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

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "bcachefs.h"
   3 #include "alloc_background.h"
   4 #include "alloc_foreground.h"
   5 #include "backpointers.h"
   6 #include "btree_cache.h"
   7 #include "btree_io.h"
   8 #include "btree_key_cache.h"
   9 #include "btree_update.h"
  10 #include "btree_update_interior.h"
  11 #include "btree_gc.h"
  12 #include "btree_write_buffer.h"
  13 #include "buckets.h"
  14 #include "buckets_waiting_for_journal.h"
  15 #include "clock.h"
  16 #include "debug.h"
  17 #include "ec.h"
  18 #include "error.h"
  19 #include "lru.h"
  20 #include "recovery.h"
  21 #include "varint.h"
  22
  23 #include <linux/kthread.h>
  24 #include <linux/math64.h>
  25 #include <linux/random.h>
  26 #include <linux/rculist.h>
  27 #include <linux/rcupdate.h>
  28 #include <linux/sched/task.h>
  29 #include <linux/sort.h>
  30 #include <trace/events/bcachefs.h>
  31
  32 /* Persistent alloc info: */
  33
  34 static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
  35 #define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8,
  36         BCH_ALLOC_FIELDS_V1()
  37 #undef x
  38 };
  39
  40 struct bkey_alloc_unpacked {
  41         u64             journal_seq;
  42         u8              gen;
  43         u8              oldest_gen;
  44         u8              data_type;
  45         bool            need_discard:1;
  46         bool            need_inc_gen:1;
  47 #define x(_name, _bits) u##_bits _name;
  48         BCH_ALLOC_FIELDS_V2()
  49 #undef  x
  50 };
  51
  52 static inline u64 alloc_field_v1_get(const struct bch_alloc *a,
  53                                      const void **p, unsigned field)
  54 {
  55         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  56         u64 v;
  57
  58         if (!(a->fields & (1 << field)))
  59                 return 0;
  60
  61         switch (bytes) {
  62         case 1:
  63                 v = *((const u8 *) *p);
  64                 break;
  65         case 2:
  66                 v = le16_to_cpup(*p);
  67                 break;
  68         case 4:
  69                 v = le32_to_cpup(*p);
  70                 break;
  71         case 8:
  72                 v = le64_to_cpup(*p);
  73                 break;
  74         default:
  75                 BUG();
  76         }
  77
  78         *p += bytes;
  79         return v;
  80 }
  81
  82 static inline void alloc_field_v1_put(struct bkey_i_alloc *a, void **p,
  83                                       unsigned field, u64 v)
  84 {
  85         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  86
  87         if (!v)
  88                 return;
  89
  90         a->v.fields |= 1 << field;
  91
  92         switch (bytes) {
  93         case 1:
  94                 *((u8 *) *p) = v;
  95                 break;
  96         case 2:
  97                 *((__le16 *) *p) = cpu_to_le16(v);
  98                 break;
  99         case 4:
 100                 *((__le32 *) *p) = cpu_to_le32(v);
 101                 break;
 102         case 8:
 103                 *((__le64 *) *p) = cpu_to_le64(v);
 104                 break;
 105         default:
 106                 BUG();
 107         }
 108
 109         *p += bytes;
 110 }
 111
 112 static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out,
 113                                  struct bkey_s_c k)
 114 {
 115         const struct bch_alloc *in = bkey_s_c_to_alloc(k).v;
 116         const void *d = in->data;
 117         unsigned idx = 0;
 118
 119         out->gen = in->gen;
 120
 121 #define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++);
 122         BCH_ALLOC_FIELDS_V1()
 123 #undef  x
 124 }
 125
 126 static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out,
 127                                 struct bkey_s_c k)
 128 {
 129         struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k);
 130         const u8 *in = a.v->data;
 131         const u8 *end = bkey_val_end(a);
 132         unsigned fieldnr = 0;
 133         int ret;
 134         u64 v;
 135
 136         out->gen        = a.v->gen;
 137         out->oldest_gen = a.v->oldest_gen;
 138         out->data_type  = a.v->data_type;
 139
 140 #define x(_name, _bits)                                                 \
 141         if (fieldnr < a.v->nr_fields) {                                 \
 142                 ret = bch2_varint_decode_fast(in, end, &v);             \
 143                 if (ret < 0)                                            \
 144                         return ret;                                     \
 145                 in += ret;                                              \
 146         } else {                                                        \
 147                 v = 0;                                                  \
 148         }                                                               \
 149         out->_name = v;                                                 \
 150         if (v != out->_name)                                            \
 151                 return -1;                                              \
 152         fieldnr++;
 153
 154         BCH_ALLOC_FIELDS_V2()
 155 #undef  x
 156         return 0;
 157 }
 158
 159 static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out,
 160                                 struct bkey_s_c k)
 161 {
 162         struct bkey_s_c_alloc_v3 a = bkey_s_c_to_alloc_v3(k);
 163         const u8 *in = a.v->data;
 164         const u8 *end = bkey_val_end(a);
 165         unsigned fieldnr = 0;
 166         int ret;
 167         u64 v;
 168
 169         out->gen        = a.v->gen;
 170         out->oldest_gen = a.v->oldest_gen;
 171         out->data_type  = a.v->data_type;
 172         out->need_discard = BCH_ALLOC_V3_NEED_DISCARD(a.v);
 173         out->need_inc_gen = BCH_ALLOC_V3_NEED_INC_GEN(a.v);
 174         out->journal_seq = le64_to_cpu(a.v->journal_seq);
 175
 176 #define x(_name, _bits)                                                 \
 177         if (fieldnr < a.v->nr_fields) {                                 \
 178                 ret = bch2_varint_decode_fast(in, end, &v);             \
 179                 if (ret < 0)                                            \
 180                         return ret;                                     \
 181                 in += ret;                                              \
 182         } else {                                                        \
 183                 v = 0;                                                  \
 184         }                                                               \
 185         out->_name = v;                                                 \
 186         if (v != out->_name)                                            \
 187                 return -1;                                              \
 188         fieldnr++;
 189
 190         BCH_ALLOC_FIELDS_V2()
 191 #undef  x
 192         return 0;
 193 }
 194
 195 static struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 196 {
 197         struct bkey_alloc_unpacked ret = { .gen = 0 };
 198
 199         switch (k.k->type) {
 200         case KEY_TYPE_alloc:
 201                 bch2_alloc_unpack_v1(&ret, k);
 202                 break;
 203         case KEY_TYPE_alloc_v2:
 204                 bch2_alloc_unpack_v2(&ret, k);
 205                 break;
 206         case KEY_TYPE_alloc_v3:
 207                 bch2_alloc_unpack_v3(&ret, k);
 208                 break;
 209         }
 210
 211         return ret;
 212 }
 213
 214 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
 215 {
 216         unsigned i, bytes = offsetof(struct bch_alloc, data);
 217
 218         for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++)
 219                 if (a->fields & (1 << i))
 220                         bytes += BCH_ALLOC_V1_FIELD_BYTES[i];
 221
 222         return DIV_ROUND_UP(bytes, sizeof(u64));
 223 }
 224
 225 int bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k,
 226                           unsigned flags, struct printbuf *err)
 227 {
 228         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 229
 230         /* allow for unknown fields */
 231         if (bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v)) {
 232                 prt_printf(err, "incorrect value size (%zu < %u)",
 233                        bkey_val_u64s(a.k), bch_alloc_v1_val_u64s(a.v));
 234                 return -BCH_ERR_invalid_bkey;
 235         }
 236
 237         return 0;
 238 }
 239
 240 int bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k,
 241                           unsigned flags, struct printbuf *err)
 242 {
 243         struct bkey_alloc_unpacked u;
 244
 245         if (bch2_alloc_unpack_v2(&u, k)) {
 246                 prt_printf(err, "unpack error");
 247                 return -BCH_ERR_invalid_bkey;
 248         }
 249
 250         return 0;
 251 }
 252
 253 int bch2_alloc_v3_invalid(const struct bch_fs *c, struct bkey_s_c k,
 254                           unsigned flags, struct printbuf *err)
 255 {
 256         struct bkey_alloc_unpacked u;
 257
 258         if (bch2_alloc_unpack_v3(&u, k)) {
 259                 prt_printf(err, "unpack error");
 260                 return -BCH_ERR_invalid_bkey;
 261         }
 262
 263         return 0;
 264 }
 265
 266 int bch2_alloc_v4_invalid(const struct bch_fs *c, struct bkey_s_c k,
 267                           unsigned flags, struct printbuf *err)
 268 {
 269         struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 270         int rw = flags & WRITE;
 271
 272         if (alloc_v4_u64s(a.v) != bkey_val_u64s(k.k)) {
 273                 prt_printf(err, "bad val size (%lu != %u)",
 274                        bkey_val_u64s(k.k), alloc_v4_u64s(a.v));
 275                 return -BCH_ERR_invalid_bkey;
 276         }
 277
 278         if (!BCH_ALLOC_V4_BACKPOINTERS_START(a.v) &&
 279             BCH_ALLOC_V4_NR_BACKPOINTERS(a.v)) {
 280                 prt_printf(err, "invalid backpointers_start");
 281                 return -BCH_ERR_invalid_bkey;
 282         }
 283
 284         if (rw == WRITE &&
 285             !(flags & BKEY_INVALID_FROM_JOURNAL) &&
 286             test_bit(BCH_FS_CHECK_BACKPOINTERS_DONE, &c->flags)) {
 287                 unsigned i, bp_len = 0;
 288
 289                 for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a.v); i++)
 290                         bp_len += alloc_v4_backpointers_c(a.v)[i].bucket_len;
 291
 292                 if (bp_len > a.v->dirty_sectors) {
 293                         prt_printf(err, "too many backpointers");
