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

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "bcachefs.h"
   3 #include "alloc_background.h"
   4 #include "alloc_foreground.h"
   5 #include "backpointers.h"
   6 #include "btree_cache.h"
   7 #include "btree_io.h"
   8 #include "btree_key_cache.h"
   9 #include "btree_update.h"
  10 #include "btree_update_interior.h"
  11 #include "btree_gc.h"
  12 #include "btree_write_buffer.h"
  13 #include "buckets.h"
  14 #include "buckets_waiting_for_journal.h"
  15 #include "clock.h"
  16 #include "debug.h"
  17 #include "ec.h"
  18 #include "error.h"
  19 #include "lru.h"
  20 #include "recovery.h"
  21 #include "trace.h"
  22 #include "varint.h"
  23
  24 #include <linux/kthread.h>
  25 #include <linux/math64.h>
  26 #include <linux/random.h>
  27 #include <linux/rculist.h>
  28 #include <linux/rcupdate.h>
  29 #include <linux/sched/task.h>
  30 #include <linux/sort.h>
  31
  32 static void bch2_discard_one_bucket_fast(struct bch_fs *c, struct bpos bucket);
  33
  34 /* Persistent alloc info: */
  35
  36 static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
  37 #define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8,
  38         BCH_ALLOC_FIELDS_V1()
  39 #undef x
  40 };
  41
  42 struct bkey_alloc_unpacked {
  43         u64             journal_seq;
  44         u8              gen;
  45         u8              oldest_gen;
  46         u8              data_type;
  47         bool            need_discard:1;
  48         bool            need_inc_gen:1;
  49 #define x(_name, _bits) u##_bits _name;
  50         BCH_ALLOC_FIELDS_V2()
  51 #undef  x
  52 };
  53
  54 static inline u64 alloc_field_v1_get(const struct bch_alloc *a,
  55                                      const void **p, unsigned field)
  56 {
  57         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  58         u64 v;
  59
  60         if (!(a->fields & (1 << field)))
  61                 return 0;
  62
  63         switch (bytes) {
  64         case 1:
  65                 v = *((const u8 *) *p);
  66                 break;
  67         case 2:
  68                 v = le16_to_cpup(*p);
  69                 break;
  70         case 4:
  71                 v = le32_to_cpup(*p);
  72                 break;
  73         case 8:
  74                 v = le64_to_cpup(*p);
  75                 break;
  76         default:
  77                 BUG();
  78         }
  79
  80         *p += bytes;
  81         return v;
  82 }
  83
  84 static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out,
  85                                  struct bkey_s_c k)
  86 {
  87         const struct bch_alloc *in = bkey_s_c_to_alloc(k).v;
  88         const void *d = in->data;
  89         unsigned idx = 0;
  90
  91         out->gen = in->gen;
  92
  93 #define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++);
  94         BCH_ALLOC_FIELDS_V1()
  95 #undef  x
  96 }
  97
  98 static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out,
  99                                 struct bkey_s_c k)
 100 {
 101         struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k);
 102         const u8 *in = a.v->data;
 103         const u8 *end = bkey_val_end(a);
 104         unsigned fieldnr = 0;
 105         int ret;
 106         u64 v;
 107
 108         out->gen        = a.v->gen;
 109         out->oldest_gen = a.v->oldest_gen;
 110         out->data_type  = a.v->data_type;
 111
 112 #define x(_name, _bits)                                                 \
 113         if (fieldnr < a.v->nr_fields) {                                 \
 114                 ret = bch2_varint_decode_fast(in, end, &v);             \
 115                 if (ret < 0)                                            \
 116                         return ret;                                     \
 117                 in += ret;                                              \
 118         } else {                                                        \
