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(c, "error reading alloc info: %s", bch2_err_str(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(c, "%s: error %s", __func__, bch2_err_str(ret));
 688
 689         return ret;
 690 }
 691
 692 int bch2_bucket_gens_read(struct bch_fs *c)
 693 {
 694         struct btree_trans trans;
 695         struct btree_iter iter;
 696         struct bkey_s_c k;
 697         const struct bch_bucket_gens *g;
 698         struct bch_dev *ca;
 699         u64 b;
 700         int ret;
 701
 702         bch2_trans_init(&trans, c, 0, 0);
 703
 704         for_each_btree_key(&trans, iter, BTREE_ID_bucket_gens, POS_MIN,
 705                            BTREE_ITER_PREFETCH, k, ret) {
 706                 u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
 707                 u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
 708
 709                 if (k.k->type != KEY_TYPE_bucket_gens)
 710                         continue;
 711
 712                 g = bkey_s_c_to_bucket_gens(k).v;
 713
 714                 /*
 715                  * Not a fsck error because this is checked/repaired by
 716                  * bch2_check_alloc_key() which runs later:
 717                  */
 718                 if (!bch2_dev_exists2(c, k.k->p.inode))
 719                         continue;
 720
 721                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 722
 723                 for (b = max_t(u64, ca->mi.first_bucket, start);
 724                      b < min_t(u64, ca->mi.nbuckets, end);
 725                      b++)
 726                         *bucket_gen(ca, b) = g->gens[b & KEY_TYPE_BUCKET_GENS_MASK];
 727         }
 728         bch2_trans_iter_exit(&trans, &iter);
 729
 730         bch2_trans_exit(&trans);
 731
 732         if (ret)
 733                 bch_err(c, "error reading alloc info: %s", bch2_err_str(ret));
 734
 735         return ret;
 736 }
 737
 738 /* Free space/discard btree: */
 739
 740 static int bch2_bucket_do_index(struct btree_trans *trans,
 741                                 struct bkey_s_c alloc_k,
 742                                 const struct bch_alloc_v4 *a,
 743                                 bool set)
 744 {
 745         struct bch_fs *c = trans->c;
 746         struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 747         struct btree_iter iter;
 748         struct bkey_s_c old;
 749         struct bkey_i *k;
 750         enum btree_id btree;
 751         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 752         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 753         struct printbuf buf = PRINTBUF;
 754         int ret;
 755
 756         if (a->data_type != BCH_DATA_free &&
 757             a->data_type != BCH_DATA_need_discard)
 758                 return 0;
 759
 760         k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
 761         if (IS_ERR(k))
 762                 return PTR_ERR(k);
 763
 764         bkey_init(&k->k);
 765         k->k.type = new_type;
 766
 767         switch (a->data_type) {
 768         case BCH_DATA_free:
 769                 btree = BTREE_ID_freespace;
 770                 k->k.p = alloc_freespace_pos(alloc_k.k->p, *a);
 771                 bch2_key_resize(&k->k, 1);
 772                 break;
 773         case BCH_DATA_need_discard:
 774                 btree = BTREE_ID_need_discard;
 775                 k->k.p = alloc_k.k->p;
 776                 break;
 777         default:
 778                 return 0;
 779         }
 780
 781         old = bch2_bkey_get_iter(trans, &iter, btree,
 782                              bkey_start_pos(&k->k),
 783                              BTREE_ITER_INTENT);
 784         ret = bkey_err(old);
 785         if (ret)
 786                 return ret;
 787
 788         if (ca->mi.freespace_initialized &&
 789             test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags) &&
 790             bch2_trans_inconsistent_on(old.k->type != old_type, trans,
 791                         "incorrect key when %s %s btree (got %s should be %s)\n"
 792                         "  for %s",
 793                         set ? "setting" : "clearing",
 794                         bch2_btree_ids[btree],
 795                         bch2_bkey_types[old.k->type],
 796                         bch2_bkey_types[old_type],
 797                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 798                 ret = -EIO;
 799                 goto err;
 800         }
 801
 802         ret = bch2_trans_update(trans, &iter, k, 0);
 803 err:
 804         bch2_trans_iter_exit(trans, &iter);
 805         printbuf_exit(&buf);
 806         return ret;
 807 }
 808
 809 static noinline int bch2_bucket_gen_update(struct btree_trans *trans,
 810                                            struct bpos bucket, u8 gen)
 811 {
 812         struct btree_iter iter;
 813         unsigned offset;
 814         struct bpos pos = alloc_gens_pos(bucket, &offset);
 815         struct bkey_i_bucket_gens *g;
 816         struct bkey_s_c k;
 817         int ret;
 818
 819         g = bch2_trans_kmalloc(trans, sizeof(*g));
 820         ret = PTR_ERR_OR_ZERO(g);
 821         if (ret)
 822                 return ret;
 823
 824         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_bucket_gens, pos,
 825                                BTREE_ITER_INTENT|
 826                                BTREE_ITER_WITH_UPDATES);
 827         ret = bkey_err(k);
 828         if (ret)
 829                 return ret;
 830
 831         if (k.k->type != KEY_TYPE_bucket_gens) {
 832                 bkey_bucket_gens_init(&g->k_i);
 833                 g->k.p = iter.pos;
 834         } else {
 835                 bkey_reassemble(&g->k_i, k);
 836         }
 837
 838         g->v.gens[offset] = gen;
 839
 840         ret = bch2_trans_update(trans, &iter, &g->k_i, 0);
 841         bch2_trans_iter_exit(trans, &iter);
 842         return ret;
 843 }
 844
 845 int bch2_trans_mark_alloc(struct btree_trans *trans,
 846                           enum btree_id btree_id, unsigned level,
 847                           struct bkey_s_c old, struct bkey_i *new,
 848                           unsigned flags)
 849 {
 850         struct bch_fs *c = trans->c;
 851         struct bch_alloc_v4 old_a_convert, *new_a;
 852         const struct bch_alloc_v4 *old_a;
 853         u64 old_lru, new_lru;
 854         int ret = 0;
 855
 856         /*
 857          * Deletion only happens in the device removal path, with
 858          * BTREE_TRIGGER_NORUN:
 859          */
 860         BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
 861
 862         old_a = bch2_alloc_to_v4(old, &old_a_convert);
 863         new_a = &bkey_i_to_alloc_v4(new)->v;
 864
 865         new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
 866
 867         if (new_a->dirty_sectors > old_a->dirty_sectors ||
 868             new_a->cached_sectors > old_a->cached_sectors) {
 869                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 870                 new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 871                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
 872                 SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
 873         }
 874
 875         if (data_type_is_empty(new_a->data_type) &&
 876             BCH_ALLOC_V4_NEED_INC_GEN(new_a) &&
 877             !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
 878                 new_a->gen++;