 294                         return -BCH_ERR_invalid_bkey;
 295                 }
 296         }
 297
 298         if (rw == WRITE) {
 299                 if (alloc_data_type(*a.v, a.v->data_type) != a.v->data_type) {
 300                         prt_printf(err, "invalid data type (got %u should be %u)",
 301                                a.v->data_type, alloc_data_type(*a.v, a.v->data_type));
 302                         return -BCH_ERR_invalid_bkey;
 303                 }
 304
 305                 switch (a.v->data_type) {
 306                 case BCH_DATA_free:
 307                 case BCH_DATA_need_gc_gens:
 308                 case BCH_DATA_need_discard:
 309                         if (a.v->dirty_sectors ||
 310                             a.v->cached_sectors ||
 311                             a.v->stripe) {
 312                                 prt_printf(err, "empty data type free but have data");
 313                                 return -BCH_ERR_invalid_bkey;
 314                         }
 315                         break;
 316                 case BCH_DATA_sb:
 317                 case BCH_DATA_journal:
 318                 case BCH_DATA_btree:
 319                 case BCH_DATA_user:
 320                 case BCH_DATA_parity:
 321                         if (!a.v->dirty_sectors) {
 322                                 prt_printf(err, "data_type %s but dirty_sectors==0",
 323                                        bch2_data_types[a.v->data_type]);
 324                                 return -BCH_ERR_invalid_bkey;
 325                         }
 326                         break;
 327                 case BCH_DATA_cached:
 328                         if (!a.v->cached_sectors ||
 329                             a.v->dirty_sectors ||
 330                             a.v->stripe) {
 331                                 prt_printf(err, "data type inconsistency");
 332                                 return -BCH_ERR_invalid_bkey;
 333                         }
 334
 335                         if (!a.v->io_time[READ] &&
 336                             test_bit(BCH_FS_CHECK_ALLOC_TO_LRU_REFS_DONE, &c->flags)) {
 337                                 prt_printf(err, "cached bucket with read_time == 0");
 338                                 return -BCH_ERR_invalid_bkey;
 339                         }
 340                         break;
 341                 case BCH_DATA_stripe:
 342                         if (!a.v->stripe) {
 343                                 prt_printf(err, "data_type %s but stripe==0",
 344                                        bch2_data_types[a.v->data_type]);
 345                                 return -BCH_ERR_invalid_bkey;
 346                         }
 347                         break;
 348                 }
 349         }
 350
 351         return 0;
 352 }
 353
 354 static inline u64 swab40(u64 x)
 355 {
 356         return (((x & 0x00000000ffULL) << 32)|
 357                 ((x & 0x000000ff00ULL) << 16)|
 358                 ((x & 0x0000ff0000ULL) >>  0)|
 359                 ((x & 0x00ff000000ULL) >> 16)|
 360                 ((x & 0xff00000000ULL) >> 32));
 361 }
 362
 363 void bch2_alloc_v4_swab(struct bkey_s k)
 364 {
 365         struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v;
 366         struct bch_backpointer *bp, *bps;
 367
 368         a->journal_seq          = swab64(a->journal_seq);
 369         a->flags                = swab32(a->flags);
 370         a->dirty_sectors        = swab32(a->dirty_sectors);
 371         a->cached_sectors       = swab32(a->cached_sectors);
 372         a->io_time[0]           = swab64(a->io_time[0]);
 373         a->io_time[1]           = swab64(a->io_time[1]);
 374         a->stripe               = swab32(a->stripe);
 375         a->nr_external_backpointers = swab32(a->nr_external_backpointers);
 376
 377         bps = alloc_v4_backpointers(a);
 378         for (bp = bps; bp < bps + BCH_ALLOC_V4_NR_BACKPOINTERS(a); bp++) {
 379                 bp->bucket_offset       = swab40(bp->bucket_offset);
 380                 bp->bucket_len          = swab32(bp->bucket_len);
 381                 bch2_bpos_swab(&bp->pos);
 382         }
 383 }
 384
 385 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 386 {
 387         struct bch_alloc_v4 _a;
 388         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &_a);
 389         unsigned i;
 390
 391         prt_newline(out);
 392         printbuf_indent_add(out, 2);
 393
 394         prt_printf(out, "gen %u oldest_gen %u data_type %s",
 395                a->gen, a->oldest_gen,
 396                a->data_type < BCH_DATA_NR
 397                ? bch2_data_types[a->data_type]
 398                : "(invalid data type)");
 399         prt_newline(out);
 400         prt_printf(out, "journal_seq       %llu",       a->journal_seq);
 401         prt_newline(out);
 402         prt_printf(out, "need_discard      %llu",       BCH_ALLOC_V4_NEED_DISCARD(a));
 403         prt_newline(out);
 404         prt_printf(out, "need_inc_gen      %llu",       BCH_ALLOC_V4_NEED_INC_GEN(a));
 405         prt_newline(out);
 406         prt_printf(out, "dirty_sectors     %u", a->dirty_sectors);
 407         prt_newline(out);
 408         prt_printf(out, "cached_sectors    %u", a->cached_sectors);
 409         prt_newline(out);
 410         prt_printf(out, "stripe            %u", a->stripe);
 411         prt_newline(out);
 412         prt_printf(out, "stripe_redundancy %u", a->stripe_redundancy);
 413         prt_newline(out);
 414         prt_printf(out, "io_time[READ]     %llu",       a->io_time[READ]);
 415         prt_newline(out);
 416         prt_printf(out, "io_time[WRITE]    %llu",       a->io_time[WRITE]);
 417         prt_newline(out);
 418         prt_printf(out, "fragmentation     %llu",       a->fragmentation_lru);
 419         prt_newline(out);
 420         prt_printf(out, "bp_start          %llu", BCH_ALLOC_V4_BACKPOINTERS_START(a));
 421         prt_newline(out);
 422
 423         if (BCH_ALLOC_V4_NR_BACKPOINTERS(a)) {
 424                 struct bkey_s_c_alloc_v4 a_raw = bkey_s_c_to_alloc_v4(k);
 425                 const struct bch_backpointer *bps = alloc_v4_backpointers_c(a_raw.v);
 426
 427                 prt_printf(out, "backpointers:     %llu", BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v));
 428                 printbuf_indent_add(out, 2);
 429
 430                 for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a_raw.v); i++) {
 431                         prt_newline(out);
 432                         bch2_backpointer_to_text(out, &bps[i]);
 433                 }
 434
 435                 printbuf_indent_sub(out, 2);
 436         }
 437
 438         printbuf_indent_sub(out, 2);
 439 }
 440
 441 void __bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out)
 442 {
 443         if (k.k->type == KEY_TYPE_alloc_v4) {
 444                 void *src, *dst;
 445
 446                 *out = *bkey_s_c_to_alloc_v4(k).v;
 447
 448                 src = alloc_v4_backpointers(out);
 449                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 450                 dst = alloc_v4_backpointers(out);
 451
 452                 if (src < dst)
 453                         memset(src, 0, dst - src);
 454         } else {
 455                 struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 456
 457                 *out = (struct bch_alloc_v4) {
 458                         .journal_seq            = u.journal_seq,
 459                         .flags                  = u.need_discard,
 460                         .gen                    = u.gen,
 461                         .oldest_gen             = u.oldest_gen,
 462                         .data_type              = u.data_type,
 463                         .stripe_redundancy      = u.stripe_redundancy,
 464                         .dirty_sectors          = u.dirty_sectors,
 465                         .cached_sectors         = u.cached_sectors,
 466                         .io_time[READ]          = u.read_time,
 467                         .io_time[WRITE]         = u.write_time,
 468                         .stripe                 = u.stripe,
 469                 };
 470
 471                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 472         }
 473 }
 474
 475 static noinline struct bkey_i_alloc_v4 *
 476 __bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 477 {
 478         struct bkey_i_alloc_v4 *ret;
 479         if (k.k->type == KEY_TYPE_alloc_v4) {
 480                 struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 481                 unsigned bytes = sizeof(struct bkey_i_alloc_v4) +
 482                         BCH_ALLOC_V4_NR_BACKPOINTERS(a.v) *
 483                         sizeof(struct bch_backpointer);
 484                 void *src, *dst;
 485
 486                 /*
 487                  * Reserve space for one more backpointer here:
 488                  * Not sketchy at doing it this way, nope...
 489                  */
 490                 ret = bch2_trans_kmalloc(trans, bytes + sizeof(struct bch_backpointer));
 491                 if (IS_ERR(ret))
 492                         return ret;
 493
 494                 bkey_reassemble(&ret->k_i, k);
 495
 496                 src = alloc_v4_backpointers(&ret->v);
 497                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(&ret->v, BCH_ALLOC_V4_U64s);
 498                 dst = alloc_v4_backpointers(&ret->v);
 499
 500                 memmove(dst, src, BCH_ALLOC_V4_NR_BACKPOINTERS(&ret->v) *
 501                         sizeof(struct bch_backpointer));
 502                 if (src < dst)
 503                         memset(src, 0, dst - src);
 504                 set_alloc_v4_u64s(ret);
 505         } else {
 506                 ret = bch2_trans_kmalloc(trans, sizeof(struct bkey_i_alloc_v4) +
 507                                          sizeof(struct bch_backpointer));
 508                 if (IS_ERR(ret))
 509                         return ret;
 510
 511                 bkey_alloc_v4_init(&ret->k_i);
 512                 ret->k.p = k.k->p;
 513                 bch2_alloc_to_v4(k, &ret->v);
 514         }
 515         return ret;
 516 }
 517
 518 static inline struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut_inlined(struct btree_trans *trans, struct bkey_s_c k)
 519 {
 520         if (likely(k.k->type == KEY_TYPE_alloc_v4) &&
 521             BCH_ALLOC_V4_BACKPOINTERS_START(bkey_s_c_to_alloc_v4(k).v) == BCH_ALLOC_V4_U64s) {
 522                 /*
 523                  * Reserve space for one more backpointer here:
 524                  * Not sketchy at doing it this way, nope...
