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

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "bcachefs.h"
   3 #include "alloc_background.h"
   4 #include "alloc_foreground.h"
   5 #include "backpointers.h"
   6 #include "btree_cache.h"
   7 #include "btree_io.h"
   8 #include "btree_key_cache.h"
   9 #include "btree_update.h"
  10 #include "btree_update_interior.h"
  11 #include "btree_gc.h"
  12 #include "btree_write_buffer.h"
  13 #include "buckets.h"
  14 #include "buckets_waiting_for_journal.h"
  15 #include "clock.h"
  16 #include "debug.h"
  17 #include "ec.h"
  18 #include "error.h"
  19 #include "lru.h"
  20 #include "recovery.h"
  21 #include "trace.h"
  22 #include "varint.h"
  23
  24 #include <linux/kthread.h>
  25 #include <linux/math64.h>
  26 #include <linux/random.h>
  27 #include <linux/rculist.h>
  28 #include <linux/rcupdate.h>
  29 #include <linux/sched/task.h>
  30 #include <linux/sort.h>
  31
  32 /* Persistent alloc info: */
  33
  34 static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
  35 #define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8,
  36         BCH_ALLOC_FIELDS_V1()
  37 #undef x
  38 };
  39
  40 struct bkey_alloc_unpacked {
  41         u64             journal_seq;
  42         u8              gen;
  43         u8              oldest_gen;
  44         u8              data_type;
  45         bool            need_discard:1;
  46         bool            need_inc_gen:1;
  47 #define x(_name, _bits) u##_bits _name;
  48         BCH_ALLOC_FIELDS_V2()
  49 #undef  x
  50 };
  51
  52 static inline u64 alloc_field_v1_get(const struct bch_alloc *a,
  53                                      const void **p, unsigned field)
  54 {
  55         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  56         u64 v;
  57
  58         if (!(a->fields & (1 << field)))
  59                 return 0;
  60
  61         switch (bytes) {
  62         case 1:
  63                 v = *((const u8 *) *p);
  64                 break;
  65         case 2:
  66                 v = le16_to_cpup(*p);
  67                 break;
  68         case 4:
  69                 v = le32_to_cpup(*p);
  70                 break;
  71         case 8:
  72                 v = le64_to_cpup(*p);
  73                 break;
  74         default:
  75                 BUG();
  76         }
  77
  78         *p += bytes;
  79         return v;
  80 }
  81
  82 static inline void alloc_field_v1_put(struct bkey_i_alloc *a, void **p,
  83                                       unsigned field, u64 v)
  84 {
  85         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  86
  87         if (!v)
  88                 return;
  89
  90         a->v.fields |= 1 << field;
  91
  92         switch (bytes) {
  93         case 1:
  94                 *((u8 *) *p) = v;
  95                 break;
  96         case 2:
  97                 *((__le16 *) *p) = cpu_to_le16(v);
  98                 break;
  99         case 4:
 100                 *((__le32 *) *p) = cpu_to_le32(v);
 101                 break;
 102         case 8:
 103                 *((__le64 *) *p) = cpu_to_le64(v);
 104                 break;
 105         default:
 106                 BUG();
 107         }
 108
 109         *p += bytes;
 110 }
 111
 112 static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out,
 113                                  struct bkey_s_c k)
 114 {
 115         const struct bch_alloc *in = bkey_s_c_to_alloc(k).v;
 116         const void *d = in->data;
 117         unsigned idx = 0;
 118
 119         out->gen = in->gen;
 120
 121 #define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++);
 122         BCH_ALLOC_FIELDS_V1()
 123 #undef  x
 124 }
 125
 126 static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out,
 127                                 struct bkey_s_c k)
 128 {
 129         struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k);
 130         const u8 *in = a.v->data;
 131         const u8 *end = bkey_val_end(a);
 132         unsigned fieldnr = 0;
 133         int ret;
 134         u64 v;
 135
 136         out->gen        = a.v->gen;
 137         out->oldest_gen = a.v->oldest_gen;
 138         out->data_type  = a.v->data_type;
 139
 140 #define x(_name, _bits)                                                 \
 141         if (fieldnr < a.v->nr_fields) {                                 \
 142                 ret = bch2_varint_decode_fast(in, end, &v);             \
 143                 if (ret < 0)                                            \
 144                         return ret;                                     \
 145                 in += ret;                                              \
 146         } else {                                                        \
 147                 v = 0;                                                  \
 148         }                                                               \
 149         out->_name = v;                                                 \
 150         if (v != out->_name)                                            \
 151                 return -1;                                              \
 152         fieldnr++;
 153
 154         BCH_ALLOC_FIELDS_V2()
 155 #undef  x
 156         return 0;
 157 }
 158
 159 static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out,
 160                                 struct bkey_s_c k)
 161 {
 162         struct bkey_s_c_alloc_v3 a = bkey_s_c_to_alloc_v3(k);
 163         const u8 *in = a.v->data;
 164         const u8 *end = bkey_val_end(a);
 165         unsigned fieldnr = 0;
 166         int ret;
 167         u64 v;
 168
 169         out->gen        = a.v->gen;
 170         out->oldest_gen = a.v->oldest_gen;
 171         out->data_type  = a.v->data_type;
 172         out->need_discard = BCH_ALLOC_V3_NEED_DISCARD(a.v);
 173         out->need_inc_gen = BCH_ALLOC_V3_NEED_INC_GEN(a.v);
 174         out->journal_seq = le64_to_cpu(a.v->journal_seq);
 175
 176 #define x(_name, _bits)                                                 \
 177         if (fieldnr < a.v->nr_fields) {                                 \
 178                 ret = bch2_varint_decode_fast(in, end, &v);             \
 179                 if (ret < 0)                                            \
 180                         return ret;                                     \
 181                 in += ret;                                              \
 182         } else {                                                        \
 183                 v = 0;                                                  \
 184         }                                                               \
 185         out->_name = v;                                                 \
 186         if (v != out->_name)                                            \
 187                 return -1;                                              \
 188         fieldnr++;
 189
 190         BCH_ALLOC_FIELDS_V2()
 191 #undef  x
 192         return 0;
 193 }
 194
 195 static struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 196 {
 197         struct bkey_alloc_unpacked ret = { .gen = 0 };
 198
 199         switch (k.k->type) {
 200         case KEY_TYPE_alloc:
 201                 bch2_alloc_unpack_v1(&ret, k);
 202                 break;
 203         case KEY_TYPE_alloc_v2:
 204                 bch2_alloc_unpack_v2(&ret, k);
 205                 break;
 206         case KEY_TYPE_alloc_v3:
 207                 bch2_alloc_unpack_v3(&ret, k);
 208                 break;
 209         }
 210
 211         return ret;
 212 }
 213
 214 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
 215 {
 216         unsigned i, bytes = offsetof(struct bch_alloc, data);
 217
 218         for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++)
 219                 if (a->fields & (1 << i))
 220                         bytes += BCH_ALLOC_V1_FIELD_BYTES[i];
 221
 222         return DIV_ROUND_UP(bytes, sizeof(u64));
 223 }
 224
 225 int bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k,
 226                           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 (%u > %lu)",
 274                        alloc_v4_u64s(a.v), bkey_val_u64s(k.k));
 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_NR_BACKPOINTERS(a.v) == 0))
 515                 return bch2_bkey_make_mut_noupdate_typed(trans, k, alloc_v4);
 516
 517         return __bch2_alloc_to_v4_mut(trans, k);
 518 }
 519
 520 struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 521 {
 522         return bch2_alloc_to_v4_mut_inlined(trans, k);
 523 }
 524
 525 struct bkey_i_alloc_v4 *
 526 bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
 527                               struct bpos pos)
 528 {
 529         struct bkey_s_c k;
 530         struct bkey_i_alloc_v4 *a;
 531         int ret;
 532
 533         k = bch2_bkey_get_iter(trans, iter, BTREE_ID_alloc, pos,
 534                              BTREE_ITER_WITH_UPDATES|
 535                              BTREE_ITER_CACHED|
 536                              BTREE_ITER_INTENT);