 879                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
 880         }
 881
 882         if (old_a->data_type != new_a->data_type ||
 883             (new_a->data_type == BCH_DATA_free &&
 884              alloc_freespace_genbits(*old_a) != alloc_freespace_genbits(*new_a))) {
 885                 ret =   bch2_bucket_do_index(trans, old, old_a, false) ?:
 886                         bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_a, true);
 887                 if (ret)
 888                         return ret;
 889         }
 890
 891         if (new_a->data_type == BCH_DATA_cached &&
 892             !new_a->io_time[READ])
 893                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 894
 895         old_lru = alloc_lru_idx_read(*old_a);
 896         new_lru = alloc_lru_idx_read(*new_a);
 897
 898         if (old_lru != new_lru) {
 899                 ret = bch2_lru_change(trans, new->k.p.inode,
 900                                       bucket_to_u64(new->k.p),
 901                                       old_lru, new_lru);
 902                 if (ret)
 903                         return ret;
 904         }
 905
 906         new_a->fragmentation_lru = alloc_lru_idx_fragmentation(*new_a,
 907                                         bch_dev_bkey_exists(c, new->k.p.inode));
 908
 909         if (old_a->fragmentation_lru != new_a->fragmentation_lru) {
 910                 ret = bch2_lru_change(trans,
 911                                 BCH_LRU_FRAGMENTATION_START,
 912                                 bucket_to_u64(new->k.p),
 913                                 old_a->fragmentation_lru, new_a->fragmentation_lru);
 914                 if (ret)
 915                         return ret;
 916         }
 917
 918         if (old_a->gen != new_a->gen) {
 919                 ret = bch2_bucket_gen_update(trans, new->k.p, new_a->gen);
 920                 if (ret)
 921                         return ret;
 922         }
 923
 924         return 0;
 925 }
 926
 927 /*
 928  * This synthesizes deleted extents for holes, similar to BTREE_ITER_SLOTS for
 929  * extents style btrees, but works on non-extents btrees:
 930  */
 931 struct bkey_s_c bch2_get_key_or_hole(struct btree_iter *iter, struct bpos end, struct bkey *hole)
 932 {
 933         struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
 934
 935         if (bkey_err(k))
 936                 return k;
 937
 938         if (k.k->type) {
 939                 return k;
 940         } else {
 941                 struct btree_iter iter2;
 942                 struct bpos next;
 943
 944                 bch2_trans_copy_iter(&iter2, iter);
 945
 946                 if (!bpos_eq(iter->path->l[0].b->key.k.p, SPOS_MAX))
 947                         end = bkey_min(end, bpos_nosnap_successor(iter->path->l[0].b->key.k.p));
 948
 949                 end = bkey_min(end, POS(iter->pos.inode, iter->pos.offset + U32_MAX - 1));
 950
 951                 /*
 952                  * btree node min/max is a closed interval, upto takes a half
 953                  * open interval:
 954                  */
 955                 k = bch2_btree_iter_peek_upto(&iter2, end);
 956                 next = iter2.pos;
 957                 bch2_trans_iter_exit(iter->trans, &iter2);
 958
 959                 BUG_ON(next.offset >= iter->pos.offset + U32_MAX);
 960
 961                 if (bkey_err(k))
 962                         return k;
 963
 964                 bkey_init(hole);
 965                 hole->p = iter->pos;
 966
 967                 bch2_key_resize(hole, next.offset - iter->pos.offset);
 968                 return (struct bkey_s_c) { hole, NULL };
 969         }
 970 }
 971
 972 static bool next_bucket(struct bch_fs *c, struct bpos *bucket)
 973 {
 974         struct bch_dev *ca;
 975         unsigned iter;
 976
 977         if (bch2_dev_bucket_exists(c, *bucket))
 978                 return true;
 979
 980         if (bch2_dev_exists2(c, bucket->inode)) {
 981                 ca = bch_dev_bkey_exists(c, bucket->inode);
 982
 983                 if (bucket->offset < ca->mi.first_bucket) {
 984                         bucket->offset = ca->mi.first_bucket;
 985                         return true;
 986                 }
 987
 988                 bucket->inode++;
 989                 bucket->offset = 0;
 990         }
 991
 992         rcu_read_lock();
 993         iter = bucket->inode;
 994         ca = __bch2_next_dev(c, &iter, NULL);
 995         if (ca)
 996                 *bucket = POS(ca->dev_idx, ca->mi.first_bucket);
 997         rcu_read_unlock();
 998
 999         return ca != NULL;
1000 }
1001
1002 struct bkey_s_c bch2_get_key_or_real_bucket_hole(struct btree_iter *iter, struct bkey *hole)
1003 {
1004         struct bch_fs *c = iter->trans->c;
1005         struct bkey_s_c k;
1006 again:
1007         k = bch2_get_key_or_hole(iter, POS_MAX, hole);
1008         if (bkey_err(k))
1009                 return k;
1010
1011         if (!k.k->type) {
1012                 struct bpos bucket = bkey_start_pos(k.k);
1013
1014                 if (!bch2_dev_bucket_exists(c, bucket)) {
1015                         if (!next_bucket(c, &bucket))
1016                                 return bkey_s_c_null;
1017
1018                         bch2_btree_iter_set_pos(iter, bucket);
1019                         goto again;
1020                 }
1021
1022                 if (!bch2_dev_bucket_exists(c, k.k->p)) {
1023                         struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
1024
1025                         bch2_key_resize(hole, ca->mi.nbuckets - bucket.offset);
1026                 }
1027         }
1028
1029         return k;
1030 }
1031
1032 static int bch2_check_alloc_key(struct btree_trans *trans,
1033                                 struct bkey_s_c alloc_k,
1034                                 struct btree_iter *alloc_iter,
1035                                 struct btree_iter *discard_iter,
1036                                 struct btree_iter *freespace_iter,
1037                                 struct btree_iter *bucket_gens_iter)
1038 {
1039         struct bch_fs *c = trans->c;
1040         struct bch_dev *ca;
1041         struct bch_alloc_v4 a_convert;
1042         const struct bch_alloc_v4 *a;
1043         unsigned discard_key_type, freespace_key_type;
1044         unsigned gens_offset;
1045         struct bkey_s_c k;
1046         struct printbuf buf = PRINTBUF;
1047         int ret;
1048
1049         if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
1050                         "alloc key for invalid device:bucket %llu:%llu",
1051                         alloc_k.k->p.inode, alloc_k.k->p.offset))
1052                 return bch2_btree_delete_at(trans, alloc_iter, 0);
1053
1054         ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
1055         if (!ca->mi.freespace_initialized)
1056                 return 0;
1057
1058         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1059
1060         discard_key_type = a->data_type == BCH_DATA_need_discard ? KEY_TYPE_set : 0;
1061         bch2_btree_iter_set_pos(discard_iter, alloc_k.k->p);
1062         k = bch2_btree_iter_peek_slot(discard_iter);
1063         ret = bkey_err(k);
1064         if (ret)
1065                 goto err;
1066
1067         if (k.k->type != discard_key_type &&
1068             (c->opts.reconstruct_alloc ||
1069              fsck_err(c, "incorrect key in need_discard btree (got %s should be %s)\n"
1070                       "  %s",