 525                  */
 526                 struct bkey_i_alloc_v4 *ret =
 527                         bch2_trans_kmalloc_nomemzero(trans, bkey_bytes(k.k) + sizeof(struct bch_backpointer));
 528                 if (!IS_ERR(ret))
 529                         bkey_reassemble(&ret->k_i, k);
 530                 return ret;
 531         }
 532
 533         return __bch2_alloc_to_v4_mut(trans, k);
 534 }
 535
 536 struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 537 {
 538         return bch2_alloc_to_v4_mut_inlined(trans, k);
 539 }
 540
 541 struct bkey_i_alloc_v4 *
 542 bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
 543                               struct bpos pos)
 544 {
 545         struct bkey_s_c k;
 546         struct bkey_i_alloc_v4 *a;
 547         int ret;
 548
 549         bch2_trans_iter_init(trans, iter, BTREE_ID_alloc, pos,
 550                              BTREE_ITER_WITH_UPDATES|
 551                              BTREE_ITER_CACHED|
 552                              BTREE_ITER_INTENT);
 553         k = bch2_btree_iter_peek_slot(iter);
 554         ret = bkey_err(k);
 555         if (unlikely(ret))
 556                 goto err;
 557
 558         a = bch2_alloc_to_v4_mut_inlined(trans, k);
 559         ret = PTR_ERR_OR_ZERO(a);
 560         if (unlikely(ret))
 561                 goto err;
 562         return a;
 563 err:
 564         bch2_trans_iter_exit(trans, iter);
 565         return ERR_PTR(ret);
 566 }
 567
 568 int bch2_alloc_read(struct bch_fs *c)
 569 {
 570         struct btree_trans trans;
 571         struct btree_iter iter;
 572         struct bkey_s_c k;
 573         struct bch_alloc_v4 a;
 574         struct bch_dev *ca;
 575         int ret;
 576
 577         bch2_trans_init(&trans, c, 0, 0);
 578
 579         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 580                            BTREE_ITER_PREFETCH, k, ret) {
 581                 /*
 582                  * Not a fsck error because this is checked/repaired by
 583                  * bch2_check_alloc_key() which runs later:
 584                  */
 585                 if (!bch2_dev_bucket_exists(c, k.k->p))
 586                         continue;
 587
 588                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 589
 590                 *bucket_gen(ca, k.k->p.offset) = bch2_alloc_to_v4(k, &a)->gen;
 591         }
 592         bch2_trans_iter_exit(&trans, &iter);
 593
 594         bch2_trans_exit(&trans);
 595
 596         if (ret)
 597                 bch_err(c, "error reading alloc info: %s", bch2_err_str(ret));
 598
 599         return ret;
 600 }
 601
 602 static struct bpos alloc_gens_pos(struct bpos pos, unsigned *offset)
 603 {
 604         *offset = pos.offset & KEY_TYPE_BUCKET_GENS_MASK;
 605
 606         pos.offset >>= KEY_TYPE_BUCKET_GENS_BITS;
 607         return pos;
 608 }
 609
 610 static struct bpos bucket_gens_pos_to_alloc(struct bpos pos, unsigned offset)
 611 {
 612         pos.offset <<= KEY_TYPE_BUCKET_GENS_BITS;
 613         pos.offset += offset;
 614         return pos;
 615 }
 616
 617 static unsigned alloc_gen(struct bkey_s_c k, unsigned offset)
 618 {
 619         return k.k->type == KEY_TYPE_bucket_gens
 620                 ? bkey_s_c_to_bucket_gens(k).v->gens[offset]
 621                 : 0;
 622 }
 623
 624 int bch2_bucket_gens_invalid(const struct bch_fs *c, struct bkey_s_c k,
 625                              unsigned flags, struct printbuf *err)
 626 {
 627         if (bkey_val_bytes(k.k) != sizeof(struct bch_bucket_gens)) {
 628                 prt_printf(err, "bad val size (%lu != %zu)",
 629                        bkey_val_bytes(k.k), sizeof(struct bch_bucket_gens));
 630                 return -BCH_ERR_invalid_bkey;
 631         }
 632
 633         return 0;
 634 }
 635
 636 void bch2_bucket_gens_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 637 {
 638         struct bkey_s_c_bucket_gens g = bkey_s_c_to_bucket_gens(k);
 639         unsigned i;
 640
 641         for (i = 0; i < ARRAY_SIZE(g.v->gens); i++) {
 642                 if (i)
 643                         prt_char(out, ' ');
 644                 prt_printf(out, "%u", g.v->gens[i]);
 645         }
 646 }
 647
 648 int bch2_bucket_gens_init(struct bch_fs *c)
 649 {
 650         struct btree_trans trans;
 651         struct btree_iter iter;
 652         struct bkey_s_c k;
 653         struct bch_alloc_v4 a;
 654         struct bkey_i_bucket_gens g;
 655         bool have_bucket_gens_key = false;
 656         unsigned offset;
 657         struct bpos pos;
 658         u8 gen;
 659         int ret;
 660
 661         bch2_trans_init(&trans, c, 0, 0);
 662
 663         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 664                            BTREE_ITER_PREFETCH, k, ret) {
 665                 /*
 666                  * Not a fsck error because this is checked/repaired by
 667                  * bch2_check_alloc_key() which runs later:
 668                  */
 669                 if (!bch2_dev_bucket_exists(c, k.k->p))
 670                         continue;
 671
 672                 gen = bch2_alloc_to_v4(k, &a)->gen;
 673                 pos = alloc_gens_pos(iter.pos, &offset);
 674
 675                 if (have_bucket_gens_key && bkey_cmp(iter.pos, pos)) {
 676                         ret = commit_do(&trans, NULL, NULL,
 677                                         BTREE_INSERT_NOFAIL|
 678                                         BTREE_INSERT_LAZY_RW,
 679                                 __bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i, 0));
 680                         if (ret)
 681                                 break;
 682                         have_bucket_gens_key = false;
 683                 }
 684
 685                 if (!have_bucket_gens_key) {
 686                         bkey_bucket_gens_init(&g.k_i);
 687                         g.k.p = pos;
 688                         have_bucket_gens_key = true;
 689                 }
 690
 691                 g.v.gens[offset] = gen;
 692         }
 693         bch2_trans_iter_exit(&trans, &iter);
 694
 695         if (have_bucket_gens_key && !ret)
 696                 ret = commit_do(&trans, NULL, NULL,
 697                                 BTREE_INSERT_NOFAIL|
 698                                 BTREE_INSERT_LAZY_RW,
 699                         __bch2_btree_insert(&trans, BTREE_ID_bucket_gens, &g.k_i, 0));
 700
 701         bch2_trans_exit(&trans);
 702
 703         if (ret)
 704                 bch_err(c, "%s: error %s", __func__, bch2_err_str(ret));
 705
 706         return ret;
 707 }
 708
 709 int bch2_bucket_gens_read(struct bch_fs *c)
 710 {
 711         struct btree_trans trans;
 712         struct btree_iter iter;
 713         struct bkey_s_c k;
 714         const struct bch_bucket_gens *g;
 715         struct bch_dev *ca;
 716         u64 b;
 717         int ret;
 718
 719         bch2_trans_init(&trans, c, 0, 0);
 720
 721         for_each_btree_key(&trans, iter, BTREE_ID_bucket_gens, POS_MIN,
 722                            BTREE_ITER_PREFETCH, k, ret) {
 723                 u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
 724                 u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
 725
 726                 if (k.k->type != KEY_TYPE_bucket_gens)
 727                         continue;
 728
 729                 g = bkey_s_c_to_bucket_gens(k).v;
 730
 731                 /*
 732                  * Not a fsck error because this is checked/repaired by
 733                  * bch2_check_alloc_key() which runs later:
 734                  */
 735                 if (!bch2_dev_exists2(c, k.k->p.inode))
 736                         continue;
 737
 738                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 739
 740                 for (b = max_t(u64, ca->mi.first_bucket, start);
 741                      b < min_t(u64, ca->mi.nbuckets, end);
 742                      b++)
 743                         *bucket_gen(ca, b) = g->gens[b & KEY_TYPE_BUCKET_GENS_MASK];
 744         }
 745         bch2_trans_iter_exit(&trans, &iter);
 746
 747         bch2_trans_exit(&trans);
 748
 749         if (ret)
 750                 bch_err(c, "error reading alloc info: %s", bch2_err_str(ret));
 751
 752         return ret;
 753 }
 754
 755 /* Free space/discard btree: */
 756
 757 static int bch2_bucket_do_index(struct btree_trans *trans,
 758                                 struct bkey_s_c alloc_k,
 759                                 const struct bch_alloc_v4 *a,
 760                                 bool set)
 761 {
 762         struct bch_fs *c = trans->c;
 763         struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 764         struct btree_iter iter;
 765         struct bkey_s_c old;
 766         struct bkey_i *k;
 767         enum btree_id btree;
 768         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 769         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 770         struct printbuf buf = PRINTBUF;
 771         int ret;
 772
 773         if (a->data_type != BCH_DATA_free &&
 774             a->data_type != BCH_DATA_need_discard)
 775                 return 0;
 776
 777         k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
 778         if (IS_ERR(k))
 779                 return PTR_ERR(k);
 780
 781         bkey_init(&k->k);
 782         k->k.type = new_type;
 783
 784         switch (a->data_type) {
 785         case BCH_DATA_free:
 786                 btree = BTREE_ID_freespace;
 787                 k->k.p = alloc_freespace_pos(alloc_k.k->p, *a);
 788                 bch2_key_resize(&k->k, 1);
 789                 break;
 790         case BCH_DATA_need_discard:
 791                 btree = BTREE_ID_need_discard;
 792                 k->k.p = alloc_k.k->p;
 793                 break;
 794         default:
 795                 return 0;
 796         }
 797
 798         bch2_trans_iter_init(trans, &iter, btree,
 799                              bkey_start_pos(&k->k),
 800                              BTREE_ITER_INTENT);
 801         old = bch2_btree_iter_peek_slot(&iter);
 802         ret = bkey_err(old);
 803         if (ret)
 804                 goto err;
 805
 806         if (ca->mi.freespace_initialized &&
 807             test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags) &&
 808             bch2_trans_inconsistent_on(old.k->type != old_type, trans,
 809                         "incorrect key when %s %s btree (got %s should be %s)\n"
 810                         "  for %s",
 811                         set ? "setting" : "clearing",
 812                         bch2_btree_ids[btree],
 813                         bch2_bkey_types[old.k->type],
 814                         bch2_bkey_types[old_type],