 537         ret = bkey_err(k);
 538         if (unlikely(ret))
 539                 return ERR_PTR(ret);
 540
 541         a = bch2_alloc_to_v4_mut_inlined(trans, k);
 542         ret = PTR_ERR_OR_ZERO(a);
 543         if (unlikely(ret))
 544                 goto err;
 545         return a;
 546 err:
 547         bch2_trans_iter_exit(trans, iter);
 548         return ERR_PTR(ret);
 549 }
 550
 551 int bch2_alloc_read(struct bch_fs *c)
 552 {
 553         struct btree_trans trans;
 554         struct btree_iter iter;
 555         struct bkey_s_c k;
 556         struct bch_alloc_v4 a;
 557         struct bch_dev *ca;
 558         int ret;
 559
 560         bch2_trans_init(&trans, c, 0, 0);
 561
 562         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 563                            BTREE_ITER_PREFETCH, k, ret) {
 564                 /*
 565                  * Not a fsck error because this is checked/repaired by
 566                  * bch2_check_alloc_key() which runs later:
 567                  */
 568                 if (!bch2_dev_bucket_exists(c, k.k->p))
 569                         continue;
 570
 571                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 572
 573                 *bucket_gen(ca, k.k->p.offset) = bch2_alloc_to_v4(k, &a)->gen;
 574         }
 575         bch2_trans_iter_exit(&trans, &iter);
 576
 577         bch2_trans_exit(&trans);
 578
 579         if (ret)
 580                 bch_err_fn(c, ret);
 581
 582         return ret;
 583 }
 584
 585 static struct bpos alloc_gens_pos(struct bpos pos, unsigned *offset)
 586 {
 587         *offset = pos.offset & KEY_TYPE_BUCKET_GENS_MASK;
 588
 589         pos.offset >>= KEY_TYPE_BUCKET_GENS_BITS;
 590         return pos;
 591 }
 592
 593 static struct bpos bucket_gens_pos_to_alloc(struct bpos pos, unsigned offset)
 594 {
 595         pos.offset <<= KEY_TYPE_BUCKET_GENS_BITS;
 596         pos.offset += offset;
 597         return pos;
 598 }
 599
 600 static unsigned alloc_gen(struct bkey_s_c k, unsigned offset)
 601 {
 602         return k.k->type == KEY_TYPE_bucket_gens
 603                 ? bkey_s_c_to_bucket_gens(k).v->gens[offset]
 604                 : 0;
 605 }
 606
 607 int bch2_bucket_gens_invalid(const struct bch_fs *c, struct bkey_s_c k,
 608                              unsigned flags, struct printbuf *err)
 609 {
 610         if (bkey_val_bytes(k.k) != sizeof(struct bch_bucket_gens)) {
 611                 prt_printf(err, "bad val size (%lu != %zu)",
 612                        bkey_val_bytes(k.k), sizeof(struct bch_bucket_gens));
 613                 return -BCH_ERR_invalid_bkey;
 614         }
 615
 616         return 0;
 617 }
 618
 619 void bch2_bucket_gens_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 620 {
 621         struct bkey_s_c_bucket_gens g = bkey_s_c_to_bucket_gens(k);
 622         unsigned i;
 623
 624         for (i = 0; i < ARRAY_SIZE(g.v->gens); i++) {
 625                 if (i)
 626                         prt_char(out, ' ');
 627                 prt_printf(out, "%u", g.v->gens[i]);
 628         }
 629 }
 630
 631 int bch2_bucket_gens_init(struct bch_fs *c)
 632 {
 633         struct btree_trans trans;
 634         struct btree_iter iter;
 635         struct bkey_s_c k;
 636         struct bch_alloc_v4 a;
 637         struct bkey_i_bucket_gens g;
 638         bool have_bucket_gens_key = false;
 639         unsigned offset;
 640         struct bpos pos;
 641         u8 gen;
 642         int ret;
 643
 644         bch2_trans_init(&trans, c, 0, 0);
 645
 646         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 647                            BTREE_ITER_PREFETCH, k, ret) {
 648                 /*
 649                  * Not a fsck error because this is checked/repaired by
 650                  * bch2_check_alloc_key() which runs later:
 651                  */
 652                 if (!bch2_dev_bucket_exists(c, k.k->p))
 653                         continue;
 654
 655                 gen = bch2_alloc_to_v4(k, &a)->gen;
 656                 pos = alloc_gens_pos(iter.pos, &offset);
 657
 658                 if (have_bucket_gens_key && bkey_cmp(iter.pos, pos)) {
 659                         ret = commit_do(&trans, NULL, NULL,
 660                                         BTREE_INSERT_NOFAIL|
 661                                         BTREE_INSERT_LAZY_RW,
 662                                 __bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i, 0));
 663                         if (ret)
 664                                 break;
 665                         have_bucket_gens_key = false;
 666                 }
 667
 668                 if (!have_bucket_gens_key) {
 669                         bkey_bucket_gens_init(&g.k_i);
 670                         g.k.p = pos;
 671                         have_bucket_gens_key = true;
 672                 }
 673
 674                 g.v.gens[offset] = gen;
 675         }
 676         bch2_trans_iter_exit(&trans, &iter);
 677
 678         if (have_bucket_gens_key && !ret)
 679                 ret = commit_do(&trans, NULL, NULL,
 680                                 BTREE_INSERT_NOFAIL|
 681                                 BTREE_INSERT_LAZY_RW,
 682                         __bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i, 0));
 683
 684         bch2_trans_exit(&trans);
 685
 686         if (ret)
 687                 bch_err_fn(c, ret);
 688         return ret;
 689 }
 690
 691 int bch2_bucket_gens_read(struct bch_fs *c)
 692 {
 693         struct btree_trans trans;
 694         struct btree_iter iter;
 695         struct bkey_s_c k;
 696         const struct bch_bucket_gens *g;
 697         struct bch_dev *ca;
 698         u64 b;
 699         int ret;
 700
 701         bch2_trans_init(&trans, c, 0, 0);
 702
 703         for_each_btree_key(&trans, iter, BTREE_ID_bucket_gens, POS_MIN,
 704                            BTREE_ITER_PREFETCH, k, ret) {
 705                 u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
 706                 u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
 707
 708                 if (k.k->type != KEY_TYPE_bucket_gens)
 709                         continue;
 710
 711                 g = bkey_s_c_to_bucket_gens(k).v;
 712
 713                 /*
 714                  * Not a fsck error because this is checked/repaired by
 715                  * bch2_check_alloc_key() which runs later:
 716                  */
 717                 if (!bch2_dev_exists2(c, k.k->p.inode))
 718                         continue;
 719
 720                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 721
 722                 for (b = max_t(u64, ca->mi.first_bucket, start);
 723                      b < min_t(u64, ca->mi.nbuckets, end);
 724                      b++)
 725                         *bucket_gen(ca, b) = g->gens[b & KEY_TYPE_BUCKET_GENS_MASK];
 726         }
 727         bch2_trans_iter_exit(&trans, &iter);
 728
 729         bch2_trans_exit(&trans);
 730
 731         if (ret)
 732                 bch_err_fn(c, ret);
 733
 734         return ret;
 735 }
 736
 737 /* Free space/discard btree: */
 738
 739 static int bch2_bucket_do_index(struct btree_trans *trans,
 740                                 struct bkey_s_c alloc_k,
 741                                 const struct bch_alloc_v4 *a,
 742                                 bool set)
 743 {
 744         struct bch_fs *c = trans->c;
 745         struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 746         struct btree_iter iter;
 747         struct bkey_s_c old;
 748         struct bkey_i *k;
 749         enum btree_id btree;
 750         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 751         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 752         struct printbuf buf = PRINTBUF;
 753         int ret;
 754
 755         if (a->data_type != BCH_DATA_free &&
 756             a->data_type != BCH_DATA_need_discard)
 757                 return 0;
 758
 759         k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
 760         if (IS_ERR(k))
 761                 return PTR_ERR(k);
 762
 763         bkey_init(&k->k);
 764         k->k.type = new_type;
 765
 766         switch (a->data_type) {
 767         case BCH_DATA_free:
 768                 btree = BTREE_ID_freespace;
 769                 k->k.p = alloc_freespace_pos(alloc_k.k->p, *a);
 770                 bch2_key_resize(&k->k, 1);
 771                 break;
 772         case BCH_DATA_need_discard:
 773                 btree = BTREE_ID_need_discard;
 774                 k->k.p = alloc_k.k->p;
 775                 break;
 776         default:
 777                 return 0;
 778         }
 779
 780         old = bch2_bkey_get_iter(trans, &iter, btree,
 781                              bkey_start_pos(&k->k),
 782                              BTREE_ITER_INTENT);
 783         ret = bkey_err(old);
 784         if (ret)
 785                 return ret;
 786
 787         if (ca->mi.freespace_initialized &&
 788             test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags) &&