1071                       bch2_bkey_types[k.k->type],
1072                       bch2_bkey_types[discard_key_type],
1073                       (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1074                 struct bkey_i *update =
1075                         bch2_trans_kmalloc(trans, sizeof(*update));
1076
1077                 ret = PTR_ERR_OR_ZERO(update);
1078                 if (ret)
1079                         goto err;
1080
1081                 bkey_init(&update->k);
1082                 update->k.type  = discard_key_type;
1083                 update->k.p     = discard_iter->pos;
1084
1085                 ret = bch2_trans_update(trans, discard_iter, update, 0);
1086                 if (ret)
1087                         goto err;
1088         }
1089
1090         freespace_key_type = a->data_type == BCH_DATA_free ? KEY_TYPE_set : 0;
1091         bch2_btree_iter_set_pos(freespace_iter, alloc_freespace_pos(alloc_k.k->p, *a));
1092         k = bch2_btree_iter_peek_slot(freespace_iter);
1093         ret = bkey_err(k);
1094         if (ret)
1095                 goto err;
1096
1097         if (k.k->type != freespace_key_type &&
1098             (c->opts.reconstruct_alloc ||
1099              fsck_err(c, "incorrect key in freespace btree (got %s should be %s)\n"
1100                       "  %s",
1101                       bch2_bkey_types[k.k->type],
1102                       bch2_bkey_types[freespace_key_type],
1103                       (printbuf_reset(&buf),
1104                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1105                 struct bkey_i *update =
1106                         bch2_trans_kmalloc(trans, sizeof(*update));
1107
1108                 ret = PTR_ERR_OR_ZERO(update);
1109                 if (ret)
1110                         goto err;
1111
1112                 bkey_init(&update->k);
1113                 update->k.type  = freespace_key_type;
1114                 update->k.p     = freespace_iter->pos;
1115                 bch2_key_resize(&update->k, 1);
1116
1117                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1118                 if (ret)
1119                         goto err;
1120         }
1121
1122         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(alloc_k.k->p, &gens_offset));
1123         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1124         ret = bkey_err(k);
1125         if (ret)
1126                 goto err;
1127
1128         if (a->gen != alloc_gen(k, gens_offset) &&
1129             (c->opts.reconstruct_alloc ||
1130              fsck_err(c, "incorrect gen in bucket_gens btree (got %u should be %u)\n"
1131                       "  %s",
1132                       alloc_gen(k, gens_offset), a->gen,
1133                       (printbuf_reset(&buf),
1134                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1135                 struct bkey_i_bucket_gens *g =
1136                         bch2_trans_kmalloc(trans, sizeof(*g));
1137
1138                 ret = PTR_ERR_OR_ZERO(g);
1139                 if (ret)
1140                         goto err;
1141
1142                 if (k.k->type == KEY_TYPE_bucket_gens) {
1143                         bkey_reassemble(&g->k_i, k);
1144                 } else {
1145                         bkey_bucket_gens_init(&g->k_i);
1146                         g->k.p = alloc_gens_pos(alloc_k.k->p, &gens_offset);
1147                 }
1148
1149                 g->v.gens[gens_offset] = a->gen;
1150
1151                 ret = bch2_trans_update(trans, bucket_gens_iter, &g->k_i, 0);
1152                 if (ret)
1153                         goto err;
1154         }
1155 err:
1156 fsck_err:
1157         printbuf_exit(&buf);
1158         return ret;
1159 }
1160
1161 static int bch2_check_alloc_hole_freespace(struct btree_trans *trans,
1162                                  struct bpos start,
1163                                  struct bpos *end,
1164                                  struct btree_iter *freespace_iter)
1165 {
1166         struct bch_fs *c = trans->c;
1167         struct bch_dev *ca;
1168         struct bkey_s_c k;
1169         struct printbuf buf = PRINTBUF;
1170         int ret;
1171
1172         ca = bch_dev_bkey_exists(c, start.inode);
1173         if (!ca->mi.freespace_initialized)
1174                 return 0;
1175
1176         bch2_btree_iter_set_pos(freespace_iter, start);
1177
1178         k = bch2_btree_iter_peek_slot(freespace_iter);
1179         ret = bkey_err(k);
1180         if (ret)
1181                 goto err;
1182
1183         *end = bkey_min(k.k->p, *end);
1184
1185         if (k.k->type != KEY_TYPE_set &&
1186             (c->opts.reconstruct_alloc ||
1187              fsck_err(c, "hole in alloc btree missing in freespace btree\n"
1188                       "  device %llu buckets %llu-%llu",
1189                       freespace_iter->pos.inode,
1190                       freespace_iter->pos.offset,
1191                       end->offset))) {
1192                 struct bkey_i *update =
1193                         bch2_trans_kmalloc(trans, sizeof(*update));
1194
1195                 ret = PTR_ERR_OR_ZERO(update);
1196                 if (ret)
1197                         goto err;
1198
1199                 bkey_init(&update->k);
1200                 update->k.type  = KEY_TYPE_set;
1201                 update->k.p     = freespace_iter->pos;
1202                 bch2_key_resize(&update->k,
1203                                 min_t(u64, U32_MAX, end->offset -
1204                                       freespace_iter->pos.offset));
1205
1206                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1207                 if (ret)
1208                         goto err;
1209         }
1210 err:
1211 fsck_err:
1212         printbuf_exit(&buf);
1213         return ret;
1214 }
1215
1216 static int bch2_check_alloc_hole_bucket_gens(struct btree_trans *trans,
1217                                  struct bpos start,
1218                                  struct bpos *end,
1219                                  struct btree_iter *bucket_gens_iter)
1220 {
1221         struct bch_fs *c = trans->c;
1222         struct bkey_s_c k;
1223         struct printbuf buf = PRINTBUF;
1224         unsigned i, gens_offset, gens_end_offset;
1225         int ret;
1226
1227         if (c->sb.version < bcachefs_metadata_version_bucket_gens &&
1228             !c->opts.version_upgrade)
1229                 return 0;
1230
1231         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(start, &gens_offset));
1232
1233         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1234         ret = bkey_err(k);
1235         if (ret)
1236                 goto err;
1237
1238         if (bkey_cmp(alloc_gens_pos(start, &gens_offset),
1239                      alloc_gens_pos(*end,  &gens_end_offset)))
1240                 gens_end_offset = KEY_TYPE_BUCKET_GENS_NR;
1241
1242         if (k.k->type == KEY_TYPE_bucket_gens) {
1243                 struct bkey_i_bucket_gens g;
1244                 bool need_update = false;
1245
1246                 bkey_reassemble(&g.k_i, k);
1247
1248                 for (i = gens_offset; i < gens_end_offset; i++) {
1249                         if (fsck_err_on(g.v.gens[i], c,
1250                                         "hole in alloc btree at %llu:%llu with nonzero gen in bucket_gens btree (%u)",
1251                                         bucket_gens_pos_to_alloc(k.k->p, i).inode,
