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 "buckets.h"
  13 #include "buckets_waiting_for_journal.h"
  14 #include "clock.h"
  15 #include "debug.h"
  16 #include "ec.h"
  17 #include "error.h"
  18 #include "lru.h"
  19 #include "recovery.h"
  20 #include "varint.h"
  21
  22 #include <linux/kthread.h>
  23 #include <linux/math64.h>
  24 #include <linux/random.h>
  25 #include <linux/rculist.h>
  26 #include <linux/rcupdate.h>
  27 #include <linux/sched/task.h>
  28 #include <linux/sort.h>
  29 #include <trace/events/bcachefs.h>
  30
  31 /* Persistent alloc info: */
  32
  33 static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
  34 #define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8,
  35         BCH_ALLOC_FIELDS_V1()
  36 #undef x
  37 };
  38
  39 struct bkey_alloc_unpacked {
  40         u64             journal_seq;
  41         u8              gen;
  42         u8              oldest_gen;
  43         u8              data_type;
  44         bool            need_discard:1;
  45         bool            need_inc_gen:1;
  46 #define x(_name, _bits) u##_bits _name;
  47         BCH_ALLOC_FIELDS_V2()
  48 #undef  x
  49 };
  50
  51 static inline u64 alloc_field_v1_get(const struct bch_alloc *a,
  52                                      const void **p, unsigned field)
  53 {
  54         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  55         u64 v;
  56
  57         if (!(a->fields & (1 << field)))
  58                 return 0;
  59
  60         switch (bytes) {
  61         case 1:
  62                 v = *((const u8 *) *p);
  63                 break;
  64         case 2:
  65                 v = le16_to_cpup(*p);
  66                 break;
  67         case 4:
  68                 v = le32_to_cpup(*p);
  69                 break;
  70         case 8:
  71                 v = le64_to_cpup(*p);
  72                 break;
  73         default:
  74                 BUG();
  75         }
  76
  77         *p += bytes;
  78         return v;
  79 }
  80
  81 static inline void alloc_field_v1_put(struct bkey_i_alloc *a, void **p,
  82                                       unsigned field, u64 v)
  83 {
  84         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  85
  86         if (!v)
  87                 return;
  88
  89         a->v.fields |= 1 << field;
  90
  91         switch (bytes) {
  92         case 1:
  93                 *((u8 *) *p) = v;
  94                 break;
  95         case 2:
  96                 *((__le16 *) *p) = cpu_to_le16(v);
  97                 break;
  98         case 4:
  99                 *((__le32 *) *p) = cpu_to_le32(v);
 100                 break;
 101         case 8:
 102                 *((__le64 *) *p) = cpu_to_le64(v);
 103                 break;
 104         default:
 105                 BUG();
 106         }
 107
 108         *p += bytes;
 109 }
 110
 111 static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out,
 112                                  struct bkey_s_c k)
 113 {
 114         const struct bch_alloc *in = bkey_s_c_to_alloc(k).v;
 115         const void *d = in->data;
 116         unsigned idx = 0;
 117
 118         out->gen = in->gen;
 119
 120 #define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++);
 121         BCH_ALLOC_FIELDS_V1()
 122 #undef  x
 123 }
 124
 125 static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out,
 126                                 struct bkey_s_c k)
 127 {
 128         struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k);
 129         const u8 *in = a.v->data;
 130         const u8 *end = bkey_val_end(a);
 131         unsigned fieldnr = 0;
 132         int ret;
 133         u64 v;
 134
 135         out->gen        = a.v->gen;
 136         out->oldest_gen = a.v->oldest_gen;
 137         out->data_type  = a.v->data_type;
 138
 139 #define x(_name, _bits)                                                 \
 140         if (fieldnr < a.v->nr_fields) {                                 \
 141                 ret = bch2_varint_decode_fast(in, end, &v);             \
 142                 if (ret < 0)                                            \
 143                         return ret;                                     \
 144                 in += ret;                                              \
 145         } else {                                                        \
 146                 v = 0;                                                  \
 147         }                                                               \
 148         out->_name = v;                                                 \
 149         if (v != out->_name)                                            \
 150                 return -1;                                              \
 151         fieldnr++;
 152
 153         BCH_ALLOC_FIELDS_V2()
 154 #undef  x
 155         return 0;
 156 }
 157
 158 static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out,
 159                                 struct bkey_s_c k)
 160 {
 161         struct bkey_s_c_alloc_v3 a = bkey_s_c_to_alloc_v3(k);
 162         const u8 *in = a.v->data;
 163         const u8 *end = bkey_val_end(a);
 164         unsigned fieldnr = 0;
 165         int ret;
 166         u64 v;
 167
 168         out->gen        = a.v->gen;
 169         out->oldest_gen = a.v->oldest_gen;
 170         out->data_type  = a.v->data_type;
 171         out->need_discard = BCH_ALLOC_V3_NEED_DISCARD(a.v);
 172         out->need_inc_gen = BCH_ALLOC_V3_NEED_INC_GEN(a.v);
 173         out->journal_seq = le64_to_cpu(a.v->journal_seq);
 174
 175 #define x(_name, _bits)                                                 \
 176         if (fieldnr < a.v->nr_fields) {                                 \
 177                 ret = bch2_varint_decode_fast(in, end, &v);             \
 178                 if (ret < 0)                                            \
 179                         return ret;                                     \
 180                 in += ret;                                              \
 181         } else {                                                        \
 182                 v = 0;                                                  \
 183         }                                                               \
 184         out->_name = v;                                                 \
 185         if (v != out->_name)                                            \
 186                 return -1;                                              \
 187         fieldnr++;
 188
 189         BCH_ALLOC_FIELDS_V2()
 190 #undef  x
 191         return 0;
 192 }
 193
 194 static struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 195 {
 196         struct bkey_alloc_unpacked ret = { .gen = 0 };
 197
 198         switch (k.k->type) {
 199         case KEY_TYPE_alloc:
 200                 bch2_alloc_unpack_v1(&ret, k);
 201                 break;
 202         case KEY_TYPE_alloc_v2:
 203                 bch2_alloc_unpack_v2(&ret, k);
 204                 break;
 205         case KEY_TYPE_alloc_v3:
 206                 bch2_alloc_unpack_v3(&ret, k);
 207                 break;
 208         }
 209
 210         return ret;
 211 }
 212
 213 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
 214 {
 215         unsigned i, bytes = offsetof(struct bch_alloc, data);
 216
 217         for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++)
 218                 if (a->fields & (1 << i))
 219                         bytes += BCH_ALLOC_V1_FIELD_BYTES[i];
 220
 221         return DIV_ROUND_UP(bytes, sizeof(u64));
 222 }
 223
 224 int bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k,
 225                           int rw, struct printbuf *err)
 226 {
 227         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 228
 229         /* allow for unknown fields */
 230         if (bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v)) {
 231                 prt_printf(err, "incorrect value size (%zu < %u)",
 232                        bkey_val_u64s(a.k), bch_alloc_v1_val_u64s(a.v));
 233                 return -BCH_ERR_invalid_bkey;
 234         }
 235
 236         return 0;
 237 }
 238
 239 int bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k,
 240                           int rw, struct printbuf *err)
 241 {
 242         struct bkey_alloc_unpacked u;
 243
 244         if (bch2_alloc_unpack_v2(&u, k)) {
 245                 prt_printf(err, "unpack error");
 246                 return -BCH_ERR_invalid_bkey;
 247         }
 248
 249         return 0;
 250 }
 251
 252 int bch2_alloc_v3_invalid(const struct bch_fs *c, struct bkey_s_c k,
 253                           int rw, struct printbuf *err)
 254 {
 255         struct bkey_alloc_unpacked u;
 256
 257         if (bch2_alloc_unpack_v3(&u, k)) {
 258                 prt_printf(err, "unpack error");
 259                 return -BCH_ERR_invalid_bkey;
 260         }
 261
 262         return 0;
 263 }
 264
 265 int bch2_alloc_v4_invalid(const struct bch_fs *c, struct bkey_s_c k,
 266                           int rw, struct printbuf *err)
 267 {
 268         struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 269
 270         if (alloc_v4_u64s(a.v) != bkey_val_u64s(k.k)) {
 271                 prt_printf(err, "bad val size (%lu != %u)",
 272                        bkey_val_u64s(k.k), alloc_v4_u64s(a.v));
 273                 return -BCH_ERR_invalid_bkey;
 274         }
 275
 276         if (!BCH_ALLOC_V4_BACKPOINTERS_START(a.v) &&
 277             BCH_ALLOC_V4_NR_BACKPOINTERS(a.v)) {
 278                 prt_printf(err, "invalid backpointers_start");
 279                 return -BCH_ERR_invalid_bkey;
 280         }
 281
 282         /*
 283          * XXX this is wrong, we'll be checking updates that happened from