 789             bch2_trans_inconsistent_on(old.k->type != old_type, trans,
 790                         "incorrect key when %s %s:%llu:%llu:0 (got %s should be %s)\n"
 791                         "  for %s",
 792                         set ? "setting" : "clearing",
 793                         bch2_btree_ids[btree],
 794                         iter.pos.inode,
 795                         iter.pos.offset,
 796                         bch2_bkey_types[old.k->type],
 797                         bch2_bkey_types[old_type],
 798                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 799                 ret = -EIO;
 800                 goto err;
 801         }
 802
 803         ret = bch2_trans_update(trans, &iter, k, 0);
 804 err:
 805         bch2_trans_iter_exit(trans, &iter);
 806         printbuf_exit(&buf);
 807         return ret;
 808 }
 809
 810 static noinline int bch2_bucket_gen_update(struct btree_trans *trans,
 811                                            struct bpos bucket, u8 gen)
 812 {
 813         struct btree_iter iter;
 814         unsigned offset;
 815         struct bpos pos = alloc_gens_pos(bucket, &offset);
 816         struct bkey_i_bucket_gens *g;
 817         struct bkey_s_c k;
 818         int ret;
 819
 820         g = bch2_trans_kmalloc(trans, sizeof(*g));
 821         ret = PTR_ERR_OR_ZERO(g);
 822         if (ret)
 823                 return ret;
 824
 825         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_bucket_gens, pos,
 826                                BTREE_ITER_INTENT|
 827                                BTREE_ITER_WITH_UPDATES);
 828         ret = bkey_err(k);
 829         if (ret)
 830                 return ret;
 831
 832         if (k.k->type != KEY_TYPE_bucket_gens) {
 833                 bkey_bucket_gens_init(&g->k_i);
 834                 g->k.p = iter.pos;
 835         } else {
 836                 bkey_reassemble(&g->k_i, k);
 837         }
 838
 839         g->v.gens[offset] = gen;
 840
 841         ret = bch2_trans_update(trans, &iter, &g->k_i, 0);
 842         bch2_trans_iter_exit(trans, &iter);
 843         return ret;
 844 }
 845
 846 int bch2_trans_mark_alloc(struct btree_trans *trans,
 847                           enum btree_id btree_id, unsigned level,
 848                           struct bkey_s_c old, struct bkey_i *new,
 849                           unsigned flags)
 850 {
 851         struct bch_fs *c = trans->c;
 852         struct bch_alloc_v4 old_a_convert, *new_a;
 853         const struct bch_alloc_v4 *old_a;
 854         u64 old_lru, new_lru;
 855         int ret = 0;
 856
 857         /*
 858          * Deletion only happens in the device removal path, with
 859          * BTREE_TRIGGER_NORUN:
 860          */
 861         BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
 862
 863         old_a = bch2_alloc_to_v4(old, &old_a_convert);
 864         new_a = &bkey_i_to_alloc_v4(new)->v;
 865
 866         new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
 867
 868         if (new_a->dirty_sectors > old_a->dirty_sectors ||
 869             new_a->cached_sectors > old_a->cached_sectors) {
 870                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 871                 new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 872                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
 873                 SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
 874         }
 875
 876         if (data_type_is_empty(new_a->data_type) &&
 877             BCH_ALLOC_V4_NEED_INC_GEN(new_a) &&
 878             !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
 879                 new_a->gen++;
 880                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
 881         }
 882
 883         if (old_a->data_type != new_a->data_type ||
 884             (new_a->data_type == BCH_DATA_free &&
 885              alloc_freespace_genbits(*old_a) != alloc_freespace_genbits(*new_a))) {
 886                 ret =   bch2_bucket_do_index(trans, old, old_a, false) ?:
 887                         bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_a, true);
 888                 if (ret)
 889                         return ret;
 890         }
 891
 892         if (new_a->data_type == BCH_DATA_cached &&
 893             !new_a->io_time[READ])
 894                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 895
 896         old_lru = alloc_lru_idx_read(*old_a);
 897         new_lru = alloc_lru_idx_read(*new_a);
 898
 899         if (old_lru != new_lru) {
 900                 ret = bch2_lru_change(trans, new->k.p.inode,
 901                                       bucket_to_u64(new->k.p),
 902                                       old_lru, new_lru);
 903                 if (ret)
 904                         return ret;
 905         }
 906
 907         new_a->fragmentation_lru = alloc_lru_idx_fragmentation(*new_a,
 908                                         bch_dev_bkey_exists(c, new->k.p.inode));
 909
 910         if (old_a->fragmentation_lru != new_a->fragmentation_lru) {
 911                 ret = bch2_lru_change(trans,
 912                                 BCH_LRU_FRAGMENTATION_START,
 913                                 bucket_to_u64(new->k.p),
 914                                 old_a->fragmentation_lru, new_a->fragmentation_lru);
 915                 if (ret)
 916                         return ret;
 917         }
 918
 919         if (old_a->gen != new_a->gen) {
 920                 ret = bch2_bucket_gen_update(trans, new->k.p, new_a->gen);
 921                 if (ret)
 922                         return ret;
 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 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                 if (!bpos_eq(iter->path->l[0].b->key.k.p, SPOS_MAX))
 948                         end = bkey_min(end, bpos_nosnap_successor(iter->path->l[0].b->key.k.p));
 949
 950                 end = bkey_min(end, POS(iter->pos.inode, iter->pos.offset + U32_MAX - 1));
 951
 952                 /*
 953                  * btree node min/max is a closed interval, upto takes a half
 954                  * open interval:
 955                  */
 956                 k = bch2_btree_iter_peek_upto(&iter2, end);
 957                 next = iter2.pos;
 958                 bch2_trans_iter_exit(iter->trans, &iter2);
 959
 960                 BUG_ON(next.offset >= iter->pos.offset + U32_MAX);
 961
 962                 if (bkey_err(k))
 963                         return k;
 964
 965                 bkey_init(hole);
 966                 hole->p = iter->pos;
 967
 968                 bch2_key_resize(hole, next.offset - iter->pos.offset);
 969                 return (struct bkey_s_c) { hole, NULL };
 970         }
 971 }
 972
 973 static bool next_bucket(struct bch_fs *c, struct bpos *bucket)
 974 {
 975         struct bch_dev *ca;
 976         unsigned iter;
 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         iter = bucket->inode;
 995         ca = __bch2_next_dev(c, &iter, NULL);
 996         if (ca)
 997                 *bucket = POS(ca->dev_idx, ca->mi.first_bucket);
 998         rcu_read_unlock();
 999
1000         return ca != NULL;
1001 }
1002
1003 struct bkey_s_c bch2_get_key_or_real_bucket_hole(struct btree_iter *iter, struct bkey *hole)
1004 {
1005         struct bch_fs *c = iter->trans->c;
1006         struct bkey_s_c k;
1007 again:
1008         k = bch2_get_key_or_hole(iter, POS_MAX, hole);
1009         if (bkey_err(k))
1010                 return k;
1011
1012         if (!k.k->type) {
1013                 struct bpos bucket = bkey_start_pos(k.k);
1014
1015                 if (!bch2_dev_bucket_exists(c, bucket)) {
1016                         if (!next_bucket(c, &bucket))
1017                                 return bkey_s_c_null;
1018
1019                         bch2_btree_iter_set_pos(iter, bucket);
1020                         goto again;
1021                 }
1022
1023                 if (!bch2_dev_bucket_exists(c, k.k->p)) {
1024                         struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
1025
1026                         bch2_key_resize(hole, ca->mi.nbuckets - bucket.offset);
1027                 }
1028         }
1029
1030         return k;
1031 }
1032
1033 static noinline_for_stack
1034 int bch2_check_alloc_key(struct btree_trans *trans,
1035                          struct bkey_s_c alloc_k,
1036                          struct btree_iter *alloc_iter,
1037                          struct btree_iter *discard_iter,
1038                          struct btree_iter *freespace_iter,
1039                          struct btree_iter *bucket_gens_iter)
1040 {
1041         struct bch_fs *c = trans->c;
1042         struct bch_dev *ca;
1043         struct bch_alloc_v4 a_convert;
1044         const struct bch_alloc_v4 *a;
1045         unsigned discard_key_type, freespace_key_type;
1046         unsigned gens_offset;
1047         struct bkey_s_c k;
1048         struct printbuf buf = PRINTBUF;
1049         int ret;
1050
1051         if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
