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 "varint.h"
  22
  23 #include <linux/kthread.h>
  24 #include <linux/math64.h>
  25 #include <linux/random.h>
  26 #include <linux/rculist.h>
  27 #include <linux/rcupdate.h>
  28 #include <linux/sched/task.h>
  29 #include <linux/sort.h>
  30 #include <trace/events/bcachefs.h>
  31
  32 /* Persistent alloc info: */
  33
  34 static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
  35 #define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8,
  36         BCH_ALLOC_FIELDS_V1()
  37 #undef x
  38 };
  39
  40 struct bkey_alloc_unpacked {
  41         u64             journal_seq;
  42         u8              gen;
  43         u8              oldest_gen;
  44         u8              data_type;
  45         bool            need_discard:1;
  46         bool            need_inc_gen:1;
  47 #define x(_name, _bits) u##_bits _name;
  48         BCH_ALLOC_FIELDS_V2()
  49 #undef  x
  50 };
  51
  52 static inline u64 alloc_field_v1_get(const struct bch_alloc *a,
  53                                      const void **p, unsigned field)
  54 {
  55         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  56         u64 v;
  57
  58         if (!(a->fields & (1 << field)))
  59                 return 0;
  60
  61         switch (bytes) {
  62         case 1:
  63                 v = *((const u8 *) *p);
  64                 break;
  65         case 2:
  66                 v = le16_to_cpup(*p);
  67                 break;
  68         case 4:
  69                 v = le32_to_cpup(*p);
  70                 break;
  71         case 8:
  72                 v = le64_to_cpup(*p);
  73                 break;
  74         default:
  75                 BUG();
  76         }
  77
  78         *p += bytes;
  79         return v;
  80 }
  81
  82 static inline void alloc_field_v1_put(struct bkey_i_alloc *a, void **p,
  83                                       unsigned field, u64 v)
  84 {
  85         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  86
  87         if (!v)
  88                 return;
  89
  90         a->v.fields |= 1 << field;
  91
  92         switch (bytes) {
  93         case 1:
  94                 *((u8 *) *p) = v;
  95                 break;
  96         case 2:
  97                 *((__le16 *) *p) = cpu_to_le16(v);
  98                 break;
  99         case 4:
 100                 *((__le32 *) *p) = cpu_to_le32(v);
 101                 break;
 102         case 8:
 103                 *((__le64 *) *p) = cpu_to_le64(v);
 104                 break;
 105         default:
 106                 BUG();
 107         }
 108
 109         *p += bytes;
 110 }
 111
 112 static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out,
 113                                  struct bkey_s_c k)
 114 {
 115         const struct bch_alloc *in = bkey_s_c_to_alloc(k).v;
 116         const void *d = in->data;
 117         unsigned idx = 0;
 118
 119         out->gen = in->gen;
 120
 121 #define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++);
 122         BCH_ALLOC_FIELDS_V1()
 123 #undef  x
 124 }
 125
 126 static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out,
 127                                 struct bkey_s_c k)
 128 {
 129         struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k);
 130         const u8 *in = a.v->data;
 131         const u8 *end = bkey_val_end(a);
 132         unsigned fieldnr = 0;
 133         int ret;
 134         u64 v;
 135
 136         out->gen        = a.v->gen;
 137         out->oldest_gen = a.v->oldest_gen;
 138         out->data_type  = a.v->data_type;
 139
 140 #define x(_name, _bits)                                                 \
 141         if (fieldnr < a.v->nr_fields) {                                 \
 142                 ret = bch2_varint_decode_fast(in, end, &v);             \
 143                 if (ret < 0)                                            \
 144                         return ret;                                     \
 145                 in += ret;                                              \
 146         } else {                                                        \
 147                 v = 0;                                                  \
 148         }                                                               \
 149         out->_name = v;                                                 \
 150         if (v != out->_name)                                            \
 151                 return -1;                                              \
 152         fieldnr++;
 153
 154         BCH_ALLOC_FIELDS_V2()
 155 #undef  x
 156         return 0;
 157 }
 158
 159 static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out,
 160                                 struct bkey_s_c k)
 161 {
 162         struct bkey_s_c_alloc_v3 a = bkey_s_c_to_alloc_v3(k);
 163         const u8 *in = a.v->data;
 164         const u8 *end = bkey_val_end(a);
 165         unsigned fieldnr = 0;
 166         int ret;
 167         u64 v;
 168
 169         out->gen        = a.v->gen;
 170         out->oldest_gen = a.v->oldest_gen;
 171         out->data_type  = a.v->data_type;
 172         out->need_discard = BCH_ALLOC_V3_NEED_DISCARD(a.v);
 173         out->need_inc_gen = BCH_ALLOC_V3_NEED_INC_GEN(a.v);
 174         out->journal_seq = le64_to_cpu(a.v->journal_seq);
 175
 176 #define x(_name, _bits)                                                 \
 177         if (fieldnr < a.v->nr_fields) {                                 \
 178                 ret = bch2_varint_decode_fast(in, end, &v);             \
 179                 if (ret < 0)                                            \
 180                         return ret;                                     \
 181                 in += ret;                                              \
 182         } else {                                                        \
 183                 v = 0;                                                  \
 184         }                                                               \
 185         out->_name = v;                                                 \
 186         if (v != out->_name)                                            \
 187                 return -1;                                              \
 188         fieldnr++;
 189
 190         BCH_ALLOC_FIELDS_V2()
 191 #undef  x
 192         return 0;
 193 }
 194
 195 static struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 196 {
 197         struct bkey_alloc_unpacked ret = { .gen = 0 };
 198
 199         switch (k.k->type) {
 200         case KEY_TYPE_alloc:
 201                 bch2_alloc_unpack_v1(&ret, k);
 202                 break;
 203         case KEY_TYPE_alloc_v2:
 204                 bch2_alloc_unpack_v2(&ret, k);
 205                 break;
 206         case KEY_TYPE_alloc_v3:
 207                 bch2_alloc_unpack_v3(&ret, k);
 208                 break;
 209         }
 210
 211         return ret;
 212 }
 213
 214 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
 215 {
 216         unsigned i, bytes = offsetof(struct bch_alloc, data);
 217
 218         for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++)
 219                 if (a->fields & (1 << i))
 220                         bytes += BCH_ALLOC_V1_FIELD_BYTES[i];
 221
 222         return DIV_ROUND_UP(bytes, sizeof(u64));
 223 }
 224
 225 int bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k,
 226                           unsigned flags, struct printbuf *err)
 227 {
 228         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 229
 230         /* allow for unknown fields */
 231         if (bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v)) {
 232                 prt_printf(err, "incorrect value size (%zu < %u)",
 233                        bkey_val_u64s(a.k), bch_alloc_v1_val_u64s(a.v));
 234                 return -BCH_ERR_invalid_bkey;
 235         }
 236
 237         return 0;
 238 }
 239
 240 int bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k,
 241                           unsigned flags, struct printbuf *err)
 242 {
 243         struct bkey_alloc_unpacked u;
 244
 245         if (bch2_alloc_unpack_v2(&u, k)) {
 246                 prt_printf(err, "unpack error");
 247                 return -BCH_ERR_invalid_bkey;
 248         }
 249
 250         return 0;
 251 }
 252
 253 int bch2_alloc_v3_invalid(const struct bch_fs *c, struct bkey_s_c k,
 254                           unsigned flags, struct printbuf *err)
 255 {
 256         struct bkey_alloc_unpacked u;
 257
 258         if (bch2_alloc_unpack_v3(&u, k)) {
 259                 prt_printf(err, "unpack error");
 260                 return -BCH_ERR_invalid_bkey;
 261         }
 262
 263         return 0;
 264 }
 265
 266 int bch2_alloc_v4_invalid(const struct bch_fs *c, struct bkey_s_c k,
 267                           unsigned flags, struct printbuf *err)
 268 {
 269         struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 270         int rw = flags & WRITE;
 271
 272         if (alloc_v4_u64s(a.v) != bkey_val_u64s(k.k)) {
 273                 prt_printf(err, "bad val size (%lu != %u)",
 274                        bkey_val_u64s(k.k), alloc_v4_u64s(a.v));
 275                 return -BCH_ERR_invalid_bkey;
 276         }
 277
 278         if (!BCH_ALLOC_V4_BACKPOINTERS_START(a.v) &&
 279             BCH_ALLOC_V4_NR_BACKPOINTERS(a.v)) {
 280                 prt_printf(err, "invalid backpointers_start");
 281                 return -BCH_ERR_invalid_bkey;
 282         }
 283
 284         if (rw == WRITE &&
 285             !(flags & BKEY_INVALID_FROM_JOURNAL) &&
 286             test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags)) {
 287                 unsigned i, bp_len = 0;
 288
 289                 for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a.v); i++)
 290                         bp_len += alloc_v4_backpointers_c(a.v)[i].bucket_len;
 291
 292                 if (bp_len > a.v->dirty_sectors) {
 293                         prt_printf(err, "too many backpointers");
