bucket %llu:%llu",
1053                         alloc_k.k->p.inode, alloc_k.k->p.offset))
1054                 return bch2_btree_delete_at(trans, alloc_iter, 0);
1055
1056         ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
1057         if (!ca->mi.freespace_initialized)
1058                 return 0;
1059
1060         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1061
1062         discard_key_type = a->data_type == BCH_DATA_need_discard ? KEY_TYPE_set : 0;
1063         bch2_btree_iter_set_pos(discard_iter, alloc_k.k->p);
1064         k = bch2_btree_iter_peek_slot(discard_iter);
1065         ret = bkey_err(k);
1066         if (ret)
1067                 goto err;
1068
1069         if (k.k->type != discard_key_type &&
1070             (c->opts.reconstruct_alloc ||
1071              fsck_err(c, need_discard_key_wrong,
1072                       "incorrect key in need_discard btree (got %s should be %s)\n"
1073                       "  %s",
1074                       bch2_bkey_types[k.k->type],
1075                       bch2_bkey_types[discard_key_type],
1076                       (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1077                 struct bkey_i *update =
1078                         bch2_trans_kmalloc(trans, sizeof(*update));
1079
1080                 ret = PTR_ERR_OR_ZERO(update);
1081                 if (ret)
1082                         goto err;
1083
1084                 bkey_init(&update->k);
1085                 update->k.type  = discard_key_type;
1086                 update->k.p     = discard_iter->pos;
1087
1088                 ret = bch2_trans_update(trans, discard_iter, update, 0);
1089                 if (ret)
1090                         goto err;
1091         }
1092
1093         freespace_key_type = a->data_type == BCH_DATA_free ? KEY_TYPE_set : 0;
1094         bch2_btree_iter_set_pos(freespace_iter, alloc_freespace_pos(alloc_k.k->p, *a));
1095         k = bch2_btree_iter_peek_slot(freespace_iter);
1096         ret = bkey_err(k);
1097         if (ret)
1098                 goto err;
1099
1100         if (k.k->type != freespace_key_type &&
1101             (c->opts.reconstruct_alloc ||
1102              fsck_err(c, freespace_key_wrong,
1103                       "incorrect key in freespace btree (got %s should be %s)\n"
1104                       "  %s",
1105                       bch2_bkey_types[k.k->type],
1106                       bch2_bkey_types[freespace_key_type],
1107                       (printbuf_reset(&buf),
1108                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1109                 struct bkey_i *update =
1110                         bch2_trans_kmalloc(trans, sizeof(*update));
1111
1112                 ret = PTR_ERR_OR_ZERO(update);
1113                 if (ret)
1114                         goto err;
1115
1116                 bkey_init(&update->k);
1117                 update->k.type  = freespace_key_type;
1118                 update->k.p     = freespace_iter->pos;
1119                 bch2_key_resize(&update->k, 1);
1120
1121                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1122                 if (ret)
1123                         goto err;
1124         }
1125
1126         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(alloc_k.k->p, &gens_offset));
1127         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1128         ret = bkey_err(k);
1129         if (ret)
1130                 goto err;
1131
1132         if (a->gen != alloc_gen(k, gens_offset) &&
1133             (c->opts.reconstruct_alloc ||
1134              fsck_err(c, bucket_gens_key_wrong,
1135                       "incorrect gen in bucket_gens btree (got %u should be %u)\n"
1136                       "  %s",
1137                       alloc_gen(k, gens_offset), a->gen,
1138                       (printbuf_reset(&buf),
1139                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1140                 struct bkey_i_bucket_gens *g =
1141                         bch2_trans_kmalloc(trans, sizeof(*g));
1142
1143                 ret = PTR_ERR_OR_ZERO(g);
1144                 if (ret)
1145                         goto err;
1146
1147                 if (k.k->type == KEY_TYPE_bucket_gens) {
1148                         bkey_reassemble(&g->k_i, k);
1149                 } else {
1150                         bkey_bucket_gens_init(&g->k_i);
1151                         g->k.p = alloc_gens_pos(alloc_k.k->p, &gens_offset);
1152                 }
1153
1154                 g->v.gens[gens_offset] = a->gen;
1155
1156                 ret = bch2_trans_update(trans, bucket_gens_iter, &g->k_i, 0);
1157                 if (ret)
1158                         goto err;
1159         }
1160 err:
1161 fsck_err:
1162         printbuf_exit(&buf);
1163         return ret;
1164 }
1165
1166 static noinline_for_stack
1167 int bch2_check_alloc_hole_freespace(struct btree_trans *trans,
1168                                     struct bpos start,
1169                                     struct bpos *end,
1170                                     struct btree_iter *freespace_iter)
1171 {
1172         struct bch_fs *c = trans->c;
1173         struct bch_dev *ca;
1174         struct bkey_s_c k;
1175         struct printbuf buf = PRINTBUF;
1176         int ret;
1177
1178         ca = bch_dev_bkey_exists(c, start.inode);
1179         if (!ca->mi.freespace_initialized)
1180                 return 0;
1181
1182         bch2_btree_iter_set_pos(freespace_iter, start);
1183
1184         k = bch2_btree_iter_peek_slot(freespace_iter);
1185         ret = bkey_err(k);
1186         if (ret)
1187                 goto err;
1188
1189         *end = bkey_min(k.k->p, *end);
1190
1191         if (k.k->type != KEY_TYPE_set &&
1192             (c->opts.reconstruct_alloc ||
1193              fsck_err(c, freespace_hole_missing,
1194                       "hole in alloc btree missing in freespace btree\n"
1195                       "  device %llu buckets %llu-%llu",
1196                       freespace_iter->pos.inode,
1197                       freespace_iter->pos.offset,
1198                       end->offset))) {
1199                 struct bkey_i *update =
1200                         bch2_trans_kmalloc(trans, sizeof(*update));