 284          * before BCH_FS_CHECK_BACKPOINTERS_DONE
 285          */
 286         if (rw == WRITE && 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
 419         if (k.k->type == KEY_TYPE_alloc_v4) {
 420                 struct bkey_s_c_alloc_v4 a_raw = bkey_s_c_to_alloc_v4(k);
 421                 const struct bch_backpointer *bps = alloc_v4_backpointers_c(a_raw.v);
 422
 423                 prt_printf(out, "bp_start          %llu", BCH_ALLOC_V4_BACKPOINTERS_START(a_raw.v));
 424                 prt_newline(out);
 425
 426                 prt_printf(out, "backpointers:     %llu", BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v));
 427                 printbuf_indent_add(out, 2);
 428
 429                 for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v); i++) {
 430                         prt_newline(out);
 431                         bch2_backpointer_to_text(out, &bps[i]);
 432                 }
 433
 434                 printbuf_indent_sub(out, 2);
 435         }
 436
 437         printbuf_indent_sub(out, 2);
 438 }
 439
 440 void __bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out)
 441 {
 442         if (k.k->type == KEY_TYPE_alloc_v4) {
 443                 void *src, *dst;
 444
 445                 *out = *bkey_s_c_to_alloc_v4(k).v;
 446
 447                 src = alloc_v4_backpointers(out);
 448                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 449                 dst = alloc_v4_backpointers(out);
 450
 451                 if (src < dst)
 452                         memset(src, 0, dst - src);
 453         } else {
 454                 struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 455
 456                 *out = (struct bch_alloc_v4) {
 457                         .journal_seq            = u.journal_seq,
 458                         .flags                  = u.need_discard,
 459                         .gen                    = u.gen,
 460                         .oldest_gen             = u.oldest_gen,
 461                         .data_type              = u.data_type,
 462                         .stripe_redundancy      = u.stripe_redundancy,
 463                         .dirty_sectors          = u.dirty_sectors,
 464                         .cached_sectors         = u.cached_sectors,
 465                         .io_time[READ]          = u.read_time,
 466                         .io_time[WRITE]         = u.write_time,
 467                         .stripe                 = u.stripe,
 468                 };
 469
 470                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 471         }
 472 }
 473
 474 static noinline struct bkey_i_alloc_v4 *
 475 __bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 476 {
 477         struct bkey_i_alloc_v4 *ret;
 478         if (k.k->type == KEY_TYPE_alloc_v4) {
 479                 struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 480                 unsigned bytes = sizeof(struct bkey_i_alloc_v4) +
 481                         BCH_ALLOC_V4_NR_BACKPOINTERS(a.v) *
 482                         sizeof(struct bch_backpointer);
 483                 void *src, *dst;
 484
 485                 /*
 486                  * Reserve space for one more backpointer here:
 487                  * Not sketchy at doing it this way, nope...
 488                  */
 489                 ret = bch2_trans_kmalloc(trans, bytes + sizeof(struct bch_backpointer));
 490                 if (IS_ERR(ret))
 491                         return ret;
 492
 493                 bkey_reassemble(&ret->k_i, k);
 494
 495                 src = alloc_v4_backpointers(&ret->v);
 496                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(&ret->v, BCH_ALLOC_V4_U64s);
 497                 dst = alloc_v4_backpointers(&ret->v);
 498
 499                 memmove(dst, src, BCH_ALLOC_V4_NR_BACKPOINTERS(&ret->v) *
 500                         sizeof(struct bch_backpointer));
 501                 if (src < dst)
 502                         memset(src, 0, dst - src);
 503                 set_alloc_v4_u64s(ret);
 504         } else {
 505                 ret = bch2_trans_kmalloc(trans, sizeof(struct bkey_i_alloc_v4) +
 506                                          sizeof(struct bch_backpointer));
 507                 if (IS_ERR(ret))
 508                         return ret;
 509
 510                 bkey_alloc_v4_init(&ret->k_i);
 511                 ret->k.p = k.k->p;
 512                 bch2_alloc_to_v4(k, &ret->v);
 513         }
 514         return ret;
 515 }
 516
 517 static inline struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut_inlined(struct btree_trans *trans, struct bkey_s_c k)
 518 {
 519         if (likely(k.k->type == KEY_TYPE_alloc_v4) &&
 520             BCH_ALLOC_V4_BACKPOINTERS_START(bkey_s_c_to_alloc_v4(k).v) == BCH_ALLOC_V4_U64s) {
 521                 /*
 522                  * Reserve space for one more backpointer here:
 523                  * Not sketchy at doing it this way, nope...
 524                  */
 525                 struct bkey_i_alloc_v4 *ret =
 526                         bch2_trans_kmalloc_nomemzero(trans, bkey_bytes(k.k) + sizeof(struct bch_backpointer));
 527                 if (!IS_ERR(ret))
 528                         bkey_reassemble(&ret->k_i, k);
 529                 return ret;
 530         }
 531
 532         return __bch2_alloc_to_v4_mut(trans, k);
 533 }
 534
 535 struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 536 {
 537         return bch2_alloc_to_v4_mut_inlined(trans, k);
 538 }
 539
 540 struct bkey_i_alloc_v4 *
 541 bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
 542                               struct bpos pos)
 543 {
 544         struct bkey_s_c k;
 545         struct bkey_i_alloc_v4 *a;
 546         int ret;
 547
 548         bch2_trans_iter_init(trans, iter, BTREE_ID_alloc, pos,
 549                              BTREE_ITER_WITH_UPDATES|
 550                              BTREE_ITER_CACHED|
 551                              BTREE_ITER_INTENT);
 552         k = bch2_btree_iter_peek_slot(iter);
 553         ret = bkey_err(k);
 554         if (unlikely(ret))
 555                 goto err;
 556
 557         a = bch2_alloc_to_v4_mut_inlined(trans, k);
 558         ret = PTR_ERR_OR_ZERO(a);
 559         if (unlikely(ret))
 560                 goto err;
 561         return a;
 562 err:
 563         bch2_trans_iter_exit(trans, iter);
 564         return ERR_PTR(ret);
 565 }
 566
 567 int bch2_alloc_read(struct bch_fs *c)
 568 {
 569         struct btree_trans trans;
 570         struct btree_iter iter;
 571         struct bkey_s_c k;
 572         struct bch_alloc_v4 a;
 573         struct bch_dev *ca;
 574         int ret;
 575
 576         bch2_trans_init(&trans, c, 0, 0);
 577
 578         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 579                            BTREE_ITER_PREFETCH, k, ret) {
 580                 /*
 581                  * Not a fsck error because this is checked/repaired by
 582                  * bch2_check_alloc_key() which runs later:
 583                  */
 584                 if (!bch2_dev_bucket_exists(c, k.k->p))
 585                         continue;
 586
 587                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 588
 589                 *bucket_gen(ca, k.k->p.offset) = bch2_alloc_to_v4(k, &a)->gen;
 590         }
 591         bch2_trans_iter_exit(&trans, &iter);
 592
 593         bch2_trans_exit(&trans);
 594
 595         if (ret)
 596                 bch_err(c, "error reading alloc info: %s", bch2_err_str(ret));
 597
 598         return ret;
 599 }
 600
 601 static struct bpos alloc_gens_pos(struct bpos pos, unsigned *offset)
 602 {
 603         *offset = pos.offset & KEY_TYPE_BUCKET_GENS_MASK;
 604
 605         pos.offset >>= KEY_TYPE_BUCKET_GENS_BITS;
 606         return pos;
 607 }
 608
 609 static struct bpos bucket_gens_pos_to_alloc(struct bpos pos, unsigned offset)
 610 {
 611         pos.offset <<= KEY_TYPE_BUCKET_GENS_BITS;
 612         pos.offset += offset;
 613         return pos;
 614 }
 615
 616 static unsigned alloc_gen(struct bkey_s_c k, unsigned offset)
 617 {
 618         return k.k->type == KEY_TYPE_bucket_gens
 619                 ? bkey_s_c_to_bucket_gens(k).v->gens[offset]
 620                 : 0;
 621 }
 622
 623 int bch2_bucket_gens_invalid(const struct bch_fs *c, struct bkey_s_c k,
 624                              int rw, struct printbuf *err)
 625 {
 626         if (bkey_val_bytes(k.k) != sizeof(struct bch_bucket_gens)) {
 627                 prt_printf(err, "bad val size (%lu != %zu)",
 628                        bkey_val_bytes(k.k), sizeof(struct bch_bucket_gens));
 629                 return -BCH_ERR_invalid_bkey;
 630         }
 631
 632         return 0;
 633 }
 634
 635 void bch2_bucket_gens_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 636 {
 637         struct bkey_s_c_bucket_gens g = bkey_s_c_to_bucket_gens(k);
 638         unsigned i;
 639
 640         for (i = 0; i < ARRAY_SIZE(g.v->gens); i++) {
 641                 if (i)
 642                         prt_char(out, ' ');
 643                 prt_printf(out, "%u", g.v->gens[i]);
 644         }
 645 }
 646
 647 int bch2_bucket_gens_init(struct bch_fs *c)
 648 {
 649         struct btree_trans trans;
 650         struct btree_iter iter;
 651         struct bkey_s_c k;
 652         struct bch_alloc_v4 a;
 653         struct bkey_i_bucket_gens g;
 654         bool have_bucket_gens_key = false;
 655         unsigned offset;
 656         struct bpos pos;
 657         u8 gen;
 658         int ret;
 659
 660         bch2_trans_init(&trans, c, 0, 0);
 661
 662         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 663                            BTREE_ITER_PREFETCH, k, ret) {
 664                 /*
 665                  * Not a fsck error because this is checked/repaired by