 119                 v = 0;                                                  \
 120         }                                                               \
 121         out->_name = v;                                                 \
 122         if (v != out->_name)                                            \
 123                 return -1;                                              \
 124         fieldnr++;
 125
 126         BCH_ALLOC_FIELDS_V2()
 127 #undef  x
 128         return 0;
 129 }
 130
 131 static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out,
 132                                 struct bkey_s_c k)
 133 {
 134         struct bkey_s_c_alloc_v3 a = bkey_s_c_to_alloc_v3(k);
 135         const u8 *in = a.v->data;
 136         const u8 *end = bkey_val_end(a);
 137         unsigned fieldnr = 0;
 138         int ret;
 139         u64 v;
 140
 141         out->gen        = a.v->gen;
 142         out->oldest_gen = a.v->oldest_gen;
 143         out->data_type  = a.v->data_type;
 144         out->need_discard = BCH_ALLOC_V3_NEED_DISCARD(a.v);
 145         out->need_inc_gen = BCH_ALLOC_V3_NEED_INC_GEN(a.v);
 146         out->journal_seq = le64_to_cpu(a.v->journal_seq);
 147
 148 #define x(_name, _bits)                                                 \
 149         if (fieldnr < a.v->nr_fields) {                                 \
 150                 ret = bch2_varint_decode_fast(in, end, &v);             \
 151                 if (ret < 0)                                            \
 152                         return ret;                                     \
 153                 in += ret;                                              \
 154         } else {                                                        \
 155                 v = 0;                                                  \
 156         }                                                               \
 157         out->_name = v;                                                 \
 158         if (v != out->_name)                                            \
 159                 return -1;                                              \
 160         fieldnr++;
 161
 162         BCH_ALLOC_FIELDS_V2()
 163 #undef  x
 164         return 0;
 165 }
 166
 167 static struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 168 {
 169         struct bkey_alloc_unpacked ret = { .gen = 0 };
 170
 171         switch (k.k->type) {
 172         case KEY_TYPE_alloc:
 173                 bch2_alloc_unpack_v1(&ret, k);
 174                 break;
 175         case KEY_TYPE_alloc_v2:
 176                 bch2_alloc_unpack_v2(&ret, k);
 177                 break;
 178         case KEY_TYPE_alloc_v3:
 179                 bch2_alloc_unpack_v3(&ret, k);
 180                 break;
 181         }
 182
 183         return ret;
 184 }
 185
 186 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
 187 {
 188         unsigned i, bytes = offsetof(struct bch_alloc, data);
 189
 190         for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++)
 191                 if (a->fields & (1 << i))
 192                         bytes += BCH_ALLOC_V1_FIELD_BYTES[i];
 193
 194         return DIV_ROUND_UP(bytes, sizeof(u64));
 195 }
 196
 197 int bch2_alloc_v1_invalid(struct bch_fs *c, struct bkey_s_c k,
 198                           enum bkey_invalid_flags flags,
 199                           struct printbuf *err)
 200 {
 201         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 202         int ret = 0;
 203
 204         /* allow for unknown fields */
 205         bkey_fsck_err_on(bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v), c, err,
 206                          alloc_v1_val_size_bad,
 207                          "incorrect value size (%zu < %u)",
 208                          bkey_val_u64s(a.k), bch_alloc_v1_val_u64s(a.v));
 209 fsck_err:
 210         return ret;
 211 }
 212
 213 int bch2_alloc_v2_invalid(struct bch_fs *c, struct bkey_s_c k,
 214                           enum bkey_invalid_flags flags,
 215                           struct printbuf *err)