1252                                         bucket_gens_pos_to_alloc(k.k->p, i).offset,
1253                                         g.v.gens[i])) {
1254                                 g.v.gens[i] = 0;
1255                                 need_update = true;
1256                         }
1257                 }
1258
1259                 if (need_update) {
1260                         struct bkey_i *k = bch2_trans_kmalloc(trans, sizeof(g));
1261
1262                         ret = PTR_ERR_OR_ZERO(k);
1263                         if (ret)
1264                                 goto err;
1265
1266                         memcpy(k, &g, sizeof(g));
1267
1268                         ret = bch2_trans_update(trans, bucket_gens_iter, k, 0);
1269                         if (ret)
1270                                 goto err;
1271                 }
1272         }
1273
1274         *end = bkey_min(*end, bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0));
1275 err:
1276 fsck_err:
1277         printbuf_exit(&buf);
1278         return ret;
1279 }
1280
1281 static int bch2_check_discard_freespace_key(struct btree_trans *trans,
1282                                             struct btree_iter *iter)
1283 {
1284         struct bch_fs *c = trans->c;
1285         struct btree_iter alloc_iter;
1286         struct bkey_s_c alloc_k;
1287         struct bch_alloc_v4 a_convert;
1288         const struct bch_alloc_v4 *a;
1289         u64 genbits;
1290         struct bpos pos;
1291         enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
1292                 ? BCH_DATA_need_discard
1293                 : BCH_DATA_free;
1294         struct printbuf buf = PRINTBUF;
1295         int ret;
1296
1297         pos = iter->pos;
1298         pos.offset &= ~(~0ULL << 56);
1299         genbits = iter->pos.offset & (~0ULL << 56);
1300
1301         alloc_k = bch2_bkey_get_iter(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
1302         ret = bkey_err(alloc_k);
1303         if (ret)
1304                 return ret;
1305
1306         if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
1307                         "entry in %s btree for nonexistant dev:bucket %llu:%llu",
1308                         bch2_btree_ids[iter->btree_id], pos.inode, pos.offset))
1309                 goto delete;
1310
1311         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1312
1313         if (fsck_err_on(a->data_type != state ||
1314                         (state == BCH_DATA_free &&
1315                          genbits != alloc_freespace_genbits(*a)), c,
1316                         "%s\n  incorrectly set in %s index (free %u, genbits %llu should be %llu)",
1317                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
1318                         bch2_btree_ids[iter->btree_id],
1319                         a->data_type == state,
1320                         genbits >> 56, alloc_freespace_genbits(*a) >> 56))
1321                 goto delete;
1322 out:
1323 fsck_err:
1324         bch2_trans_iter_exit(trans, &alloc_iter);
1325         printbuf_exit(&buf);
1326         return ret;
1327 delete:
1328         ret = bch2_btree_delete_extent_at(trans, iter,
1329                         iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0);
1330         goto out;
1331 }
1332
1333 /*
1334  * We've already checked that generation numbers in the bucket_gens btree are
1335  * valid for buckets that exist; this just checks for keys for nonexistent
1336  * buckets.
1337  */
1338 static int bch2_check_bucket_gens_key(struct btree_trans *trans,
1339                                       struct btree_iter *iter,
1340                                       struct bkey_s_c k)
1341 {
1342         struct bch_fs *c = trans->c;
1343         struct bkey_i_bucket_gens g;
1344         struct bch_dev *ca;
1345         u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
1346         u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
1347         u64 b;
1348         bool need_update = false, dev_exists;
1349         struct printbuf buf = PRINTBUF;
1350         int ret = 0;
1351
1352         BUG_ON(k.k->type != KEY_TYPE_bucket_gens);
1353         bkey_reassemble(&g.k_i, k);
1354
1355         /* if no bch_dev, skip out whether we repair or not */
1356         dev_exists = bch2_dev_exists2(c, k.k->p.inode);
1357         if (!dev_exists) {
1358                 if (fsck_err_on(!dev_exists, c,
1359                                 "bucket_gens key for invalid device:\n  %s",
1360                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1361                         ret = bch2_btree_delete_at(trans, iter, 0);
1362                 }
1363                 goto out;
1364         }
1365
1366         ca = bch_dev_bkey_exists(c, k.k->p.inode);
1367         if (fsck_err_on(end <= ca->mi.first_bucket ||
1368                         start >= ca->mi.nbuckets, c,
1369                         "bucket_gens key for invalid buckets:\n  %s",
1370                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1371                 ret = bch2_btree_delete_at(trans, iter, 0);
1372                 goto out;
1373         }
1374
1375         for (b = start; b < ca->mi.first_bucket; b++)
1376                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1377                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1378                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1379                         need_update = true;
1380                 }
1381
1382         for (b = ca->mi.nbuckets; b < end; b++)
1383                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1384                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1385                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1386                         need_update = true;
1387                 }
1388
1389         if (need_update) {
1390                 struct bkey_i *k;
1391
1392                 k = bch2_trans_kmalloc(trans, sizeof(g));
1393                 ret = PTR_ERR_OR_ZERO(k);
1394                 if (ret)
1395                         goto out;
1396
1397                 memcpy(k, &g, sizeof(g));
1398                 ret = bch2_trans_update(trans, iter, k, 0);
1399         }
1400 out:
1401 fsck_err:
1402         printbuf_exit(&buf);
1403         return ret;
1404 }
1405
1406 int bch2_check_alloc_info(struct bch_fs *c)
1407 {
1408         struct btree_trans trans;
1409         struct btree_iter iter, discard_iter, freespace_iter, bucket_gens_iter;
1410         struct bkey hole;
1411         struct bkey_s_c k;
1412         int ret = 0;
1413
1414         bch2_trans_init(&trans, c, 0, 0);
1415
1416         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc, POS_MIN,
1417                              BTREE_ITER_PREFETCH);
1418         bch2_trans_iter_init(&trans, &discard_iter, BTREE_ID_need_discard, POS_MIN,
1419                              BTREE_ITER_PREFETCH);
1420         bch2_trans_iter_init(&trans, &freespace_iter, BTREE_ID_freespace, POS_MIN,
1421                              BTREE_ITER_PREFETCH);
1422         bch2_trans_iter_init(&trans, &bucket_gens_iter, BTREE_ID_bucket_gens, POS_MIN,
1423                              BTREE_ITER_PREFETCH);
1424
1425         while (1) {
1426                 struct bpos next;
1427
1428                 bch2_trans_begin(&trans);
1429
1430                 k = bch2_get_key_or_real_bucket_hole(&iter, &hole);
1431                 ret = bkey_err(k);
1432                 if (ret)
1433                         goto bkey_err;