 294                         return -BCH_ERR_invalid_bkey;
 295                 }
 296         }
 297
 298         if (rw == WRITE) {
 299                 if (alloc_data_type(*a.v, a.v->data_type) != a.v->data_type) {
 300                         prt_printf(err, "invalid data type (got %u should be %u)",
 301                                a.v->data_type, alloc_data_type(*a.v, a.v->data_type));
 302                         return -BCH_ERR_invalid_bkey;
 303                 }
 304
 305                 switch (a.v->data_type) {
 306                 case BCH_DATA_free:
 307                 case BCH_DATA_need_gc_gens:
 308                 case BCH_DATA_need_discard:
 309                         if (a.v->dirty_sectors ||
 310                             a.v->cached_sectors ||
 311                             a.v->stripe) {
 312                                 prt_printf(err, "empty data type free but have data");
 313                                 return -BCH_ERR_invalid_bkey;
 314                         }
 315                         break;
 316                 case BCH_DATA_sb:
 317                 case BCH_DATA_journal:
 318                 case BCH_DATA_btree:
 319                 case BCH_DATA_user:
 320                 case BCH_DATA_parity:
 321                         if (!a.v->dirty_sectors) {
 322                                 prt_printf(err, "data_type %s but dirty_sectors==0",
 323                                        bch2_data_types[a.v->data_type]);
 324                                 return -BCH_ERR_invalid_bkey;
 325                         }
 326                         break;
 327                 case BCH_DATA_cached:
 328                         if (!a.v->cached_sectors ||
 329                             a.v->dirty_sectors ||
 330                             a.v->stripe) {
 331                                 prt_printf(err, "data type inconsistency");
 332                                 return -BCH_ERR_invalid_bkey;
 333                         }
 334
 335                         if (!a.v->io_time[READ] &&
 336                             test_bit(BCH_FS_CHECK_ALLOC_TO_LRU_REFS_DONE, &c->flags)) {
 337                                 prt_printf(err, "cached bucket with read_time == 0");
 338                                 return -BCH_ERR_invalid_bkey;
 339                         }
 340                         break;
 341                 case BCH_DATA_stripe:
 342                         if (!a.v->stripe) {
 343                                 prt_printf(err, "data_type %s but stripe==0",
 344                                        bch2_data_types[a.v->data_type]);
 345                                 return -BCH_ERR_invalid_bkey;
 346                         }
 347                         break;
 348                 }
 349         }
 350
 351         return 0;
 352 }
 353
 354 static inline u64 swab40(u64 x)
 355 {
 356         return (((x & 0x00000000ffULL) << 32)|
 357                 ((x & 0x000000ff00ULL) << 16)|
 358                 ((x & 0x0000ff0000ULL) >>  0)|
 359                 ((x & 0x00ff000000ULL) >> 16)|
 360                 ((x & 0xff00000000ULL) >> 32));
 361 }
 362
 363 void bch2_alloc_v4_swab(struct bkey_s k)
 364 {
 365         struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v;
 366         struct bch_backpointer *bp, *bps;
 367
 368         a->journal_seq          = swab64(a->journal_seq);
 369         a->flags                = swab32(a->flags);
 370         a->dirty_sectors        = swab32(a->dirty_sectors);
 371         a->cached_sectors       = swab32(a->cached_sectors);
 372         a->io_time[0]           = swab64(a->io_time[0]);
 373         a->io_time[1]           = swab64(a->io_time[1]);
 374         a->stripe               = swab32(a->stripe);
 375         a->nr_external_backpointers = swab32(a->nr_external_backpointers);
 376
 377         bps = alloc_v4_backpointers(a);
 378         for (bp = bps; bp < bps + BCH_ALLOC_V4_NR_BACKPOINTERS(a); bp++) {
 379                 bp->bucket_offset       = swab40(bp->bucket_offset);
 380                 bp->bucket_len          = swab32(bp->bucket_len);
 381                 bch2_bpos_swab(&bp->pos);
 382         }
 383 }
 384
 385 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 386 {
 387         struct bch_alloc_v4 _a;
 388         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &_a);
 389         unsigned i;
 390
 391         prt_newline(out);
 392         printbuf_indent_add(out, 2);
 393
 394         prt_printf(out, "gen %u oldest_gen %u data_type %s",
 395                a->gen, a->oldest_gen,
 396                a->data_type < BCH_DATA_NR
 397                ? bch2_data_types[a->data_type]
 398                : "(invalid data type)");
 399         prt_newline(out);
 400         prt_printf(out, "journal_seq       %llu",       a->journal_seq);
 401         prt_newline(out);
 402         prt_printf(out, "need_discard      %llu",       BCH_ALLOC_V4_NEED_DISCARD(a));
 403         prt_newline(out);
 404         prt_printf(out, "need_inc_gen      %llu",       BCH_ALLOC_V4_NEED_INC_GEN(a));
 405         prt_newline(out);
 406         prt_printf(out, "dirty_sectors     %u", a->dirty_sectors);
 407         prt_newline(out);
 408         prt_printf(out, "cached_sectors    %u", a->cached_sectors);
 409         prt_newline(out);
 410         prt_printf(out, "stripe            %u", a->stripe);
 411         prt_newline(out);
 412         prt_printf(out, "stripe_redundancy %u", a->stripe_redundancy);
 413         prt_newline(out);
 414         prt_printf(out, "io_time[READ]     %llu",       a->io_time[READ]);
 415         prt_newline(out);
 416         prt_printf(out, "io_time[WRITE]    %llu",       a->io_time[WRITE]);
 417         prt_newline(out);
 418         prt_printf(out, "fragmentation     %llu",       a->fragmentation_lru);
 419         prt_newline(out);
 420         prt_printf(out, "bp_start          %llu", BCH_ALLOC_V4_BACKPOINTERS_START(a));
 421         prt_newline(out);
 422
 423         if (BCH_ALLOC_V4_NR_BACKPOINTERS(a)) {
 424                 struct bkey_s_c_alloc_v4 a_raw = bkey_s_c_to_alloc_v4(k);
 425                 const struct bch_backpointer *bps = alloc_v4_backpointers_c(a_raw.v);
 426
 427                 prt_printf(out, "backpointers:     %llu", BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v));
 428                 printbuf_indent_add(out, 2);
 429
 430                 for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v); i++) {
 431                         prt_newline(out);
 432                         bch2_backpointer_to_text(out, &bps[i]);
 433                 }
 434
 435                 printbuf_indent_sub(out, 2);
 436         }
 437
 438         printbuf_indent_sub(out, 2);
 439 }
 440
 441 void __bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out)
 442 {
 443         if (k.k->type == KEY_TYPE_alloc_v4) {
 444                 void *src, *dst;
 445
 446                 *out = *bkey_s_c_to_alloc_v4(k).v;
 447
 448                 src = alloc_v4_backpointers(out);
 449                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 450                 dst = alloc_v4_backpointers(out);
 451
 452                 if (src < dst)
 453                         memset(src, 0, dst - src);
 454
 455                 SET_BCH_ALLOC_V4_NR_BACKPOINTERS(out, 0);
 456         } else {
 457                 struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 458
 459                 *out = (struct bch_alloc_v4) {
 460                         .journal_seq            = u.journal_seq,
 461                         .flags                  = u.need_discard,
 462                         .gen                    = u.gen,
 463                         .oldest_gen             = u.oldest_gen,
 464                         .data_type              = u.data_type,
 465                         .stripe_redundancy      = u.stripe_redundancy,
 466                         .dirty_sectors          = u.dirty_sectors,
 467                         .cached_sectors         = u.cached_sectors,
 468                         .io_time[READ]          = u.read_time,
 469                         .io_time[WRITE]         = u.write_time,
 470                         .stripe                 = u.stripe,
 471                 };
 472
 473                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 474         }
 475 }
 476
 477 static noinline struct bkey_i_alloc_v4 *
 478 __bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 479 {
 480         struct bkey_i_alloc_v4 *ret;
 481
 482         ret = bch2_trans_kmalloc(trans, max(bkey_bytes(k.k), sizeof(struct bkey_i_alloc_v4)));
 483         if (IS_ERR(ret))
 484                 return ret;
 485
 486         if (k.k->type == KEY_TYPE_alloc_v4) {
 487                 void *src, *dst;
 488
 489                 bkey_reassemble(&ret->k_i, k);
 490
 491                 src = alloc_v4_backpointers(&ret->v);
 492                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(&ret->v, BCH_ALLOC_V4_U64s);
 493                 dst = alloc_v4_backpointers(&ret->v);
 494
 495                 if (src < dst)
 496                         memset(src, 0, dst - src);
 497
 498                 SET_BCH_ALLOC_V4_NR_BACKPOINTERS(&ret->v, 0);
 499                 set_alloc_v4_u64s(ret);
 500         } else {
 501                 bkey_alloc_v4_init(&ret->k_i);
 502                 ret->k.p = k.k->p;
 503                 bch2_alloc_to_v4(k, &ret->v);
 504         }
 505         return ret;
 506 }
 507
 508 static inline struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut_inlined(struct btree_trans *trans, struct bkey_s_c k)
 509 {
 510         struct bkey_s_c_alloc_v4 a;
 511
 512         if (likely(k.k->type == KEY_TYPE_alloc_v4) &&
 513             ((a = bkey_s_c_to_alloc_v4(k), true) &&
 514              BCH_ALLOC_V4_BACKPOINTERS_START(a.v) == BCH_ALLOC_V4_U64s &&
 515              BCH_ALLOC_V4_NR_BACKPOINTERS(a.v) == 0)) {
 516                 /*
 517                  * Reserve space for one more backpointer here:
 518                  * Not sketchy at doing it this way, nope...