1201
1202                 ret = PTR_ERR_OR_ZERO(update);
1203                 if (ret)
1204                         goto err;
1205
1206                 bkey_init(&update->k);
1207                 update->k.type  = KEY_TYPE_set;
1208                 update->k.p     = freespace_iter->pos;
1209                 bch2_key_resize(&update->k,
1210                                 min_t(u64, U32_MAX, end->offset -
1211                                       freespace_iter->pos.offset));
1212
1213                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1214                 if (ret)
1215                         goto err;
1216         }
1217 err:
1218 fsck_err:
1219         printbuf_exit(&buf);
1220         return ret;
1221 }
1222
1223 static noinline_for_stack
1224 int bch2_check_alloc_hole_bucket_gens(struct btree_trans *trans,
1225                                       struct bpos start,
1226                                       struct bpos *end,
1227                                       struct btree_iter *bucket_gens_iter)
1228 {
1229         struct bch_fs *c = trans->c;
1230         struct bkey_s_c k;
1231         struct printbuf buf = PRINTBUF;
1232         unsigned i, gens_offset, gens_end_offset;
1233         int ret;
1234
1235         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(start, &gens_offset));
1236
1237         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1238         ret = bkey_err(k);
1239         if (ret)
1240                 goto err;
1241
1242         if (bkey_cmp(alloc_gens_pos(start, &gens_offset),
1243                      alloc_gens_pos(*end,  &gens_end_offset)))
1244                 gens_end_offset = KEY_TYPE_BUCKET_GENS_NR;
1245
1246         if (k.k->type == KEY_TYPE_bucket_gens) {
1247                 struct bkey_i_bucket_gens g;
1248                 bool need_update = false;
1249
1250                 bkey_reassemble(&g.k_i, k);
1251
1252                 for (i = gens_offset; i < gens_end_offset; i++) {
1253                         if (fsck_err_on(g.v.gens[i], c,
1254                                         bucket_gens_hole_wrong,
1255                                         "hole in alloc btree at %llu:%llu with nonzero gen in bucket_gens btree (%u)",
1256                                         bucket_gens_pos_to_alloc(k.k->p, i).inode,
1257                                         bucket_gens_pos_to_alloc(k.k->p, i).offset,
1258                                         g.v.gens[i])) {
1259                                 g.v.gens[i] = 0;
1260                                 need_update = true;
1261                         }
1262                 }
1263
1264                 if (need_update) {
1265                         struct bkey_i *u = bch2_trans_kmalloc(trans, sizeof(g));
1266
1267                         ret = PTR_ERR_OR_ZERO(u);
1268                         if (ret)
1269                                 goto err;
1270
1271                         memcpy(u, &g, sizeof(g));
1272
1273                         ret = bch2_trans_update(trans, bucket_gens_iter, u, 0);
1274                         if (ret)
1275                                 goto err;
1276                 }
1277         }
1278
1279         *end = bkey_min(*end, bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0));
1280 err:
1281 fsck_err:
1282         printbuf_exit(&buf);
1283         return ret;
1284 }
1285
1286 static noinline_for_stack int bch2_check_discard_freespace_key(struct btree_trans *trans,
1287                                               struct btree_iter *iter)
1288 {
1289         struct bch_fs *c = trans->c;
1290         struct btree_iter alloc_iter;
1291         struct bkey_s_c alloc_k;
1292         struct bch_alloc_v4 a_convert;
1293         const struct bch_alloc_v4 *a;
1294         u64 genbits;
1295         struct bpos pos;
1296         enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
1297                 ? BCH_DATA_need_discard
1298                 : BCH_DATA_free;
1299         struct printbuf buf = PRINTBUF;
1300         int ret;
1301
1302         pos = iter->pos;
1303         pos.offset &= ~(~0ULL << 56);
1304         genbits = iter->pos.offset & (~0ULL << 56);
1305
1306         alloc_k = bch2_bkey_get_iter(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
1307         ret = bkey_err(alloc_k);
1308         if (ret)
1309                 return ret;
1310
1311         if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
1312                         need_discard_freespace_key_to_invalid_dev_bucket,
1313                         "entry in %s btree for nonexistant dev:bucket %llu:%llu",
1314                         bch2_btree_id_str(iter->btree_id), pos.inode, pos.offset))
1315                 goto delete;
1316
1317         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1318
1319         if (fsck_err_on(a->data_type != state ||
1320                         (state == BCH_DATA_free &&
1321                          genbits != alloc_freespace_genbits(*a)), c,
1322                         need_discard_freespace_key_bad,
1323                         "%s\n  incorrectly set at %s:%llu:%llu:0 (free %u, genbits %llu should be %llu)",
1324                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
1325                         bch2_btree_id_str(iter->btree_id),
1326                         iter->pos.inode,
1327                         iter->pos.offset,
1328                         a->data_type == state,
1329                         genbits >> 56, alloc_freespace_genbits(*a) >> 56))
1330                 goto delete;
1331 out:
1332 fsck_err:
1333         set_btree_iter_dontneed(&alloc_iter);
1334         bch2_trans_iter_exit(trans, &alloc_iter);
1335         printbuf_exit(&buf);
1336         return ret;
1337 delete:
1338         ret =   bch2_btree_delete_extent_at(trans, iter,
1339                         iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0) ?:
1340                 bch2_trans_commit(trans, NULL, NULL,
1341                         BCH_TRANS_COMMIT_no_enospc);
1342         goto out;
1343 }
1344
1345 /*
1346  * We've already checked that generation numbers in the bucket_gens btree are
1347  * valid for buckets that exist; this just checks for keys for nonexistent
1348  * buckets.