1434
1435                 if (!k.k)
1436                         break;
1437
1438                 if (k.k->type) {
1439                         next = bpos_nosnap_successor(k.k->p);
1440
1441                         ret = bch2_check_alloc_key(&trans,
1442                                                    k, &iter,
1443                                                    &discard_iter,
1444                                                    &freespace_iter,
1445                                                    &bucket_gens_iter);
1446                         if (ret)
1447                                 goto bkey_err;
1448                 } else {
1449                         next = k.k->p;
1450
1451                         ret = bch2_check_alloc_hole_freespace(&trans,
1452                                                     bkey_start_pos(k.k),
1453                                                     &next,
1454                                                     &freespace_iter) ?:
1455                                 bch2_check_alloc_hole_bucket_gens(&trans,
1456                                                     bkey_start_pos(k.k),
1457                                                     &next,
1458                                                     &bucket_gens_iter);
1459                         if (ret)
1460                                 goto bkey_err;
1461                 }
1462
1463                 ret = bch2_trans_commit(&trans, NULL, NULL,
1464                                         BTREE_INSERT_NOFAIL|
1465                                         BTREE_INSERT_LAZY_RW);
1466                 if (ret)
1467                         goto bkey_err;
1468
1469                 bch2_btree_iter_set_pos(&iter, next);
1470 bkey_err:
1471                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1472                         continue;
1473                 if (ret)
1474                         break;
1475         }
1476         bch2_trans_iter_exit(&trans, &bucket_gens_iter);
1477         bch2_trans_iter_exit(&trans, &freespace_iter);
1478         bch2_trans_iter_exit(&trans, &discard_iter);
1479         bch2_trans_iter_exit(&trans, &iter);
1480
1481         if (ret < 0)
1482                 goto err;
1483
1484         ret = for_each_btree_key_commit(&trans, iter,
1485                         BTREE_ID_need_discard, POS_MIN,
1486                         BTREE_ITER_PREFETCH, k,
1487                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1488                 bch2_check_discard_freespace_key(&trans, &iter)) ?:
1489               for_each_btree_key_commit(&trans, iter,
1490                         BTREE_ID_freespace, POS_MIN,
1491                         BTREE_ITER_PREFETCH, k,
1492                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1493                 bch2_check_discard_freespace_key(&trans, &iter)) ?:
1494               for_each_btree_key_commit(&trans, iter,
1495                         BTREE_ID_bucket_gens, POS_MIN,
1496                         BTREE_ITER_PREFETCH, k,
1497                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1498                 bch2_check_bucket_gens_key(&trans, &iter, k));
1499 err:
1500         bch2_trans_exit(&trans);
1501         return ret < 0 ? ret : 0;
1502 }
1503
1504 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
1505                                        struct btree_iter *alloc_iter)
1506 {
1507         struct bch_fs *c = trans->c;
1508         struct btree_iter lru_iter;
1509         struct bch_alloc_v4 a_convert;
1510         const struct bch_alloc_v4 *a;
1511         struct bkey_s_c alloc_k, lru_k;
1512         struct printbuf buf = PRINTBUF;
1513         int ret;
1514
1515         alloc_k = bch2_btree_iter_peek(alloc_iter);
1516         if (!alloc_k.k)
1517                 return 0;
1518
1519         ret = bkey_err(alloc_k);
1520         if (ret)
1521                 return ret;
1522
1523         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1524
1525         if (a->data_type != BCH_DATA_cached)
1526                 return 0;
1527
1528         lru_k = bch2_bkey_get_iter(trans, &lru_iter, BTREE_ID_lru,
1529                              lru_pos(alloc_k.k->p.inode,
1530                                      bucket_to_u64(alloc_k.k->p),
1531                                      a->io_time[READ]), 0);
1532         ret = bkey_err(lru_k);
1533         if (ret)
1534                 return ret;
1535
1536         if (fsck_err_on(!a->io_time[READ], c,
1537                         "cached bucket with read_time 0\n"
1538                         "  %s",
1539                 (printbuf_reset(&buf),
1540                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) ||
1541             fsck_err_on(lru_k.k->type != KEY_TYPE_set, c,
1542                         "missing lru entry\n"
1543                         "  %s",
1544                         (printbuf_reset(&buf),
1545                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
1546                 u64 read_time = a->io_time[READ] ?:
1547                         atomic64_read(&c->io_clock[READ].now);
1548
1549                 ret = bch2_lru_set(trans,
1550                                    alloc_k.k->p.inode,
1551                                    bucket_to_u64(alloc_k.k->p),
1552                                    read_time);
1553                 if (ret)
1554                         goto err;
1555
1556                 if (a->io_time[READ] != read_time) {
1557                         struct bkey_i_alloc_v4 *a_mut =
1558                                 bch2_alloc_to_v4_mut(trans, alloc_k);
1559                         ret = PTR_ERR_OR_ZERO(a_mut);
1560                         if (ret)
1561                                 goto err;
1562
1563                         a_mut->v.io_time[READ] = read_time;
1564                         ret = bch2_trans_update(trans, alloc_iter,
1565                                                 &a_mut->k_i, BTREE_TRIGGER_NORUN);
1566                         if (ret)
1567                                 goto err;
1568                 }
1569         }
1570 err:
1571 fsck_err:
1572         bch2_trans_iter_exit(trans, &lru_iter);
1573         printbuf_exit(&buf);
1574         return ret;
1575 }
1576
1577 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
1578 {
1579         struct btree_trans trans;
1580         struct btree_iter iter;
1581         struct bkey_s_c k;
1582         int ret = 0;
1583
1584         bch2_trans_init(&trans, c, 0, 0);
1585
1586         for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1587                         POS_MIN, BTREE_ITER_PREFETCH, k,
1588                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1589                 bch2_check_alloc_to_lru_ref(&trans, &iter));
1590
1591         bch2_trans_exit(&trans);
1592         return ret < 0 ? ret : 0;
1593 }
1594
1595 static int bch2_discard_one_bucket(struct btree_trans *trans,
1596                                    struct btree_iter *need_discard_iter,
1597                                    struct bpos *discard_pos_done,
1598                                    u64 *seen,
1599                                    u64 *open,
1600                                    u64 *need_journal_commit,
1601                                    u64 *discarded)
1602 {
1603         struct bch_fs *c = trans->c;
1604         struct bpos pos = need_discard_iter->pos;
1605         struct btree_iter iter = { NULL };