 519                  */
 520                 struct bkey_i_alloc_v4 *ret =
 521                         bch2_trans_kmalloc_nomemzero(trans, bkey_bytes(k.k) + sizeof(struct bch_backpointer));
 522                 if (!IS_ERR(ret))
 523                         bkey_reassemble(&ret->k_i, k);
 524                 return ret;
 525         }
 526
 527         return __bch2_alloc_to_v4_mut(trans, k);
 528 }
 529
 530 struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 531 {
 532         return bch2_alloc_to_v4_mut_inlined(trans, k);
 533 }
 534
 535 struct bkey_i_alloc_v4 *
 536 bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
 537                               struct bpos pos)
 538 {
 539         struct bkey_s_c k;
 540         struct bkey_i_alloc_v4 *a;
 541         int ret;
 542
 543         bch2_trans_iter_init(trans, iter, BTREE_ID_alloc, pos,
 544                              BTREE_ITER_WITH_UPDATES|
 545                              BTREE_ITER_CACHED|
 546                              BTREE_ITER_INTENT);
 547         k = bch2_btree_iter_peek_slot(iter);
 548         ret = bkey_err(k);
 549         if (unlikely(ret))
 550                 goto err;
 551
 552         a = bch2_alloc_to_v4_mut_inlined(trans, k);
 553         ret = PTR_ERR_OR_ZERO(a);
 554         if (unlikely(ret))
 555                 goto err;
 556         return a;
 557 err:
 558         bch2_trans_iter_exit(trans, iter);
 559         return ERR_PTR(ret);
 560 }
 561
 562 int bch2_alloc_read(struct bch_fs *c)
 563 {
 564         struct btree_trans trans;
 565         struct btree_iter iter;
 566         struct bkey_s_c k;
 567         struct bch_alloc_v4 a;
 568         struct bch_dev *ca;
 569         int ret;
 570
 571         bch2_trans_init(&trans, c, 0, 0);
 572
 573         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 574                            BTREE_ITER_PREFETCH, k, ret) {
 575                 /*
 576                  * Not a fsck error because this is checked/repaired by
 577                  * bch2_check_alloc_key() which runs later:
 578                  */
 579                 if (!bch2_dev_bucket_exists(c, k.k->p))
 580                         continue;
 581
 582                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 583
 584                 *bucket_gen(ca, k.k->p.offset) = bch2_alloc_to_v4(k, &a)->gen;
 585         }
 586         bch2_trans_iter_exit(&trans, &iter);
 587
 588         bch2_trans_exit(&trans);
 589
 590         if (ret)
 591                 bch_err(c, "error reading alloc info: %s", bch2_err_str(ret));
 592
 593         return ret;
 594 }
 595
 596 static struct bpos alloc_gens_pos(struct bpos pos, unsigned *offset)
 597 {
 598         *offset = pos.offset & KEY_TYPE_BUCKET_GENS_MASK;
 599
 600         pos.offset >>= KEY_TYPE_BUCKET_GENS_BITS;
 601         return pos;
 602 }
 603
 604 static struct bpos bucket_gens_pos_to_alloc(struct bpos pos, unsigned offset)
 605 {
 606         pos.offset <<= KEY_TYPE_BUCKET_GENS_BITS;
 607         pos.offset += offset;
 608         return pos;
 609 }
 610
 611 static unsigned alloc_gen(struct bkey_s_c k, unsigned offset)
 612 {
 613         return k.k->type == KEY_TYPE_bucket_gens
 614                 ? bkey_s_c_to_bucket_gens(k).v->gens[offset]
 615                 : 0;
 616 }
 617
 618 int bch2_bucket_gens_invalid(const struct bch_fs *c, struct bkey_s_c k,
 619                              unsigned flags, struct printbuf *err)
 620 {
 621         if (bkey_val_bytes(k.k) != sizeof(struct bch_bucket_gens)) {
 622                 prt_printf(err, "bad val size (%lu != %zu)",
 623                        bkey_val_bytes(k.k), sizeof(struct bch_bucket_gens));
 624                 return -BCH_ERR_invalid_bkey;
 625         }
 626
 627         return 0;
 628 }
 629
 630 void bch2_bucket_gens_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 631 {
 632         struct bkey_s_c_bucket_gens g = bkey_s_c_to_bucket_gens(k);
 633         unsigned i;
 634
 635         for (i = 0; i < ARRAY_SIZE(g.v->gens); i++) {
 636                 if (i)
 637                         prt_char(out, ' ');
 638                 prt_printf(out, "%u", g.v->gens[i]);
 639         }
 640 }
 641
 642 int bch2_bucket_gens_init(struct bch_fs *c)
 643 {
 644         struct btree_trans trans;
 645         struct btree_iter iter;
 646         struct bkey_s_c k;
 647         struct bch_alloc_v4 a;
 648         struct bkey_i_bucket_gens g;
 649         bool have_bucket_gens_key = false;
 650         unsigned offset;
 651         struct bpos pos;
 652         u8 gen;
 653         int ret;
 654
 655         bch2_trans_init(&trans, c, 0, 0);
 656
 657         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 658                            BTREE_ITER_PREFETCH, k, ret) {
 659                 /*
 660                  * Not a fsck error because this is checked/repaired by
 661                  * bch2_check_alloc_key() which runs later:
 662                  */
 663                 if (!bch2_dev_bucket_exists(c, k.k->p))
 664                         continue;
 665
 666                 gen = bch2_alloc_to_v4(k, &a)->gen;
 667                 pos = alloc_gens_pos(iter.pos, &offset);
 668
 669                 if (have_bucket_gens_key && bkey_cmp(iter.pos, pos)) {
 670                         ret = commit_do(&trans, NULL, NULL,
 671                                         BTREE_INSERT_NOFAIL|
 672                                         BTREE_INSERT_LAZY_RW,
 673                                 __bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i, 0));
 674                         if (ret)
 675                                 break;
 676                         have_bucket_gens_key = false;
 677                 }
 678
 679                 if (!have_bucket_gens_key) {
 680                         bkey_bucket_gens_init(&g.k_i);
 681                         g.k.p = pos;
 682                         have_bucket_gens_key = true;
 683                 }
 684
 685                 g.v.gens[offset] = gen;
 686         }
 687         bch2_trans_iter_exit(&trans, &iter);
 688
 689         if (have_bucket_gens_key && !ret)
 690                 ret = commit_do(&trans, NULL, NULL,
 691                                 BTREE_INSERT_NOFAIL|
 692                                 BTREE_INSERT_LAZY_RW,
 693                         __bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i, 0));
 694
 695         bch2_trans_exit(&trans);
 696
 697         if (ret)
 698                 bch_err(c, "%s: error %s", __func__, bch2_err_str(ret));
 699
 700         return ret;
 701 }
 702
 703 int bch2_bucket_gens_read(struct bch_fs *c)
 704 {
 705         struct btree_trans trans;
 706         struct btree_iter iter;
 707         struct bkey_s_c k;
 708         const struct bch_bucket_gens *g;
 709         struct bch_dev *ca;
 710         u64 b;
 711         int ret;
 712
 713         bch2_trans_init(&trans, c, 0, 0);
 714
 715         for_each_btree_key(&trans, iter, BTREE_ID_bucket_gens, POS_MIN,
 716                            BTREE_ITER_PREFETCH, k, ret) {
 717                 u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
 718                 u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
 719
 720                 if (k.k->type != KEY_TYPE_bucket_gens)
 721                         continue;
 722
 723                 g = bkey_s_c_to_bucket_gens(k).v;
 724
 725                 /*
 726                  * Not a fsck error because this is checked/repaired by
 727                  * bch2_check_alloc_key() which runs later:
 728                  */
 729                 if (!bch2_dev_exists2(c, k.k->p.inode))
 730                         continue;
 731
 732                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 733
 734                 for (b = max_t(u64, ca->mi.first_bucket, start);
 735                      b < min_t(u64, ca->mi.nbuckets, end);
 736                      b++)
 737                         *bucket_gen(ca, b) = g->gens[b & KEY_TYPE_BUCKET_GENS_MASK];
 738         }
 739         bch2_trans_iter_exit(&trans, &iter);
 740
 741         bch2_trans_exit(&trans);
 742
 743         if (ret)
 744                 bch_err(c, "error reading alloc info: %s", bch2_err_str(ret));
 745
 746         return ret;
 747 }
 748
 749 /* Free space/discard btree: */
 750
 751 static int bch2_bucket_do_index(struct btree_trans *trans,
 752                                 struct bkey_s_c alloc_k,
 753                                 const struct bch_alloc_v4 *a,
 754                                 bool set)
 755 {
 756         struct bch_fs *c = trans->c;
 757         struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 758         struct btree_iter iter;
 759         struct bkey_s_c old;
 760         struct bkey_i *k;
 761         enum btree_id btree;
 762         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 763         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 764         struct printbuf buf = PRINTBUF;
 765         int ret;
 766
 767         if (a->data_type != BCH_DATA_free &&
 768             a->data_type != BCH_DATA_need_discard)
 769                 return 0;
 770
 771         k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
 772         if (IS_ERR(k))
 773                 return PTR_ERR(k);
 774
 775         bkey_init(&k->k);
 776         k->k.type = new_type;
 777
 778         switch (a->data_type) {
 779         case BCH_DATA_free:
 780                 btree = BTREE_ID_freespace;
 781                 k->k.p = alloc_freespace_pos(alloc_k.k->p, *a);
 782                 bch2_key_resize(&k->k, 1);
 783                 break;
 784         case BCH_DATA_need_discard:
 785                 btree = BTREE_ID_need_discard;
 786                 k->k.p = alloc_k.k->p;
 787                 break;
 788         default:
 789                 return 0;
 790         }
 791
 792         bch2_trans_iter_init(trans, &iter, btree,
 793                              bkey_start_pos(&k->k),
 794                              BTREE_ITER_INTENT);
 795         old = bch2_btree_iter_peek_slot(&iter);
 796         ret = bkey_err(old);
 797         if (ret)
 798                 goto err;
 799
 800         if (ca->mi.freespace_initialized &&
 801             test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags) &&
 802             bch2_trans_inconsistent_on(old.k->type != old_type, trans,
 803                         "incorrect key when %s %s btree (got %s should be %s)\n"
 804                         "  for %s",
 805                         set ? "setting" : "clearing",
 806                         bch2_btree_ids[btree],
 807                         bch2_bkey_types[old.k->type],
 808                         bch2_bkey_types[old_type],
 809                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 810                 ret = -EIO;