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, "incorrect key in need_discard btree (got %s should be %s)\n"
1072                       "  %s",
1073                       bch2_bkey_types[k.k->type],
1074                       bch2_bkey_types[discard_key_type],
1075                       (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1076                 struct bkey_i *update =
1077                         bch2_trans_kmalloc(trans, sizeof(*update));
1078
1079                 ret = PTR_ERR_OR_ZERO(update);
1080                 if (ret)
1081                         goto err;
1082
1083                 bkey_init(&update->k);
1084                 update->k.type  = discard_key_type;
1085                 update->k.p     = discard_iter->pos;
1086
1087                 ret = bch2_trans_update(trans, discard_iter, update, 0);
1088                 if (ret)
1089                         goto err;
1090         }
1091
1092         freespace_key_type = a->data_type == BCH_DATA_free ? KEY_TYPE_set : 0;
1093         bch2_btree_iter_set_pos(freespace_iter, alloc_freespace_pos(alloc_k.k->p, *a));
1094         k = bch2_btree_iter_peek_slot(freespace_iter);
1095         ret = bkey_err(k);
1096         if (ret)
1097                 goto err;
1098
1099         if (k.k->type != freespace_key_type &&
1100             (c->opts.reconstruct_alloc ||
1101              fsck_err(c, "incorrect key in freespace btree (got %s should be %s)\n"
1102                       "  %s",
1103                       bch2_bkey_types[k.k->type],
1104                       bch2_bkey_types[freespace_key_type],
1105                       (printbuf_reset(&buf),
1106                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1107                 struct bkey_i *update =
1108                         bch2_trans_kmalloc(trans, sizeof(*update));
1109
1110                 ret = PTR_ERR_OR_ZERO(update);
1111                 if (ret)
1112                         goto err;
1113
1114                 bkey_init(&update->k);
1115                 update->k.type  = freespace_key_type;
1116                 update->k.p     = freespace_iter->pos;
1117                 bch2_key_resize(&update->k, 1);
1118
1119                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1120                 if (ret)
1121                         goto err;
1122         }
1123
1124         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(alloc_k.k->p, &gens_offset));
1125         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1126         ret = bkey_err(k);
1127         if (ret)
1128                 goto err;
1129
1130         if (a->gen != alloc_gen(k, gens_offset) &&
1131             (c->opts.reconstruct_alloc ||
1132              fsck_err(c, "incorrect gen in bucket_gens btree (got %u should be %u)\n"
1133                       "  %s",
1134                       alloc_gen(k, gens_offset), a->gen,
1135                       (printbuf_reset(&buf),
1136                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1137                 struct bkey_i_bucket_gens *g =
1138                         bch2_trans_kmalloc(trans, sizeof(*g));
1139
1140                 ret = PTR_ERR_OR_ZERO(g);
1141                 if (ret)
1142                         goto err;
1143
1144                 if (k.k->type == KEY_TYPE_bucket_gens) {
1145                         bkey_reassemble(&g->k_i, k);
1146                 } else {
1147                         bkey_bucket_gens_init(&g->k_i);
1148                         g->k.p = alloc_gens_pos(alloc_k.k->p, &gens_offset);
1149                 }
1150
1151                 g->v.gens[gens_offset] = a->gen;
1152
1153                 ret = bch2_trans_update(trans, bucket_gens_iter, &g->k_i, 0);
1154                 if (ret)
1155                         goto err;
1156         }
1157 err:
1158 fsck_err:
1159         printbuf_exit(&buf);
1160         return ret;
1161 }
1162
1163 static noinline_for_stack
1164 int bch2_check_alloc_hole_freespace(struct btree_trans *trans,
1165                                     struct bpos start,
1166                                     struct bpos *end,
1167                                     struct btree_iter *freespace_iter)
1168 {
1169         struct bch_fs *c = trans->c;
1170         struct bch_dev *ca;
1171         struct bkey_s_c k;
1172         struct printbuf buf = PRINTBUF;
1173         int ret;
1174
1175         ca = bch_dev_bkey_exists(c, start.inode);
1176         if (!ca->mi.freespace_initialized)
1177                 return 0;
1178
1179         bch2_btree_iter_set_pos(freespace_iter, start);
1180
1181         k = bch2_btree_iter_peek_slot(freespace_iter);
1182         ret = bkey_err(k);
1183         if (ret)
1184                 goto err;
1185
1186         *end = bkey_min(k.k->p, *end);
1187
1188         if (k.k->type != KEY_TYPE_set &&
1189             (c->opts.reconstruct_alloc ||
1190              fsck_err(c, "hole in alloc btree missing in freespace btree\n"
1191                       "  device %llu buckets %llu-%llu",
1192                       freespace_iter->pos.inode,
1193                       freespace_iter->pos.offset,
1194                       end->offset))) {
1195                 struct bkey_i *update =
1196                         bch2_trans_kmalloc(trans, sizeof(*update));
1197
1198                 ret = PTR_ERR_OR_ZERO(update);
1199                 if (ret)
1200                         goto err;
1201
1202                 bkey_init(&update->k);
1203                 update->k.type  = KEY_TYPE_set;
1204                 update->k.p     = freespace_iter->pos;
1205                 bch2_key_resize(&update->k,
1206                                 min_t(u64, U32_MAX, end->offset -
1207                                       freespace_iter->pos.offset));
1208
1209                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1210                 if (ret)
1211                         goto err;
1212         }
1213 err:
1214 fsck_err:
1215         printbuf_exit(&buf);
1216         return ret;
1217 }
1218
1219 static noinline_for_stack
1220 int bch2_check_alloc_hole_bucket_gens(struct btree_trans *trans,
1221                                       struct bpos start,
1222                                       struct bpos *end,
1223                                       struct btree_iter *bucket_gens_iter)
1224 {
1225         struct bch_fs *c = trans->c;
1226         struct bkey_s_c k;
1227         struct printbuf buf = PRINTBUF;
1228         unsigned i, gens_offset, gens_end_offset;
1229         int ret;
1230
1231         if (c->sb.version < bcachefs_metadata_version_bucket_gens &&
1232             !c->opts.version_upgrade)
1233                 return 0;
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                                         "hole in alloc btree at %llu:%llu with nonzero gen in bucket_gens btree (%u)",
1255                                         bucket_gens_pos_to_alloc(k.k->p, i).inode,
1256                                         bucket_gens_pos_to_alloc(k.k->p, i).offset,
1257                                         g.v.gens[i])) {
1258                                 g.v.gens[i] = 0;
1259                                 need_update = true;
1260                         }
1261                 }
1262
1263                 if (need_update) {
1264                         struct bkey_i *k = bch2_trans_kmalloc(trans, sizeof(g));
1265
1266                         ret = PTR_ERR_OR_ZERO(k);
1267                         if (ret)
1268                                 goto err;
1269
1270                         memcpy(k, &g, sizeof(g));
1271
1272                         ret = bch2_trans_update(trans, bucket_gens_iter, k, 0);
1273                         if (ret)
1274                                 goto err;
1275                 }
1276         }
1277
1278         *end = bkey_min(*end, bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0));
1279 err:
1280 fsck_err:
1281         printbuf_exit(&buf);
1282         return ret;
1283 }
1284
1285 static noinline_for_stack int __bch2_check_discard_freespace_key(struct btree_trans *trans,
1286                                               struct btree_iter *iter)
1287 {
1288         struct bch_fs *c = trans->c;
1289         struct btree_iter alloc_iter;
1290         struct bkey_s_c alloc_k;
1291         struct bch_alloc_v4 a_convert;
1292         const struct bch_alloc_v4 *a;
1293         u64 genbits;
1294         struct bpos pos;
1295         enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
1296                 ? BCH_DATA_need_discard
1297                 : BCH_DATA_free;
1298         struct printbuf buf = PRINTBUF;
1299         int ret;
1300
1301         pos = iter->pos;
1302         pos.offset &= ~(~0ULL << 56);
1303         genbits = iter->pos.offset & (~0ULL << 56);
1304
1305         alloc_k = bch2_bkey_get_iter(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
1306         ret = bkey_err(alloc_k);
1307         if (ret)
1308                 return ret;
1309
1310         if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
1311                         "entry in %s btree for nonexistant dev:bucket %llu:%llu",