1606         struct bkey_s_c k;
1607         struct bch_dev *ca;
1608         struct bkey_i_alloc_v4 *a;
1609         struct printbuf buf = PRINTBUF;
1610         int ret = 0;
1611
1612         ca = bch_dev_bkey_exists(c, pos.inode);
1613         if (!percpu_ref_tryget(&ca->io_ref)) {
1614                 bch2_btree_iter_set_pos(need_discard_iter, POS(pos.inode + 1, 0));
1615                 return 0;
1616         }
1617
1618         if (bch2_bucket_is_open_safe(c, pos.inode, pos.offset)) {
1619                 (*open)++;
1620                 goto out;
1621         }
1622
1623         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
1624                         c->journal.flushed_seq_ondisk,
1625                         pos.inode, pos.offset)) {
1626                 (*need_journal_commit)++;
1627                 goto out;
1628         }
1629
1630         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_alloc,
1631                                need_discard_iter->pos,
1632                                BTREE_ITER_CACHED);
1633         ret = bkey_err(k);
1634         if (ret)
1635                 goto out;
1636
1637         a = bch2_alloc_to_v4_mut(trans, k);
1638         ret = PTR_ERR_OR_ZERO(a);
1639         if (ret)
1640                 goto out;
1641
1642         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
1643                 a->v.gen++;
1644                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1645                 goto write;
1646         }
1647
1648         if (a->v.journal_seq > c->journal.flushed_seq_ondisk) {
1649                 if (test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
1650                         bch2_trans_inconsistent(trans,
1651                                 "clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
1652                                 "%s",
1653                                 a->v.journal_seq,
1654                                 c->journal.flushed_seq_ondisk,
1655                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1656                         ret = -EIO;
1657                 }
1658                 goto out;
1659         }
1660
1661         if (a->v.data_type != BCH_DATA_need_discard) {
1662                 if (test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
1663                         bch2_trans_inconsistent(trans,
1664                                 "bucket incorrectly set in need_discard btree\n"
1665                                 "%s",
1666                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1667                         ret = -EIO;
1668                 }
1669
1670                 goto out;
1671         }
1672
1673         if (!bkey_eq(*discard_pos_done, iter.pos) &&
1674             ca->mi.discard && !c->opts.nochanges) {
1675                 /*
1676                  * This works without any other locks because this is the only
1677                  * thread that removes items from the need_discard tree
1678                  */
1679                 bch2_trans_unlock(trans);
1680                 blkdev_issue_discard(ca->disk_sb.bdev,
1681                                      k.k->p.offset * ca->mi.bucket_size,
1682                                      ca->mi.bucket_size,
1683                                      GFP_KERNEL);
1684                 *discard_pos_done = iter.pos;
1685
1686                 ret = bch2_trans_relock_notrace(trans);
1687                 if (ret)
1688                         goto out;
1689         }
1690
1691         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
1692         a->v.data_type = alloc_data_type(a->v, a->v.data_type);
1693 write:
1694         ret =   bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1695                 bch2_trans_commit(trans, NULL, NULL,
1696                                   BTREE_INSERT_USE_RESERVE|BTREE_INSERT_NOFAIL);
1697         if (ret)
1698                 goto out;
1699
1700         this_cpu_inc(c->counters[BCH_COUNTER_bucket_discard]);
1701         (*discarded)++;
1702 out:
1703         (*seen)++;
1704         bch2_trans_iter_exit(trans, &iter);
1705         percpu_ref_put(&ca->io_ref);
1706         printbuf_exit(&buf);
1707         return ret;
1708 }
1709
1710 static void bch2_do_discards_work(struct work_struct *work)
1711 {
1712         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
1713         struct btree_trans trans;
1714         struct btree_iter iter;
1715         struct bkey_s_c k;
1716         u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
1717         struct bpos discard_pos_done = POS_MAX;
1718         int ret;
1719
1720         bch2_trans_init(&trans, c, 0, 0);
1721
1722         /*
1723          * We're doing the commit in bch2_discard_one_bucket instead of using
1724          * for_each_btree_key_commit() so that we can increment counters after
1725          * successful commit:
1726          */
1727         ret = for_each_btree_key2(&trans, iter,
1728                         BTREE_ID_need_discard, POS_MIN, 0, k,
1729                 bch2_discard_one_bucket(&trans, &iter, &discard_pos_done,
1730                                         &seen,
1731                                         &open,
1732                                         &need_journal_commit,
1733                                         &discarded));
1734
1735         bch2_trans_exit(&trans);
1736
1737         if (need_journal_commit * 2 > seen)
1738                 bch2_journal_flush_async(&c->journal, NULL);
1739
1740         bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1741
1742         trace_discard_buckets(c, seen, open, need_journal_commit, discarded,
1743                               bch2_err_str(ret));
1744 }
1745
1746 void bch2_do_discards(struct bch_fs *c)
1747 {
1748         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_discard) &&
1749             !queue_work(c->write_ref_wq, &c->discard_work))
1750                 bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1751 }
1752
1753 static int invalidate_one_bucket(struct btree_trans *trans,
1754                                  struct btree_iter *lru_iter,
1755                                  struct bkey_s_c lru_k,
1756                                  s64 *nr_to_invalidate)
1757 {
1758         struct bch_fs *c = trans->c;
1759         struct btree_iter alloc_iter = { NULL };
1760         struct bkey_i_alloc_v4 *a = NULL;
1761         struct printbuf buf = PRINTBUF;
1762         struct bpos bucket = u64_to_bucket(lru_k.k->p.offset);
1763         unsigned cached_sectors;
1764         int ret = 0;
1765
1766         if (*nr_to_invalidate <= 0)
1767                 return 1;
1768
1769         if (!bch2_dev_bucket_exists(c, bucket)) {
1770                 prt_str(&buf, "lru entry points to invalid bucket");
1771                 goto err;
1772         }
1773
1774         if (bch2_bucket_is_open_safe(c, bucket.inode, bucket.offset))
1775                 return 0;
1776
1777         a = bch2_trans_start_alloc_update(trans, &alloc_iter, bucket);
1778         ret = PTR_ERR_OR_ZERO(a);
1779         if (ret)
1780                 goto out;
1781
1782         /* We expect harmless races here due to the btree write buffer: */
1783         if (lru_pos_time(lru_iter->pos) != alloc_lru_idx_read(a->v))
1784                 goto out;
1785
1786         BUG_ON(a->v.data_type != BCH_DATA_cached);
1787
1788         if (!a->v.cached_sectors)
1789                 bch_err(c, "invalidating empty bucket, confused");
1790