 666                  * bch2_check_alloc_key() which runs later:
 667                  */
 668                 if (!bch2_dev_bucket_exists(c, k.k->p))
 669                         continue;
 670
 671                 gen = bch2_alloc_to_v4(k, &a)->gen;
 672                 pos = alloc_gens_pos(iter.pos, &offset);
 673
 674                 if (have_bucket_gens_key && bkey_cmp(iter.pos, pos)) {
 675                         ret = commit_do(&trans, NULL, NULL,
 676                                         BTREE_INSERT_NOFAIL|
 677                                         BTREE_INSERT_LAZY_RW,
 678                                 __bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i));
 679                         if (ret)
 680                                 break;
 681                         have_bucket_gens_key = false;
 682                 }
 683
 684                 if (!have_bucket_gens_key) {
 685                         bkey_bucket_gens_init(&g.k_i);
 686                         g.k.p = pos;
 687                         have_bucket_gens_key = true;
 688                 }
 689
 690                 g.v.gens[offset] = gen;
 691         }
 692         bch2_trans_iter_exit(&trans, &iter);
 693
 694         if (have_bucket_gens_key && !ret)
 695                 ret = commit_do(&trans, NULL, NULL,
 696                                 BTREE_INSERT_NOFAIL|
 697                                 BTREE_INSERT_LAZY_RW,
 698                         __bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i));
 699
 700         bch2_trans_exit(&trans);
 701
 702         if (ret)
 703                 bch_err(c, "%s: error %s", __func__, bch2_err_str(ret));
 704
 705         return ret;
 706 }
 707
 708 int bch2_bucket_gens_read(struct bch_fs *c)
 709 {
 710         struct btree_trans trans;
 711         struct btree_iter iter;
 712         struct bkey_s_c k;
 713         const struct bch_bucket_gens *g;
 714         struct bch_dev *ca;
 715         u64 b;
 716         int ret;
 717
 718         bch2_trans_init(&trans, c, 0, 0);
 719
 720         for_each_btree_key(&trans, iter, BTREE_ID_bucket_gens, POS_MIN,
 721                            BTREE_ITER_PREFETCH, k, ret) {
 722                 u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
 723                 u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
 724
 725                 if (k.k->type != KEY_TYPE_bucket_gens)
 726                         continue;
 727
 728                 g = bkey_s_c_to_bucket_gens(k).v;
 729
 730                 /*
 731                  * Not a fsck error because this is checked/repaired by
 732                  * bch2_check_alloc_key() which runs later:
 733                  */
 734                 if (!bch2_dev_exists2(c, k.k->p.inode))
 735                         continue;
 736
 737                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 738
 739                 for (b = max_t(u64, ca->mi.first_bucket, start);
 740                      b < min_t(u64, ca->mi.nbuckets, end);
 741                      b++)
 742                         *bucket_gen(ca, b) = g->gens[b & KEY_TYPE_BUCKET_GENS_MASK];
 743         }
 744         bch2_trans_iter_exit(&trans, &iter);
 745
 746         bch2_trans_exit(&trans);
 747
 748         if (ret)
 749                 bch_err(c, "error reading alloc info: %s", bch2_err_str(ret));
 750
 751         return ret;
 752 }
 753
 754 /* Free space/discard btree: */
 755
 756 static int bch2_bucket_do_index(struct btree_trans *trans,
 757                                 struct bkey_s_c alloc_k,
 758                                 const struct bch_alloc_v4 *a,
 759                                 bool set)
 760 {
 761         struct bch_fs *c = trans->c;
 762         struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 763         struct btree_iter iter;
 764         struct bkey_s_c old;
 765         struct bkey_i *k;
 766         enum btree_id btree;
 767         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 768         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 769         struct printbuf buf = PRINTBUF;
 770         int ret;
 771
 772         if (a->data_type != BCH_DATA_free &&
 773             a->data_type != BCH_DATA_need_discard)
 774                 return 0;
 775
 776         k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
 777         if (IS_ERR(k))
 778                 return PTR_ERR(k);
 779
 780         bkey_init(&k->k);
 781         k->k.type = new_type;
 782
 783         switch (a->data_type) {
 784         case BCH_DATA_free:
 785                 btree = BTREE_ID_freespace;
 786                 k->k.p = alloc_freespace_pos(alloc_k.k->p, *a);
 787                 bch2_key_resize(&k->k, 1);
 788                 break;
 789         case BCH_DATA_need_discard:
 790                 btree = BTREE_ID_need_discard;
 791                 k->k.p = alloc_k.k->p;
 792                 break;
 793         default:
 794                 return 0;
 795         }
 796
 797         bch2_trans_iter_init(trans, &iter, btree,
 798                              bkey_start_pos(&k->k),
 799                              BTREE_ITER_INTENT);
 800         old = bch2_btree_iter_peek_slot(&iter);
 801         ret = bkey_err(old);
 802         if (ret)
 803                 goto err;
 804
 805         if (ca->mi.freespace_initialized &&
 806             test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags) &&
 807             bch2_trans_inconsistent_on(old.k->type != old_type, trans,
 808                         "incorrect key when %s %s btree (got %s should be %s)\n"
 809                         "  for %s",
 810                         set ? "setting" : "clearing",
 811                         bch2_btree_ids[btree],
 812                         bch2_bkey_types[old.k->type],
 813                         bch2_bkey_types[old_type],
 814                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 815                 ret = -EIO;
 816                 goto err;
 817         }
 818
 819         ret = bch2_trans_update(trans, &iter, k, 0);
 820 err:
 821         bch2_trans_iter_exit(trans, &iter);
 822         printbuf_exit(&buf);
 823         return ret;
 824 }
 825
 826 static noinline int bch2_bucket_gen_update(struct btree_trans *trans,
 827                                            struct bpos bucket, u8 gen)
 828 {
 829         struct btree_iter iter;
 830         unsigned offset;
 831         struct bpos pos = alloc_gens_pos(bucket, &offset);
 832         struct bkey_i_bucket_gens *g;
 833         struct bkey_s_c k;
 834         int ret;
 835
 836         g = bch2_trans_kmalloc(trans, sizeof(*g));
 837         ret = PTR_ERR_OR_ZERO(g);
 838         if (ret)
 839                 return ret;
 840
 841         bch2_trans_iter_init(trans, &iter, BTREE_ID_bucket_gens, pos,
 842                              BTREE_ITER_INTENT|
 843                              BTREE_ITER_WITH_UPDATES);
 844         k = bch2_btree_iter_peek_slot(&iter);
 845         ret = bkey_err(k);
 846         if (ret)
 847                 goto err;
 848
 849         if (k.k->type != KEY_TYPE_bucket_gens) {
 850                 bkey_bucket_gens_init(&g->k_i);
 851                 g->k.p = iter.pos;
 852         } else {
 853                 bkey_reassemble(&g->k_i, k);
 854         }
 855
 856         g->v.gens[offset] = gen;
 857
 858         ret = bch2_trans_update(trans, &iter, &g->k_i, 0);
 859 err:
 860         bch2_trans_iter_exit(trans, &iter);
 861         return ret;
 862 }
 863
 864 int bch2_trans_mark_alloc(struct btree_trans *trans,
 865                           enum btree_id btree_id, unsigned level,
 866                           struct bkey_s_c old, struct bkey_i *new,
 867                           unsigned flags)
 868 {
 869         struct bch_fs *c = trans->c;
 870         struct bch_alloc_v4 old_a_convert, *new_a;
 871         const struct bch_alloc_v4 *old_a;
 872         u64 old_lru, new_lru;
 873         int ret = 0;
 874
 875         /*
 876          * Deletion only happens in the device removal path, with
 877          * BTREE_TRIGGER_NORUN:
 878          */
 879         BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
 880
 881         old_a = bch2_alloc_to_v4(old, &old_a_convert);
 882         new_a = &bkey_i_to_alloc_v4(new)->v;
 883
 884         new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
 885
 886         if (new_a->dirty_sectors > old_a->dirty_sectors ||
 887             new_a->cached_sectors > old_a->cached_sectors) {
 888                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 889                 new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 890                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
 891                 SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
 892         }
 893
 894         if (data_type_is_empty(new_a->data_type) &&
 895             BCH_ALLOC_V4_NEED_INC_GEN(new_a) &&
 896             !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
 897                 new_a->gen++;
 898                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
 899         }
 900
 901         if (old_a->data_type != new_a->data_type ||
 902             (new_a->data_type == BCH_DATA_free &&
 903              alloc_freespace_genbits(*old_a) != alloc_freespace_genbits(*new_a))) {
 904                 ret =   bch2_bucket_do_index(trans, old, old_a, false) ?:
 905                         bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_a, true);
 906                 if (ret)
 907                         return ret;
 908         }
 909
 910         if (new_a->data_type == BCH_DATA_cached &&
 911             !new_a->io_time[READ])
 912                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 913
 914         old_lru = alloc_lru_idx(*old_a);