1312                         bch2_btree_ids[iter->btree_id], pos.inode, pos.offset))
1313                 goto delete;
1314
1315         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1316
1317         if (fsck_err_on(a->data_type != state ||
1318                         (state == BCH_DATA_free &&
1319                          genbits != alloc_freespace_genbits(*a)), c,
1320                         "%s\n  incorrectly set at %s:%llu:%llu:0 (free %u, genbits %llu should be %llu)",
1321                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
1322                         bch2_btree_ids[iter->btree_id],
1323                         iter->pos.inode,
1324                         iter->pos.offset,
1325                         a->data_type == state,
1326                         genbits >> 56, alloc_freespace_genbits(*a) >> 56))
1327                 goto delete;
1328 out:
1329 fsck_err:
1330         set_btree_iter_dontneed(&alloc_iter);
1331         bch2_trans_iter_exit(trans, &alloc_iter);
1332         printbuf_exit(&buf);
1333         return ret;
1334 delete:
1335         ret =   bch2_btree_delete_extent_at(trans, iter,
1336                         iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0) ?:
1337                 bch2_trans_commit(trans, NULL, NULL,
1338                         BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW);
1339         goto out;
1340 }
1341
1342 static int bch2_check_discard_freespace_key(struct btree_trans *trans,
1343                                             struct btree_iter *iter,
1344                                             struct bpos end)
1345 {
1346         if (!btree_node_type_is_extents(iter->btree_id)) {
1347                 return __bch2_check_discard_freespace_key(trans, iter);
1348         } else {
1349                 int ret;
1350
1351                 while (!bkey_eq(iter->pos, end) &&
1352                        !(ret = btree_trans_too_many_iters(trans) ?:
1353                                __bch2_check_discard_freespace_key(trans, iter)))
1354                         bch2_btree_iter_set_pos(iter, bpos_nosnap_successor(iter->pos));
1355
1356                 return ret;
1357         }
1358 }
1359
1360 /*
1361  * We've already checked that generation numbers in the bucket_gens btree are
1362  * valid for buckets that exist; this just checks for keys for nonexistent
1363  * buckets.
1364  */
1365 static noinline_for_stack
1366 int bch2_check_bucket_gens_key(struct btree_trans *trans,
1367                                struct btree_iter *iter,
1368                                struct bkey_s_c k)
1369 {
1370         struct bch_fs *c = trans->c;
1371         struct bkey_i_bucket_gens g;
1372         struct bch_dev *ca;
1373         u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
1374         u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
1375         u64 b;
1376         bool need_update = false, dev_exists;
1377         struct printbuf buf = PRINTBUF;
1378         int ret = 0;
1379
1380         BUG_ON(k.k->type != KEY_TYPE_bucket_gens);
1381         bkey_reassemble(&g.k_i, k);
1382
1383         /* if no bch_dev, skip out whether we repair or not */
1384         dev_exists = bch2_dev_exists2(c, k.k->p.inode);
1385         if (!dev_exists) {
1386                 if (fsck_err_on(!dev_exists, c,
1387                                 "bucket_gens key for invalid device:\n  %s",
1388                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1389                         ret = bch2_btree_delete_at(trans, iter, 0);
1390                 }
1391                 goto out;
1392         }
1393
1394         ca = bch_dev_bkey_exists(c, k.k->p.inode);
1395         if (fsck_err_on(end <= ca->mi.first_bucket ||
1396                         start >= ca->mi.nbuckets, c,
1397                         "bucket_gens key for invalid buckets:\n  %s",
1398                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1399                 ret = bch2_btree_delete_at(trans, iter, 0);
1400                 goto out;
1401         }
1402
1403         for (b = start; b < ca->mi.first_bucket; b++)
1404                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1405                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1406                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1407                         need_update = true;
1408                 }
1409
1410         for (b = ca->mi.nbuckets; b < end; b++)
1411                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1412                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1413                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1414                         need_update = true;
1415                 }
1416
1417         if (need_update) {
1418                 struct bkey_i *k;
1419
1420                 k = bch2_trans_kmalloc(trans, sizeof(g));
1421                 ret = PTR_ERR_OR_ZERO(k);
1422                 if (ret)
1423                         goto out;
1424
1425                 memcpy(k, &g, sizeof(g));
1426                 ret = bch2_trans_update(trans, iter, k, 0);
1427         }
1428 out:
1429 fsck_err:
1430         printbuf_exit(&buf);
1431         return ret;
1432 }
1433
1434 int bch2_check_alloc_info(struct bch_fs *c)
1435 {
1436         struct btree_trans trans;
1437         struct btree_iter iter, discard_iter, freespace_iter, bucket_gens_iter;
1438         struct bkey hole;
1439         struct bkey_s_c k;
1440         int ret = 0;
1441
1442         bch2_trans_init(&trans, c, 0, 0);
1443
1444         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc, POS_MIN,
1445                              BTREE_ITER_PREFETCH);
1446         bch2_trans_iter_init(&trans, &discard_iter, BTREE_ID_need_discard, POS_MIN,
1447                              BTREE_ITER_PREFETCH);
1448         bch2_trans_iter_init(&trans, &freespace_iter, BTREE_ID_freespace, POS_MIN,
1449                              BTREE_ITER_PREFETCH);
1450         bch2_trans_iter_init(&trans, &bucket_gens_iter, BTREE_ID_bucket_gens, POS_MIN,
1451                              BTREE_ITER_PREFETCH);
1452
1453         while (1) {
1454                 struct bpos next;
1455
1456                 bch2_trans_begin(&trans);
1457
1458                 k = bch2_get_key_or_real_bucket_hole(&iter, &hole);
1459                 ret = bkey_err(k);
1460                 if (ret)
1461                         goto bkey_err;
1462
1463                 if (!k.k)
1464                         break;
1465
1466                 if (k.k->type) {
1467                         next = bpos_nosnap_successor(k.k->p);
1468
1469                         ret = bch2_check_alloc_key(&trans,
1470                                                    k, &iter,
1471                                                    &discard_iter,
1472                                                    &freespace_iter,
1473                                                    &bucket_gens_iter);
1474                         if (ret)
1475                                 goto bkey_err;
1476                 } else {
1477                         next = k.k->p;
1478
1479                         ret = bch2_check_alloc_hole_freespace(&trans,
1480                                                     bkey_start_pos(k.k),
1481                                                     &next,
1482                                                     &freespace_iter) ?:
1483                                 bch2_check_alloc_hole_bucket_gens(&trans,
1484                                                     bkey_start_pos(k.k),
1485                                                     &next,
1486                                                     &bucket_gens_iter);
1487                         if (ret)
1488                                 goto bkey_err;
1489                 }
1490
1491                 ret = bch2_trans_commit(&trans, NULL, NULL,
1492                                         BTREE_INSERT_NOFAIL|
1493                                         BTREE_INSERT_LAZY_RW);
1494                 if (ret)
1495                         goto bkey_err;
1496
1497                 bch2_btree_iter_set_pos(&iter, next);
1498 bkey_err:
1499                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1500                         continue;
1501                 if (ret)
1502                         break;
1503         }
1504         bch2_trans_iter_exit(&trans, &bucket_gens_iter);
1505         bch2_trans_iter_exit(&trans, &freespace_iter);
1506         bch2_trans_iter_exit(&trans, &discard_iter);
1507         bch2_trans_iter_exit(&trans, &iter);
1508
1509         if (ret < 0)
1510                 goto err;
1511
1512         ret = for_each_btree_key2(&trans, iter,
1513                         BTREE_ID_need_discard, POS_MIN,
1514                         BTREE_ITER_PREFETCH, k,
1515                 bch2_check_discard_freespace_key(&trans, &iter, k.k->p)) ?:
1516               for_each_btree_key2(&trans, iter,
1517                         BTREE_ID_freespace, POS_MIN,
1518                         BTREE_ITER_PREFETCH, k,