1791         cached_sectors = a->v.cached_sectors;
1792
1793         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1794         a->v.gen++;
1795         a->v.data_type          = 0;
1796         a->v.dirty_sectors      = 0;
1797         a->v.cached_sectors     = 0;
1798         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1799         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1800
1801         ret =   bch2_trans_update(trans, &alloc_iter, &a->k_i,
1802                                 BTREE_TRIGGER_BUCKET_INVALIDATE) ?:
1803                 bch2_trans_commit(trans, NULL, NULL,
1804                                   BTREE_INSERT_USE_RESERVE|BTREE_INSERT_NOFAIL);
1805         if (ret)
1806                 goto out;
1807
1808         trace_and_count(c, bucket_invalidate, c, bucket.inode, bucket.offset, cached_sectors);
1809         --*nr_to_invalidate;
1810 out:
1811         bch2_trans_iter_exit(trans, &alloc_iter);
1812         printbuf_exit(&buf);
1813         return ret;
1814 err:
1815         prt_str(&buf, "\n  lru key: ");
1816         bch2_bkey_val_to_text(&buf, c, lru_k);
1817
1818         prt_str(&buf, "\n  lru entry: ");
1819         bch2_lru_pos_to_text(&buf, lru_iter->pos);
1820
1821         prt_str(&buf, "\n  alloc key: ");
1822         if (!a)
1823                 bch2_bpos_to_text(&buf, bucket);
1824         else
1825                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
1826
1827         bch_err(c, "%s", buf.buf);
1828         if (test_bit(BCH_FS_CHECK_LRUS_DONE, &c->flags)) {
1829                 bch2_inconsistent_error(c);
1830                 ret = -EINVAL;
1831         }
1832
1833         goto out;
1834 }
1835
1836 static void bch2_do_invalidates_work(struct work_struct *work)
1837 {
1838         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1839         struct bch_dev *ca;
1840         struct btree_trans trans;
1841         struct btree_iter iter;
1842         struct bkey_s_c k;
1843         unsigned i;
1844         int ret = 0;
1845
1846         bch2_trans_init(&trans, c, 0, 0);
1847
1848         ret = bch2_btree_write_buffer_flush(&trans);
1849         if (ret)
1850                 goto err;
1851
1852         for_each_member_device(ca, c, i) {
1853                 s64 nr_to_invalidate =
1854                         should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
1855
1856                 ret = for_each_btree_key2_upto(&trans, iter, BTREE_ID_lru,
1857                                 lru_pos(ca->dev_idx, 0, 0),
1858                                 lru_pos(ca->dev_idx, U64_MAX, LRU_TIME_MAX),
1859                                 BTREE_ITER_INTENT, k,
1860                         invalidate_one_bucket(&trans, &iter, k, &nr_to_invalidate));
1861
1862                 if (ret < 0) {
1863                         percpu_ref_put(&ca->ref);
1864                         break;
1865                 }
1866         }
1867 err:
1868         bch2_trans_exit(&trans);
1869         bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1870 }
1871
1872 void bch2_do_invalidates(struct bch_fs *c)
1873 {
1874         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_invalidate) &&
1875             !queue_work(c->write_ref_wq, &c->invalidate_work))
1876                 bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1877 }
1878
1879 static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca,
1880                                    unsigned long *last_updated)
1881 {
1882         struct btree_trans trans;
1883         struct btree_iter iter;
1884         struct bkey_s_c k;
1885         struct bkey hole;
1886         struct bpos end = POS(ca->dev_idx, ca->mi.nbuckets);
1887         struct bch_member *m;
1888         int ret;
1889
1890         bch2_trans_init(&trans, c, 0, 0);
1891
1892         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc,
1893                              POS(ca->dev_idx, ca->mi.first_bucket),
1894                              BTREE_ITER_PREFETCH);
1895         /*
1896          * Scan the alloc btree for every bucket on @ca, and add buckets to the
1897          * freespace/need_discard/need_gc_gens btrees as needed:
1898          */
1899         while (1) {
1900                 if (*last_updated + HZ * 10 < jiffies) {
1901                         bch_info(ca, "%s: currently at %llu/%llu",
1902                                  __func__, iter.pos.offset, ca->mi.nbuckets);
1903                         *last_updated = jiffies;
1904                 }
1905
1906                 bch2_trans_begin(&trans);
1907
1908                 if (bkey_ge(iter.pos, end)) {
1909                         ret = 0;
1910                         break;
1911                 }
1912
1913                 k = bch2_get_key_or_hole(&iter, end, &hole);
1914                 ret = bkey_err(k);
1915                 if (ret)
1916                         goto bkey_err;
1917
1918                 if (k.k->type) {
1919                         /*
1920                          * We process live keys in the alloc btree one at a
1921                          * time:
1922                          */
1923                         struct bch_alloc_v4 a_convert;
1924                         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1925
1926                         ret =   bch2_bucket_do_index(&trans, k, a, true) ?:
1927                                 bch2_trans_commit(&trans, NULL, NULL,
1928                                                   BTREE_INSERT_LAZY_RW|
1929                                                   BTREE_INSERT_NOFAIL);
1930                         if (ret)
1931                                 goto bkey_err;
1932
1933                         bch2_btree_iter_advance(&iter);
1934                 } else {
1935                         struct bkey_i *freespace;
1936
1937                         freespace = bch2_trans_kmalloc(&trans, sizeof(*freespace));
1938                         ret = PTR_ERR_OR_ZERO(freespace);
1939                         if (ret)
1940                                 goto bkey_err;
1941
1942                         bkey_init(&freespace->k);
1943                         freespace->k.type       = KEY_TYPE_set;
1944                         freespace->k.p          = k.k->p;
1945                         freespace->k.size       = k.k->size;
1946
1947                         ret = __bch2_btree_insert(&trans, BTREE_ID_freespace, freespace, 0) ?:
1948                                 bch2_trans_commit(&trans, NULL, NULL,
1949                                                   BTREE_INSERT_LAZY_RW|
1950                                                   BTREE_INSERT_NOFAIL);
1951                         if (ret)
1952                                 goto bkey_err;
1953
1954                         bch2_btree_iter_set_pos(&iter, k.k->p);
1955                 }
1956 bkey_err:
1957                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1958                         continue;
1959                 if (ret)
1960                         break;
1961         }
1962
1963         bch2_trans_iter_exit(&trans, &iter);
1964         bch2_trans_exit(&trans);
1965
1966         if (ret < 0) {
1967                 bch_err(ca, "error initializing free space: %s", bch2_err_str(ret));
1968                 return ret;
1969         }
1970
1971         mutex_lock(&c->sb_lock);
1972         m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx;
1973         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1974         mutex_unlock(&c->sb_lock);