 915         new_lru = alloc_lru_idx(*new_a);
 916
 917         if (old_lru != new_lru) {
 918                 ret = bch2_lru_change(trans, new->k.p.inode,
 919                                       bucket_to_u64(new->k.p),
 920                                       old_lru, new_lru);
 921                 if (ret)
 922                         return ret;
 923
 924                 if (new_a->data_type == BCH_DATA_cached)
 925                         new_a->io_time[READ] = new_lru;
 926         }
 927
 928         if (old_a->gen != new_a->gen) {
 929                 ret = bch2_bucket_gen_update(trans, new->k.p, new_a->gen);
 930                 if (ret)
 931                         return ret;
 932         }
 933
 934         return 0;
 935 }
 936
 937 /*
 938  * This synthesizes deleted extents for holes, similar to BTREE_ITER_SLOTS for
 939  * extents style btrees, but works on non-extents btrees:
 940  */
 941 struct bkey_s_c bch2_get_key_or_hole(struct btree_iter *iter, struct bpos end, struct bkey *hole)
 942 {
 943         struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
 944
 945         if (bkey_err(k))
 946                 return k;
 947
 948         if (k.k->type) {
 949                 return k;
 950         } else {
 951                 struct btree_iter iter2;
 952                 struct bpos next;
 953
 954                 bch2_trans_copy_iter(&iter2, iter);
 955                 k = bch2_btree_iter_peek_upto(&iter2,
 956                                 bkey_min(bkey_min(end,
 957                                                   iter->path->l[0].b->key.k.p),
 958                                                   POS(iter->pos.inode, iter->pos.offset + U32_MAX - 1)));
 959                 next = iter2.pos;
 960                 bch2_trans_iter_exit(iter->trans, &iter2);
 961
 962                 BUG_ON(next.offset >= iter->pos.offset + U32_MAX);
 963
 964                 if (bkey_err(k))
 965                         return k;
 966
 967                 bkey_init(hole);
 968                 hole->p = iter->pos;
 969
 970                 bch2_key_resize(hole, next.offset - iter->pos.offset);
 971                 return (struct bkey_s_c) { hole, NULL };
 972         }
 973 }
 974
 975 static bool next_bucket(struct bch_fs *c, struct bpos *bucket)
 976 {
 977         struct bch_dev *ca;
 978         unsigned iter;
 979
 980         if (bch2_dev_bucket_exists(c, *bucket))
 981                 return true;
 982
 983         if (bch2_dev_exists2(c, bucket->inode)) {
 984                 ca = bch_dev_bkey_exists(c, bucket->inode);
 985
 986                 if (bucket->offset < ca->mi.first_bucket) {
 987                         bucket->offset = ca->mi.first_bucket;
 988                         return true;
 989                 }
 990
 991                 bucket->inode++;
 992                 bucket->offset = 0;
 993         }
 994
 995         rcu_read_lock();
 996         iter = bucket->inode;
 997         ca = __bch2_next_dev(c, &iter, NULL);
 998         if (ca)
 999                 bucket->offset = ca->mi.first_bucket;
1000         rcu_read_unlock();
1001
1002         return ca != NULL;
1003 }
1004
1005 struct bkey_s_c bch2_get_key_or_real_bucket_hole(struct btree_iter *iter, struct bkey *hole)
1006 {
1007         struct bch_fs *c = iter->trans->c;
1008         struct bkey_s_c k;
1009 again:
1010         k = bch2_get_key_or_hole(iter, POS_MAX, hole);
1011         if (bkey_err(k))
1012                 return k;
1013
1014         if (!k.k->type) {
1015                 struct bpos bucket = bkey_start_pos(k.k);
1016
1017                 if (!bch2_dev_bucket_exists(c, bucket)) {
1018                         if (!next_bucket(c, &bucket))
1019                                 return bkey_s_c_null;
1020
1021                         bch2_btree_iter_set_pos(iter, bucket);
1022                         goto again;
1023                 }
1024
1025                 if (!bch2_dev_bucket_exists(c, k.k->p)) {
1026                         struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
1027
1028                         bch2_key_resize(hole, ca->mi.nbuckets - bucket.offset);
1029                 }
1030         }
1031
1032         return k;
1033 }
1034
1035 static int bch2_check_alloc_key(struct btree_trans *trans,
1036                                 struct bkey_s_c alloc_k,
1037                                 struct btree_iter *alloc_iter,
1038                                 struct btree_iter *discard_iter,
1039                                 struct btree_iter *freespace_iter,
1040                                 struct btree_iter *bucket_gens_iter)
1041 {
1042         struct bch_fs *c = trans->c;
1043         struct bch_dev *ca;
1044         struct bch_alloc_v4 a_convert;
1045         const struct bch_alloc_v4 *a;
1046         unsigned discard_key_type, freespace_key_type;
1047         unsigned gens_offset;
1048         struct bkey_s_c k;
1049         struct printbuf buf = PRINTBUF;
1050         int ret;
1051
1052         if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
1053                         "alloc key for invalid device:bucket %llu:%llu",
1054                         alloc_k.k->p.inode, alloc_k.k->p.offset))
1055                 return bch2_btree_delete_at(trans, alloc_iter, 0);
1056
1057         ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
1058         if (!ca->mi.freespace_initialized)
1059                 return 0;
1060
1061         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1062
1063         discard_key_type = a->data_type == BCH_DATA_need_discard ? KEY_TYPE_set : 0;
1064         bch2_btree_iter_set_pos(discard_iter, alloc_k.k->p);
1065         k = bch2_btree_iter_peek_slot(discard_iter);
1066         ret = bkey_err(k);
1067         if (ret)
1068                 goto err;
1069
1070         if (k.k->type != discard_key_type &&
1071             (c->opts.reconstruct_alloc ||
1072              fsck_err(c, "incorrect key in need_discard btree (got %s should be %s)\n"
1073                       "  %s",
1074                       bch2_bkey_types[k.k->type],
1075                       bch2_bkey_types[discard_key_type],
1076                       (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1077                 struct bkey_i *update =
1078                         bch2_trans_kmalloc(trans, sizeof(*update));
1079
1080                 ret = PTR_ERR_OR_ZERO(update);
1081                 if (ret)
1082                         goto err;
1083
1084                 bkey_init(&update->k);
1085                 update->k.type  = discard_key_type;
1086                 update->k.p     = discard_iter->pos;
1087
1088                 ret = bch2_trans_update(trans, discard_iter, update, 0);
1089                 if (ret)
1090                         goto err;
1091         }
1092
1093         freespace_key_type = a->data_type == BCH_DATA_free ? KEY_TYPE_set : 0;
1094         bch2_btree_iter_set_pos(freespace_iter, alloc_freespace_pos(alloc_k.k->p, *a));
1095         k = bch2_btree_iter_peek_slot(freespace_iter);
1096         ret = bkey_err(k);
1097         if (ret)
1098                 goto err;
1099
1100         if (k.k->type != freespace_key_type &&
1101             (c->opts.reconstruct_alloc ||
1102              fsck_err(c, "incorrect key in freespace btree (got %s should be %s)\n"
1103                       "  %s",
1104                       bch2_bkey_types[k.k->type],
1105                       bch2_bkey_types[freespace_key_type],
1106                       (printbuf_reset(&buf),
1107                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1108                 struct bkey_i *update =
1109                         bch2_trans_kmalloc(trans, sizeof(*update));
1110
1111                 ret = PTR_ERR_OR_ZERO(update);
1112                 if (ret)
1113                         goto err;
1114
1115                 bkey_init(&update->k);
1116                 update->k.type  = freespace_key_type;
1117                 update->k.p     = freespace_iter->pos;
1118                 bch2_key_resize(&update->k, 1);
1119
1120                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1121                 if (ret)
1122                         goto err;
1123         }
1124
1125         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(alloc_k.k->p, &gens_offset));
1126         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1127         ret = bkey_err(k);
1128         if (ret)
1129                 goto err;
1130
1131         if (a->gen != alloc_gen(k, gens_offset) &&
1132             (c->opts.reconstruct_alloc ||
1133              fsck_err(c, "incorrect gen in bucket_gens btree (got %u should be %u)\n"
1134                       "  %s",
1135                       alloc_gen(k, gens_offset), a->gen,
1136                       (printbuf_reset(&buf),
1137                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1138                 struct bkey_i_bucket_gens *g =
1139                         bch2_trans_kmalloc(trans, sizeof(*g));
1140
1141                 ret = PTR_ERR_OR_ZERO(g);
1142                 if (ret)
1143                         goto err;
1144
1145                 if (k.k->type == KEY_TYPE_bucket_gens) {
1146                         bkey_reassemble(&g->k_i, k);
1147                 } else {
1148                         bkey_bucket_gens_init(&g->k_i);
1149                         g->k.p = alloc_gens_pos(alloc_k.k->p, &gens_offset);
1150                 }
1151
1152                 g->v.gens[gens_offset] = a->gen;
1153
1154                 ret = bch2_trans_update(trans, bucket_gens_iter, &g->k_i, 0);
1155                 if (ret)
1156                         goto err;
1157         }
1158 err:
1159 fsck_err:
1160         printbuf_exit(&buf);
1161         return ret;
1162 }
1163