 216 {
 217         struct bkey_alloc_unpacked u;
 218         int ret = 0;
 219
 220         bkey_fsck_err_on(bch2_alloc_unpack_v2(&u, k), c, err,
 221                          alloc_v2_unpack_error,
 222                          "unpack error");
 223 fsck_err:
 224         return ret;
 225 }
 226
 227 int bch2_alloc_v3_invalid(struct bch_fs *c, struct bkey_s_c k,
 228                           enum bkey_invalid_flags flags,
 229                           struct printbuf *err)
 230 {
 231         struct bkey_alloc_unpacked u;
 232         int ret = 0;
 233
 234         bkey_fsck_err_on(bch2_alloc_unpack_v3(&u, k), c, err,
 235                          alloc_v2_unpack_error,
 236                          "unpack error");
 237 fsck_err:
 238         return ret;
 239 }
 240
 241 int bch2_alloc_v4_invalid(struct bch_fs *c, struct bkey_s_c k,
 242                           enum bkey_invalid_flags flags, struct printbuf *err)
 243 {
 244         struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 245         int ret = 0;
 246
 247         bkey_fsck_err_on(alloc_v4_u64s(a.v) > bkey_val_u64s(k.k), c, err,
 248                          alloc_v4_val_size_bad,
 249                          "bad val size (%u > %zu)",
 250                          alloc_v4_u64s(a.v), bkey_val_u64s(k.k));
 251
 252         bkey_fsck_err_on(!BCH_ALLOC_V4_BACKPOINTERS_START(a.v) &&
 253                          BCH_ALLOC_V4_NR_BACKPOINTERS(a.v), c, err,
 254                          alloc_v4_backpointers_start_bad,
 255                          "invalid backpointers_start");
 256
 257         bkey_fsck_err_on(alloc_data_type(*a.v, a.v->data_type) != a.v->data_type, c, err,
 258                          alloc_key_data_type_bad,
 259                          "invalid data type (got %u should be %u)",
 260                          a.v->data_type, alloc_data_type(*a.v, a.v->data_type));
 261
 262         switch (a.v->data_type) {
 263         case BCH_DATA_free:
 264         case BCH_DATA_need_gc_gens:
 265         case BCH_DATA_need_discard:
 266                 bkey_fsck_err_on(bch2_bucket_sectors(*a.v) || a.v->stripe,
 267                                  c, err, alloc_key_empty_but_have_data,
 268                                  "empty data type free but have data");
 269                 break;
 270         case BCH_DATA_sb:
 271         case BCH_DATA_journal:
 272         case BCH_DATA_btree:
 273         case BCH_DATA_user:
 274         case BCH_DATA_parity:
 275                 bkey_fsck_err_on(!bch2_bucket_sectors_dirty(*a.v),
 276                                  c, err, alloc_key_dirty_sectors_0,
 277                                  "data_type %s but dirty_sectors==0",
 278                                  bch2_data_type_str(a.v->data_type));
 279                 break;
 280         case BCH_DATA_cached:
 281                 bkey_fsck_err_on(!a.v->cached_sectors ||
 282                                  bch2_bucket_sectors_dirty(*a.v) ||
 283                                  a.v->stripe,
 284                                  c, err, alloc_key_cached_inconsistency,
 285                                  "data type inconsistency");
 286
 287                 bkey_fsck_err_on(!a.v->io_time[READ] &&
 288                                  c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_to_lru_refs,
 289                                  c, err, alloc_key_cached_but_read_time_zero,
 290                                  "cached bucket with read_time == 0");
 291                 break;
 292         case BCH_DATA_stripe:
 293                 break;
 294         }
 295 fsck_err:
 296         return ret;
 297 }
 298
 299 void bch2_alloc_v4_swab(struct bkey_s k)
 300 {
 301         struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v;
 302         struct bch_backpointer *bp, *bps;
 303
 304         a->journal_seq          = swab64(a->journal_seq);
 305         a->flags                = swab32(a->flags);
 306         a->dirty_sectors        = swab32(a->dirty_sectors);