1349  */
1350 static noinline_for_stack
1351 int bch2_check_bucket_gens_key(struct btree_trans *trans,
1352                                struct btree_iter *iter,
1353                                struct bkey_s_c k)
1354 {
1355         struct bch_fs *c = trans->c;
1356         struct bkey_i_bucket_gens g;
1357         struct bch_dev *ca;
1358         u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
1359         u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
1360         u64 b;
1361         bool need_update = false, dev_exists;
1362         struct printbuf buf = PRINTBUF;
1363         int ret = 0;
1364
1365         BUG_ON(k.k->type != KEY_TYPE_bucket_gens);
1366         bkey_reassemble(&g.k_i, k);
1367
1368         /* if no bch_dev, skip out whether we repair or not */
1369         dev_exists = bch2_dev_exists2(c, k.k->p.inode);
1370         if (!dev_exists) {
1371                 if (fsck_err_on(!dev_exists, c,
1372                                 bucket_gens_to_invalid_dev,
1373                                 "bucket_gens key for invalid device:\n  %s",
1374                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1375                         ret = bch2_btree_delete_at(trans, iter, 0);
1376                 }
1377                 goto out;
1378         }
1379
1380         ca = bch_dev_bkey_exists(c, k.k->p.inode);
1381         if (fsck_err_on(end <= ca->mi.first_bucket ||
1382                         start >= ca->mi.nbuckets, c,
1383                         bucket_gens_to_invalid_buckets,
1384                         "bucket_gens key for invalid buckets:\n  %s",
1385                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1386                 ret = bch2_btree_delete_at(trans, iter, 0);
1387                 goto out;
1388         }
1389
1390         for (b = start; b < ca->mi.first_bucket; b++)
1391                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1392                                 bucket_gens_nonzero_for_invalid_buckets,
1393                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1394                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1395                         need_update = true;
1396                 }
1397
1398         for (b = ca->mi.nbuckets; b < end; b++)
1399                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1400                                 bucket_gens_nonzero_for_invalid_buckets,
1401                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1402                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1403                         need_update = true;
1404                 }
1405
1406         if (need_update) {
1407                 struct bkey_i *u = bch2_trans_kmalloc(trans, sizeof(g));
1408
1409                 ret = PTR_ERR_OR_ZERO(u);
1410                 if (ret)
1411                         goto out;
1412
1413                 memcpy(u, &g, sizeof(g));
1414                 ret = bch2_trans_update(trans, iter, u, 0);
1415         }
1416 out:
1417 fsck_err:
1418         printbuf_exit(&buf);
1419         return ret;
1420 }
1421
1422 int bch2_check_alloc_info(struct bch_fs *c)
1423 {
1424         struct btree_trans *trans = bch2_trans_get(c);
1425         struct btree_iter iter, discard_iter, freespace_iter, bucket_gens_iter;
1426         struct bkey hole;
1427         struct bkey_s_c k;
1428         int ret = 0;
1429
1430         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, POS_MIN,
1431                              BTREE_ITER_PREFETCH);
1432         bch2_trans_iter_init(trans, &discard_iter, BTREE_ID_need_discard, POS_MIN,
1433                              BTREE_ITER_PREFETCH);
1434         bch2_trans_iter_init(trans, &freespace_iter, BTREE_ID_freespace, POS_MIN,
1435                              BTREE_ITER_PREFETCH);
1436         bch2_trans_iter_init(trans, &bucket_gens_iter, BTREE_ID_bucket_gens, POS_MIN,
1437                              BTREE_ITER_PREFETCH);
1438
1439         while (1) {
1440                 struct bpos next;
1441
1442                 bch2_trans_begin(trans);
1443
1444                 k = bch2_get_key_or_real_bucket_hole(&iter, &hole);
1445                 ret = bkey_err(k);
1446                 if (ret)
1447                         goto bkey_err;
1448
1449                 if (!k.k)
1450                         break;
1451
1452                 if (k.k->type) {
1453                         next = bpos_nosnap_successor(k.k->p);
1454
1455                         ret = bch2_check_alloc_key(trans,
1456                                                    k, &iter,
1457                                                    &discard_iter,
1458                                                    &freespace_iter,
1459                                                    &bucket_gens_iter);
1460                         if (ret)
1461                                 goto bkey_err;
1462                 } else {
1463                         next = k.k->p;
1464
1465                         ret = bch2_check_alloc_hole_freespace(trans,
1466                                                     bkey_start_pos(k.k),
1467                                                     &next,
1468                                                     &freespace_iter) ?:
1469                                 bch2_check_alloc_hole_bucket_gens(trans,
1470                                                     bkey_start_pos(k.k),
1471                                                     &next,
1472                                                     &bucket_gens_iter);
1473                         if (ret)
1474                                 goto bkey_err;
1475                 }
1476
1477                 ret = bch2_trans_commit(trans, NULL, NULL,
1478                                         BCH_TRANS_COMMIT_no_enospc);
1479                 if (ret)
1480                         goto bkey_err;
1481
1482                 bch2_btree_iter_set_pos(&iter, next);
1483 bkey_err:
1484                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1485                         continue;
1486                 if (ret)
1487                         break;
1488         }
1489         bch2_trans_iter_exit(trans, &bucket_gens_iter);
1490         bch2_trans_iter_exit(trans, &freespace_iter);
1491         bch2_trans_iter_exit(trans, &discard_iter);
1492         bch2_trans_iter_exit(trans, &iter);
1493
1494         if (ret < 0)
1495                 goto err;
1496
1497         ret = for_each_btree_key(trans, iter,
1498                         BTREE_ID_need_discard, POS_MIN,
1499                         BTREE_ITER_PREFETCH, k,
1500                 bch2_check_discard_freespace_key(trans, &iter));
1501         if (ret)
1502                 goto err;
1503
1504         bch2_trans_iter_init(trans, &iter, BTREE_ID_freespace, POS_MIN,
1505                              BTREE_ITER_PREFETCH);
1506         while (1) {
1507                 bch2_trans_begin(trans);
1508                 k = bch2_btree_iter_peek(&iter);
1509                 if (!k.k)
1510                         break;
1511
1512                 ret = bkey_err(k) ?:
1513                         bch2_check_discard_freespace_key(trans, &iter);
1514                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
1515                         ret = 0;
1516                         continue;
1517                 }
1518                 if (ret) {
1519                         struct printbuf buf = PRINTBUF;
1520                         bch2_bkey_val_to_text(&buf, c, k);
1521
1522                         bch_err(c, "while checking %s", buf.buf);
1523                         printbuf_exit(&buf);
1524                         break;
1525                 }
1526
1527                 bch2_btree_iter_set_pos(&iter, bpos_nosnap_successor(iter.pos));
1528         }
1529         bch2_trans_iter_exit(trans, &iter);
1530         if (ret)
1531                 goto err;
1532
1533         ret = for_each_btree_key_commit(trans, iter,
1534                         BTREE_ID_bucket_gens, POS_MIN,
1535                         BTREE_ITER_PREFETCH, k,
1536                         NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1537                 bch2_check_bucket_gens_key(trans, &iter, k));
1538 err:
1539         bch2_trans_put(trans);
1540         bch_err_fn(c, ret);
1541         return ret;
1542 }
1543
1544 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
1545                                        struct btree_iter *alloc_iter)
1546 {
1547         struct bch_fs *c = trans->c;
1548         struct btree_iter lru_iter;
1549         struct bch_alloc_v4 a_convert;
1550         const struct bch_alloc_v4 *a;
1551         struct bkey_s_c alloc_k, lru_k;
1552         struct printbuf buf = PRINTBUF;
1553         int ret;
1554
1555         alloc_k = bch2_btree_iter_peek(alloc_iter);
1556         if (!alloc_k.k)
1557                 return 0;
1558
1559         ret = bkey_err(alloc_k);
1560         if (ret)
1561                 return ret;
1562
1563         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1564
1565         if (a->data_type != BCH_DATA_cached)
1566                 return 0;
1567
1568         if (fsck_err_on(!a->io_time[READ], c,
1569                         alloc_key_cached_but_read_time_zero,
1570                         "cached bucket with read_time 0\n"
1571                         "  %s",
1572                 (printbuf_reset(&buf),
1573                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
1574                 struct bkey_i_alloc_v4 *a_mut =
1575                         bch2_alloc_to_v4_mut(trans, alloc_k);
1576                 ret = PTR_ERR_OR_ZERO(a_mut);
1577                 if (ret)
1578                         goto err;
1579
1580                 a_mut->v.io_time[READ] = atomic64_read(&c->io_clock[READ].now);
1581                 ret = bch2_trans_update(trans, alloc_iter,
1582                                         &a_mut->k_i, BTREE_TRIGGER_NORUN);
1583                 if (ret)
1584                         goto err;
1585
1586                 a = &a_mut->v;
1587         }
1588
1589         lru_k = bch2_bkey_get_iter(trans, &lru_iter, BTREE_ID_lru,
1590                              lru_pos(alloc_k.k->p.inode,
1591                                      bucket_to_u64(alloc_k.k->p),
1592                                      a->io_time[READ]), 0);
1593         ret = bkey_err(lru_k);
1594         if (ret)
1595                 return ret;
1596
1597         if (fsck_err_on(lru_k.k->type != KEY_TYPE_set, c,
1598                         alloc_key_to_missing_lru_entry,
1599                         "missing lru entry\n"
1600                         "  %s",
1601                         (printbuf_reset(&buf),
1602                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
1603                 ret = bch2_lru_set(trans,
1604                                    alloc_k.k->p.inode,
1605                                    bucket_to_u64(alloc_k.k->p),
1606                                    a->io_time[READ]);
1607                 if (ret)
1608                         goto err;
1609         }
1610 err:
1611 fsck_err:
1612         bch2_trans_iter_exit(trans, &lru_iter);
1613         printbuf_exit(&buf);
1614         return ret;
1615 }
1616
1617 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
1618 {
1619         int ret = bch2_trans_run(c,
1620                 for_each_btree_key_commit(trans, iter, BTREE_ID_alloc,
1621                                 POS_MIN, BTREE_ITER_PREFETCH, k,
1622                                 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1623                         bch2_check_alloc_to_lru_ref(trans, &iter)));
1624         bch_err_fn(c, ret);
1625         return ret;
1626 }
1627
1628 static int bch2_discard_one_bucket(struct btree_trans *trans,
1629                                    struct btree_iter *need_discard_iter,
1630                                    struct bpos *discard_pos_done,
1631                                    u64 *seen,
1632                                    u64 *open,