1164 static 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 int bch2_check_alloc_hole_bucket_gens(struct btree_trans *trans,
1220                                  struct bpos start,
1221                                  struct bpos *end,
1222                                  struct btree_iter *bucket_gens_iter)
1223 {
1224         struct bch_fs *c = trans->c;
1225         struct bkey_s_c k;
1226         struct printbuf buf = PRINTBUF;
1227         unsigned i, gens_offset, gens_end_offset;
1228         int ret;
1229
1230         if (c->sb.version < bcachefs_metadata_version_bucket_gens &&
1231             !c->opts.version_upgrade)
1232                 return 0;
1233
1234         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(start, &gens_offset));
1235
1236         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1237         ret = bkey_err(k);
1238         if (ret)
1239                 goto err;
1240
1241         if (bkey_cmp(alloc_gens_pos(start, &gens_offset),
1242                      alloc_gens_pos(*end,  &gens_end_offset)))
1243                 gens_end_offset = KEY_TYPE_BUCKET_GENS_NR;
1244
1245         if (k.k->type == KEY_TYPE_bucket_gens) {
1246                 struct bkey_i_bucket_gens g;
1247                 bool need_update = false;
1248
1249                 bkey_reassemble(&g.k_i, k);
1250
1251                 for (i = gens_offset; i < gens_end_offset; i++) {
1252                         if (fsck_err_on(g.v.gens[i], c,
1253                                         "hole in alloc btree at %llu:%llu with nonzero gen in bucket_gens btree (%u)",
1254                                         bucket_gens_pos_to_alloc(k.k->p, i).inode,
1255                                         bucket_gens_pos_to_alloc(k.k->p, i).offset,
1256                                         g.v.gens[i])) {
1257                                 g.v.gens[i] = 0;
1258                                 need_update = true;
1259                         }
1260                 }
1261
1262                 if (need_update) {
1263                         struct bkey_i *k = bch2_trans_kmalloc(trans, sizeof(g));
1264
1265                         ret = PTR_ERR_OR_ZERO(k);
1266                         if (ret)
1267                                 goto err;
1268
1269                         memcpy(k, &g, sizeof(g));
1270
1271                         ret = bch2_trans_update(trans, bucket_gens_iter, k, 0);
1272                         if (ret)
1273                                 goto err;
1274                 }
1275         }
1276
1277         *end = bkey_min(*end, bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0));
1278 err:
1279 fsck_err:
1280         printbuf_exit(&buf);
1281         return ret;
1282 }
1283
1284 static int bch2_check_discard_freespace_key(struct btree_trans *trans,
1285                                             struct btree_iter *iter)
1286 {
1287         struct bch_fs *c = trans->c;
1288         struct btree_iter alloc_iter;
1289         struct bkey_s_c alloc_k;
1290         struct bch_alloc_v4 a_convert;
1291         const struct bch_alloc_v4 *a;
1292         u64 genbits;
1293         struct bpos pos;
1294         enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
1295                 ? BCH_DATA_need_discard
1296                 : BCH_DATA_free;
1297         struct printbuf buf = PRINTBUF;
1298         int ret;
1299
1300         pos = iter->pos;
1301         pos.offset &= ~(~0ULL << 56);
1302         genbits = iter->pos.offset & (~0ULL << 56);
1303
1304         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
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         alloc_k = bch2_btree_iter_peek_slot(&alloc_iter);
1312         ret = bkey_err(alloc_k);
1313         if (ret)
1314                 goto err;
1315
1316         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1317
1318         if (fsck_err_on(a->data_type != state ||
1319                         (state == BCH_DATA_free &&
1320                          genbits != alloc_freespace_genbits(*a)), c,
1321                         "%s\n  incorrectly set in %s index (free %u, genbits %llu should be %llu)",
1322                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
1323                         bch2_btree_ids[iter->btree_id],
1324                         a->data_type == state,
1325                         genbits >> 56, alloc_freespace_genbits(*a) >> 56))
1326                 goto delete;
1327 out:
1328 err:
1329 fsck_err:
1330         bch2_trans_iter_exit(trans, &alloc_iter);
1331         printbuf_exit(&buf);
1332         return ret;
1333 delete:
1334         ret = bch2_btree_delete_extent_at(trans, iter,
1335                         iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0);
1336         goto out;
1337 }
1338
1339 /*
1340  * We've already checked that generation numbers in the bucket_gens btree are
1341  * valid for buckets that exist; this just checks for keys for nonexistent
1342  * buckets.
1343  */
1344 static int bch2_check_bucket_gens_key(struct btree_trans *trans,
1345                                       struct btree_iter *iter,
1346                                       struct bkey_s_c k)
1347 {
1348         struct bch_fs *c = trans->c;
1349         struct bkey_i_bucket_gens g;
1350         struct bch_dev *ca;
1351         u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
1352         u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
1353         u64 b;
1354         bool need_update = false;
1355         struct printbuf buf = PRINTBUF;
1356         int ret = 0;
1357
1358         BUG_ON(k.k->type != KEY_TYPE_bucket_gens);
1359         bkey_reassemble(&g.k_i, k);
1360
1361         if (fsck_err_on(!bch2_dev_exists2(c, k.k->p.inode), c,
1362                         "bucket_gens key for invalid device:\n  %s",
1363                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1364                 ret = bch2_btree_delete_at(trans, iter, 0);
1365                 goto out;
1366         }
1367
1368         ca = bch_dev_bkey_exists(c, k.k->p.inode);
1369         if (fsck_err_on(end <= ca->mi.first_bucket ||
1370                         start >= ca->mi.nbuckets, c,
1371                         "bucket_gens key for invalid buckets:\n  %s",
1372                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1373                 ret = bch2_btree_delete_at(trans, iter, 0);
1374                 goto out;
1375         }
1376
1377         for (b = start; b < ca->mi.first_bucket; b++)
1378                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1379                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1380                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1381                         need_update = true;
1382                 }
1383
1384         for (b = ca->mi.nbuckets; b < end; b++)
1385                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1386                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1387                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1388                         need_update = true;
1389                 }
1390
1391         if (need_update) {
1392                 struct bkey_i *k;
1393
1394                 k = bch2_trans_kmalloc(trans, sizeof(g));
1395                 ret = PTR_ERR_OR_ZERO(k);
1396                 if (ret)
1397                         goto out;
1398
1399                 memcpy(k, &g, sizeof(g));
1400                 ret = bch2_trans_update(trans, iter, k, 0);
1401         }
1402 out:
1403 fsck_err:
1404         printbuf_exit(&buf);
1405         return ret;
1406 }
1407
1408 int bch2_check_alloc_info(struct bch_fs *c)
1409 {
1410         struct btree_trans trans;
1411         struct btree_iter iter, discard_iter, freespace_iter, bucket_gens_iter;
1412         struct bkey hole;
1413         struct bkey_s_c k;
1414         int ret = 0;
1415
1416         bch2_trans_init(&trans, c, 0, 0);
1417
1418         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc, POS_MIN,
1419                              BTREE_ITER_PREFETCH);
1420         bch2_trans_iter_init(&trans, &discard_iter, BTREE_ID_need_discard, POS_MIN,
1421                              BTREE_ITER_PREFETCH);
1422         bch2_trans_iter_init(&trans, &freespace_iter, BTREE_ID_freespace, POS_MIN,
1423                              BTREE_ITER_PREFETCH);
1424         bch2_trans_iter_init(&trans, &bucket_gens_iter, BTREE_ID_bucket_gens, POS_MIN,
1425                              BTREE_ITER_PREFETCH);
1426
1427         while (1) {
1428                 struct bpos next;
1429
1430                 bch2_trans_begin(&trans);
1431
1432                 k = bch2_get_key_or_real_bucket_hole(&iter, &hole);
1433                 ret = bkey_err(k);
1434                 if (ret)
1435                         goto bkey_err;
1436
1437                 if (!k.k)
1438                         break;
1439
1440                 if (k.k->type) {
1441                         next = bpos_nosnap_successor(k.k->p);
1442
1443                         ret = bch2_check_alloc_key(&trans,
1444                                                    k, &iter,
1445                                                    &discard_iter,
1446                                                    &freespace_iter,
1447                                                    &bucket_gens_iter);
1448                         if (ret)
1449                                 goto bkey_err;
1450                 } else {
1451                         next = k.k->p;
1452
1453                         ret = bch2_check_alloc_hole_freespace(&trans,
1454                                                     bkey_start_pos(k.k),
1455                                                     &next,
1456                                                     &freespace_iter) ?:
1457                                 bch2_check_alloc_hole_bucket_gens(&trans,