 811                 goto err;
 812         }
 813
 814         ret = bch2_trans_update(trans, &iter, k, 0);
 815 err:
 816         bch2_trans_iter_exit(trans, &iter);
 817         printbuf_exit(&buf);
 818         return ret;
 819 }
 820
 821 static noinline int bch2_bucket_gen_update(struct btree_trans *trans,
 822                                            struct bpos bucket, u8 gen)
 823 {
 824         struct btree_iter iter;
 825         unsigned offset;
 826         struct bpos pos = alloc_gens_pos(bucket, &offset);
 827         struct bkey_i_bucket_gens *g;
 828         struct bkey_s_c k;
 829         int ret;
 830
 831         g = bch2_trans_kmalloc(trans, sizeof(*g));
 832         ret = PTR_ERR_OR_ZERO(g);
 833         if (ret)
 834                 return ret;
 835
 836         bch2_trans_iter_init(trans, &iter, BTREE_ID_bucket_gens, pos,
 837                              BTREE_ITER_INTENT|
 838                              BTREE_ITER_WITH_UPDATES);
 839         k = bch2_btree_iter_peek_slot(&iter);
 840         ret = bkey_err(k);
 841         if (ret)
 842                 goto err;
 843
 844         if (k.k->type != KEY_TYPE_bucket_gens) {
 845                 bkey_bucket_gens_init(&g->k_i);
 846                 g->k.p = iter.pos;
 847         } else {
 848                 bkey_reassemble(&g->k_i, k);
 849         }
 850
 851         g->v.gens[offset] = gen;
 852
 853         ret = bch2_trans_update(trans, &iter, &g->k_i, 0);
 854 err:
 855         bch2_trans_iter_exit(trans, &iter);
 856         return ret;
 857 }
 858
 859 int bch2_trans_mark_alloc(struct btree_trans *trans,
 860                           enum btree_id btree_id, unsigned level,
 861                           struct bkey_s_c old, struct bkey_i *new,
 862                           unsigned flags)
 863 {
 864         struct bch_fs *c = trans->c;
 865         struct bch_alloc_v4 old_a_convert, *new_a;
 866         const struct bch_alloc_v4 *old_a;
 867         u64 old_lru, new_lru;
 868         int ret = 0;
 869
 870         /*
 871          * Deletion only happens in the device removal path, with
 872          * BTREE_TRIGGER_NORUN:
 873          */
 874         BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
 875
 876         old_a = bch2_alloc_to_v4(old, &old_a_convert);
 877         new_a = &bkey_i_to_alloc_v4(new)->v;
 878
 879         new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
 880
 881         if (new_a->dirty_sectors > old_a->dirty_sectors ||
 882             new_a->cached_sectors > old_a->cached_sectors) {
 883                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 884                 new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 885                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
 886                 SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
 887         }
 888
 889         if (data_type_is_empty(new_a->data_type) &&
 890             BCH_ALLOC_V4_NEED_INC_GEN(new_a) &&
 891             !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
 892                 new_a->gen++;
 893                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
 894         }
 895
 896         if (old_a->data_type != new_a->data_type ||
 897             (new_a->data_type == BCH_DATA_free &&
 898              alloc_freespace_genbits(*old_a) != alloc_freespace_genbits(*new_a))) {
 899                 ret =   bch2_bucket_do_index(trans, old, old_a, false) ?:
 900                         bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_a, true);
 901                 if (ret)
 902                         return ret;
 903         }
 904
 905         if (new_a->data_type == BCH_DATA_cached &&
 906             !new_a->io_time[READ])
 907                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 908
 909         old_lru = alloc_lru_idx_read(*old_a);
 910         new_lru = alloc_lru_idx_read(*new_a);
 911
 912         if (old_lru != new_lru) {
 913                 ret = bch2_lru_change(trans, new->k.p.inode,
 914                                       bucket_to_u64(new->k.p),
 915                                       old_lru, new_lru);
 916                 if (ret)
 917                         return ret;
 918         }
 919
 920         new_a->fragmentation_lru = alloc_lru_idx_fragmentation(*new_a,
 921                                         bch_dev_bkey_exists(c, new->k.p.inode));
 922
 923         if (old_a->fragmentation_lru != new_a->fragmentation_lru) {
 924                 ret = bch2_lru_change(trans,
 925                                 BCH_LRU_FRAGMENTATION_START,
 926                                 bucket_to_u64(new->k.p),
 927                                 old_a->fragmentation_lru, new_a->fragmentation_lru);
 928                 if (ret)
 929                         return ret;
 930         }
 931
 932         if (old_a->gen != new_a->gen) {
 933                 ret = bch2_bucket_gen_update(trans, new->k.p, new_a->gen);
 934                 if (ret)
 935                         return ret;
 936         }
 937
 938         return 0;
 939 }
 940
 941 /*
 942  * This synthesizes deleted extents for holes, similar to BTREE_ITER_SLOTS for
 943  * extents style btrees, but works on non-extents btrees:
 944  */
 945 struct bkey_s_c bch2_get_key_or_hole(struct btree_iter *iter, struct bpos end, struct bkey *hole)
 946 {
 947         struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
 948
 949         if (bkey_err(k))
 950                 return k;
 951
 952         if (k.k->type) {
 953                 return k;
 954         } else {
 955                 struct btree_iter iter2;
 956                 struct bpos next;
 957
 958                 bch2_trans_copy_iter(&iter2, iter);
 959
 960                 if (!bpos_eq(iter->path->l[0].b->key.k.p, SPOS_MAX))
 961                         end = bkey_min(end, bpos_nosnap_successor(iter->path->l[0].b->key.k.p));
 962
 963                 end = bkey_min(end, POS(iter->pos.inode, iter->pos.offset + U32_MAX - 1));
 964
 965                 /*
 966                  * btree node min/max is a closed interval, upto takes a half
 967                  * open interval:
 968                  */
 969                 k = bch2_btree_iter_peek_upto(&iter2, end);
 970                 next = iter2.pos;
 971                 bch2_trans_iter_exit(iter->trans, &iter2);
 972
 973                 BUG_ON(next.offset >= iter->pos.offset + U32_MAX);
 974
 975                 if (bkey_err(k))
 976                         return k;
 977
 978                 bkey_init(hole);
 979                 hole->p = iter->pos;
 980
 981                 bch2_key_resize(hole, next.offset - iter->pos.offset);
 982                 return (struct bkey_s_c) { hole, NULL };
 983         }
 984 }
 985
 986 static bool next_bucket(struct bch_fs *c, struct bpos *bucket)
 987 {
 988         struct bch_dev *ca;
 989         unsigned iter;
 990
 991         if (bch2_dev_bucket_exists(c, *bucket))
 992                 return true;
 993
 994         if (bch2_dev_exists2(c, bucket->inode)) {
 995                 ca = bch_dev_bkey_exists(c, bucket->inode);
 996
 997                 if (bucket->offset < ca->mi.first_bucket) {
 998                         bucket->offset = ca->mi.first_bucket;
 999                         return true;
1000                 }
1001
1002                 bucket->inode++;
1003                 bucket->offset = 0;
1004         }
1005
1006         rcu_read_lock();
1007         iter = bucket->inode;
1008         ca = __bch2_next_dev(c, &iter, NULL);
1009         if (ca)
1010                 *bucket = POS(ca->dev_idx, ca->mi.first_bucket);
1011         rcu_read_unlock();
1012
1013         return ca != NULL;
1014 }
1015
1016 struct bkey_s_c bch2_get_key_or_real_bucket_hole(struct btree_iter *iter, struct bkey *hole)
1017 {
1018         struct bch_fs *c = iter->trans->c;
1019         struct bkey_s_c k;
1020 again:
1021         k = bch2_get_key_or_hole(iter, POS_MAX, hole);
1022         if (bkey_err(k))
1023                 return k;
1024
1025         if (!k.k->type) {
1026                 struct bpos bucket = bkey_start_pos(k.k);
1027
1028                 if (!bch2_dev_bucket_exists(c, bucket)) {
1029                         if (!next_bucket(c, &bucket))
1030                                 return bkey_s_c_null;
1031
1032                         bch2_btree_iter_set_pos(iter, bucket);
1033                         goto again;
1034                 }
1035
1036                 if (!bch2_dev_bucket_exists(c, k.k->p)) {
1037                         struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
1038
1039                         bch2_key_resize(hole, ca->mi.nbuckets - bucket.offset);
1040                 }
1041         }
1042
1043         return k;
1044 }
1045
1046 static int bch2_check_alloc_key(struct btree_trans *trans,
1047                                 struct bkey_s_c alloc_k,
1048                                 struct btree_iter *alloc_iter,
1049                                 struct btree_iter *discard_iter,
1050                                 struct btree_iter *freespace_iter,
1051                                 struct btree_iter *bucket_gens_iter)
1052 {
1053         struct bch_fs *c = trans->c;
1054         struct bch_dev *ca;
1055         struct bch_alloc_v4 a_convert;
1056         const struct bch_alloc_v4 *a;
1057         unsigned discard_key_type, freespace_key_type;
1058         unsigned gens_offset;
1059         struct bkey_s_c k;
1060         struct printbuf buf = PRINTBUF;
1061         int ret;
1062
1063         if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
1064                         "alloc key for invalid device:bucket %llu:%llu",
1065                         alloc_k.k->p.inode, alloc_k.k->p.offset))
1066                 return bch2_btree_delete_at(trans, alloc_iter, 0);
1067
1068         ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
1069         if (!ca->mi.freespace_initialized)
1070                 return 0;
1071
1072         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1073
1074         discard_key_type = a->data_type == BCH_DATA_need_discard ? KEY_TYPE_set : 0;
1075         bch2_btree_iter_set_pos(discard_iter, alloc_k.k->p);
1076         k = bch2_btree_iter_peek_slot(discard_iter);
1077         ret = bkey_err(k);
1078         if (ret)
1079                 goto err;
1080
1081         if (k.k->type != discard_key_type &&
1082             (c->opts.reconstruct_alloc ||
1083              fsck_err(c, "incorrect key in need_discard btree (got %s should be %s)\n"
1084                       "  %s",
1085                       bch2_bkey_types[k.k->type],
1086                       bch2_bkey_types[discard_key_type],
1087                       (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1088                 struct bkey_i *update =