1633                                    u64 *need_journal_commit,
1634                                    u64 *discarded)
1635 {
1636         struct bch_fs *c = trans->c;
1637         struct bpos pos = need_discard_iter->pos;
1638         struct btree_iter iter = { NULL };
1639         struct bkey_s_c k;
1640         struct bch_dev *ca;
1641         struct bkey_i_alloc_v4 *a;
1642         struct printbuf buf = PRINTBUF;
1643         int ret = 0;
1644
1645         ca = bch_dev_bkey_exists(c, pos.inode);
1646         if (!percpu_ref_tryget(&ca->io_ref)) {
1647                 bch2_btree_iter_set_pos(need_discard_iter, POS(pos.inode + 1, 0));
1648                 return 0;
1649         }
1650
1651         if (bch2_bucket_is_open_safe(c, pos.inode, pos.offset)) {
1652                 (*open)++;
1653                 goto out;
1654         }
1655
1656         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
1657                         c->journal.flushed_seq_ondisk,
1658                         pos.inode, pos.offset)) {
1659                 (*need_journal_commit)++;
1660                 goto out;
1661         }
1662
1663         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_alloc,
1664                                need_discard_iter->pos,
1665                                BTREE_ITER_CACHED);
1666         ret = bkey_err(k);
1667         if (ret)
1668                 goto out;
1669
1670         a = bch2_alloc_to_v4_mut(trans, k);
1671         ret = PTR_ERR_OR_ZERO(a);
1672         if (ret)
1673                 goto out;
1674
1675         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
1676                 a->v.gen++;
1677                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1678                 goto write;
1679         }
1680
1681         if (a->v.journal_seq > c->journal.flushed_seq_ondisk) {
1682                 if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
1683                         bch2_trans_inconsistent(trans,
1684                                 "clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
1685                                 "%s",
1686                                 a->v.journal_seq,
1687                                 c->journal.flushed_seq_ondisk,
1688                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1689                         ret = -EIO;
1690                 }
1691                 goto out;
1692         }
1693
1694         if (a->v.data_type != BCH_DATA_need_discard) {
1695                 if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
1696                         bch2_trans_inconsistent(trans,
1697                                 "bucket incorrectly set in need_discard btree\n"
1698                                 "%s",
1699                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1700                         ret = -EIO;
1701                 }
1702
1703                 goto out;
1704         }
1705
1706         if (!bkey_eq(*discard_pos_done, iter.pos) &&
1707             ca->mi.discard && !c->opts.nochanges) {
1708                 /*
1709                  * This works without any other locks because this is the only
1710                  * thread that removes items from the need_discard tree
1711                  */
1712                 bch2_trans_unlock(trans);
1713                 blkdev_issue_discard(ca->disk_sb.bdev,
1714                                      k.k->p.offset * ca->mi.bucket_size,
1715                                      ca->mi.bucket_size,
1716                                      GFP_KERNEL);
1717                 *discard_pos_done = iter.pos;
1718
1719                 ret = bch2_trans_relock_notrace(trans);
1720                 if (ret)
1721                         goto out;
1722         }
1723
1724         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
1725         a->v.data_type = alloc_data_type(a->v, a->v.data_type);
1726 write:
1727         ret =   bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1728                 bch2_trans_commit(trans, NULL, NULL,
1729                                   BCH_WATERMARK_btree|
1730                                   BCH_TRANS_COMMIT_no_enospc);
1731         if (ret)
1732                 goto out;
1733
1734         count_event(c, bucket_discard);
1735         (*discarded)++;
1736 out:
1737         (*seen)++;
1738         bch2_trans_iter_exit(trans, &iter);
1739         percpu_ref_put(&ca->io_ref);
1740         printbuf_exit(&buf);
1741         return ret;
1742 }
1743
1744 static void bch2_do_discards_work(struct work_struct *work)
1745 {
1746         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
1747         u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
1748         struct bpos discard_pos_done = POS_MAX;
1749         int ret;
1750
1751         /*
1752          * We're doing the commit in bch2_discard_one_bucket instead of using
1753          * for_each_btree_key_commit() so that we can increment counters after
1754          * successful commit:
1755          */
1756         ret = bch2_trans_run(c,
1757                 for_each_btree_key(trans, iter,
1758                                    BTREE_ID_need_discard, POS_MIN, 0, k,
1759                         bch2_discard_one_bucket(trans, &iter, &discard_pos_done,
1760                                                 &seen,
1761                                                 &open,
1762                                                 &need_journal_commit,
1763                                                 &discarded)));
1764
1765         if (need_journal_commit * 2 > seen)
1766                 bch2_journal_flush_async(&c->journal, NULL);
1767
1768         bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1769
1770         trace_discard_buckets(c, seen, open, need_journal_commit, discarded,
1771                               bch2_err_str(ret));
1772 }
1773
1774 void bch2_do_discards(struct bch_fs *c)
1775 {
1776         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_discard) &&