 307         a->cached_sectors       = swab32(a->cached_sectors);
 308         a->io_time[0]           = swab64(a->io_time[0]);
 309         a->io_time[1]           = swab64(a->io_time[1]);
 310         a->stripe               = swab32(a->stripe);
 311         a->nr_external_backpointers = swab32(a->nr_external_backpointers);
 312         a->fragmentation_lru    = swab64(a->fragmentation_lru);
 313
 314         bps = alloc_v4_backpointers(a);
 315         for (bp = bps; bp < bps + BCH_ALLOC_V4_NR_BACKPOINTERS(a); bp++) {
 316                 bp->bucket_offset       = swab40(bp->bucket_offset);
 317                 bp->bucket_len          = swab32(bp->bucket_len);
 318                 bch2_bpos_swab(&bp->pos);
 319         }
 320 }
 321
 322 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 323 {
 324         struct bch_alloc_v4 _a;
 325         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &_a);
 326
 327         prt_newline(out);
 328         printbuf_indent_add(out, 2);
 329
 330         prt_printf(out, "gen %u oldest_gen %u data_type ", a->gen, a->oldest_gen);
 331         bch2_prt_data_type(out, a->data_type);
 332         prt_newline(out);
 333         prt_printf(out, "journal_seq       %llu",       a->journal_seq);
 334         prt_newline(out);
 335         prt_printf(out, "need_discard      %llu",       BCH_ALLOC_V4_NEED_DISCARD(a));
 336         prt_newline(out);
 337         prt_printf(out, "need_inc_gen      %llu",       BCH_ALLOC_V4_NEED_INC_GEN(a));
 338         prt_newline(out);
 339         prt_printf(out, "dirty_sectors     %u", a->dirty_sectors);
 340         prt_newline(out);
 341         prt_printf(out, "cached_sectors    %u", a->cached_sectors);
 342         prt_newline(out);
 343         prt_printf(out, "stripe            %u", a->stripe);
 344         prt_newline(out);
 345         prt_printf(out, "stripe_redundancy %u", a->stripe_redundancy);
 346         prt_newline(out);
 347         prt_printf(out, "io_time[READ]     %llu",       a->io_time[READ]);
 348         prt_newline(out);
 349         prt_printf(out, "io_time[WRITE]    %llu",       a->io_time[WRITE]);
 350         prt_newline(out);
 351         prt_printf(out, "fragmentation     %llu",       a->fragmentation_lru);
 352         prt_newline(out);
 353         prt_printf(out, "bp_start          %llu", BCH_ALLOC_V4_BACKPOINTERS_START(a));
 354         printbuf_indent_sub(out, 2);
 355 }
 356
 357 void __bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out)
 358 {
 359         if (k.k->type == KEY_TYPE_alloc_v4) {
 360                 void *src, *dst;
 361
 362                 *out = *bkey_s_c_to_alloc_v4(k).v;
 363
 364                 src = alloc_v4_backpointers(out);
 365                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 366                 dst = alloc_v4_backpointers(out);
 367
 368                 if (src < dst)
 369                         memset(src, 0, dst - src);
 370
 371                 SET_BCH_ALLOC_V4_NR_BACKPOINTERS(out, 0);
 372         } else {
 373                 struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 374
 375                 *out = (struct bch_alloc_v4) {
 376                         .journal_seq            = u.journal_seq,
 377                         .flags                  = u.need_discard,
 378                         .gen                    = u.gen,
 379                         .oldest_gen             = u.oldest_gen,
 380                         .data_type              = u.data_type,
 381                         .stripe_redundancy      = u.stripe_redundancy,
 382                         .dirty_sectors          = u.dirty_sectors,
 383                         .cached_sectors         = u.cached_sectors,
 384                         .io_time[READ]          = u.read_time,
 385                         .io_time[WRITE]         = u.write_time,
 386                         .stripe                 = u.stripe,
 387                 };
 388
 389                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 390         }
 391 }
 392