1519                 bch2_check_discard_freespace_key(&trans, &iter, k.k->p)) ?:
1520               for_each_btree_key_commit(&trans, iter,
1521                         BTREE_ID_bucket_gens, POS_MIN,
1522                         BTREE_ITER_PREFETCH, k,
1523                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1524                 bch2_check_bucket_gens_key(&trans, &iter, k));
1525 err:
1526         bch2_trans_exit(&trans);
1527         if (ret)
1528                 bch_err_fn(c, ret);
1529         return ret;
1530 }
1531
1532 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
1533                                        struct btree_iter *alloc_iter)
1534 {
1535         struct bch_fs *c = trans->c;
1536         struct btree_iter lru_iter;
1537         struct bch_alloc_v4 a_convert;
1538         const struct bch_alloc_v4 *a;
1539         struct bkey_s_c alloc_k, lru_k;
1540         struct printbuf buf = PRINTBUF;
1541         int ret;
1542
1543         alloc_k = bch2_btree_iter_peek(alloc_iter);
1544         if (!alloc_k.k)
1545                 return 0;
1546
1547         ret = bkey_err(alloc_k);
1548         if (ret)
1549                 return ret;
1550
1551         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1552
1553         if (a->data_type != BCH_DATA_cached)
1554                 return 0;
1555
1556         lru_k = bch2_bkey_get_iter(trans, &lru_iter, BTREE_ID_lru,
1557                              lru_pos(alloc_k.k->p.inode,
1558                                      bucket_to_u64(alloc_k.k->p),
1559                                      a->io_time[READ]), 0);
1560         ret = bkey_err(lru_k);
1561         if (ret)
1562                 return ret;
1563
1564         if (fsck_err_on(!a->io_time[READ], c,
1565                         "cached bucket with read_time 0\n"
1566                         "  %s",
1567                 (printbuf_reset(&buf),
1568                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) ||
1569             fsck_err_on(lru_k.k->type != KEY_TYPE_set, c,
1570                         "missing lru entry\n"
1571                         "  %s",
1572                         (printbuf_reset(&buf),
1573                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
1574                 u64 read_time = a->io_time[READ] ?:
1575                         atomic64_read(&c->io_clock[READ].now);
1576
1577                 ret = bch2_lru_set(trans,
1578                                    alloc_k.k->p.inode,
1579                                    bucket_to_u64(alloc_k.k->p),
1580                                    read_time);
1581                 if (ret)
1582                         goto err;
1583
1584                 if (a->io_time[READ] != read_time) {
1585                         struct bkey_i_alloc_v4 *a_mut =
1586                                 bch2_alloc_to_v4_mut(trans, alloc_k);
1587                         ret = PTR_ERR_OR_ZERO(a_mut);
1588                         if (ret)
1589                                 goto err;
1590
1591                         a_mut->v.io_time[READ] = read_time;
1592                         ret = bch2_trans_update(trans, alloc_iter,
1593                                                 &a_mut->k_i, BTREE_TRIGGER_NORUN);
1594                         if (ret)
1595                                 goto err;
1596                 }
1597         }
1598 err:
1599 fsck_err:
1600         bch2_trans_iter_exit(trans, &lru_iter);
1601         printbuf_exit(&buf);
1602         return ret;
1603 }
1604
1605 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
1606 {
1607         struct btree_iter iter;
1608         struct bkey_s_c k;
1609         int ret = 0;
1610
1611         ret = bch2_trans_run(c,
1612                 for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1613                                 POS_MIN, BTREE_ITER_PREFETCH, k,
1614                                 NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1615                         bch2_check_alloc_to_lru_ref(&trans, &iter)));
1616         if (ret)
1617                 bch_err_fn(c, ret);
1618         return ret;
1619 }
1620
1621 static int bch2_discard_one_bucket(struct btree_trans *trans,
1622                                    struct btree_iter *need_discard_iter,
1623                                    struct bpos *discard_pos_done,
1624                                    u64 *seen,
1625                                    u64 *open,
1626                                    u64 *need_journal_commit,
1627                                    u64 *discarded)
1628 {
1629         struct bch_fs *c = trans->c;
1630         struct bpos pos = need_discard_iter->pos;
1631         struct btree_iter iter = { NULL };
1632         struct bkey_s_c k;
1633         struct bch_dev *ca;
1634         struct bkey_i_alloc_v4 *a;
1635         struct printbuf buf = PRINTBUF;
1636         int ret = 0;
1637
1638         ca = bch_dev_bkey_exists(c, pos.inode);
1639         if (!percpu_ref_tryget(&ca->io_ref)) {
1640                 bch2_btree_iter_set_pos(need_discard_iter, POS(pos.inode + 1, 0));
1641                 return 0;
1642         }
1643
1644         if (bch2_bucket_is_open_safe(c, pos.inode, pos.offset)) {
1645                 (*open)++;
1646                 goto out;
1647         }
1648
1649         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
1650                         c->journal.flushed_seq_ondisk,
1651                         pos.inode, pos.offset)) {
1652                 (*need_journal_commit)++;
1653                 goto out;
1654         }
1655
1656         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_alloc,
1657                                need_discard_iter->pos,
1658                                BTREE_ITER_CACHED);
1659         ret = bkey_err(k);
1660         if (ret)
1661                 goto out;
1662
1663         a = bch2_alloc_to_v4_mut(trans, k);
1664         ret = PTR_ERR_OR_ZERO(a);
1665         if (ret)
1666                 goto out;
1667
1668         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
1669                 a->v.gen++;
1670                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1671                 goto write;
1672         }
1673
1674         if (a->v.journal_seq > c->journal.flushed_seq_ondisk) {
1675                 if (test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
1676                         bch2_trans_inconsistent(trans,
1677                                 "clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
1678                                 "%s",
1679                                 a->v.journal_seq,
1680                                 c->journal.flushed_seq_ondisk,
1681                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1682                         ret = -EIO;
1683                 }
1684                 goto out;
1685         }
1686
1687         if (a->v.data_type != BCH_DATA_need_discard) {
1688                 if (test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
1689                         bch2_trans_inconsistent(trans,
1690                                 "bucket incorrectly set in need_discard btree\n"
1691                                 "%s",
1692                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1693                         ret = -EIO;
1694                 }
1695
1696                 goto out;
1697         }
1698
1699         if (!bkey_eq(*discard_pos_done, iter.pos) &&
1700             ca->mi.discard && !c->opts.nochanges) {
1701                 /*
1702                  * This works without any other locks because this is the only
1703                  * thread that removes items from the need_discard tree
1704                  */
1705                 bch2_trans_unlock(trans);
1706                 blkdev_issue_discard(ca->disk_sb.bdev,
1707                                      k.k->p.offset * ca->mi.bucket_size,
1708                                      ca->mi.bucket_size,
1709                                      GFP_KERNEL);
1710                 *discard_pos_done = iter.pos;
1711
1712                 ret = bch2_trans_relock_notrace(trans);
1713                 if (ret)
1714                         goto out;
1715         }
1716
1717         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
1718         a->v.data_type = alloc_data_type(a->v, a->v.data_type);
1719 write:
1720         ret =   bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1721                 bch2_trans_commit(trans, NULL, NULL,
1722                                   BTREE_INSERT_USE_RESERVE|BTREE_INSERT_NOFAIL);
1723         if (ret)
1724                 goto out;
1725
1726         this_cpu_inc(c->counters[BCH_COUNTER_bucket_discard]);
1727         (*discarded)++;
1728 out:
1729         (*seen)++;
1730         bch2_trans_iter_exit(trans, &iter);
1731         percpu_ref_put(&ca->io_ref);
1732         printbuf_exit(&buf);
1733         return ret;
1734 }
1735
1736 static void bch2_do_discards_work(struct work_struct *work)
1737 {
1738         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