1777             !queue_work(c->write_ref_wq, &c->discard_work))
1778                 bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1779 }
1780
1781 static int invalidate_one_bucket(struct btree_trans *trans,
1782                                  struct btree_iter *lru_iter,
1783                                  struct bkey_s_c lru_k,
1784                                  s64 *nr_to_invalidate)
1785 {
1786         struct bch_fs *c = trans->c;
1787         struct btree_iter alloc_iter = { NULL };
1788         struct bkey_i_alloc_v4 *a = NULL;
1789         struct printbuf buf = PRINTBUF;
1790         struct bpos bucket = u64_to_bucket(lru_k.k->p.offset);
1791         unsigned cached_sectors;
1792         int ret = 0;
1793
1794         if (*nr_to_invalidate <= 0)
1795                 return 1;
1796
1797         if (!bch2_dev_bucket_exists(c, bucket)) {
1798                 prt_str(&buf, "lru entry points to invalid bucket");
1799                 goto err;
1800         }
1801
1802         if (bch2_bucket_is_open_safe(c, bucket.inode, bucket.offset))
1803                 return 0;
1804
1805         a = bch2_trans_start_alloc_update(trans, &alloc_iter, bucket);
1806         ret = PTR_ERR_OR_ZERO(a);
1807         if (ret)
1808                 goto out;
1809
1810         /* We expect harmless races here due to the btree write buffer: */
1811         if (lru_pos_time(lru_iter->pos) != alloc_lru_idx_read(a->v))
1812                 goto out;
1813
1814         BUG_ON(a->v.data_type != BCH_DATA_cached);
1815
1816         if (!a->v.cached_sectors)
1817                 bch_err(c, "invalidating empty bucket, confused");
1818
1819         cached_sectors = a->v.cached_sectors;
1820
1821         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1822         a->v.gen++;
1823         a->v.data_type          = 0;
1824         a->v.dirty_sectors      = 0;
1825         a->v.cached_sectors     = 0;
1826         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1827         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1828
1829         ret =   bch2_trans_update(trans, &alloc_iter, &a->k_i,
1830                                 BTREE_TRIGGER_BUCKET_INVALIDATE) ?:
1831                 bch2_trans_commit(trans, NULL, NULL,
1832                                   BCH_WATERMARK_btree|
1833                                   BCH_TRANS_COMMIT_no_enospc);
1834         if (ret)
1835                 goto out;
1836
1837         trace_and_count(c, bucket_invalidate, c, bucket.inode, bucket.offset, cached_sectors);
1838         --*nr_to_invalidate;
1839 out:
1840         bch2_trans_iter_exit(trans, &alloc_iter);
1841         printbuf_exit(&buf);
1842         return ret;
1843 err:
1844         prt_str(&buf, "\n  lru key: ");
1845         bch2_bkey_val_to_text(&buf, c, lru_k);
1846
1847         prt_str(&buf, "\n  lru entry: ");
1848         bch2_lru_pos_to_text(&buf, lru_iter->pos);
1849
1850         prt_str(&buf, "\n  alloc key: ");
1851         if (!a)
1852                 bch2_bpos_to_text(&buf, bucket);
1853         else
1854                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
1855
1856         bch_err(c, "%s", buf.buf);
1857         if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_lrus) {
1858                 bch2_inconsistent_error(c);
1859                 ret = -EINVAL;
1860         }
1861
1862         goto out;
1863 }
1864
1865 static void bch2_do_invalidates_work(struct work_struct *work)
1866 {
1867         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1868         struct btree_trans *trans = bch2_trans_get(c);
1869         int ret = 0;
1870
1871         ret = bch2_btree_write_buffer_tryflush(trans);
1872         if (ret)
1873                 goto err;
1874
1875         for_each_member_device(c, ca) {
1876                 s64 nr_to_invalidate =
1877                         should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
1878
1879                 ret = for_each_btree_key_upto(trans, iter, BTREE_ID_lru,
1880                                 lru_pos(ca->dev_idx, 0, 0),
1881                                 lru_pos(ca->dev_idx, U64_MAX, LRU_TIME_MAX),
1882                                 BTREE_ITER_INTENT, k,
1883                         invalidate_one_bucket(trans, &iter, k, &nr_to_invalidate));
1884
1885                 if (ret < 0) {
1886                         percpu_ref_put(&ca->ref);
1887                         break;
1888                 }
1889         }
1890 err:
1891         bch2_trans_put(trans);
1892         bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1893 }
1894
1895 void bch2_do_invalidates(struct bch_fs *c)
1896 {
1897         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_invalidate) &&
1898             !queue_work(c->write_ref_wq, &c->invalidate_work))
1899                 bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1900 }
1901
1902 int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca,
1903                             u64 bucket_start, u64 bucket_end)
1904 {
1905         struct btree_trans *trans = bch2_trans_get(c);
1906         struct btree_iter iter;
1907         struct bkey_s_c k;
1908         struct bkey hole;
1909         struct bpos end = POS(ca->dev_idx, bucket_end);
1910         struct bch_member *m;
1911         unsigned long last_updated = jiffies;
1912         int ret;
1913
1914         BUG_ON(bucket_start > bucket_end);
1915         BUG_ON(bucket_end > ca->mi.nbuckets);
1916
1917         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
1918                 POS(ca->dev_idx, max_t(u64, ca->mi.first_bucket, bucket_start)),
1919                 BTREE_ITER_PREFETCH);
1920         /*
1921          * Scan the alloc btree for every bucket on @ca, and add buckets to the
1922          * freespace/need_discard/need_gc_gens btrees as needed:
1923          */
1924         while (1) {
1925                 if (last_updated + HZ * 10 < jiffies) {
1926                         bch_info(ca, "%s: currently at %llu/%llu",