 815                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 816                 ret = -EIO;
 817                 goto err;
 818         }
 819
 820         ret = bch2_trans_update(trans, &iter, k, 0);
 821 err:
 822         bch2_trans_iter_exit(trans, &iter);
 823         printbuf_exit(&buf);
 824         return ret;
 825 }
 826
 827 static noinline int bch2_bucket_gen_update(struct btree_trans *trans,
 828                                            struct bpos bucket, u8 gen)
 829 {
 830         struct btree_iter iter;
 831         unsigned offset;
 832         struct bpos pos = alloc_gens_pos(bucket, &offset);
 833         struct bkey_i_bucket_gens *g;
 834         struct bkey_s_c k;
 835         int ret;
 836
 837         g = bch2_trans_kmalloc(trans, sizeof(*g));
 838         ret = PTR_ERR_OR_ZERO(g);
 839         if (ret)
 840                 return ret;
 841
 842         bch2_trans_iter_init(trans, &iter, BTREE_ID_bucket_gens, pos,
 843                              BTREE_ITER_INTENT|
 844                              BTREE_ITER_WITH_UPDATES);
 845         k = bch2_btree_iter_peek_slot(&iter);
 846         ret = bkey_err(k);
 847         if (ret)
 848                 goto err;
 849
 850         if (k.k->type != KEY_TYPE_bucket_gens) {
 851                 bkey_bucket_gens_init(&g->k_i);
 852                 g->k.p = iter.pos;
 853         } else {
 854                 bkey_reassemble(&g->k_i, k);
 855         }
 856
 857         g->v.gens[offset] = gen;
 858
 859         ret = bch2_trans_update(trans, &iter, &g->k_i, 0);
 860 err:
 861         bch2_trans_iter_exit(trans, &iter);
 862         return ret;
 863 }
 864
 865 int bch2_trans_mark_alloc(struct btree_trans *trans,
 866                           enum btree_id btree_id, unsigned level,
 867                           struct bkey_s_c old, struct bkey_i *new,
 868                           unsigned flags)
 869 {
 870         struct bch_fs *c = trans->c;
 871         struct bch_alloc_v4 old_a_convert, *new_a;
 872         const struct bch_alloc_v4 *old_a;
 873         u64 old_lru, new_lru;
 874         int ret = 0;
 875
 876         /*
 877          * Deletion only happens in the device removal path, with
 878          * BTREE_TRIGGER_NORUN:
 879          */
 880         BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
 881
 882         old_a = bch2_alloc_to_v4(old, &old_a_convert);
 883         new_a = &bkey_i_to_alloc_v4(new)->v;
 884
 885         new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
 886
 887         if (new_a->dirty_sectors > old_a->dirty_sectors ||
 888             new_a->cached_sectors > old_a->cached_sectors) {
 889                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 890                 new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 891                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
 892                 SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
 893         }
 894
 895         if (data_type_is_empty(new_a->data_type) &&
 896             BCH_ALLOC_V4_NEED_INC_GEN(new_a) &&
 897             !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
 898                 new_a->gen++;
 899                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
 900         }
 901
 902         if (old_a->data_type != new_a->data_type ||
 903             (new_a->data_type == BCH_DATA_free &&
 904              alloc_freespace_genbits(*old_a) != alloc_freespace_genbits(*new_a))) {
 905                 ret =   bch2_bucket_do_index(trans, old, old_a, false) ?:
 906                         bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_a, true);
 907                 if (ret)
 908                         return ret;
 909         }
 910
 911         if (new_a->data_type == BCH_DATA_cached &&
 912             !new_a->io_time[READ])
 913                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 914
 915         old_lru = alloc_lru_idx_read(*old_a);
 916         new_lru = alloc_lru_idx_read(*new_a);
 917
 918         if (old_lru != new_lru) {
 919                 ret = bch2_lru_change(trans, new->k.p.inode,
 920                                       bucket_to_u64(new->k.p),
 921                                       old_lru, new_lru);
 922                 if (ret)
 923                         return ret;
 924         }
 925
 926         new_a->fragmentation_lru = alloc_lru_idx_fragmentation(*new_a,
 927                                         bch_dev_bkey_exists(c, new->k.p.inode));
 928
 929         if (old_a->fragmentation_lru != new_a->fragmentation_lru) {
 930                 ret = bch2_lru_change(trans,
 931                                 BCH_LRU_FRAGMENTATION_START,
 932                                 bucket_to_u64(new->k.p),
 933                                 old_a->fragmentation_lru, new_a->fragmentation_lru);
 934                 if (ret)
 935                         return ret;
 936         }
 937
 938         if (old_a->gen != new_a->gen) {
 939                 ret = bch2_bucket_gen_update(trans, new->k.p, new_a->gen);
 940                 if (ret)
 941                         return ret;
 942         }
 943
 944         return 0;
 945 }
 946
 947 /*
 948  * This synthesizes deleted extents for holes, similar to BTREE_ITER_SLOTS for
 949  * extents style btrees, but works on non-extents btrees:
 950  */
 951 struct bkey_s_c bch2_get_key_or_hole(struct btree_iter *iter, struct bpos end, struct bkey *hole)
 952 {
 953         struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
 954
 955         if (bkey_err(k))
 956                 return k;
 957
 958         if (k.k->type) {
 959                 return k;
 960         } else {
 961                 struct btree_iter iter2;
 962                 struct bpos next;
 963
 964                 bch2_trans_copy_iter(&iter2, iter);
 965                 k = bch2_btree_iter_peek_upto(&iter2,
 966                                 bkey_min(bkey_min(end,
 967                                                   iter->path->l[0].b->key.k.p),
 968                                                   POS(iter->pos.inode, iter->pos.offset + U32_MAX - 1)));
 969                 next = iter2.pos;
 970                 bch2_trans_iter_exit(iter->trans, &iter2);
 971
 972                 BUG_ON(next.offset >= iter->pos.offset + U32_MAX);
 973
 974                 if (bkey_err(k))
 975                         return k;
 976
 977                 bkey_init(hole);
 978                 hole->p = iter->pos;
 979
 980                 bch2_key_resize(hole, next.offset - iter->pos.offset);
 981                 return (struct bkey_s_c) { hole, NULL };
 982         }
 983 }
 984
 985 static bool next_bucket(struct bch_fs *c, struct bpos *bucket)
 986 {
 987         struct bch_dev *ca;
 988         unsigned iter;
 989
 990         if (bch2_dev_bucket_exists(c, *bucket))
 991                 return true;
 992
 993         if (bch2_dev_exists2(c, bucket->inode)) {
 994                 ca = bch_dev_bkey_exists(c, bucket->inode);
 995
 996                 if (bucket->offset < ca->mi.first_bucket) {
 997                         bucket->offset = ca->mi.first_bucket;
 998                         return true;
 999                 }
1000
1001                 bucket->inode++;
1002                 bucket->offset = 0;
1003         }
1004
1005         rcu_read_lock();
1006         iter = bucket->inode;
1007         ca = __bch2_next_dev(c, &iter, NULL);
1008         if (ca)
1009                 bucket->offset = ca->mi.first_bucket;
1010         rcu_read_unlock();
1011
1012         return ca != NULL;
1013 }
1014
1015 struct bkey_s_c bch2_get_key_or_real_bucket_hole(struct btree_iter *iter, struct bkey *hole)
1016 {
1017         struct bch_fs *c = iter->trans->c;
1018         struct bkey_s_c k;
1019 again:
1020         k = bch2_get_key_or_hole(iter, POS_MAX, hole);
1021         if (bkey_err(k))
1022                 return k;
1023
1024         if (!k.k->type) {
1025                 struct bpos bucket = bkey_start_pos(k.k);
1026
1027                 if (!bch2_dev_bucket_exists(c, bucket)) {
1028                         if (!next_bucket(c, &bucket))
1029                                 return bkey_s_c_null;
1030
1031                         bch2_btree_iter_set_pos(iter, bucket);
1032                         goto again;
1033                 }
1034
1035                 if (!bch2_dev_bucket_exists(c, k.k->p)) {
1036                         struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
1037
1038                         bch2_key_resize(hole, ca->mi.nbuckets - bucket.offset);
1039                 }
1040         }
1041
1042         return k;
1043 }
1044
1045 static int bch2_check_alloc_key(struct btree_trans *trans,
1046                                 struct bkey_s_c alloc_k,
1047                                 struct btree_iter *alloc_iter,
1048                                 struct btree_iter *discard_iter,
1049                                 struct btree_iter *freespace_iter,
1050                                 struct btree_iter *bucket_gens_iter)
1051 {
1052         struct bch_fs *c = trans->c;
1053         struct bch_dev *ca;
1054         struct bch_alloc_v4 a_convert;
1055         const struct bch_alloc_v4 *a;
1056         unsigned discard_key_type, freespace_key_type;
1057         unsigned gens_offset;
1058         struct bkey_s_c k;
1059         struct printbuf buf = PRINTBUF;
1060         int ret;
1061
1062         if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
1063                         "alloc key for invalid device:bucket %llu:%llu",
1064                         alloc_k.k->p.inode, alloc_k.k->p.offset))
1065                 return bch2_btree_delete_at(trans, alloc_iter, 0);
1066
1067         ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
1068         if (!ca->mi.freespace_initialized)
1069                 return 0;
1070
1071         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1072
1073         discard_key_type = a->data_type == BCH_DATA_need_discard ? KEY_TYPE_set : 0;
1074         bch2_btree_iter_set_pos(discard_iter, alloc_k.k->p);
1075         k = bch2_btree_iter_peek_slot(discard_iter);
1076         ret = bkey_err(k);
1077         if (ret)
1078                 goto err;
1079
1080         if (k.k->type != discard_key_type &&
1081             (c->opts.reconstruct_alloc ||
1082              fsck_err(c, "incorrect key in need_discard btree (got %s should be %s)\n"
1083                       "  %s",
1084                       bch2_bkey_types[k.k->type],
1085                       bch2_bkey_types[discard_key_type],
1086                       (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1087                 struct bkey_i *update =