1089                         bch2_trans_kmalloc(trans, sizeof(*update));
1090
1091                 ret = PTR_ERR_OR_ZERO(update);
1092                 if (ret)
1093                         goto err;
1094
1095                 bkey_init(&update->k);
1096                 update->k.type  = discard_key_type;
1097                 update->k.p     = discard_iter->pos;
1098
1099                 ret = bch2_trans_update(trans, discard_iter, update, 0);
1100                 if (ret)
1101                         goto err;
1102         }
1103
1104         freespace_key_type = a->data_type == BCH_DATA_free ? KEY_TYPE_set : 0;
1105         bch2_btree_iter_set_pos(freespace_iter, alloc_freespace_pos(alloc_k.k->p, *a));
1106         k = bch2_btree_iter_peek_slot(freespace_iter);
1107         ret = bkey_err(k);
1108         if (ret)
1109                 goto err;
1110
1111         if (k.k->type != freespace_key_type &&
1112             (c->opts.reconstruct_alloc ||
1113              fsck_err(c, "incorrect key in freespace btree (got %s should be %s)\n"
1114                       "  %s",
1115                       bch2_bkey_types[k.k->type],
1116                       bch2_bkey_types[freespace_key_type],
1117                       (printbuf_reset(&buf),
1118                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1119                 struct bkey_i *update =
1120                         bch2_trans_kmalloc(trans, sizeof(*update));
1121
1122                 ret = PTR_ERR_OR_ZERO(update);
1123                 if (ret)
1124                         goto err;
1125
1126                 bkey_init(&update->k);
1127                 update->k.type  = freespace_key_type;
1128                 update->k.p     = freespace_iter->pos;
1129                 bch2_key_resize(&update->k, 1);
1130
1131                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1132                 if (ret)
1133                         goto err;
1134         }
1135
1136         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(alloc_k.k->p, &gens_offset));
1137         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1138         ret = bkey_err(k);
1139         if (ret)
1140                 goto err;
1141
1142         if (a->gen != alloc_gen(k, gens_offset) &&
1143             (c->opts.reconstruct_alloc ||
1144              fsck_err(c, "incorrect gen in bucket_gens btree (got %u should be %u)\n"
1145                       "  %s",
1146                       alloc_gen(k, gens_offset), a->gen,
1147                       (printbuf_reset(&buf),
1148                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1149                 struct bkey_i_bucket_gens *g =
1150                         bch2_trans_kmalloc(trans, sizeof(*g));
1151
1152                 ret = PTR_ERR_OR_ZERO(g);
1153                 if (ret)
1154                         goto err;
1155
1156                 if (k.k->type == KEY_TYPE_bucket_gens) {
1157                         bkey_reassemble(&g->k_i, k);
1158                 } else {
1159                         bkey_bucket_gens_init(&g->k_i);
1160                         g->k.p = alloc_gens_pos(alloc_k.k->p, &gens_offset);
1161                 }
1162
1163                 g->v.gens[gens_offset] = a->gen;
1164
1165                 ret = bch2_trans_update(trans, bucket_gens_iter, &g->k_i, 0);
1166                 if (ret)
1167                         goto err;
1168         }
1169 err:
1170 fsck_err:
1171         printbuf_exit(&buf);
1172         return ret;
1173 }
1174
1175 static int bch2_check_alloc_hole_freespace(struct btree_trans *trans,
1176                                  struct bpos start,
1177                                  struct bpos *end,
1178                                  struct btree_iter *freespace_iter)
1179 {
1180         struct bch_fs *c = trans->c;
1181         struct bch_dev *ca;
1182         struct bkey_s_c k;
1183         struct printbuf buf = PRINTBUF;
1184         int ret;
1185
1186         ca = bch_dev_bkey_exists(c, start.inode);
1187         if (!ca->mi.freespace_initialized)
1188                 return 0;
1189
1190         bch2_btree_iter_set_pos(freespace_iter, start);
1191
1192         k = bch2_btree_iter_peek_slot(freespace_iter);
1193         ret = bkey_err(k);
1194         if (ret)
1195                 goto err;
1196
1197         *end = bkey_min(k.k->p, *end);
1198
1199         if (k.k->type != KEY_TYPE_set &&
1200             (c->opts.reconstruct_alloc ||
1201              fsck_err(c, "hole in alloc btree missing in freespace btree\n"
1202                       "  device %llu buckets %llu-%llu",
1203                       freespace_iter->pos.inode,
1204                       freespace_iter->pos.offset,
1205                       end->offset))) {
1206                 struct bkey_i *update =
1207                         bch2_trans_kmalloc(trans, sizeof(*update));
1208
1209                 ret = PTR_ERR_OR_ZERO(update);
1210                 if (ret)
1211                         goto err;
1212
1213                 bkey_init(&update->k);
1214                 update->k.type  = KEY_TYPE_set;
1215                 update->k.p     = freespace_iter->pos;
1216                 bch2_key_resize(&update->k,
1217                                 min_t(u64, U32_MAX, end->offset -
1218                                       freespace_iter->pos.offset));
1219
1220                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1221                 if (ret)
1222                         goto err;
1223         }
1224 err:
1225 fsck_err:
1226         printbuf_exit(&buf);
1227         return ret;
1228 }
1229
1230 static int bch2_check_alloc_hole_bucket_gens(struct btree_trans *trans,
1231                                  struct bpos start,
1232                                  struct bpos *end,
1233                                  struct btree_iter *bucket_gens_iter)
1234 {
1235         struct bch_fs *c = trans->c;
1236         struct bkey_s_c k;
1237         struct printbuf buf = PRINTBUF;
1238         unsigned i, gens_offset, gens_end_offset;
1239         int ret;
1240
1241         if (c->sb.version < bcachefs_metadata_version_bucket_gens &&
1242             !c->opts.version_upgrade)
1243                 return 0;
1244
1245         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(start, &gens_offset));
1246
1247         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1248         ret = bkey_err(k);
1249         if (ret)
1250                 goto err;
1251
1252         if (bkey_cmp(alloc_gens_pos(start, &gens_offset),
1253                      alloc_gens_pos(*end,  &gens_end_offset)))
1254                 gens_end_offset = KEY_TYPE_BUCKET_GENS_NR;
1255
1256         if (k.k->type == KEY_TYPE_bucket_gens) {
1257                 struct bkey_i_bucket_gens g;
1258                 bool need_update = false;
1259
1260                 bkey_reassemble(&g.k_i, k);
1261
1262                 for (i = gens_offset; i < gens_end_offset; i++) {
1263                         if (fsck_err_on(g.v.gens[i], c,
1264                                         "hole in alloc btree at %llu:%llu with nonzero gen in bucket_gens btree (%u)",
1265                                         bucket_gens_pos_to_alloc(k.k->p, i).inode,
1266                                         bucket_gens_pos_to_alloc(k.k->p, i).offset,
1267                                         g.v.gens[i])) {
1268                                 g.v.gens[i] = 0;
1269                                 need_update = true;
1270                         }
1271                 }
1272
1273                 if (need_update) {
1274                         struct bkey_i *k = bch2_trans_kmalloc(trans, sizeof(g));
1275
1276                         ret = PTR_ERR_OR_ZERO(k);
1277                         if (ret)
1278                                 goto err;
1279
1280                         memcpy(k, &g, sizeof(g));
1281
1282                         ret = bch2_trans_update(trans, bucket_gens_iter, k, 0);
1283                         if (ret)
1284                                 goto err;
1285                 }
1286         }
1287
1288         *end = bkey_min(*end, bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0));
1289 err:
1290 fsck_err:
1291         printbuf_exit(&buf);
1292         return ret;
1293 }
1294
1295 static int bch2_check_discard_freespace_key(struct btree_trans *trans,
1296                                             struct btree_iter *iter)
1297 {
1298         struct bch_fs *c = trans->c;
1299         struct btree_iter alloc_iter;
1300         struct bkey_s_c alloc_k;
1301         struct bch_alloc_v4 a_convert;
1302         const struct bch_alloc_v4 *a;
1303         u64 genbits;
1304         struct bpos pos;
1305         enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
1306                 ? BCH_DATA_need_discard
1307                 : BCH_DATA_free;
1308         struct printbuf buf = PRINTBUF;
1309         int ret;
1310
1311         pos = iter->pos;
1312         pos.offset &= ~(~0ULL << 56);
1313         genbits = iter->pos.offset & (~0ULL << 56);
1314
1315         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
1316
1317         if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
1318                         "entry in %s btree for nonexistant dev:bucket %llu:%llu",
1319                         bch2_btree_ids[iter->btree_id], pos.inode, pos.offset))
1320                 goto delete;
1321
1322         alloc_k = bch2_btree_iter_peek_slot(&alloc_iter);
1323         ret = bkey_err(alloc_k);
1324         if (ret)
1325                 goto err;
1326
1327         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1328
1329         if (fsck_err_on(a->data_type != state ||
1330                         (state == BCH_DATA_free &&
1331                          genbits != alloc_freespace_genbits(*a)), c,
1332                         "%s\n  incorrectly set in %s index (free %u, genbits %llu should be %llu)",
1333                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
1334                         bch2_btree_ids[iter->btree_id],
1335                         a->data_type == state,
1336                         genbits >> 56, alloc_freespace_genbits(*a) >> 56))
1337                 goto delete;
1338 out:
1339 err:
1340 fsck_err:
1341         bch2_trans_iter_exit(trans, &alloc_iter);
1342         printbuf_exit(&buf);
1343         return ret;
1344 delete:
1345         ret = bch2_btree_delete_extent_at(trans, iter,
1346                         iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0);
1347         goto out;
1348 }
1349
1350 /*
1351  * We've already checked that generation numbers in the bucket_gens btree are
1352  * valid for buckets that exist; this just checks for keys for nonexistent
1353  * buckets.