1927                                  __func__, iter.pos.offset, ca->mi.nbuckets);
1928                         last_updated = jiffies;
1929                 }
1930
1931                 bch2_trans_begin(trans);
1932
1933                 if (bkey_ge(iter.pos, end)) {
1934                         ret = 0;
1935                         break;
1936                 }
1937
1938                 k = bch2_get_key_or_hole(&iter, end, &hole);
1939                 ret = bkey_err(k);
1940                 if (ret)
1941                         goto bkey_err;
1942
1943                 if (k.k->type) {
1944                         /*
1945                          * We process live keys in the alloc btree one at a
1946                          * time:
1947                          */
1948                         struct bch_alloc_v4 a_convert;
1949                         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1950
1951                         ret =   bch2_bucket_do_index(trans, k, a, true) ?:
1952                                 bch2_trans_commit(trans, NULL, NULL,
1953                                                   BCH_TRANS_COMMIT_no_enospc);
1954                         if (ret)
1955                                 goto bkey_err;
1956
1957                         bch2_btree_iter_advance(&iter);
1958                 } else {
1959                         struct bkey_i *freespace;
1960
1961                         freespace = bch2_trans_kmalloc(trans, sizeof(*freespace));
1962                         ret = PTR_ERR_OR_ZERO(freespace);
1963                         if (ret)
1964                                 goto bkey_err;
1965
1966                         bkey_init(&freespace->k);
1967                         freespace->k.type       = KEY_TYPE_set;
1968                         freespace->k.p          = k.k->p;
1969                         freespace->k.size       = k.k->size;
1970
1971                         ret = bch2_btree_insert_trans(trans, BTREE_ID_freespace, freespace, 0) ?:
1972                                 bch2_trans_commit(trans, NULL, NULL,
1973                                                   BCH_TRANS_COMMIT_no_enospc);
1974                         if (ret)
1975                                 goto bkey_err;
1976
1977                         bch2_btree_iter_set_pos(&iter, k.k->p);
1978                 }
1979 bkey_err:
1980                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1981                         continue;
1982                 if (ret)
1983                         break;
1984         }
1985
1986         bch2_trans_iter_exit(trans, &iter);
1987         bch2_trans_put(trans);
1988
1989         if (ret < 0) {
1990                 bch_err_msg(ca, ret, "initializing free space");
1991                 return ret;
1992         }
1993
1994         mutex_lock(&c->sb_lock);
1995         m = bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx);
1996         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1997         mutex_unlock(&c->sb_lock);
1998
1999         return 0;
2000 }
2001
2002 int bch2_fs_freespace_init(struct bch_fs *c)
2003 {
2004         int ret = 0;
2005         bool doing_init = false;
2006
2007         /*
2008          * We can crash during the device add path, so we need to check this on
2009          * every mount:
2010          */
2011
2012         for_each_member_device(c, ca) {
2013                 if (ca->mi.freespace_initialized)
2014                         continue;
2015
2016                 if (!doing_init) {
2017                         bch_info(c, "initializing freespace");
2018                         doing_init = true;
2019                 }
2020
2021                 ret = bch2_dev_freespace_init(c, ca, 0, ca->mi.nbuckets);
2022                 if (ret) {
2023                         percpu_ref_put(&ca->ref);
2024                         bch_err_fn(c, ret);
2025                         return ret;
2026                 }
2027         }
2028
2029         if (doing_init) {
2030                 mutex_lock(&c->sb_lock);
2031                 bch2_write_super(c);
2032                 mutex_unlock(&c->sb_lock);
2033                 bch_verbose(c, "done initializing freespace");
2034         }
2035
2036         return 0;
2037 }
2038
2039 /* Bucket IO clocks: */
2040
2041 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
2042                               size_t bucket_nr, int rw)
2043 {
2044         struct bch_fs *c = trans->c;
2045         struct btree_iter iter;
2046         struct bkey_i_alloc_v4 *a;
2047         u64 now;
2048         int ret = 0;
2049
2050         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
2051         ret = PTR_ERR_OR_ZERO(a);
2052         if (ret)
2053                 return ret;
2054
2055         now = atomic64_read(&c->io_clock[rw].now);
2056         if (a->v.io_time[rw] == now)
2057                 goto out;
2058
2059         a->v.io_time[rw] = now;
2060
2061         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
2062                 bch2_trans_commit(trans, NULL, NULL, 0);
2063 out:
2064         bch2_trans_iter_exit(trans, &iter);
2065         return ret;
2066 }
2067
2068 /* Startup/shutdown (ro/rw): */
2069
2070 void bch2_recalc_capacity(struct bch_fs *c)
2071 {
2072         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
2073         unsigned bucket_size_max = 0;
2074         unsigned long ra_pages = 0;
2075
2076         lockdep_assert_held(&c->state_lock);
2077
2078         for_each_online_member(c, ca) {
2079                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
2080
2081                 ra_pages += bdi->ra_pages;
2082         }
2083
2084         bch2_set_ra_pages(c, ra_pages);
2085
2086         for_each_rw_member(c, ca) {
2087                 u64 dev_reserve = 0;
2088
2089                 /*
2090                  * We need to reserve buckets (from the number
2091                  * of currently available buckets) against
2092                  * foreground writes so that mainly copygc can
2093                  * make forward progress.