1088                         bch2_trans_kmalloc(trans, sizeof(*update));
1089
1090                 ret = PTR_ERR_OR_ZERO(update);
1091                 if (ret)
1092                         goto err;
1093
1094                 bkey_init(&update->k);
1095                 update->k.type  = discard_key_type;
1096                 update->k.p     = discard_iter->pos;
1097
1098                 ret = bch2_trans_update(trans, discard_iter, update, 0);
1099                 if (ret)
1100                         goto err;
1101         }
1102
1103         freespace_key_type = a->data_type == BCH_DATA_free ? KEY_TYPE_set : 0;
1104         bch2_btree_iter_set_pos(freespace_iter, alloc_freespace_pos(alloc_k.k->p, *a));
1105         k = bch2_btree_iter_peek_slot(freespace_iter);
1106         ret = bkey_err(k);
1107         if (ret)
1108                 goto err;
1109
1110         if (k.k->type != freespace_key_type &&
1111             (c->opts.reconstruct_alloc ||
1112              fsck_err(c, "incorrect key in freespace btree (got %s should be %s)\n"
1113                       "  %s",
1114                       bch2_bkey_types[k.k->type],
1115                       bch2_bkey_types[freespace_key_type],
1116                       (printbuf_reset(&buf),
1117                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1118                 struct bkey_i *update =
1119                         bch2_trans_kmalloc(trans, sizeof(*update));
1120
1121                 ret = PTR_ERR_OR_ZERO(update);
1122                 if (ret)
1123                         goto err;
1124
1125                 bkey_init(&update->k);
1126                 update->k.type  = freespace_key_type;
1127                 update->k.p     = freespace_iter->pos;
1128                 bch2_key_resize(&update->k, 1);
1129
1130                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1131                 if (ret)
1132                         goto err;
1133         }
1134
1135         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(alloc_k.k->p, &gens_offset));
1136         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1137         ret = bkey_err(k);
1138         if (ret)
1139                 goto err;
1140
1141         if (a->gen != alloc_gen(k, gens_offset) &&
1142             (c->opts.reconstruct_alloc ||
1143              fsck_err(c, "incorrect gen in bucket_gens btree (got %u should be %u)\n"
1144                       "  %s",
1145                       alloc_gen(k, gens_offset), a->gen,
1146                       (printbuf_reset(&buf),
1147                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1148                 struct bkey_i_bucket_gens *g =
1149                         bch2_trans_kmalloc(trans, sizeof(*g));
1150
1151                 ret = PTR_ERR_OR_ZERO(g);
1152                 if (ret)
1153                         goto err;
1154
1155                 if (k.k->type == KEY_TYPE_bucket_gens) {
1156                         bkey_reassemble(&g->k_i, k);
1157                 } else {
1158                         bkey_bucket_gens_init(&g->k_i);
1159                         g->k.p = alloc_gens_pos(alloc_k.k->p, &gens_offset);
1160                 }
1161
1162                 g->v.gens[gens_offset] = a->gen;
1163
1164                 ret = bch2_trans_update(trans, bucket_gens_iter, &g->k_i, 0);
1165                 if (ret)
1166                         goto err;
1167         }
1168 err:
1169 fsck_err:
1170         printbuf_exit(&buf);
1171         return ret;
1172 }
1173
1174 static int bch2_check_alloc_hole_freespace(struct btree_trans *trans,
1175                                  struct bpos start,
1176                                  struct bpos *end,
1177                                  struct btree_iter *freespace_iter)
1178 {
1179         struct bch_fs *c = trans->c;
1180         struct bch_dev *ca;
1181         struct bkey_s_c k;
1182         struct printbuf buf = PRINTBUF;
1183         int ret;
1184
1185         ca = bch_dev_bkey_exists(c, start.inode);
1186         if (!ca->mi.freespace_initialized)
1187                 return 0;
1188
1189         bch2_btree_iter_set_pos(freespace_iter, start);
1190
1191         k = bch2_btree_iter_peek_slot(freespace_iter);
1192         ret = bkey_err(k);
1193         if (ret)
1194                 goto err;
1195
1196         *end = bkey_min(k.k->p, *end);
1197
1198         if (k.k->type != KEY_TYPE_set &&
1199             (c->opts.reconstruct_alloc ||
1200              fsck_err(c, "hole in alloc btree missing in freespace btree\n"
1201                       "  device %llu buckets %llu-%llu",
1202                       freespace_iter->pos.inode,
1203                       freespace_iter->pos.offset,
1204                       end->offset))) {
1205                 struct bkey_i *update =
1206                         bch2_trans_kmalloc(trans, sizeof(*update));
1207
1208                 ret = PTR_ERR_OR_ZERO(update);
1209                 if (ret)
1210                         goto err;
1211
1212                 bkey_init(&update->k);
1213                 update->k.type  = KEY_TYPE_set;
1214                 update->k.p     = freespace_iter->pos;
1215                 bch2_key_resize(&update->k,
1216                                 min_t(u64, U32_MAX, end->offset -
1217                                       freespace_iter->pos.offset));
1218
1219                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1220                 if (ret)
1221                         goto err;
1222         }
1223 err:
1224 fsck_err:
1225         printbuf_exit(&buf);
1226         return ret;
1227 }
1228
1229 static int bch2_check_alloc_hole_bucket_gens(struct btree_trans *trans,
1230                                  struct bpos start,
1231                                  struct bpos *end,
1232                                  struct btree_iter *bucket_gens_iter)
1233 {
1234         struct bch_fs *c = trans->c;
1235         struct bkey_s_c k;
1236         struct printbuf buf = PRINTBUF;
1237         unsigned i, gens_offset, gens_end_offset;
1238         int ret;
1239
1240         if (c->sb.version < bcachefs_metadata_version_bucket_gens &&
1241             !c->opts.version_upgrade)
1242                 return 0;
1243
1244         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(start, &gens_offset));
1245
1246         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1247         ret = bkey_err(k);
1248         if (ret)
1249                 goto err;
1250
1251         if (bkey_cmp(alloc_gens_pos(start, &gens_offset),
1252                      alloc_gens_pos(*end,  &gens_end_offset)))
1253                 gens_end_offset = KEY_TYPE_BUCKET_GENS_NR;
1254
1255         if (k.k->type == KEY_TYPE_bucket_gens) {
1256                 struct bkey_i_bucket_gens g;
1257                 bool need_update = false;
1258
1259                 bkey_reassemble(&g.k_i, k);
1260
1261                 for (i = gens_offset; i < gens_end_offset; i++) {
1262                         if (fsck_err_on(g.v.gens[i], c,
1263                                         "hole in alloc btree at %llu:%llu with nonzero gen in bucket_gens btree (%u)",
1264                                         bucket_gens_pos_to_alloc(k.k->p, i).inode,
1265                                         bucket_gens_pos_to_alloc(k.k->p, i).offset,
1266                                         g.v.gens[i])) {
1267                                 g.v.gens[i] = 0;
1268                                 need_update = true;
1269                         }
1270                 }
1271
1272                 if (need_update) {
1273                         struct bkey_i *k = bch2_trans_kmalloc(trans, sizeof(g));
1274
1275                         ret = PTR_ERR_OR_ZERO(k);
1276                         if (ret)
1277                                 goto err;
1278
1279                         memcpy(k, &g, sizeof(g));
1280
1281                         ret = bch2_trans_update(trans, bucket_gens_iter, k, 0);
1282                         if (ret)
1283                                 goto err;
1284                 }
1285         }
1286
1287         *end = bkey_min(*end, bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0));
1288 err:
1289 fsck_err:
1290         printbuf_exit(&buf);
1291         return ret;
1292 }
1293
1294 static int bch2_check_discard_freespace_key(struct btree_trans *trans,
1295                                             struct btree_iter *iter)
1296 {
1297         struct bch_fs *c = trans->c;
1298         struct btree_iter alloc_iter;
1299         struct bkey_s_c alloc_k;
1300         struct bch_alloc_v4 a_convert;
1301         const struct bch_alloc_v4 *a;
1302         u64 genbits;
1303         struct bpos pos;
1304         enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
1305                 ? BCH_DATA_need_discard
1306                 : BCH_DATA_free;
1307         struct printbuf buf = PRINTBUF;
1308         int ret;
1309
1310         pos = iter->pos;
1311         pos.offset &= ~(~0ULL << 56);
1312         genbits = iter->pos.offset & (~0ULL << 56);
1313
1314         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
1315
1316         if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
1317                         "entry in %s btree for nonexistant dev:bucket %llu:%llu",
1318                         bch2_btree_ids[iter->btree_id], pos.inode, pos.offset))
1319                 goto delete;
1320
1321         alloc_k = bch2_btree_iter_peek_slot(&alloc_iter);
1322         ret = bkey_err(alloc_k);
1323         if (ret)
1324                 goto err;
1325
1326         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1327
1328         if (fsck_err_on(a->data_type != state ||
1329                         (state == BCH_DATA_free &&
1330                          genbits != alloc_freespace_genbits(*a)), c,
1331                         "%s\n  incorrectly set in %s index (free %u, genbits %llu should be %llu)",
1332                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
1333                         bch2_btree_ids[iter->btree_id],
1334                         a->data_type == state,
1335                         genbits >> 56, alloc_freespace_genbits(*a) >> 56))
1336                 goto delete;
1337 out:
1338 err:
1339 fsck_err:
1340         bch2_trans_iter_exit(trans, &alloc_iter);
1341         printbuf_exit(&buf);
1342         return ret;
1343 delete:
1344         ret = bch2_btree_delete_extent_at(trans, iter,
1345                         iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0);
1346         goto out;
1347 }
1348
1349 /*
1350  * We've already checked that generation numbers in the bucket_gens btree are
1351  * valid for buckets that exist; this just checks for keys for nonexistent
1352  * buckets.