1458                                                     bkey_start_pos(k.k),
1459                                                     &next,
1460                                                     &bucket_gens_iter);
1461                         if (ret)
1462                                 goto bkey_err;
1463                 }
1464
1465                 ret = bch2_trans_commit(&trans, NULL, NULL,
1466                                         BTREE_INSERT_NOFAIL|
1467                                         BTREE_INSERT_LAZY_RW);
1468                 if (ret)
1469                         goto bkey_err;
1470
1471                 bch2_btree_iter_set_pos(&iter, next);
1472 bkey_err:
1473                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1474                         continue;
1475                 if (ret)
1476                         break;
1477         }
1478         bch2_trans_iter_exit(&trans, &bucket_gens_iter);
1479         bch2_trans_iter_exit(&trans, &freespace_iter);
1480         bch2_trans_iter_exit(&trans, &discard_iter);
1481         bch2_trans_iter_exit(&trans, &iter);
1482
1483         if (ret < 0)
1484                 goto err;
1485
1486         ret = for_each_btree_key_commit(&trans, iter,
1487                         BTREE_ID_need_discard, POS_MIN,
1488                         BTREE_ITER_PREFETCH, k,
1489                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1490                 bch2_check_discard_freespace_key(&trans, &iter)) ?:
1491               for_each_btree_key_commit(&trans, iter,
1492                         BTREE_ID_freespace, POS_MIN,
1493                         BTREE_ITER_PREFETCH, k,
1494                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1495                 bch2_check_discard_freespace_key(&trans, &iter)) ?:
1496               for_each_btree_key_commit(&trans, iter,
1497                         BTREE_ID_bucket_gens, POS_MIN,
1498                         BTREE_ITER_PREFETCH, k,
1499                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1500                 bch2_check_bucket_gens_key(&trans, &iter, k));
1501 err:
1502         bch2_trans_exit(&trans);
1503         return ret < 0 ? ret : 0;
1504 }
1505
1506 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
1507                                        struct btree_iter *alloc_iter)
1508 {
1509         struct bch_fs *c = trans->c;
1510         struct btree_iter lru_iter;
1511         struct bch_alloc_v4 a_convert;
1512         const struct bch_alloc_v4 *a;
1513         struct bkey_s_c alloc_k, k;
1514         struct printbuf buf = PRINTBUF;
1515         struct printbuf buf2 = PRINTBUF;
1516         int ret;
1517
1518         alloc_k = bch2_btree_iter_peek(alloc_iter);
1519         if (!alloc_k.k)
1520                 return 0;
1521
1522         ret = bkey_err(alloc_k);
1523         if (ret)
1524                 return ret;
1525
1526         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1527
1528         if (a->data_type != BCH_DATA_cached)
1529                 return 0;
1530
1531         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
1532                              POS(alloc_k.k->p.inode, a->io_time[READ]), 0);
1533
1534         k = bch2_btree_iter_peek_slot(&lru_iter);
1535         ret = bkey_err(k);
1536         if (ret)
1537                 goto err;
1538
1539         if (fsck_err_on(!a->io_time[READ], c,
1540                         "cached bucket with read_time 0\n"
1541                         "  %s",
1542                 (printbuf_reset(&buf),
1543                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) ||
1544             fsck_err_on(k.k->type != KEY_TYPE_set ||
1545                         le64_to_cpu(bkey_s_c_to_lru(k).v->idx) != alloc_k.k->p.offset, c,
1546                         "incorrect/missing lru entry\n"
1547                         "  %s\n"
1548                         "  %s",
1549                         (printbuf_reset(&buf),
1550                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
1551                         (bch2_bkey_val_to_text(&buf2, c, k), buf2.buf))) {
1552                 u64 read_time = a->io_time[READ] ?:
1553                         atomic64_read(&c->io_clock[READ].now);
1554
1555                 ret = bch2_lru_set(trans,
1556                                    alloc_k.k->p.inode,
1557                                    bucket_to_u64(alloc_k.k->p),
1558                                    read_time);
1559                 if (ret)
1560                         goto err;
1561
1562                 if (a->io_time[READ] != read_time) {
1563                         struct bkey_i_alloc_v4 *a_mut =
1564                                 bch2_alloc_to_v4_mut(trans, alloc_k);
1565                         ret = PTR_ERR_OR_ZERO(a_mut);
1566                         if (ret)
1567                                 goto err;
1568
1569                         a_mut->v.io_time[READ] = read_time;
1570                         ret = bch2_trans_update(trans, alloc_iter,
1571                                                 &a_mut->k_i, BTREE_TRIGGER_NORUN);
1572                         if (ret)
1573                                 goto err;
1574                 }
1575         }
1576 err:
1577 fsck_err:
1578         bch2_trans_iter_exit(trans, &lru_iter);
1579         printbuf_exit(&buf2);
1580         printbuf_exit(&buf);
1581         return ret;
1582 }
1583
1584 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
1585 {
1586         struct btree_trans trans;
1587         struct btree_iter iter;
1588         struct bkey_s_c k;
1589         int ret = 0;
1590
1591         bch2_trans_init(&trans, c, 0, 0);
1592
1593         for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1594                         POS_MIN, BTREE_ITER_PREFETCH, k,
1595                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1596                 bch2_check_alloc_to_lru_ref(&trans, &iter));
1597
1598         bch2_trans_exit(&trans);
1599         return ret < 0 ? ret : 0;
1600 }
1601
1602 static int bch2_discard_one_bucket(struct btree_trans *trans,
1603                                    struct btree_iter *need_discard_iter,
1604                                    struct bpos *discard_pos_done,
1605                                    u64 *seen,
1606                                    u64 *open,
1607                                    u64 *need_journal_commit,
1608                                    u64 *discarded)
1609 {
1610         struct bch_fs *c = trans->c;
1611         struct bpos pos = need_discard_iter->pos;
1612         struct btree_iter iter = { NULL };
1613         struct bkey_s_c k;
1614         struct bch_dev *ca;
1615         struct bkey_i_alloc_v4 *a;
1616         struct printbuf buf = PRINTBUF;
1617         bool did_discard = false;
1618         int ret = 0;
1619
1620         ca = bch_dev_bkey_exists(c, pos.inode);
1621         if (!percpu_ref_tryget(&ca->io_ref)) {
1622                 bch2_btree_iter_set_pos(need_discard_iter, POS(pos.inode + 1, 0));
1623                 return 0;
1624         }
1625
1626         if (bch2_bucket_is_open_safe(c, pos.inode, pos.offset)) {
1627                 (*open)++;
1628                 goto out;
1629         }
1630
1631         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
1632                         c->journal.flushed_seq_ondisk,
1633                         pos.inode, pos.offset)) {
1634                 (*need_journal_commit)++;
1635                 goto out;
1636         }
1637
1638         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
1639                              need_discard_iter->pos,
1640                              BTREE_ITER_CACHED);
1641         k = bch2_btree_iter_peek_slot(&iter);
1642         ret = bkey_err(k);
1643         if (ret)
1644                 goto out;
1645
1646         a = bch2_alloc_to_v4_mut(trans, k);
1647         ret = PTR_ERR_OR_ZERO(a);
1648         if (ret)
1649                 goto out;
1650
1651         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
1652                 a->v.gen++;
1653                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1654                 goto write;
1655         }
1656
1657         if (a->v.journal_seq > c->journal.flushed_seq_ondisk) {
1658                 if (test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
1659                         bch2_trans_inconsistent(trans,
1660                                 "clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
1661                                 "%s",
1662                                 a->v.journal_seq,
1663                                 c->journal.flushed_seq_ondisk,
1664                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1665                         ret = -EIO;
1666                 }
1667                 goto out;
1668         }
1669
1670         if (a->v.data_type != BCH_DATA_need_discard) {
1671                 if (test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
1672                         bch2_trans_inconsistent(trans,
1673                                 "bucket incorrectly set in need_discard btree\n"
1674                                 "%s",
1675                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1676                         ret = -EIO;
1677                 }
1678
1679                 goto out;
1680         }
1681
1682         if (!bkey_eq(*discard_pos_done, iter.pos) &&
1683             ca->mi.discard && !c->opts.nochanges) {
1684                 /*
1685                  * This works without any other locks because this is the only
1686                  * thread that removes items from the need_discard tree
1687                  */
1688                 bch2_trans_unlock(trans);