2094                  *
2095                  * We need enough to refill the various reserves
2096                  * from scratch - copygc will use its entire
2097                  * reserve all at once, then run against when
2098                  * its reserve is refilled (from the formerly
2099                  * available buckets).
2100                  *
2101                  * This reserve is just used when considering if
2102                  * allocations for foreground writes must wait -
2103                  * not -ENOSPC calculations.
2104                  */
2105
2106                 dev_reserve += ca->nr_btree_reserve * 2;
2107                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
2108
2109                 dev_reserve += 1;       /* btree write point */
2110                 dev_reserve += 1;       /* copygc write point */
2111                 dev_reserve += 1;       /* rebalance write point */
2112
2113                 dev_reserve *= ca->mi.bucket_size;
2114
2115                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
2116                                              ca->mi.first_bucket);
2117
2118                 reserved_sectors += dev_reserve * 2;
2119
2120                 bucket_size_max = max_t(unsigned, bucket_size_max,
2121                                         ca->mi.bucket_size);
2122         }
2123
2124         gc_reserve = c->opts.gc_reserve_bytes
2125                 ? c->opts.gc_reserve_bytes >> 9
2126                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
2127
2128         reserved_sectors = max(gc_reserve, reserved_sectors);
2129
2130         reserved_sectors = min(reserved_sectors, capacity);
2131
2132         c->capacity = capacity - reserved_sectors;
2133
2134         c->bucket_size_max = bucket_size_max;
2135
2136         /* Wake up case someone was waiting for buckets */
2137         closure_wake_up(&c->freelist_wait);
2138 }
2139
2140 u64 bch2_min_rw_member_capacity(struct bch_fs *c)
2141 {
2142         u64 ret = U64_MAX;
2143
2144         for_each_rw_member(c, ca)
2145                 ret = min(ret, ca->mi.nbuckets * ca->mi.bucket_size);
2146         return ret;
2147 }
2148
2149 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
2150 {
2151         struct open_bucket *ob;
2152         bool ret = false;
2153
2154         for (ob = c->open_buckets;
2155              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
2156              ob++) {
2157                 spin_lock(&ob->lock);
2158                 if (ob->valid && !ob->on_partial_list &&
2159                     ob->dev == ca->dev_idx)
2160                         ret = true;
2161                 spin_unlock(&ob->lock);
2162         }
2163
2164         return ret;
2165 }
2166
2167 /* device goes ro: */
2168 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
2169 {
2170         unsigned i;
2171
2172         /* First, remove device from allocation groups: */
2173
2174         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2175                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
2176
2177         /*
2178          * Capacity is calculated based off of devices in allocation groups:
2179          */
2180         bch2_recalc_capacity(c);
2181
2182         bch2_open_buckets_stop(c, ca, false);
2183
2184         /*
2185          * Wake up threads that were blocked on allocation, so they can notice
2186          * the device can no longer be removed and the capacity has changed:
2187          */
2188         closure_wake_up(&c->freelist_wait);
2189
2190         /*
2191          * journal_res_get() can block waiting for free space in the journal -
2192          * it needs to notice there may not be devices to allocate from anymore:
2193          */
2194         wake_up(&c->journal.wait);
2195
2196         /* Now wait for any in flight writes: */
2197
2198         closure_wait_event(&c->open_buckets_wait,
2199                            !bch2_dev_has_open_write_point(c, ca));
2200 }
2201
2202 /* device goes rw: */
2203 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
2204 {
2205         unsigned i;
2206
2207         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2208                 if (ca->mi.data_allowed & (1 << i))
2209                         set_bit(ca->dev_idx, c->rw_devs[i].d);
2210 }
2211
2212 void bch2_fs_allocator_background_init(struct bch_fs *c)
2213 {
2214         spin_lock_init(&c->freelist_lock);
2215         INIT_WORK(&c->discard_work, bch2_do_discards_work);
2216         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
2217 }