1353  */
1354 static int bch2_check_bucket_gens_key(struct btree_trans *trans,
1355                                       struct btree_iter *iter,
1356                                       struct bkey_s_c k)
1357 {
1358         struct bch_fs *c = trans->c;
1359         struct bkey_i_bucket_gens g;
1360         struct bch_dev *ca;
1361         u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
1362         u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
1363         u64 b;
1364         bool need_update = false;
1365         struct printbuf buf = PRINTBUF;
1366         int ret = 0;
1367
1368         BUG_ON(k.k->type != KEY_TYPE_bucket_gens);
1369         bkey_reassemble(&g.k_i, k);
1370
1371         if (fsck_err_on(!bch2_dev_exists2(c, k.k->p.inode), c,
1372                         "bucket_gens key for invalid device:\n  %s",
1373                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1374                 ret = bch2_btree_delete_at(trans, iter, 0);
1375                 goto out;
1376         }
1377
1378         ca = bch_dev_bkey_exists(c, k.k->p.inode);
1379         if (fsck_err_on(end <= ca->mi.first_bucket ||
1380                         start >= ca->mi.nbuckets, c,
1381                         "bucket_gens key for invalid buckets:\n  %s",
1382                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1383                 ret = bch2_btree_delete_at(trans, iter, 0);
1384                 goto out;
1385         }
1386
1387         for (b = start; b < ca->mi.first_bucket; b++)
1388                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1389                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1390                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1391                         need_update = true;
1392                 }
1393
1394         for (b = ca->mi.nbuckets; b < end; b++)
1395                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1396                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1397                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1398                         need_update = true;
1399                 }
1400
1401         if (need_update) {
1402                 struct bkey_i *k;
1403
1404                 k = bch2_trans_kmalloc(trans, sizeof(g));
1405                 ret = PTR_ERR_OR_ZERO(k);
1406                 if (ret)
1407                         goto out;
1408
1409                 memcpy(k, &g, sizeof(g));
1410                 ret = bch2_trans_update(trans, iter, k, 0);
1411         }
1412 out:
1413 fsck_err:
1414         printbuf_exit(&buf);
1415         return ret;
1416 }
1417
1418 int bch2_check_alloc_info(struct bch_fs *c)
1419 {
1420         struct btree_trans trans;
1421         struct btree_iter iter, discard_iter, freespace_iter, bucket_gens_iter;
1422         struct bkey hole;
1423         struct bkey_s_c k;
1424         int ret = 0;
1425
1426         bch2_trans_init(&trans, c, 0, 0);
1427
1428         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc, POS_MIN,
1429                              BTREE_ITER_PREFETCH);
1430         bch2_trans_iter_init(&trans, &discard_iter, BTREE_ID_need_discard, POS_MIN,
1431                              BTREE_ITER_PREFETCH);
1432         bch2_trans_iter_init(&trans, &freespace_iter, BTREE_ID_freespace, POS_MIN,
1433                              BTREE_ITER_PREFETCH);
1434         bch2_trans_iter_init(&trans, &bucket_gens_iter, BTREE_ID_bucket_gens, POS_MIN,
1435                              BTREE_ITER_PREFETCH);
1436
1437         while (1) {
1438                 struct bpos next;
1439
1440                 bch2_trans_begin(&trans);
1441
1442                 k = bch2_get_key_or_real_bucket_hole(&iter, &hole);
1443                 ret = bkey_err(k);
1444                 if (ret)
1445                         goto bkey_err;
1446
1447                 if (!k.k)
1448                         break;
1449
1450                 if (k.k->type) {
1451                         next = bpos_nosnap_successor(k.k->p);
1452
1453                         ret = bch2_check_alloc_key(&trans,
1454                                                    k, &iter,
1455                                                    &discard_iter,
1456                                                    &freespace_iter,
1457                                                    &bucket_gens_iter);
1458                         if (ret)
1459                                 goto bkey_err;
1460                 } else {
1461                         next = k.k->p;
1462
1463                         ret = bch2_check_alloc_hole_freespace(&trans,
1464                                                     bkey_start_pos(k.k),
1465                                                     &next,
1466                                                     &freespace_iter) ?:
1467                                 bch2_check_alloc_hole_bucket_gens(&trans,
1468                                                     bkey_start_pos(k.k),
1469                                                     &next,
1470                                                     &bucket_gens_iter);
1471                         if (ret)
1472                                 goto bkey_err;
1473                 }
1474
1475                 ret = bch2_trans_commit(&trans, NULL, NULL,
1476                                         BTREE_INSERT_NOFAIL|
1477                                         BTREE_INSERT_LAZY_RW);
1478                 if (ret)
1479                         goto bkey_err;
1480
1481                 bch2_btree_iter_set_pos(&iter, next);
1482 bkey_err:
1483                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1484                         continue;
1485                 if (ret)
1486                         break;
1487         }
1488         bch2_trans_iter_exit(&trans, &bucket_gens_iter);
1489         bch2_trans_iter_exit(&trans, &freespace_iter);
1490         bch2_trans_iter_exit(&trans, &discard_iter);
1491         bch2_trans_iter_exit(&trans, &iter);
1492
1493         if (ret < 0)
1494                 goto err;
1495
1496         ret = for_each_btree_key_commit(&trans, iter,
1497                         BTREE_ID_need_discard, POS_MIN,
1498                         BTREE_ITER_PREFETCH, k,
1499                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1500                 bch2_check_discard_freespace_key(&trans, &iter)) ?:
1501               for_each_btree_key_commit(&trans, iter,
1502                         BTREE_ID_freespace, POS_MIN,
1503                         BTREE_ITER_PREFETCH, k,
1504                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1505                 bch2_check_discard_freespace_key(&trans, &iter)) ?:
1506               for_each_btree_key_commit(&trans, iter,
1507                         BTREE_ID_bucket_gens, POS_MIN,
1508                         BTREE_ITER_PREFETCH, k,
1509                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1510                 bch2_check_bucket_gens_key(&trans, &iter, k));
1511 err:
1512         bch2_trans_exit(&trans);
1513         return ret < 0 ? ret : 0;
1514 }
1515
1516 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
1517                                        struct btree_iter *alloc_iter)
1518 {
1519         struct bch_fs *c = trans->c;
1520         struct btree_iter lru_iter;
1521         struct bch_alloc_v4 a_convert;
1522         const struct bch_alloc_v4 *a;
1523         struct bkey_s_c alloc_k, k;
1524         struct printbuf buf = PRINTBUF;
1525         int ret;
1526
1527         alloc_k = bch2_btree_iter_peek(alloc_iter);
1528         if (!alloc_k.k)
1529                 return 0;
1530
1531         ret = bkey_err(alloc_k);
1532         if (ret)
1533                 return ret;
1534
1535         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1536
1537         if (a->data_type != BCH_DATA_cached)
1538                 return 0;
1539
1540         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
1541                              lru_pos(alloc_k.k->p.inode,
1542                                      bucket_to_u64(alloc_k.k->p),
1543                                      a->io_time[READ]), 0);
1544         k = bch2_btree_iter_peek_slot(&lru_iter);
1545         ret = bkey_err(k);
1546         if (ret)
1547                 goto err;
1548
1549         if (fsck_err_on(!a->io_time[READ], c,
1550                         "cached bucket with read_time 0\n"
1551                         "  %s",
1552                 (printbuf_reset(&buf),
1553                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) ||
1554             fsck_err_on(k.k->type != KEY_TYPE_set, c,
1555                         "missing lru entry\n"
1556                         "  %s",
1557                         (printbuf_reset(&buf),
1558                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
1559                 u64 read_time = a->io_time[READ] ?:
1560                         atomic64_read(&c->io_clock[READ].now);
1561
1562                 ret = bch2_lru_set(trans,
1563                                    alloc_k.k->p.inode,
1564                                    bucket_to_u64(alloc_k.k->p),
1565                                    read_time);
1566                 if (ret)
1567                         goto err;
1568
1569                 if (a->io_time[READ] != read_time) {
1570                         struct bkey_i_alloc_v4 *a_mut =
1571                                 bch2_alloc_to_v4_mut(trans, alloc_k);
1572                         ret = PTR_ERR_OR_ZERO(a_mut);
1573                         if (ret)
1574                                 goto err;
1575
1576                         a_mut->v.io_time[READ] = read_time;
1577                         ret = bch2_trans_update(trans, alloc_iter,
1578                                                 &a_mut->k_i, BTREE_TRIGGER_NORUN);
1579                         if (ret)
1580                                 goto err;
1581                 }
1582         }
1583 err:
1584 fsck_err:
1585         bch2_trans_iter_exit(trans, &lru_iter);
1586         printbuf_exit(&buf);
1587         return ret;
1588 }
1589
1590 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
1591 {
1592         struct btree_trans trans;
1593         struct btree_iter iter;
1594         struct bkey_s_c k;
1595         int ret = 0;
1596
1597         bch2_trans_init(&trans, c, 0, 0);
1598
1599         for_each_btree_key_commit(&trans, iter, BTREE_ID_alloc,
1600                         POS_MIN, BTREE_ITER_PREFETCH, k,
1601                         NULL, NULL, BTREE_INSERT_NOFAIL|BTREE_INSERT_LAZY_RW,
1602                 bch2_check_alloc_to_lru_ref(&trans, &iter));
1603
1604         bch2_trans_exit(&trans);
1605         return ret < 0 ? ret : 0;
1606 }
1607
1608 static int bch2_discard_one_bucket(struct btree_trans *trans,
1609                                    struct btree_iter *need_discard_iter,
1610                                    struct bpos *discard_pos_done,
1611                                    u64 *seen,
1612                                    u64 *open,
1613                                    u64 *need_journal_commit,
1614                                    u64 *discarded)
1615 {
1616         struct bch_fs *c = trans->c;
1617         struct bpos pos = need_discard_iter->pos;
1618         struct btree_iter iter = { NULL };
1619         struct bkey_s_c k;
1620         struct bch_dev *ca;
1621         struct bkey_i_alloc_v4 *a;
1622         struct printbuf buf = PRINTBUF;
1623         int ret = 0;
1624
1625         ca = bch_dev_bkey_exists(c, pos.inode);
1626         if (!percpu_ref_tryget(&ca->io_ref)) {