1739         struct btree_trans trans;
1740         struct btree_iter iter;
1741         struct bkey_s_c k;
1742         u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
1743         struct bpos discard_pos_done = POS_MAX;
1744         int ret;
1745
1746         bch2_trans_init(&trans, c, 0, 0);
1747
1748         /*
1749          * We're doing the commit in bch2_discard_one_bucket instead of using
1750          * for_each_btree_key_commit() so that we can increment counters after
1751          * successful commit:
1752          */
1753         ret = for_each_btree_key2(&trans, iter,
1754                         BTREE_ID_need_discard, POS_MIN, 0, k,
1755                 bch2_discard_one_bucket(&trans, &iter, &discard_pos_done,
1756                                         &seen,
1757                                         &open,
1758                                         &need_journal_commit,
1759                                         &discarded));
1760
1761         bch2_trans_exit(&trans);
1762
1763         if (need_journal_commit * 2 > seen)
1764                 bch2_journal_flush_async(&c->journal, NULL);
1765
1766         bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1767
1768         trace_discard_buckets(c, seen, open, need_journal_commit, discarded,
1769                               bch2_err_str(ret));
1770 }
1771
1772 void bch2_do_discards(struct bch_fs *c)
1773 {
1774         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_discard) &&
1775             !queue_work(c->write_ref_wq, &c->discard_work))
1776                 bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1777 }
1778
1779 static int invalidate_one_bucket(struct btree_trans *trans,
1780                                  struct btree_iter *lru_iter,
1781                                  struct bkey_s_c lru_k,
1782                                  s64 *nr_to_invalidate)
1783 {
1784         struct bch_fs *c = trans->c;
1785         struct btree_iter alloc_iter = { NULL };
1786         struct bkey_i_alloc_v4 *a = NULL;
1787         struct printbuf buf = PRINTBUF;
1788         struct bpos bucket = u64_to_bucket(lru_k.k->p.offset);
1789         unsigned cached_sectors;
1790         int ret = 0;
1791
1792         if (*nr_to_invalidate <= 0)
1793                 return 1;
1794
1795         if (!bch2_dev_bucket_exists(c, bucket)) {
1796                 prt_str(&buf, "lru entry points to invalid bucket");
1797                 goto err;
1798         }
1799
1800         if (bch2_bucket_is_open_safe(c, bucket.inode, bucket.offset))
1801                 return 0;
1802
1803         a = bch2_trans_start_alloc_update(trans, &alloc_iter, bucket);
1804         ret = PTR_ERR_OR_ZERO(a);
1805         if (ret)
1806                 goto out;
1807
1808         /* We expect harmless races here due to the btree write buffer: */
1809         if (lru_pos_time(lru_iter->pos) != alloc_lru_idx_read(a->v))
1810                 goto out;
1811
1812         BUG_ON(a->v.data_type != BCH_DATA_cached);
1813
1814         if (!a->v.cached_sectors)
1815                 bch_err(c, "invalidating empty bucket, confused");
1816
1817         cached_sectors = a->v.cached_sectors;
1818
1819         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1820         a->v.gen++;
1821         a->v.data_type          = 0;
1822         a->v.dirty_sectors      = 0;
1823         a->v.cached_sectors     = 0;
1824         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1825         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1826
1827         ret =   bch2_trans_update(trans, &alloc_iter, &a->k_i,
1828                                 BTREE_TRIGGER_BUCKET_INVALIDATE) ?:
1829                 bch2_trans_commit(trans, NULL, NULL,
1830                                   BTREE_INSERT_USE_RESERVE|BTREE_INSERT_NOFAIL);
1831         if (ret)
1832                 goto out;
1833
1834         trace_and_count(c, bucket_invalidate, c, bucket.inode, bucket.offset, cached_sectors);
1835         --*nr_to_invalidate;
1836 out:
1837         bch2_trans_iter_exit(trans, &alloc_iter);
1838         printbuf_exit(&buf);
1839         return ret;
1840 err:
1841         prt_str(&buf, "\n  lru key: ");
1842         bch2_bkey_val_to_text(&buf, c, lru_k);
1843
1844         prt_str(&buf, "\n  lru entry: ");
1845         bch2_lru_pos_to_text(&buf, lru_iter->pos);
1846
1847         prt_str(&buf, "\n  alloc key: ");
1848         if (!a)
1849                 bch2_bpos_to_text(&buf, bucket);
1850         else
1851                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
1852
1853         bch_err(c, "%s", buf.buf);
1854         if (test_bit(BCH_FS_CHECK_LRUS_DONE, &c->flags)) {
1855                 bch2_inconsistent_error(c);
1856                 ret = -EINVAL;
1857         }
1858
1859         goto out;
1860 }
1861
1862 static void bch2_do_invalidates_work(struct work_struct *work)
1863 {
1864         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1865         struct bch_dev *ca;
1866         struct btree_trans trans;
1867         struct btree_iter iter;
1868         struct bkey_s_c k;
1869         unsigned i;
1870         int ret = 0;
1871
1872         bch2_trans_init(&trans, c, 0, 0);
1873
1874         ret = bch2_btree_write_buffer_flush(&trans);
1875         if (ret)
1876                 goto err;
1877
1878         for_each_member_device(ca, c, i) {
1879                 s64 nr_to_invalidate =
1880                         should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
1881
1882                 ret = for_each_btree_key2_upto(&trans, iter, BTREE_ID_lru,
1883                                 lru_pos(ca->dev_idx, 0, 0),
1884                                 lru_pos(ca->dev_idx, U64_MAX, LRU_TIME_MAX),
1885                                 BTREE_ITER_INTENT, k,
1886                         invalidate_one_bucket(&trans, &iter, k, &nr_to_invalidate));
1887
1888                 if (ret < 0) {
1889                         percpu_ref_put(&ca->ref);
1890                         break;
1891                 }
1892         }
1893 err:
1894         bch2_trans_exit(&trans);
1895         bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1896 }
1897
1898 void bch2_do_invalidates(struct bch_fs *c)
1899 {
1900         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_invalidate) &&
1901             !queue_work(c->write_ref_wq, &c->invalidate_work))
1902                 bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1903 }
1904
1905 static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca,
1906                                    unsigned long *last_updated)
1907 {
1908         struct btree_trans trans;
1909         struct btree_iter iter;
1910         struct bkey_s_c k;
1911         struct bkey hole;
1912         struct bpos end = POS(ca->dev_idx, ca->mi.nbuckets);
1913         struct bch_member *m;
1914         int ret;
1915
1916         bch2_trans_init(&trans, c, 0, 0);
1917
1918         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc,
1919                              POS(ca->dev_idx, ca->mi.first_bucket),
1920                              BTREE_ITER_PREFETCH);
1921         /*
1922          * Scan the alloc btree for every bucket on @ca, and add buckets to the
1923          * freespace/need_discard/need_gc_gens btrees as needed:
1924          */
1925         while (1) {
1926                 if (*last_updated + HZ * 10 < jiffies) {
1927                         bch_info(ca, "%s: currently at %llu/%llu",
1928                                  __func__, iter.pos.offset, ca->mi.nbuckets);
1929                         *last_updated = jiffies;
1930                 }
1931
1932                 bch2_trans_begin(&trans);
1933
1934                 if (bkey_ge(iter.pos, end)) {
1935                         ret = 0;
1936                         break;
1937                 }
1938
1939                 k = bch2_get_key_or_hole(&iter, end, &hole);
1940                 ret = bkey_err(k);
1941                 if (ret)
1942                         goto bkey_err;
1943
1944                 if (k.k->type) {
1945                         /*
1946                          * We process live keys in the alloc btree one at a
1947                          * time:
1948                          */
1949                         struct bch_alloc_v4 a_convert;
1950                         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1951
1952                         ret =   bch2_bucket_do_index(&trans, k, a, true) ?:
1953                                 bch2_trans_commit(&trans, NULL, NULL,
1954                                                   BTREE_INSERT_LAZY_RW|
1955                                                   BTREE_INSERT_NOFAIL);
1956                         if (ret)
1957                                 goto bkey_err;
1958
1959                         bch2_btree_iter_advance(&iter);
1960                 } else {
1961                         struct bkey_i *freespace;
1962
1963                         freespace = bch2_trans_kmalloc(&trans, sizeof(*freespace));
1964                         ret = PTR_ERR_OR_ZERO(freespace);