1354  */
1355 static int bch2_check_bucket_gens_key(struct btree_trans *trans,
1356                                       struct btree_iter *iter,
1357                                       struct bkey_s_c k)
1358 {
1359         struct bch_fs *c = trans->c;
1360         struct bkey_i_bucket_gens g;
1361         struct bch_dev *ca;
1362         u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
1363         u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
1364         u64 b;
1365         bool need_update = false, dev_exists;
1366         struct printbuf buf = PRINTBUF;
1367         int ret = 0;
1368
1369         BUG_ON(k.k->type != KEY_TYPE_bucket_gens);
1370         bkey_reassemble(&g.k_i, k);
1371
1372         /* if no bch_dev, skip out whether we repair or not */
1373         dev_exists = bch2_dev_exists2(c, k.k->p.inode);
1374         if (!dev_exists) {
1375                 if (fsck_err_on(!dev_exists, c,
1376                                 "bucket_gens key for invalid device:\n  %s",
1377                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1378                         ret = bch2_btree_delete_at(trans, iter, 0);
1379                 }
1380                 goto out;
1381         }
1382
1383         ca = bch_dev_bkey_exists(c, k.k->p.inode);
1384         if (fsck_err_on(end <= ca->mi.first_bucket ||
1385                         start >= ca->mi.nbuckets, c,
1386                         "bucket_gens key for invalid buckets:\n  %s",
1387                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1388                 ret = bch2_btree_delete_at(trans, iter, 0);
1389                 goto out;
1390         }
1391
1392         for (b = start; b < ca->mi.first_bucket; b++)
1393                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1394                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1395                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1396                         need_update = true;
1397                 }
1398
1399         for (b = ca->mi.nbuckets; b < end; b++)
1400                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
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 *k;
1408
1409                 k = bch2_trans_kmalloc(trans, sizeof(g));
1410                 ret = PTR_ERR_OR_ZERO(k);
1411                 if (ret)
1412                         goto out;
1413
1414                 memcpy(k, &g, sizeof(g));
1415                 ret = bch2_trans_update(trans, iter, k, 0);
1416         }
1417 out:
1418 fsck_err:
1419         printbuf_exit(&buf);
1420         return ret;
1421 }
1422
1423 int bch2_check_alloc_info(struct bch_fs *c)
1424 {
1425         struct btree_trans trans;
1426         struct btree_iter iter, discard_iter, freespace_iter, bucket_gens_iter;
1427         struct bkey hole;
1428         struct bkey_s_c k;
1429         int ret = 0;
1430
1431         bch2_trans_init(&trans, c, 0, 0);
1432
1433         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc, POS_MIN,
1434                              BTREE_ITER_PREFETCH);
1435         bch2_trans_iter_init(&trans, &discard_iter, BTREE_ID_need_discard, POS_MIN,
1436                              BTREE_ITER_PREFETCH);
1437         bch2_trans_iter_init(&trans, &freespace_iter, BTREE_ID_freespace, POS_MIN,
1438                              BTREE_ITER_PREFETCH);
1439         bch2_trans_iter_init(&trans, &bucket_gens_iter, BTREE_ID_bucket_gens, POS_MIN,
1440                              BTREE_ITER_PREFETCH);
1441
1442         while (1) {
1443                 struct bpos next;
1444
1445                 bch2_trans_begin(&trans);
1446
1447                 k = bch2_get_key_or_real_bucket_hole(&iter, &hole);
1448                 ret = bkey_err(k);
1449                 if (ret)
1450                         goto bkey_err;
1451
1452                 if (!k.k)
1453                         break;
1454
1455                 if (k.k->type) {
1456                         next = bpos_nosnap_successor(k.k->p);
1457
1458                         ret = bch2_check_alloc_key(&trans,
1459                                                    k, &iter,
1460                                                    &discard_iter,
1461                                                    &freespace_iter,
1462                                                    &bucket_gens_iter);
1463                         if (ret)
1464                                 goto bkey_err;
1465                 } else {
1466                         next = k.k->p;
1467
1468                         ret = bch2_check_alloc_hole_freespace(&trans,
1469                                                     bkey_start_pos(k.k),
1470                                                     &next,
1471                                                     &freespace_iter) ?:
1472                                 bch2_check_alloc_hole_bucket_gens(&trans,
1473                                                     bkey_start_pos(k.k),
1474                                                     &next,
1475                                                     &bucket_gens_iter);
1476                         if (ret)
1477                                 goto bkey_err;
1478                 }
1479
1480                 ret = bch2_trans_commit(&trans, NULL, NULL,
1481                                         BTREE_INSERT_NOFAIL|
1482                                         BTREE_INSERT_LAZY_RW);
1483                 if (ret)
1484                         goto bkey_err;
1485
1486                 bch2_btree_iter_set_pos(&iter, next);
1487 bkey_err:
1488                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1489                         continue;
1490                 if (ret)
1491                         break;
1492         }
1493         bch2_trans_iter_exit(&trans, &bucket_gens_iter);
1494         bch2_trans_iter_exit(&trans, &freespace_iter);
1495         bch2_trans_iter_exit(&trans, &discard_iter);
1496         bch2_trans_iter_exit(&trans, &iter);
1497
1498         if (ret < 0)
1499                 goto err;
1500
1501         ret = for_each_btree_key_commit(&trans, iter,
1502                         BTREE_ID_need_discard, POS_MIN,
1503                         BTREE_ITER_PREFETCH, k,
1504                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1505                 bch2_check_discard_freespace_key(&trans, &iter)) ?:
1506               for_each_btree_key_commit(&trans, iter,
1507                         BTREE_ID_freespace, POS_MIN,
1508                         BTREE_ITER_PREFETCH, k,
1509                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1510                 bch2_check_discard_freespace_key(&trans, &iter)) ?:
1511               for_each_btree_key_commit(&trans, iter,
1512                         BTREE_ID_bucket_gens, POS_MIN,
1513                         BTREE_ITER_PREFETCH, k,
1514                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1515                 bch2_check_bucket_gens_key(&trans, &iter, k));
1516 err:
1517         bch2_trans_exit(&trans);
1518         return ret < 0 ? ret : 0;
1519 }
1520
1521 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
1522                                        struct btree_iter *alloc_iter)
1523 {
1524         struct bch_fs *c = trans->c;
1525         struct btree_iter lru_iter;
1526         struct bch_alloc_v4 a_convert;
1527         const struct bch_alloc_v4 *a;
1528         struct bkey_s_c alloc_k, k;
1529         struct printbuf buf = PRINTBUF;
1530         int ret;
1531
1532         alloc_k = bch2_btree_iter_peek(alloc_iter);
1533         if (!alloc_k.k)
1534                 return 0;
1535
1536         ret = bkey_err(alloc_k);
1537         if (ret)
1538                 return ret;
1539
1540         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1541
1542         if (a->data_type != BCH_DATA_cached)
1543                 return 0;
1544
1545         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
1546                              lru_pos(alloc_k.k->p.inode,
1547                                      bucket_to_u64(alloc_k.k->p),
1548                                      a->io_time[READ]), 0);
1549         k = bch2_btree_iter_peek_slot(&lru_iter);
1550         ret = bkey_err(k);
1551         if (ret)
1552                 goto err;
1553
1554         if (fsck_err_on(!a->io_time[READ], c,
1555                         "cached bucket with read_time 0\n"
1556                         "  %s",
1557                 (printbuf_reset(&buf),
1558                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) ||
1559             fsck_err_on(k.k->type != KEY_TYPE_set, c,
1560                         "missing lru entry\n"
1561                         "  %s",
1562                         (printbuf_reset(&buf),
1563                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
1564                 u64 read_time = a->io_time[READ] ?:
1565                         atomic64_read(&c->io_clock[READ].now);
1566
1567                 ret = bch2_lru_set(trans,
1568                                    alloc_k.k->p.inode,
1569                                    bucket_to_u64(alloc_k.k->p),
1570                                    read_time);
1571                 if (ret)
1572                         goto err;
1573
1574                 if (a->io_time[READ] != read_time) {
1575                         struct bkey_i_alloc_v4 *a_mut =
1576                                 bch2_alloc_to_v4_mut(trans, alloc_k);
1577                         ret = PTR_ERR_OR_ZERO(a_mut);
1578                         if (ret)
1579                                 goto err;
1580
1581                         a_mut->v.io_time[READ] = read_time;
1582                         ret = bch2_trans_update(trans, alloc_iter,
1583                                                 &a_mut->k_i, BTREE_TRIGGER_NORUN);
1584                         if (ret)
1585                                 goto err;
1586                 }
1587         }
1588 err:
1589 fsck_err:
1590         bch2_trans_iter_exit(trans, &lru_iter);
1591         printbuf_exit(&buf);
1592         return ret;
1593 }
1594
1595 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
1596 {
1597         struct btree_trans trans;
1598         struct btree_iter iter;
1599         struct bkey_s_c k;
1600         int ret = 0;
1601
1602         bch2_trans_init(&trans, c, 0, 0);
1603
1604         for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1605                         POS_MIN, BTREE_ITER_PREFETCH, k,
1606                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1607                 bch2_check_alloc_to_lru_ref(&trans, &iter));
1608
1609         bch2_trans_exit(&trans);
1610         return ret < 0 ? ret : 0;
1611 }
1612
1613 static int bch2_discard_one_bucket(struct btree_trans *trans,
1614                                    struct btree_iter *need_discard_iter,
1615                                    struct bpos *discard_pos_done,
1616                                    u64 *seen,
1617                                    u64 *open,
1618                                    u64 *need_journal_commit,
1619                                    u64 *discarded)
1620 {
1621         struct bch_fs *c = trans->c;
1622         struct bpos pos = need_discard_iter->pos;