1627                 bch2_btree_iter_set_pos(need_discard_iter, POS(pos.inode + 1, 0));
1628                 return 0;
1629         }
1630
1631         if (bch2_bucket_is_open_safe(c, pos.inode, pos.offset)) {
1632                 (*open)++;
1633                 goto out;
1634         }
1635
1636         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
1637                         c->journal.flushed_seq_ondisk,
1638                         pos.inode, pos.offset)) {
1639                 (*need_journal_commit)++;
1640                 goto out;
1641         }
1642
1643         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
1644                              need_discard_iter->pos,
1645                              BTREE_ITER_CACHED);
1646         k = bch2_btree_iter_peek_slot(&iter);
1647         ret = bkey_err(k);
1648         if (ret)
1649                 goto out;
1650
1651         a = bch2_alloc_to_v4_mut(trans, k);
1652         ret = PTR_ERR_OR_ZERO(a);
1653         if (ret)
1654                 goto out;
1655
1656         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
1657                 a->v.gen++;
1658                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1659                 goto write;
1660         }
1661
1662         if (a->v.journal_seq > c->journal.flushed_seq_ondisk) {
1663                 if (test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
1664                         bch2_trans_inconsistent(trans,
1665                                 "clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
1666                                 "%s",
1667                                 a->v.journal_seq,
1668                                 c->journal.flushed_seq_ondisk,
1669                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1670                         ret = -EIO;
1671                 }
1672                 goto out;
1673         }
1674
1675         if (a->v.data_type != BCH_DATA_need_discard) {
1676                 if (test_bit(BCH_FS_CHECK_ALLOC_DONE, &c->flags)) {
1677                         bch2_trans_inconsistent(trans,
1678                                 "bucket incorrectly set in need_discard btree\n"
1679                                 "%s",
1680                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1681                         ret = -EIO;
1682                 }
1683
1684                 goto out;
1685         }
1686
1687         if (!bkey_eq(*discard_pos_done, iter.pos) &&
1688             ca->mi.discard && !c->opts.nochanges) {
1689                 /*
1690                  * This works without any other locks because this is the only
1691                  * thread that removes items from the need_discard tree
1692                  */
1693                 bch2_trans_unlock(trans);
1694                 blkdev_issue_discard(ca->disk_sb.bdev,
1695                                      k.k->p.offset * ca->mi.bucket_size,
1696                                      ca->mi.bucket_size,
1697                                      GFP_KERNEL);
1698                 *discard_pos_done = iter.pos;
1699
1700                 ret = bch2_trans_relock_notrace(trans);
1701                 if (ret)
1702                         goto out;
1703         }
1704
1705         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
1706         a->v.data_type = alloc_data_type(a->v, a->v.data_type);
1707 write:
1708         ret =   bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1709                 bch2_trans_commit(trans, NULL, NULL,
1710                                   BTREE_INSERT_USE_RESERVE|BTREE_INSERT_NOFAIL);
1711         if (ret)
1712                 goto out;
1713
1714         this_cpu_inc(c->counters[BCH_COUNTER_bucket_discard]);
1715         (*discarded)++;
1716 out:
1717         (*seen)++;
1718         bch2_trans_iter_exit(trans, &iter);
1719         percpu_ref_put(&ca->io_ref);
1720         printbuf_exit(&buf);
1721         return ret;
1722 }
1723
1724 static void bch2_do_discards_work(struct work_struct *work)
1725 {
1726         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
1727         struct btree_trans trans;
1728         struct btree_iter iter;
1729         struct bkey_s_c k;
1730         u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
1731         struct bpos discard_pos_done = POS_MAX;
1732         int ret;
1733
1734         bch2_trans_init(&trans, c, 0, 0);
1735
1736         /*
1737          * We're doing the commit in bch2_discard_one_bucket instead of using
1738          * for_each_btree_key_commit() so that we can increment counters after
1739          * successful commit:
1740          */
1741         ret = for_each_btree_key2(&trans, iter,
1742                         BTREE_ID_need_discard, POS_MIN, 0, k,
1743                 bch2_discard_one_bucket(&trans, &iter, &discard_pos_done,
1744                                         &seen,
1745                                         &open,
1746                                         &need_journal_commit,
1747                                         &discarded));
1748
1749         bch2_trans_exit(&trans);
1750
1751         if (need_journal_commit * 2 > seen)
1752                 bch2_journal_flush_async(&c->journal, NULL);
1753
1754         bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1755
1756         trace_discard_buckets(c, seen, open, need_journal_commit, discarded,
1757                               bch2_err_str(ret));
1758 }
1759
1760 void bch2_do_discards(struct bch_fs *c)
1761 {
1762         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_discard) &&
1763             !queue_work(system_long_wq, &c->discard_work))
1764                 bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1765 }
1766
1767 static int invalidate_one_bucket(struct btree_trans *trans,
1768                                  struct btree_iter *lru_iter,
1769                                  struct bkey_s_c lru_k,
1770                                  s64 *nr_to_invalidate)
1771 {
1772         struct bch_fs *c = trans->c;
1773         struct btree_iter alloc_iter = { NULL };
1774         struct bkey_i_alloc_v4 *a = NULL;
1775         struct printbuf buf = PRINTBUF;
1776         struct bpos bucket = u64_to_bucket(lru_k.k->p.offset);
1777         unsigned cached_sectors;
1778         int ret = 0;
1779
1780         if (*nr_to_invalidate <= 0)
1781                 return 1;
1782
1783         if (!bch2_dev_bucket_exists(c, bucket)) {
1784                 prt_str(&buf, "lru entry points to invalid bucket");
1785                 goto err;
1786         }
1787
1788         if (bch2_bucket_is_open_safe(c, bucket.inode, bucket.offset))
1789                 return 0;
1790
1791         a = bch2_trans_start_alloc_update(trans, &alloc_iter, bucket);
1792         ret = PTR_ERR_OR_ZERO(a);
1793         if (ret)
1794                 goto out;
1795
1796         /* We expect harmless races here due to the btree write buffer: */
1797         if (lru_pos_time(lru_iter->pos) != alloc_lru_idx_read(a->v))
1798                 goto out;
1799
1800         BUG_ON(a->v.data_type != BCH_DATA_cached);
1801
1802         if (!a->v.cached_sectors)
1803                 bch_err(c, "invalidating empty bucket, confused");
1804
1805         cached_sectors = a->v.cached_sectors;
1806
1807         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1808         a->v.gen++;
1809         a->v.data_type          = 0;
1810         a->v.dirty_sectors      = 0;
1811         a->v.cached_sectors     = 0;
1812         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1813         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1814
1815         ret =   bch2_trans_update(trans, &alloc_iter, &a->k_i,
1816                                 BTREE_TRIGGER_BUCKET_INVALIDATE) ?:
1817                 bch2_trans_commit(trans, NULL, NULL,
1818                                   BTREE_INSERT_USE_RESERVE|BTREE_INSERT_NOFAIL);
1819         if (ret)
1820                 goto out;
1821
1822         trace_and_count(c, bucket_invalidate, c, bucket.inode, bucket.offset, cached_sectors);
1823         --*nr_to_invalidate;
1824 out:
1825         bch2_trans_iter_exit(trans, &alloc_iter);
1826         printbuf_exit(&buf);
1827         return ret;
1828 err:
1829         prt_str(&buf, "\n  lru key: ");
1830         bch2_bkey_val_to_text(&buf, c, lru_k);
1831
1832         prt_str(&buf, "\n  lru entry: ");
1833         bch2_lru_pos_to_text(&buf, lru_iter->pos);
1834
1835         prt_str(&buf, "\n  alloc key: ");
1836         if (!a)
1837                 bch2_bpos_to_text(&buf, bucket);
1838         else
1839                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
1840
1841         bch_err(c, "%s", buf.buf);
1842         if (test_bit(BCH_FS_CHECK_LRUS_DONE, &c->flags)) {
1843                 bch2_inconsistent_error(c);
1844                 ret = -EINVAL;
1845         }
1846
1847         goto out;
1848 }
1849
1850 static void bch2_do_invalidates_work(struct work_struct *work)
1851 {
1852         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1853         struct bch_dev *ca;
1854         struct btree_trans trans;
1855         struct btree_iter iter;
1856         struct bkey_s_c k;
1857         unsigned i;
1858         int ret = 0;
1859
1860         bch2_trans_init(&trans, c, 0, 0);
1861
1862         ret = bch2_btree_write_buffer_flush(&trans);
1863         if (ret)
1864                 goto err;
1865
1866         for_each_member_device(ca, c, i) {
1867                 s64 nr_to_invalidate =
1868                         should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
1869
1870                 ret = for_each_btree_key2_upto(&trans, iter, BTREE_ID_lru,
1871                                 lru_pos(ca->dev_idx, 0, 0),
1872                                 lru_pos(ca->dev_idx, U64_MAX, LRU_TIME_MAX),
1873                                 BTREE_ITER_INTENT, k,
1874                         invalidate_one_bucket(&trans, &iter, k, &nr_to_invalidate));
1875
1876                 if (ret < 0) {
1877                         percpu_ref_put(&ca->ref);
1878                         break;
1879                 }
1880         }
1881 err:
1882         bch2_trans_exit(&trans);
1883         bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1884 }
1885
1886 void bch2_do_invalidates(struct bch_fs *c)
1887 {
1888         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_invalidate) &&
1889             !queue_work(system_long_wq, &c->invalidate_work))
1890                 bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
1891 }
1892
1893 static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca)
1894 {
1895         struct btree_trans trans;
1896         struct btree_iter iter;
1897         struct bkey_s_c k;
1898         struct bkey hole;
1899         struct bpos end = POS(ca->dev_idx, ca->mi.nbuckets);
1900         struct bch_member *m;
1901         int ret;
1902
1903         bch2_trans_init(&trans, c, 0, 0);
1904
1905         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc,
1906                              POS(ca->dev_idx, ca->mi.first_bucket),
1907                              BTREE_ITER_PREFETCH);
1908         /*
1909          * Scan the alloc btree for every bucket on @ca, and add buckets to the
1910          * freespace/need_discard/need_gc_gens btrees as needed:
1911          */
1912         while (1) {
1913                 bch2_trans_begin(&trans);
1914
1915                 if (bkey_ge(iter.pos, end)) {
1916                         ret = 0;
1917                         break;
1918                 }
1919
1920                 k = bch2_get_key_or_hole(&iter, end, &hole);
1921                 ret = bkey_err(k);
1922                 if (ret)