 393 static noinline struct bkey_i_alloc_v4 *
 394 __bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 395 {
 396         struct bkey_i_alloc_v4 *ret;
 397
 398         ret = bch2_trans_kmalloc(trans, max(bkey_bytes(k.k), sizeof(struct bkey_i_alloc_v4)));
 399         if (IS_ERR(ret))
 400                 return ret;
 401
 402         if (k.k->type == KEY_TYPE_alloc_v4) {
 403                 void *src, *dst;
 404
 405                 bkey_reassemble(&ret->k_i, k);
 406
 407                 src = alloc_v4_backpointers(&ret->v);
 408                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(&ret->v, BCH_ALLOC_V4_U64s);
 409                 dst = alloc_v4_backpointers(&ret->v);
 410
 411                 if (src < dst)
 412                         memset(src, 0, dst - src);
 413
 414                 SET_BCH_ALLOC_V4_NR_BACKPOINTERS(&ret->v, 0);
 415                 set_alloc_v4_u64s(ret);
 416         } else {
 417                 bkey_alloc_v4_init(&ret->k_i);
 418                 ret->k.p = k.k->p;
 419                 bch2_alloc_to_v4(k, &ret->v);
 420         }
 421         return ret;
 422 }
 423
 424 static inline struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut_inlined(struct btree_trans *trans, struct bkey_s_c k)
 425 {
 426         struct bkey_s_c_alloc_v4 a;
 427
 428         if (likely(k.k->type == KEY_TYPE_alloc_v4) &&
 429             ((a = bkey_s_c_to_alloc_v4(k), true) &&
 430              BCH_ALLOC_V4_NR_BACKPOINTERS(a.v) == 0))
 431                 return bch2_bkey_make_mut_noupdate_typed(trans, k, alloc_v4);
 432
 433         return __bch2_alloc_to_v4_mut(trans, k);
 434 }
 435
 436 struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 437 {
 438         return bch2_alloc_to_v4_mut_inlined(trans, k);
 439 }
 440
 441 struct bkey_i_alloc_v4 *
 442 bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
 443                               struct bpos pos)
 444 {
 445         struct bkey_s_c k;
 446         struct bkey_i_alloc_v4 *a;
 447         int ret;
 448
 449         k = bch2_bkey_get_iter(trans, iter, BTREE_ID_alloc, pos,
 450                              BTREE_ITER_WITH_UPDATES|
 451                              BTREE_ITER_CACHED|
 452                              BTREE_ITER_INTENT);
 453         ret = bkey_err(k);
 454         if (unlikely(ret))
 455                 return ERR_PTR(ret);
 456
 457         a = bch2_alloc_to_v4_mut_inlined(trans, k);
 458         ret = PTR_ERR_OR_ZERO(a);
 459         if (unlikely(ret))
 460                 goto err;
 461         return a;
 462 err:
 463         bch2_trans_iter_exit(trans, iter);
 464         return ERR_PTR(ret);
 465 }
 466
 467 static struct bpos alloc_gens_pos(struct bpos pos, unsigned *offset)
 468 {
 469         *offset = pos.offset & KEY_TYPE_BUCKET_GENS_MASK;
 470
 471         pos.offset >>= KEY_TYPE_BUCKET_GENS_BITS;
 472         return pos;
 473 }
 474
 475 static struct bpos bucket_gens_pos_to_alloc(struct bpos pos, unsigned offset)
 476 {
 477         pos.offset <<= KEY_TYPE_BUCKET_GENS_BITS;
 478         pos.offset += offset;
 479         return pos;
 480 }
 481
 482 static unsigned alloc_gen(struct bkey_s_c k, unsigned offset)
 483 {
 484         return k.k->type == KEY_TYPE_bucket_gens
 485                 ? bkey_s_c_to_bucket_gens(k).v->gens[offset]
 486                 : 0;
 487 }
 488
 489 int bch2_bucket_gens_invalid(struct bch_fs *c, struct bkey_s_c k,
 490                              enum bkey_invalid_flags flags,
 491                              struct printbuf *err)
 492 {
 493         int ret = 0;
 494
 495         bkey_fsck_err_on(bkey_val_bytes(k.k) != sizeof(struct bch_bucket_gens), c, err,
 496                          bucket_gens_val_size_bad,
 497                          "bad val size (%zu != %zu)",
 498                          bkey_val_bytes(k.k), sizeof(struct bch_bucket_gens));
 499 fsck_err:
 500         return ret;
 501 }
 502
 503 void bch2_bucket_gens_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 504 {
 505         struct bkey_s_c_bucket_gens g = bkey_s_c_to_bucket_gens(k);