1689                 blkdev_issue_discard(ca->disk_sb.bdev,
1690                                      k.k->p.offset * ca->mi.bucket_size,
1691                                      ca->mi.bucket_size,
1692                                      GFP_KERNEL);
1693
1694                 ret = bch2_trans_relock(trans);
1695                 if (ret)
1696                         goto out;
1697         }
1698
1699         *discard_pos_done = iter.pos;
1700         did_discard = true;
1701
1702         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
1703         a->v.data_type = alloc_data_type(a->v, a->v.data_type);
1704 write:
1705         ret =   bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1706                 bch2_trans_commit(trans, NULL, NULL,
1707                                   BTREE_INSERT_USE_RESERVE|BTREE_INSERT_NOFAIL);
1708         if (ret)
1709                 goto out;
1710
1711         if (did_discard) {
1712                 this_cpu_inc(c->counters[BCH_COUNTER_bucket_discard]);
1713                 (*discarded)++;
1714         }
1715 out:
1716         bch2_trans_iter_exit(trans, &iter);
1717         percpu_ref_put(&ca->io_ref);
1718         printbuf_exit(&buf);
1719         return ret;
1720 }
1721
1722 static void bch2_do_discards_work(struct work_struct *work)
1723 {
1724         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
1725         struct btree_trans trans;
1726         struct btree_iter iter;
1727         struct bkey_s_c k;
1728         u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
1729         struct bpos discard_pos_done = POS_MAX;
1730         int ret;
1731
1732         bch2_trans_init(&trans, c, 0, 0);
1733
1734         /*
1735          * We're doing the commit in bch2_discard_one_bucket instead of using
1736          * for_each_btree_key_commit() so that we can increment counters after
1737          * successful commit:
1738          */
1739         ret = for_each_btree_key2(&trans, iter,
1740                         BTREE_ID_need_discard, POS_MIN, 0, k,
1741                 bch2_discard_one_bucket(&trans, &iter, &discard_pos_done,
1742                                         &seen,
1743                                         &open,
1744                                         &need_journal_commit,
1745                                         &discarded));
1746
1747         bch2_trans_exit(&trans);
1748
1749         if (need_journal_commit * 2 > seen)
1750                 bch2_journal_flush_async(&c->journal, NULL);
1751
1752         percpu_ref_put(&c->writes);
1753
1754         trace_discard_buckets(c, seen, open, need_journal_commit, discarded,
1755                               bch2_err_str(ret));
1756 }
1757
1758 void bch2_do_discards(struct bch_fs *c)
1759 {
1760         if (percpu_ref_tryget_live(&c->writes) &&
1761             !queue_work(system_long_wq, &c->discard_work))
1762                 percpu_ref_put(&c->writes);
1763 }
1764
1765 static int invalidate_one_bucket(struct btree_trans *trans,
1766                                  struct btree_iter *lru_iter,
1767                                  struct bpos bucket,
1768                                  s64 *nr_to_invalidate)
1769 {
1770         struct bch_fs *c = trans->c;
1771         struct btree_iter alloc_iter = { NULL };
1772         struct bkey_i_alloc_v4 *a;
1773         struct printbuf buf = PRINTBUF;
1774         unsigned cached_sectors;
1775         int ret = 0;
1776
1777         if (*nr_to_invalidate <= 0)
1778                 return 1;
1779
1780         a = bch2_trans_start_alloc_update(trans, &alloc_iter, bucket);
1781         ret = PTR_ERR_OR_ZERO(a);
1782         if (ret)
1783                 goto out;
1784
1785         if (lru_pos_time(lru_iter->pos) != alloc_lru_idx(a->v)) {
1786                 prt_printf(&buf, "alloc key does not point back to lru entry when invalidating bucket:\n  ");
1787                 bch2_bpos_to_text(&buf, lru_iter->pos);
1788                 prt_printf(&buf, "\n  ");
1789                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
1790
1791                 bch_err(c, "%s", buf.buf);
1792                 if (test_bit(BCH_FS_CHECK_LRUS_DONE, &c->flags)) {
1793                         bch2_inconsistent_error(c);
1794                         ret = -EINVAL;
1795                 }
1796
1797                 goto out;
1798         }
1799
1800         if (!a->v.cached_sectors)
1801                 bch_err(c, "invalidating empty bucket, confused");
1802
1803         cached_sectors = a->v.cached_sectors;
1804
1805         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1806         a->v.gen++;
1807         a->v.data_type          = 0;
1808         a->v.dirty_sectors      = 0;
1809         a->v.cached_sectors     = 0;
1810         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1811         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1812
1813         ret =   bch2_trans_update(trans, &alloc_iter, &a->k_i,
1814                                 BTREE_TRIGGER_BUCKET_INVALIDATE) ?:
1815                 bch2_trans_commit(trans, NULL, NULL,
1816                                   BTREE_INSERT_USE_RESERVE|BTREE_INSERT_NOFAIL);
1817         if (ret)
1818                 goto out;
1819
1820         trace_and_count(c, bucket_invalidate, c, bucket.inode, bucket.offset, cached_sectors);
1821         --*nr_to_invalidate;
1822 out:
1823         bch2_trans_iter_exit(trans, &alloc_iter);
1824         printbuf_exit(&buf);
1825         return ret;
1826 }
1827
1828 static void bch2_do_invalidates_work(struct work_struct *work)
1829 {
1830         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1831         struct bch_dev *ca;
1832         struct btree_trans trans;
1833         struct btree_iter iter;
1834         struct bkey_s_c k;
1835         unsigned i;
1836         int ret = 0;
1837
1838         bch2_trans_init(&trans, c, 0, 0);
1839
1840         for_each_member_device(ca, c, i) {
1841                 s64 nr_to_invalidate =
1842                         should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
1843
1844                 ret = for_each_btree_key2_upto(&trans, iter, BTREE_ID_lru,
1845                                 lru_pos(ca->dev_idx, 0, 0),
1846                                 lru_pos(ca->dev_idx, U64_MAX, LRU_TIME_MAX),
1847                                 BTREE_ITER_INTENT, k,
1848                         invalidate_one_bucket(&trans, &iter,
1849                                               u64_to_bucket(k.k->p.offset),
1850                                               &nr_to_invalidate));
1851
1852                 if (ret < 0) {
1853                         percpu_ref_put(&ca->ref);
1854                         break;
1855                 }
1856         }
1857
1858         bch2_trans_exit(&trans);
1859         percpu_ref_put(&c->writes);
1860 }
1861
1862 void bch2_do_invalidates(struct bch_fs *c)
1863 {
1864         if (percpu_ref_tryget_live(&c->writes) &&
1865             !queue_work(system_long_wq, &c->invalidate_work))
1866                 percpu_ref_put(&c->writes);
1867 }
1868
1869 static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca)
1870 {
1871         struct btree_trans trans;
1872         struct btree_iter iter;
1873         struct bkey_s_c k;
1874         struct bkey hole;
1875         struct bpos end = POS(ca->dev_idx, ca->mi.nbuckets);
1876         struct bch_member *m;
1877         int ret;
1878
1879         bch2_trans_init(&trans, c, 0, 0);
1880
1881         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc,
1882                              POS(ca->dev_idx, ca->mi.first_bucket),
1883                              BTREE_ITER_PREFETCH);
1884         /*
1885          * Scan the alloc btree for every bucket on @ca, and add buckets to the
1886          * freespace/need_discard/need_gc_gens btrees as needed:
1887          */
1888         while (1) {
1889                 bch2_trans_begin(&trans);
1890
1891                 if (bkey_ge(iter.pos, end)) {
1892                         ret = 0;
1893                         break;
1894                 }
1895
1896                 k = bch2_get_key_or_hole(&iter, end, &hole);
1897                 ret = bkey_err(k);
1898                 if (ret)
1899                         goto bkey_err;
1900
1901                 if (k.k->type) {
1902                         /*
1903                          * We process live keys in the alloc btree one at a
1904                          * time:
1905                          */
1906                         struct bch_alloc_v4 a_convert;
1907                         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1908
1909                         ret =   bch2_bucket_do_index(&trans, k, a, true) ?:
1910                                 bch2_trans_commit(&trans, NULL, NULL,
1911                                                   BTREE_INSERT_LAZY_RW|
1912                                                   BTREE_INSERT_NOFAIL);
1913                         if (ret)
1914                                 goto bkey_err;
1915
1916                         bch2_btree_iter_advance(&iter);
1917                 } else {
1918                         struct bkey_i *freespace;
1919
1920                         freespace = bch2_trans_kmalloc(&trans, sizeof(*freespace));
1921                         ret = PTR_ERR_OR_ZERO(freespace);
1922                         if (ret)
1923                                 goto bkey_err;
1924
1925                         bkey_init(&freespace->k);
1926                         freespace->k.type       = KEY_TYPE_set;
1927                         freespace->k.p          = k.k->p;
1928                         freespace->k.size       = k.k->size;
1929
1930                         ret = __bch2_btree_insert(&trans, BTREE_ID_freespace, freespace) ?:
1931                                 bch2_trans_commit(&trans, NULL, NULL,