1623         struct btree_iter iter = { NULL };
1624         struct bkey_s_c k;
1625         struct bch_dev *ca;
1626         struct bkey_i_alloc_v4 *a;
1627         struct printbuf buf = PRINTBUF;
1628         int ret = 0;
1629
1630         ca = bch_dev_bkey_exists(c, pos.inode);
1631         if (!percpu_ref_tryget(&ca->io_ref)) {
1632                 bch2_btree_iter_set_pos(need_discard_iter, POS(pos.inode + 1, 0));
1633                 return 0;
1634         }
1635
1636         if (bch2_bucket_is_open_safe(c, pos.inode, pos.offset)) {
1637                 (*open)++;
1638                 goto out;
1639         }
1640
1641         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
1642                         c->journal.flushed_seq_ondisk,
1643                         pos.inode, pos.offset)) {
1644                 (*need_journal_commit)++;
1645                 goto out;
1646         }
1647
1648         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
1649                              need_discard_iter->pos,
1650                              BTREE_ITER_CACHED);
1651         k = bch2_btree_iter_peek_slot(&iter);
1652         ret = bkey_err(k);
1653         if (ret)
1654                 goto out;
1655
1656         a = bch2_alloc_to_v4_mut(trans, k);
1657         ret = PTR_ERR_OR_ZERO(a);
1658         if (ret)
1659                 goto out;
1660
1661         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
1662                 a->v.gen++;
1663                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1664                 goto write;
1665         }
1666
1667         if (a->v.journal_seq > c->journal.flushed_seq_ondisk) {
1668                 if (test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
1669                         bch2_trans_inconsistent(trans,
1670                                 "clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
1671                                 "%s",
1672                                 a->v.journal_seq,
1673                                 c->journal.flushed_seq_ondisk,
1674                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1675                         ret = -EIO;
1676                 }
1677                 goto out;
1678         }
1679
1680         if (a->v.data_type != BCH_DATA_need_discard) {
1681                 if (test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
1682                         bch2_trans_inconsistent(trans,
1683                                 "bucket incorrectly set in need_discard btree\n"
1684                                 "%s",
1685                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1686                         ret = -EIO;
1687                 }
1688
1689                 goto out;
1690         }
1691
1692         if (!bkey_eq(*discard_pos_done, iter.pos) &&
1693             ca->mi.discard && !c->opts.nochanges) {
1694                 /*
1695                  * This works without any other locks because this is the only
1696                  * thread that removes items from the need_discard tree
1697                  */
1698                 bch2_trans_unlock(trans);
1699                 blkdev_issue_discard(ca->disk_sb.bdev,
1700                                      k.k->p.offset * ca->mi.bucket_size,
1701                                      ca->mi.bucket_size,
1702                                      GFP_KERNEL);
1703                 *discard_pos_done = iter.pos;
1704
1705                 ret = bch2_trans_relock_notrace(trans);
1706                 if (ret)
1707                         goto out;
1708         }
1709
1710         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
1711         a->v.data_type = alloc_data_type(a->v, a->v.data_type);
1712 write:
1713         ret =   bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1714                 bch2_trans_commit(trans, NULL, NULL,
1715                                   BTREE_INSERT_USE_RESERVE|BTREE_INSERT_NOFAIL);
1716         if (ret)
1717                 goto out;
1718
1719         this_cpu_inc(c->counters[BCH_COUNTER_bucket_discard]);
1720         (*discarded)++;
1721 out:
1722         (*seen)++;
1723         bch2_trans_iter_exit(trans, &iter);
1724         percpu_ref_put(&ca->io_ref);
1725         printbuf_exit(&buf);
1726         return ret;
1727 }
1728
1729 static void bch2_do_discards_work(struct work_struct *work)
1730 {
1731         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
1732         struct btree_trans trans;
1733         struct btree_iter iter;
1734         struct bkey_s_c k;
1735         u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
1736         struct bpos discard_pos_done = POS_MAX;
1737         int ret;
1738
1739         bch2_trans_init(&trans, c, 0, 0);
1740
1741         /*
1742          * We're doing the commit in bch2_discard_one_bucket instead of using
1743          * for_each_btree_key_commit() so that we can increment counters after
1744          * successful commit:
1745          */
1746         ret = for_each_btree_key2(&trans, iter,
1747                         BTREE_ID_need_discard, POS_MIN, 0, k,
1748                 bch2_discard_one_bucket(&trans, &iter, &discard_pos_done,
1749                                         &seen,
1750                                         &open,
1751                                         &need_journal_commit,
1752                                         &discarded));
1753
1754         bch2_trans_exit(&trans);
1755
1756         if (need_journal_commit * 2 > seen)
1757                 bch2_journal_flush_async(&c->journal, NULL);
1758
1759         bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1760
1761         trace_discard_buckets(c, seen, open, need_journal_commit, discarded,
1762                               bch2_err_str(ret));
1763 }
1764
1765 void bch2_do_discards(struct bch_fs *c)
1766 {
1767         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_discard) &&
1768             !queue_work(c->write_ref_wq, &c->discard_work))
1769                 bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1770 }
1771
1772 static int invalidate_one_bucket(struct btree_trans *trans,
1773                                  struct btree_iter *lru_iter,
1774                                  struct bkey_s_c lru_k,
1775                                  s64 *nr_to_invalidate)
1776 {
1777         struct bch_fs *c = trans->c;
1778         struct btree_iter alloc_iter = { NULL };
1779         struct bkey_i_alloc_v4 *a = NULL;
1780         struct printbuf buf = PRINTBUF;
1781         struct bpos bucket = u64_to_bucket(lru_k.k->p.offset);
1782         unsigned cached_sectors;
1783         int ret = 0;
1784
1785         if (*nr_to_invalidate <= 0)
1786                 return 1;
1787
1788         if (!bch2_dev_bucket_exists(c, bucket)) {
1789                 prt_str(&buf, "lru entry points to invalid bucket");
1790                 goto err;
1791         }
1792
1793         if (bch2_bucket_is_open_safe(c, bucket.inode, bucket.offset))
1794                 return 0;
1795
1796         a = bch2_trans_start_alloc_update(trans, &alloc_iter, bucket);
1797         ret = PTR_ERR_OR_ZERO(a);
1798         if (ret)
1799                 goto out;
1800
1801         /* We expect harmless races here due to the btree write buffer: */
1802         if (lru_pos_time(lru_iter->pos) != alloc_lru_idx_read(a->v))
1803                 goto out;
1804
1805         BUG_ON(a->v.data_type != BCH_DATA_cached);
1806
1807         if (!a->v.cached_sectors)
1808                 bch_err(c, "invalidating empty bucket, confused");
1809
1810         cached_sectors = a->v.cached_sectors;
1811
1812         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1813         a->v.gen++;
1814         a->v.data_type          = 0;
1815         a->v.dirty_sectors      = 0;
1816         a->v.cached_sectors     = 0;
1817         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1818         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1819
1820         ret =   bch2_trans_update(trans, &alloc_iter, &a->k_i,
1821                                 BTREE_TRIGGER_BUCKET_INVALIDATE) ?:
1822                 bch2_trans_commit(trans, NULL, NULL,
1823                                   BTREE_INSERT_USE_RESERVE|BTREE_INSERT_NOFAIL);
1824         if (ret)
1825                 goto out;
1826
1827         trace_and_count(c, bucket_invalidate, c, bucket.inode, bucket.offset, cached_sectors);
1828         --*nr_to_invalidate;
1829 out:
1830         bch2_trans_iter_exit(trans, &alloc_iter);
1831         printbuf_exit(&buf);
1832         return ret;
1833 err:
1834         prt_str(&buf, "\n  lru key: ");
1835         bch2_bkey_val_to_text(&buf, c, lru_k);
1836
1837         prt_str(&buf, "\n  lru entry: ");
1838         bch2_lru_pos_to_text(&buf, lru_iter->pos);
1839
1840         prt_str(&buf, "\n  alloc key: ");
1841         if (!a)
1842                 bch2_bpos_to_text(&buf, bucket);
1843         else
1844                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
1845
1846         bch_err(c, "%s", buf.buf);
1847         if (test_bit(BCH_FS_CHECK_LRUS_DONE, &c->flags)) {
1848                 bch2_inconsistent_error(c);
1849                 ret = -EINVAL;
1850         }
1851
1852         goto out;
1853 }
1854
1855 static void bch2_do_invalidates_work(struct work_struct *work)
1856 {
1857         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1858         struct bch_dev *ca;
1859         struct btree_trans trans;
1860         struct btree_iter iter;
1861         struct bkey_s_c k;
1862         unsigned i;
1863         int ret = 0;
1864
1865         bch2_trans_init(&trans, c, 0, 0);
1866
1867         ret = bch2_btree_write_buffer_flush(&trans);
1868         if (ret)
1869                 goto err;
1870
1871         for_each_member_device(ca, c, i) {
1872                 s64 nr_to_invalidate =
1873                         should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
1874
1875                 ret = for_each_btree_key2_upto(&trans, iter, BTREE_ID_lru,
1876                                 lru_pos(ca->dev_idx, 0, 0),
1877                                 lru_pos(ca->dev_idx, U64_MAX, LRU_TIME_MAX),
1878                                 BTREE_ITER_INTENT, k,
1879                         invalidate_one_bucket(&trans, &iter, k, &nr_to_invalidate));
1880
1881                 if (ret < 0) {
1882                         percpu_ref_put(&ca->ref);
1883                         break;
1884                 }
1885         }
1886 err:
1887         bch2_trans_exit(&trans);
1888         bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1889 }
1890
1891 void bch2_do_invalidates(struct bch_fs *c)
1892 {
1893         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_invalidate) &&
1894             !queue_work(c->write_ref_wq, &c->invalidate_work))
1895                 bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1896 }
1897
1898 static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca,
1899                                    unsigned long *last_updated)
1900 {
1901         struct btree_trans trans;
1902         struct btree_iter iter;
1903         struct bkey_s_c k;
1904         struct bkey hole;
1905         struct bpos end = POS(ca->dev_idx, ca->mi.nbuckets);
1906         struct bch_member *m;