1965                         if (ret)
1966                                 goto bkey_err;
1967
1968                         bkey_init(&freespace->k);
1969                         freespace->k.type       = KEY_TYPE_set;
1970                         freespace->k.p          = k.k->p;
1971                         freespace->k.size       = k.k->size;
1972
1973                         ret = __bch2_btree_insert(&trans, BTREE_ID_freespace, freespace, 0) ?:
1974                                 bch2_trans_commit(&trans, NULL, NULL,
1975                                                   BTREE_INSERT_LAZY_RW|
1976                                                   BTREE_INSERT_NOFAIL);
1977                         if (ret)
1978                                 goto bkey_err;
1979
1980                         bch2_btree_iter_set_pos(&iter, k.k->p);
1981                 }
1982 bkey_err:
1983                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1984                         continue;
1985                 if (ret)
1986                         break;
1987         }
1988
1989         bch2_trans_iter_exit(&trans, &iter);
1990         bch2_trans_exit(&trans);
1991
1992         if (ret < 0) {
1993                 bch_err(ca, "error initializing free space: %s", bch2_err_str(ret));
1994                 return ret;
1995         }
1996
1997         mutex_lock(&c->sb_lock);
1998         m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx;
1999         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
2000         mutex_unlock(&c->sb_lock);
2001
2002         return 0;
2003 }
2004
2005 int bch2_fs_freespace_init(struct bch_fs *c)
2006 {
2007         struct bch_dev *ca;
2008         unsigned i;
2009         int ret = 0;
2010         bool doing_init = false;
2011         unsigned long last_updated = jiffies;
2012
2013         /*
2014          * We can crash during the device add path, so we need to check this on
2015          * every mount:
2016          */
2017
2018         for_each_member_device(ca, c, i) {
2019                 if (ca->mi.freespace_initialized)
2020                         continue;
2021
2022                 if (!doing_init) {
2023                         bch_info(c, "initializing freespace");
2024                         doing_init = true;
2025                 }
2026
2027                 ret = bch2_dev_freespace_init(c, ca, &last_updated);
2028                 if (ret) {
2029                         percpu_ref_put(&ca->ref);
2030                         bch_err_fn(c, ret);
2031                         return ret;
2032                 }
2033         }
2034
2035         if (doing_init) {
2036                 mutex_lock(&c->sb_lock);
2037                 bch2_write_super(c);
2038                 mutex_unlock(&c->sb_lock);
2039                 bch_verbose(c, "done initializing freespace");
2040         }
2041
2042         return 0;
2043 }
2044
2045 /* Bucket IO clocks: */
2046
2047 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
2048                               size_t bucket_nr, int rw)
2049 {
2050         struct bch_fs *c = trans->c;
2051         struct btree_iter iter;
2052         struct bkey_i_alloc_v4 *a;
2053         u64 now;
2054         int ret = 0;
2055
2056         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
2057         ret = PTR_ERR_OR_ZERO(a);
2058         if (ret)
2059                 return ret;
2060
2061         now = atomic64_read(&c->io_clock[rw].now);
2062         if (a->v.io_time[rw] == now)
2063                 goto out;
2064
2065         a->v.io_time[rw] = now;
2066
2067         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
2068                 bch2_trans_commit(trans, NULL, NULL, 0);
2069 out:
2070         bch2_trans_iter_exit(trans, &iter);
2071         return ret;
2072 }
2073
2074 /* Startup/shutdown (ro/rw): */
2075
2076 void bch2_recalc_capacity(struct bch_fs *c)
2077 {
2078         struct bch_dev *ca;
2079         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
2080         unsigned bucket_size_max = 0;
2081         unsigned long ra_pages = 0;
2082         unsigned i;
2083
2084         lockdep_assert_held(&c->state_lock);
2085
2086         for_each_online_member(ca, c, i) {
2087                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
2088
2089                 ra_pages += bdi->ra_pages;
2090         }
2091
2092         bch2_set_ra_pages(c, ra_pages);
2093
2094         for_each_rw_member(ca, c, i) {
2095                 u64 dev_reserve = 0;
2096
2097                 /*
2098                  * We need to reserve buckets (from the number
2099                  * of currently available buckets) against
2100                  * foreground writes so that mainly copygc can
2101                  * make forward progress.
2102                  *
2103                  * We need enough to refill the various reserves
2104                  * from scratch - copygc will use its entire
2105                  * reserve all at once, then run against when
2106                  * its reserve is refilled (from the formerly
2107                  * available buckets).
2108                  *
2109                  * This reserve is just used when considering if
2110                  * allocations for foreground writes must wait -
2111                  * not -ENOSPC calculations.
2112                  */
2113
2114                 dev_reserve += ca->nr_btree_reserve * 2;
2115                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
2116
2117                 dev_reserve += 1;       /* btree write point */
2118                 dev_reserve += 1;       /* copygc write point */
2119                 dev_reserve += 1;       /* rebalance write point */
2120
2121                 dev_reserve *= ca->mi.bucket_size;
2122
2123                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
2124                                              ca->mi.first_bucket);
2125
2126                 reserved_sectors += dev_reserve * 2;
2127
2128                 bucket_size_max = max_t(unsigned, bucket_size_max,
2129                                         ca->mi.bucket_size);
2130         }
2131
2132         gc_reserve = c->opts.gc_reserve_bytes
2133                 ? c->opts.gc_reserve_bytes >> 9
2134                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
2135
2136         reserved_sectors = max(gc_reserve, reserved_sectors);
2137
2138         reserved_sectors = min(reserved_sectors, capacity);
2139
2140         c->capacity = capacity - reserved_sectors;
2141
2142         c->bucket_size_max = bucket_size_max;
2143
2144         /* Wake up case someone was waiting for buckets */
2145         closure_wake_up(&c->freelist_wait);
2146 }
2147
2148 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
2149 {
2150         struct open_bucket *ob;
2151         bool ret = false;
2152
2153         for (ob = c->open_buckets;
2154              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
2155              ob++) {
2156                 spin_lock(&ob->lock);
2157                 if (ob->valid && !ob->on_partial_list &&
2158                     ob->dev == ca->dev_idx)
2159                         ret = true;
2160                 spin_unlock(&ob->lock);
2161         }
2162
2163         return ret;
2164 }
2165
2166 /* device goes ro: */
2167 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
2168 {
2169         unsigned i;
2170
2171         /* First, remove device from allocation groups: */
2172
2173         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2174                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
2175
2176         /*
2177          * Capacity is calculated based off of devices in allocation groups:
2178          */
2179         bch2_recalc_capacity(c);
2180
2181         bch2_open_buckets_stop(c, ca, false);
2182
2183         /*
2184          * Wake up threads that were blocked on allocation, so they can notice
2185          * the device can no longer be removed and the capacity has changed:
2186          */
2187         closure_wake_up(&c->freelist_wait);
2188
2189         /*
2190          * journal_res_get() can block waiting for free space in the journal -
2191          * it needs to notice there may not be devices to allocate from anymore:
2192          */
2193         wake_up(&c->journal.wait);
2194
2195         /* Now wait for any in flight writes: */
2196
2197         closure_wait_event(&c->open_buckets_wait,
2198                            !bch2_dev_has_open_write_point(c, ca));
2199 }
2200
2201 /* device goes rw: */
2202 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
2203 {
2204         unsigned i;
2205
2206         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2207                 if (ca->mi.data_allowed & (1 << i))
2208                         set_bit(ca->dev_idx, c->rw_devs[i].d);
2209 }
2210
2211 void bch2_fs_allocator_background_init(struct bch_fs *c)
2212 {
2213         spin_lock_init(&c->freelist_lock);
2214         INIT_WORK(&c->discard_work, bch2_do_discards_work);
2215         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
2216 }