1932                                                   BTREE_INSERT_LAZY_RW|
1933                                                   BTREE_INSERT_NOFAIL);
1934                         if (ret)
1935                                 goto bkey_err;
1936
1937                         bch2_btree_iter_set_pos(&iter, k.k->p);
1938                 }
1939 bkey_err:
1940                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1941                         continue;
1942                 if (ret)
1943                         break;
1944         }
1945
1946         bch2_trans_iter_exit(&trans, &iter);
1947         bch2_trans_exit(&trans);
1948
1949         if (ret < 0) {
1950                 bch_err(ca, "error initializing free space: %s", bch2_err_str(ret));
1951                 return ret;
1952         }
1953
1954         mutex_lock(&c->sb_lock);
1955         m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx;
1956         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1957         mutex_unlock(&c->sb_lock);
1958
1959         return 0;
1960 }
1961
1962 int bch2_fs_freespace_init(struct bch_fs *c)
1963 {
1964         struct bch_dev *ca;
1965         unsigned i;
1966         int ret = 0;
1967         bool doing_init = false;
1968
1969         /*
1970          * We can crash during the device add path, so we need to check this on
1971          * every mount:
1972          */
1973
1974         for_each_member_device(ca, c, i) {
1975                 if (ca->mi.freespace_initialized)
1976                         continue;
1977
1978                 if (!doing_init) {
1979                         bch_info(c, "initializing freespace");
1980                         doing_init = true;
1981                 }
1982
1983                 ret = bch2_dev_freespace_init(c, ca);
1984                 if (ret) {
1985                         percpu_ref_put(&ca->ref);
1986                         return ret;
1987                 }
1988         }
1989
1990         if (doing_init) {
1991                 mutex_lock(&c->sb_lock);
1992                 bch2_write_super(c);
1993                 mutex_unlock(&c->sb_lock);
1994
1995                 bch_verbose(c, "done initializing freespace");
1996         }
1997
1998         return ret;
1999 }
2000
2001 /* Bucket IO clocks: */
2002
2003 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
2004                               size_t bucket_nr, int rw)
2005 {
2006         struct bch_fs *c = trans->c;
2007         struct btree_iter iter;
2008         struct bkey_i_alloc_v4 *a;
2009         u64 now;
2010         int ret = 0;
2011
2012         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
2013         ret = PTR_ERR_OR_ZERO(a);
2014         if (ret)
2015                 return ret;
2016
2017         now = atomic64_read(&c->io_clock[rw].now);
2018         if (a->v.io_time[rw] == now)
2019                 goto out;
2020
2021         a->v.io_time[rw] = now;
2022
2023         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
2024                 bch2_trans_commit(trans, NULL, NULL, 0);
2025 out:
2026         bch2_trans_iter_exit(trans, &iter);
2027         return ret;
2028 }
2029
2030 /* Startup/shutdown (ro/rw): */
2031
2032 void bch2_recalc_capacity(struct bch_fs *c)
2033 {
2034         struct bch_dev *ca;
2035         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
2036         unsigned bucket_size_max = 0;
2037         unsigned long ra_pages = 0;
2038         unsigned i;
2039
2040         lockdep_assert_held(&c->state_lock);
2041
2042         for_each_online_member(ca, c, i) {
2043                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
2044
2045                 ra_pages += bdi->ra_pages;
2046         }
2047
2048         bch2_set_ra_pages(c, ra_pages);
2049
2050         for_each_rw_member(ca, c, i) {
2051                 u64 dev_reserve = 0;
2052
2053                 /*
2054                  * We need to reserve buckets (from the number
2055                  * of currently available buckets) against
2056                  * foreground writes so that mainly copygc can
2057                  * make forward progress.
2058                  *
2059                  * We need enough to refill the various reserves
2060                  * from scratch - copygc will use its entire
2061                  * reserve all at once, then run against when
2062                  * its reserve is refilled (from the formerly
2063                  * available buckets).
2064                  *
2065                  * This reserve is just used when considering if
2066                  * allocations for foreground writes must wait -
2067                  * not -ENOSPC calculations.
2068                  */
2069
2070                 dev_reserve += ca->nr_btree_reserve * 2;
2071                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
2072
2073                 dev_reserve += 1;       /* btree write point */
2074                 dev_reserve += 1;       /* copygc write point */
2075                 dev_reserve += 1;       /* rebalance write point */
2076
2077                 dev_reserve *= ca->mi.bucket_size;
2078
2079                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
2080                                              ca->mi.first_bucket);
2081
2082                 reserved_sectors += dev_reserve * 2;
2083
2084                 bucket_size_max = max_t(unsigned, bucket_size_max,
2085                                         ca->mi.bucket_size);
2086         }
2087
2088         gc_reserve = c->opts.gc_reserve_bytes
2089                 ? c->opts.gc_reserve_bytes >> 9
2090                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
2091
2092         reserved_sectors = max(gc_reserve, reserved_sectors);
2093
2094         reserved_sectors = min(reserved_sectors, capacity);
2095
2096         c->capacity = capacity - reserved_sectors;
2097
2098         c->bucket_size_max = bucket_size_max;
2099
2100         /* Wake up case someone was waiting for buckets */
2101         closure_wake_up(&c->freelist_wait);
2102 }
2103
2104 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
2105 {
2106         struct open_bucket *ob;
2107         bool ret = false;
2108
2109         for (ob = c->open_buckets;
2110              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
2111              ob++) {
2112                 spin_lock(&ob->lock);
2113                 if (ob->valid && !ob->on_partial_list &&
2114                     ob->dev == ca->dev_idx)
2115                         ret = true;
2116                 spin_unlock(&ob->lock);
2117         }
2118
2119         return ret;
2120 }
2121
2122 /* device goes ro: */
2123 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
2124 {
2125         unsigned i;
2126
2127         /* First, remove device from allocation groups: */
2128
2129         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2130                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
2131
2132         /*
2133          * Capacity is calculated based off of devices in allocation groups:
2134          */
2135         bch2_recalc_capacity(c);
2136
2137         /* Next, close write points that point to this device... */
2138         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
2139                 bch2_writepoint_stop(c, ca, &c->write_points[i]);
2140
2141         bch2_writepoint_stop(c, ca, &c->copygc_write_point);
2142         bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
2143         bch2_writepoint_stop(c, ca, &c->btree_write_point);
2144
2145         mutex_lock(&c->btree_reserve_cache_lock);
2146         while (c->btree_reserve_cache_nr) {
2147                 struct btree_alloc *a =
2148                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
2149
2150                 bch2_open_buckets_put(c, &a->ob);
2151         }
2152         mutex_unlock(&c->btree_reserve_cache_lock);
2153
2154         while (1) {
2155                 struct open_bucket *ob;
2156
2157                 spin_lock(&c->freelist_lock);
2158                 if (!ca->open_buckets_partial_nr) {
2159                         spin_unlock(&c->freelist_lock);
2160                         break;
2161                 }
2162                 ob = c->open_buckets +
2163                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
2164                 ob->on_partial_list = false;
2165                 spin_unlock(&c->freelist_lock);
2166
2167                 bch2_open_bucket_put(c, ob);
2168         }
2169
2170         bch2_ec_stop_dev(c, ca);
2171
2172         /*
2173          * Wake up threads that were blocked on allocation, so they can notice
2174          * the device can no longer be removed and the capacity has changed:
2175          */
2176         closure_wake_up(&c->freelist_wait);
2177
2178         /*
2179          * journal_res_get() can block waiting for free space in the journal -
2180          * it needs to notice there may not be devices to allocate from anymore:
2181          */
2182         wake_up(&c->journal.wait);
2183
2184         /* Now wait for any in flight writes: */
2185
2186         closure_wait_event(&c->open_buckets_wait,
2187                            !bch2_dev_has_open_write_point(c, ca));
2188 }
2189
2190 /* device goes rw: */
2191 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
2192 {
2193         unsigned i;
2194
2195         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2196                 if (ca->mi.data_allowed & (1 << i))
2197                         set_bit(ca->dev_idx, c->rw_devs[i].d);
2198 }
2199
2200 void bch2_fs_allocator_background_init(struct bch_fs *c)
2201 {
2202         spin_lock_init(&c->freelist_lock);
2203         INIT_WORK(&c->discard_work, bch2_do_discards_work);
2204         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
2205 }