1923                         goto bkey_err;
1924
1925                 if (k.k->type) {
1926                         /*
1927                          * We process live keys in the alloc btree one at a
1928                          * time:
1929                          */
1930                         struct bch_alloc_v4 a_convert;
1931                         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
1932
1933                         ret =   bch2_bucket_do_index(&trans, k, a, true) ?:
1934                                 bch2_trans_commit(&trans, NULL, NULL,
1935                                                   BTREE_INSERT_LAZY_RW|
1936                                                   BTREE_INSERT_NOFAIL);
1937                         if (ret)
1938                                 goto bkey_err;
1939
1940                         bch2_btree_iter_advance(&iter);
1941                 } else {
1942                         struct bkey_i *freespace;
1943
1944                         freespace = bch2_trans_kmalloc(&trans, sizeof(*freespace));
1945                         ret = PTR_ERR_OR_ZERO(freespace);
1946                         if (ret)
1947                                 goto bkey_err;
1948
1949                         bkey_init(&freespace->k);
1950                         freespace->k.type       = KEY_TYPE_set;
1951                         freespace->k.p          = k.k->p;
1952                         freespace->k.size       = k.k->size;
1953
1954                         ret = __bch2_btree_insert(&trans, BTREE_ID_freespace, freespace, 0) ?:
1955                                 bch2_trans_commit(&trans, NULL, NULL,
1956                                                   BTREE_INSERT_LAZY_RW|
1957                                                   BTREE_INSERT_NOFAIL);
1958                         if (ret)
1959                                 goto bkey_err;
1960
1961                         bch2_btree_iter_set_pos(&iter, k.k->p);
1962                 }
1963 bkey_err:
1964                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1965                         continue;
1966                 if (ret)
1967                         break;
1968         }
1969
1970         bch2_trans_iter_exit(&trans, &iter);
1971         bch2_trans_exit(&trans);
1972
1973         if (ret < 0) {
1974                 bch_err(ca, "error initializing free space: %s", bch2_err_str(ret));
1975                 return ret;
1976         }
1977
1978         mutex_lock(&c->sb_lock);
1979         m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx;
1980         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1981         mutex_unlock(&c->sb_lock);
1982
1983         return 0;
1984 }
1985
1986 int bch2_fs_freespace_init(struct bch_fs *c)
1987 {
1988         struct bch_dev *ca;
1989         unsigned i;
1990         int ret = 0;
1991         bool doing_init = false;
1992
1993         /*
1994          * We can crash during the device add path, so we need to check this on
1995          * every mount:
1996          */
1997
1998         for_each_member_device(ca, c, i) {
1999                 if (ca->mi.freespace_initialized)
2000                         continue;
2001
2002                 if (!doing_init) {
2003                         bch_info(c, "initializing freespace");
2004                         doing_init = true;
2005                 }
2006
2007                 ret = bch2_dev_freespace_init(c, ca);
2008                 if (ret) {
2009                         percpu_ref_put(&ca->ref);
2010                         return ret;
2011                 }
2012         }
2013
2014         if (doing_init) {
2015                 mutex_lock(&c->sb_lock);
2016                 bch2_write_super(c);
2017                 mutex_unlock(&c->sb_lock);
2018
2019                 bch_verbose(c, "done initializing freespace");
2020         }
2021
2022         return ret;
2023 }
2024
2025 /* Bucket IO clocks: */
2026
2027 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
2028                               size_t bucket_nr, int rw)
2029 {
2030         struct bch_fs *c = trans->c;
2031         struct btree_iter iter;
2032         struct bkey_i_alloc_v4 *a;
2033         u64 now;
2034         int ret = 0;
2035
2036         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
2037         ret = PTR_ERR_OR_ZERO(a);
2038         if (ret)
2039                 return ret;
2040
2041         now = atomic64_read(&c->io_clock[rw].now);
2042         if (a->v.io_time[rw] == now)
2043                 goto out;
2044
2045         a->v.io_time[rw] = now;
2046
2047         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
2048                 bch2_trans_commit(trans, NULL, NULL, 0);
2049 out:
2050         bch2_trans_iter_exit(trans, &iter);
2051         return ret;
2052 }
2053
2054 /* Startup/shutdown (ro/rw): */
2055
2056 void bch2_recalc_capacity(struct bch_fs *c)
2057 {
2058         struct bch_dev *ca;
2059         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
2060         unsigned bucket_size_max = 0;
2061         unsigned long ra_pages = 0;
2062         unsigned i;
2063
2064         lockdep_assert_held(&c->state_lock);
2065
2066         for_each_online_member(ca, c, i) {
2067                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
2068
2069                 ra_pages += bdi->ra_pages;
2070         }
2071
2072         bch2_set_ra_pages(c, ra_pages);
2073
2074         for_each_rw_member(ca, c, i) {
2075                 u64 dev_reserve = 0;
2076
2077                 /*
2078                  * We need to reserve buckets (from the number
2079                  * of currently available buckets) against
2080                  * foreground writes so that mainly copygc can
2081                  * make forward progress.
2082                  *
2083                  * We need enough to refill the various reserves
2084                  * from scratch - copygc will use its entire
2085                  * reserve all at once, then run against when
2086                  * its reserve is refilled (from the formerly
2087                  * available buckets).
2088                  *
2089                  * This reserve is just used when considering if
2090                  * allocations for foreground writes must wait -
2091                  * not -ENOSPC calculations.
2092                  */
2093
2094                 dev_reserve += ca->nr_btree_reserve * 2;
2095                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
2096
2097                 dev_reserve += 1;       /* btree write point */
2098                 dev_reserve += 1;       /* copygc write point */
2099                 dev_reserve += 1;       /* rebalance write point */
2100
2101                 dev_reserve *= ca->mi.bucket_size;
2102
2103                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
2104                                              ca->mi.first_bucket);
2105
2106                 reserved_sectors += dev_reserve * 2;
2107
2108                 bucket_size_max = max_t(unsigned, bucket_size_max,
2109                                         ca->mi.bucket_size);
2110         }
2111
2112         gc_reserve = c->opts.gc_reserve_bytes
2113                 ? c->opts.gc_reserve_bytes >> 9
2114                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
2115
2116         reserved_sectors = max(gc_reserve, reserved_sectors);
2117
2118         reserved_sectors = min(reserved_sectors, capacity);
2119
2120         c->capacity = capacity - reserved_sectors;
2121
2122         c->bucket_size_max = bucket_size_max;
2123
2124         /* Wake up case someone was waiting for buckets */
2125         closure_wake_up(&c->freelist_wait);
2126 }
2127
2128 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
2129 {
2130         struct open_bucket *ob;
2131         bool ret = false;
2132
2133         for (ob = c->open_buckets;
2134              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
2135              ob++) {
2136                 spin_lock(&ob->lock);
2137                 if (ob->valid && !ob->on_partial_list &&
2138                     ob->dev == ca->dev_idx)
2139                         ret = true;
2140                 spin_unlock(&ob->lock);
2141         }
2142
2143         return ret;
2144 }
2145
2146 /* device goes ro: */
2147 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
2148 {
2149         unsigned i;
2150
2151         /* First, remove device from allocation groups: */
2152
2153         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2154                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
2155
2156         /*
2157          * Capacity is calculated based off of devices in allocation groups:
2158          */
2159         bch2_recalc_capacity(c);
2160
2161         /* Next, close write points that point to this device... */
2162         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
2163                 bch2_writepoint_stop(c, ca, &c->write_points[i]);
2164
2165         bch2_writepoint_stop(c, ca, &c->copygc_write_point);
2166         bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
2167         bch2_writepoint_stop(c, ca, &c->btree_write_point);
2168
2169         mutex_lock(&c->btree_reserve_cache_lock);
2170         while (c->btree_reserve_cache_nr) {
2171                 struct btree_alloc *a =
2172                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
2173
2174                 bch2_open_buckets_put(c, &a->ob);
2175         }
2176         mutex_unlock(&c->btree_reserve_cache_lock);
2177
2178         spin_lock(&c->freelist_lock);
2179         i = 0;
2180         while (i < c->open_buckets_partial_nr) {
2181                 struct open_bucket *ob =
2182                         c->open_buckets + c->open_buckets_partial[i];
2183
2184                 if (ob->dev == ca->dev_idx) {
2185                         swap(c->open_buckets_partial[i],
2186                              c->open_buckets_partial[--c->open_buckets_partial_nr]);
2187                         ob->on_partial_list = false;
2188                         spin_unlock(&c->freelist_lock);
2189                         bch2_open_bucket_put(c, ob);
2190                         spin_lock(&c->freelist_lock);
2191                 } else {
2192                         i++;
2193                 }
2194         }
2195         spin_unlock(&c->freelist_lock);
2196
2197         bch2_ec_stop_dev(c, ca);
2198
2199         /*
2200          * Wake up threads that were blocked on allocation, so they can notice
2201          * the device can no longer be removed and the capacity has changed:
2202          */
2203         closure_wake_up(&c->freelist_wait);
2204
2205         /*
2206          * journal_res_get() can block waiting for free space in the journal -
2207          * it needs to notice there may not be devices to allocate from anymore:
2208          */
2209         wake_up(&c->journal.wait);
2210
2211         /* Now wait for any in flight writes: */
2212
2213         closure_wait_event(&c->open_buckets_wait,
2214                            !bch2_dev_has_open_write_point(c, ca));
2215 }
2216
2217 /* device goes rw: */
2218 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
2219 {
2220         unsigned i;
2221
2222         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2223                 if (ca->mi.data_allowed & (1 << i))
2224                         set_bit(ca->dev_idx, c->rw_devs[i].d);
2225 }
2226
2227 void bch2_fs_allocator_background_init(struct bch_fs *c)
2228 {
2229         spin_lock_init(&c->freelist_lock);
2230         INIT_WORK(&c->discard_work, bch2_do_discards_work);
2231         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
2232 }