1975
1976         return 0;
1977 }
1978
1979 int bch2_fs_freespace_init(struct bch_fs *c)
1980 {
1981         struct bch_dev *ca;
1982         unsigned i;
1983         int ret = 0;
1984         bool doing_init = false;
1985         unsigned long last_updated = jiffies;
1986
1987         /*
1988          * We can crash during the device add path, so we need to check this on
1989          * every mount:
1990          */
1991
1992         for_each_member_device(ca, c, i) {
1993                 if (ca->mi.freespace_initialized)
1994                         continue;
1995
1996                 if (!doing_init) {
1997                         bch_info(c, "initializing freespace");
1998                         doing_init = true;
1999                 }
2000
2001                 ret = bch2_dev_freespace_init(c, ca, &last_updated);
2002                 if (ret) {
2003                         percpu_ref_put(&ca->ref);
2004                         return ret;
2005                 }
2006         }
2007
2008         if (doing_init) {
2009                 mutex_lock(&c->sb_lock);
2010                 bch2_write_super(c);
2011                 mutex_unlock(&c->sb_lock);
2012
2013                 bch_verbose(c, "done initializing freespace");
2014         }
2015
2016         return ret;
2017 }
2018
2019 /* Bucket IO clocks: */
2020
2021 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
2022                               size_t bucket_nr, int rw)
2023 {
2024         struct bch_fs *c = trans->c;
2025         struct btree_iter iter;
2026         struct bkey_i_alloc_v4 *a;
2027         u64 now;
2028         int ret = 0;
2029
2030         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
2031         ret = PTR_ERR_OR_ZERO(a);
2032         if (ret)
2033                 return ret;
2034
2035         now = atomic64_read(&c->io_clock[rw].now);
2036         if (a->v.io_time[rw] == now)
2037                 goto out;
2038
2039         a->v.io_time[rw] = now;
2040
2041         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
2042                 bch2_trans_commit(trans, NULL, NULL, 0);
2043 out:
2044         bch2_trans_iter_exit(trans, &iter);
2045         return ret;
2046 }
2047
2048 /* Startup/shutdown (ro/rw): */
2049
2050 void bch2_recalc_capacity(struct bch_fs *c)
2051 {
2052         struct bch_dev *ca;
2053         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
2054         unsigned bucket_size_max = 0;
2055         unsigned long ra_pages = 0;
2056         unsigned i;
2057
2058         lockdep_assert_held(&c->state_lock);
2059
2060         for_each_online_member(ca, c, i) {
2061                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
2062
2063                 ra_pages += bdi->ra_pages;
2064         }
2065
2066         bch2_set_ra_pages(c, ra_pages);
2067
2068         for_each_rw_member(ca, c, i) {
2069                 u64 dev_reserve = 0;
2070
2071                 /*
2072                  * We need to reserve buckets (from the number
2073                  * of currently available buckets) against
2074                  * foreground writes so that mainly copygc can
2075                  * make forward progress.
2076                  *
2077                  * We need enough to refill the various reserves
2078                  * from scratch - copygc will use its entire
2079                  * reserve all at once, then run against when
2080                  * its reserve is refilled (from the formerly
2081                  * available buckets).
2082                  *
2083                  * This reserve is just used when considering if
2084                  * allocations for foreground writes must wait -
2085                  * not -ENOSPC calculations.
2086                  */
2087
2088                 dev_reserve += ca->nr_btree_reserve * 2;
2089                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
2090
2091                 dev_reserve += 1;       /* btree write point */
2092                 dev_reserve += 1;       /* copygc write point */
2093                 dev_reserve += 1;       /* rebalance write point */
2094
2095                 dev_reserve *= ca->mi.bucket_size;
2096
2097                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
2098                                              ca->mi.first_bucket);
2099
2100                 reserved_sectors += dev_reserve * 2;
2101
2102                 bucket_size_max = max_t(unsigned, bucket_size_max,
2103                                         ca->mi.bucket_size);
2104         }
2105
2106         gc_reserve = c->opts.gc_reserve_bytes
2107                 ? c->opts.gc_reserve_bytes >> 9
2108                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
2109
2110         reserved_sectors = max(gc_reserve, reserved_sectors);
2111
2112         reserved_sectors = min(reserved_sectors, capacity);
2113
2114         c->capacity = capacity - reserved_sectors;
2115
2116         c->bucket_size_max = bucket_size_max;
2117
2118         /* Wake up case someone was waiting for buckets */
2119         closure_wake_up(&c->freelist_wait);
2120 }
2121
2122 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
2123 {
2124         struct open_bucket *ob;
2125         bool ret = false;
2126
2127         for (ob = c->open_buckets;
2128              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
2129              ob++) {
2130                 spin_lock(&ob->lock);
2131                 if (ob->valid && !ob->on_partial_list &&
2132                     ob->dev == ca->dev_idx)
2133                         ret = true;
2134                 spin_unlock(&ob->lock);
2135         }
2136
2137         return ret;
2138 }
2139
2140 /* device goes ro: */
2141 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
2142 {
2143         unsigned i;
2144
2145         /* First, remove device from allocation groups: */
2146
2147         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2148                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
2149
2150         /*
2151          * Capacity is calculated based off of devices in allocation groups:
2152          */
2153         bch2_recalc_capacity(c);
2154
2155         bch2_open_buckets_stop(c, ca, false);
2156
2157         /*
2158          * Wake up threads that were blocked on allocation, so they can notice
2159          * the device can no longer be removed and the capacity has changed:
2160          */
2161         closure_wake_up(&c->freelist_wait);
2162
2163         /*
2164          * journal_res_get() can block waiting for free space in the journal -
2165          * it needs to notice there may not be devices to allocate from anymore:
2166          */
2167         wake_up(&c->journal.wait);
2168
2169         /* Now wait for any in flight writes: */
2170
2171         closure_wait_event(&c->open_buckets_wait,
2172                            !bch2_dev_has_open_write_point(c, ca));
2173 }
2174
2175 /* device goes rw: */
2176 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
2177 {
2178         unsigned i;
2179
2180         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2181                 if (ca->mi.data_allowed & (1 << i))
2182                         set_bit(ca->dev_idx, c->rw_devs[i].d);
2183 }
2184
2185 void bch2_fs_allocator_background_init(struct bch_fs *c)
2186 {
2187         spin_lock_init(&c->freelist_lock);
2188         INIT_WORK(&c->discard_work, bch2_do_discards_work);
2189         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
2190 }