1907         int ret;
1908
1909         bch2_trans_init(&trans, c, 0, 0);
1910
1911         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc,
1912                              POS(ca->dev_idx, ca->mi.first_bucket),
1913                              BTREE_ITER_PREFETCH);
1914         /*
1915          * Scan the alloc btree for every bucket on @ca, and add buckets to the
1916          * freespace/need_discard/need_gc_gens btrees as needed:
1917          */
1918         while (1) {
1919                 if (*last_updated + HZ * 10 < jiffies) {
1920                         bch_info(ca, "%s: currently at %llu/%llu",
1921                                  __func__, iter.pos.offset, ca->mi.nbuckets);
1922                         *last_updated = jiffies;
1923                 }
1924
1925                 bch2_trans_begin(&trans);
1926
1927                 if (bkey_ge(iter.pos, end)) {
1928                         ret = 0;
1929                         break;
1930                 }
1931
1932                 k = bch2_get_key_or_hole(&iter, end, &hole);
1933                 ret = bkey_err(k);
1934                 if (ret)
1935                         goto bkey_err;
1936
1937                 if (k.k->type) {
1938                         /*
1939                          * We process live keys in the alloc btree one at a
1940                          * time:
1941                          */
1942                         struct bch_alloc_v4 a_convert;
1943                         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1944
1945                         ret =   bch2_bucket_do_index(&trans, k, a, true) ?:
1946                                 bch2_trans_commit(&trans, NULL, NULL,
1947                                                   BTREE_INSERT_LAZY_RW|
1948                                                   BTREE_INSERT_NOFAIL);
1949                         if (ret)
1950                                 goto bkey_err;
1951
1952                         bch2_btree_iter_advance(&iter);
1953                 } else {
1954                         struct bkey_i *freespace;
1955
1956                         freespace = bch2_trans_kmalloc(&trans, sizeof(*freespace));
1957                         ret = PTR_ERR_OR_ZERO(freespace);
1958                         if (ret)
1959                                 goto bkey_err;
1960
1961                         bkey_init(&freespace->k);
1962                         freespace->k.type       = KEY_TYPE_set;
1963                         freespace->k.p          = k.k->p;
1964                         freespace->k.size       = k.k->size;
1965
1966                         ret = __bch2_btree_insert(&trans, BTREE_ID_freespace, freespace, 0) ?:
1967                                 bch2_trans_commit(&trans, NULL, NULL,
1968                                                   BTREE_INSERT_LAZY_RW|
1969                                                   BTREE_INSERT_NOFAIL);
1970                         if (ret)
1971                                 goto bkey_err;
1972
1973                         bch2_btree_iter_set_pos(&iter, k.k->p);
1974                 }
1975 bkey_err:
1976                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1977                         continue;
1978                 if (ret)
1979                         break;
1980         }
1981
1982         bch2_trans_iter_exit(&trans, &iter);
1983         bch2_trans_exit(&trans);
1984
1985         if (ret < 0) {
1986                 bch_err(ca, "error initializing free space: %s", bch2_err_str(ret));
1987                 return ret;
1988         }
1989
1990         mutex_lock(&c->sb_lock);
1991         m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx;
1992         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1993         mutex_unlock(&c->sb_lock);
1994
1995         return 0;
1996 }
1997
1998 int bch2_fs_freespace_init(struct bch_fs *c)
1999 {
2000         struct bch_dev *ca;
2001         unsigned i;
2002         int ret = 0;
2003         bool doing_init = false;
2004         unsigned long last_updated = jiffies;
2005
2006         /*
2007          * We can crash during the device add path, so we need to check this on
2008          * every mount:
2009          */
2010
2011         for_each_member_device(ca, c, i) {
2012                 if (ca->mi.freespace_initialized)
2013                         continue;
2014
2015                 if (!doing_init) {
2016                         bch_info(c, "initializing freespace");
2017                         doing_init = true;
2018                 }
2019
2020                 ret = bch2_dev_freespace_init(c, ca, &last_updated);
2021                 if (ret) {
2022                         percpu_ref_put(&ca->ref);
2023                         return ret;
2024                 }
2025         }
2026
2027         if (doing_init) {
2028                 mutex_lock(&c->sb_lock);
2029                 bch2_write_super(c);
2030                 mutex_unlock(&c->sb_lock);
2031
2032                 bch_verbose(c, "done initializing freespace");
2033         }
2034
2035         return ret;
2036 }
2037
2038 /* Bucket IO clocks: */
2039
2040 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
2041                               size_t bucket_nr, int rw)
2042 {
2043         struct bch_fs *c = trans->c;
2044         struct btree_iter iter;
2045         struct bkey_i_alloc_v4 *a;
2046         u64 now;
2047         int ret = 0;
2048
2049         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
2050         ret = PTR_ERR_OR_ZERO(a);
2051         if (ret)
2052                 return ret;
2053
2054         now = atomic64_read(&c->io_clock[rw].now);
2055         if (a->v.io_time[rw] == now)
2056                 goto out;
2057
2058         a->v.io_time[rw] = now;
2059
2060         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
2061                 bch2_trans_commit(trans, NULL, NULL, 0);
2062 out:
2063         bch2_trans_iter_exit(trans, &iter);
2064         return ret;
2065 }
2066
2067 /* Startup/shutdown (ro/rw): */
2068
2069 void bch2_recalc_capacity(struct bch_fs *c)
2070 {
2071         struct bch_dev *ca;
2072         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
2073         unsigned bucket_size_max = 0;
2074         unsigned long ra_pages = 0;
2075         unsigned i;
2076
2077         lockdep_assert_held(&c->state_lock);
2078
2079         for_each_online_member(ca, c, i) {
2080                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
2081
2082                 ra_pages += bdi->ra_pages;
2083         }
2084
2085         bch2_set_ra_pages(c, ra_pages);
2086
2087         for_each_rw_member(ca, c, i) {
2088                 u64 dev_reserve = 0;
2089
2090                 /*
2091                  * We need to reserve buckets (from the number
2092                  * of currently available buckets) against
2093                  * foreground writes so that mainly copygc can
2094                  * make forward progress.
2095                  *
2096                  * We need enough to refill the various reserves
2097                  * from scratch - copygc will use its entire
2098                  * reserve all at once, then run against when
2099                  * its reserve is refilled (from the formerly
2100                  * available buckets).
2101                  *
2102                  * This reserve is just used when considering if
2103                  * allocations for foreground writes must wait -
2104                  * not -ENOSPC calculations.
2105                  */
2106
2107                 dev_reserve += ca->nr_btree_reserve * 2;
2108                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
2109
2110                 dev_reserve += 1;       /* btree write point */
2111                 dev_reserve += 1;       /* copygc write point */
2112                 dev_reserve += 1;       /* rebalance write point */
2113
2114                 dev_reserve *= ca->mi.bucket_size;
2115
2116                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
2117                                              ca->mi.first_bucket);
2118
2119                 reserved_sectors += dev_reserve * 2;
2120
2121                 bucket_size_max = max_t(unsigned, bucket_size_max,
2122                                         ca->mi.bucket_size);
2123         }
2124
2125         gc_reserve = c->opts.gc_reserve_bytes
2126                 ? c->opts.gc_reserve_bytes >> 9
2127                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
2128
2129         reserved_sectors = max(gc_reserve, reserved_sectors);
2130
2131         reserved_sectors = min(reserved_sectors, capacity);
2132
2133         c->capacity = capacity - reserved_sectors;
2134
2135         c->bucket_size_max = bucket_size_max;
2136
2137         /* Wake up case someone was waiting for buckets */
2138         closure_wake_up(&c->freelist_wait);
2139 }
2140
2141 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
2142 {
2143         struct open_bucket *ob;
2144         bool ret = false;
2145
2146         for (ob = c->open_buckets;
2147              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
2148              ob++) {
2149                 spin_lock(&ob->lock);
2150                 if (ob->valid && !ob->on_partial_list &&
2151                     ob->dev == ca->dev_idx)
2152                         ret = true;
2153                 spin_unlock(&ob->lock);
2154         }
2155
2156         return ret;
2157 }
2158
2159 /* device goes ro: */
2160 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
2161 {
2162         unsigned i;
2163
2164         /* First, remove device from allocation groups: */
2165
2166         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2167                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
2168
2169         /*
2170          * Capacity is calculated based off of devices in allocation groups:
2171          */
2172         bch2_recalc_capacity(c);
2173
2174         bch2_open_buckets_stop(c, ca, false);
2175
2176         /*
2177          * Wake up threads that were blocked on allocation, so they can notice
2178          * the device can no longer be removed and the capacity has changed:
2179          */
2180         closure_wake_up(&c->freelist_wait);
2181
2182         /*
2183          * journal_res_get() can block waiting for free space in the journal -
2184          * it needs to notice there may not be devices to allocate from anymore:
2185          */
2186         wake_up(&c->journal.wait);
2187
2188         /* Now wait for any in flight writes: */
2189
2190         closure_wait_event(&c->open_buckets_wait,
2191                            !bch2_dev_has_open_write_point(c, ca));
2192 }
2193
2194 /* device goes rw: */
2195 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
2196 {
2197         unsigned i;
2198
2199         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2200                 if (ca->mi.data_allowed & (1 << i))
2201                         set_bit(ca->dev_idx, c->rw_devs[i].d);
2202 }
2203
2204 void bch2_fs_allocator_background_init(struct bch_fs *c)
2205 {
2206         spin_lock_init(&c->freelist_lock);
2207         INIT_WORK(&c->discard_work, bch2_do_discards_work);
2208         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
2209 }