 506         unsigned i;
 507
 508         for (i = 0; i < ARRAY_SIZE(g.v->gens); i++) {
 509                 if (i)
 510                         prt_char(out, ' ');
 511                 prt_printf(out, "%u", g.v->gens[i]);
 512         }
 513 }
 514
 515 int bch2_bucket_gens_init(struct bch_fs *c)
 516 {
 517         struct btree_trans *trans = bch2_trans_get(c);
 518         struct bkey_i_bucket_gens g;
 519         bool have_bucket_gens_key = false;
 520         int ret;
 521
 522         ret = for_each_btree_key(trans, iter, BTREE_ID_alloc, POS_MIN,
 523                                  BTREE_ITER_PREFETCH, k, ({
 524                 /*
 525                  * Not a fsck error because this is checked/repaired by
 526                  * bch2_check_alloc_key() which runs later:
 527                  */
 528                 if (!bch2_dev_bucket_exists(c, k.k->p))
 529                         continue;
 530
 531                 struct bch_alloc_v4 a;
 532                 u8 gen = bch2_alloc_to_v4(k, &a)->gen;
 533                 unsigned offset;
 534                 struct bpos pos = alloc_gens_pos(iter.pos, &offset);
 535
 536                 if (have_bucket_gens_key && bkey_cmp(iter.pos, pos)) {
 537                         ret = commit_do(trans, NULL, NULL,
 538                                         BCH_TRANS_COMMIT_no_enospc,
 539                                 bch2_btree_insert_trans(trans, BTREE_ID_bucket_gens, &g.k_i, 0));
 540                         if (ret)
 541                                 break;
 542                         have_bucket_gens_key = false;
 543                 }
 544
 545                 if (!have_bucket_gens_key) {
 546                         bkey_bucket_gens_init(&g.k_i);
 547                         g.k.p = pos;
 548                         have_bucket_gens_key = true;
 549                 }
 550
 551                 g.v.gens[offset] = gen;
 552                 0;
 553         }));
 554
 555         if (have_bucket_gens_key && !ret)
 556                 ret = commit_do(trans, NULL, NULL,
 557                                 BCH_TRANS_COMMIT_no_enospc,
 558                         bch2_btree_insert_trans(trans, BTREE_ID_bucket_gens, &g.k_i, 0));
 559
 560         bch2_trans_put(trans);
 561
 562         bch_err_fn(c, ret);
 563         return ret;
 564 }
 565
 566 int bch2_alloc_read(struct bch_fs *c)
 567 {
 568         struct btree_trans *trans = bch2_trans_get(c);
 569         int ret;
 570
 571         down_read(&c->gc_lock);
 572
 573         if (c->sb.version_upgrade_complete >= bcachefs_metadata_version_bucket_gens) {
 574                 ret = for_each_btree_key(trans, iter, BTREE_ID_bucket_gens, POS_MIN,
 575                                          BTREE_ITER_PREFETCH, k, ({
 576                         u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
 577                         u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
 578
 579                         if (k.k->type != KEY_TYPE_bucket_gens)
 580                                 continue;
 581
 582                         const struct bch_bucket_gens *g = bkey_s_c_to_bucket_gens(k).v;
 583
 584                         /*
 585                          * Not a fsck error because this is checked/repaired by
 586                          * bch2_check_alloc_key() which runs later:
 587                          */
 588                         if (!bch2_dev_exists2(c, k.k->p.inode))
 589                                 continue;
 590
 591                         struct bch_dev *ca = bch_dev_bkey_exists(c, k.k->p.inode);
 592
 593                         for (u64 b = max_t(u64, ca->mi.first_bucket, start);
 594                              b < min_t(u64, ca->mi.nbuckets, end);
 595                              b++)
 596                                 *bucket_gen(ca, b) = g->gens[b & KEY_TYPE_BUCKET_GENS_MASK];
 597                         0;
 598                 }));
 599         } else {
 600                 ret = for_each_btree_key(trans, iter, BTREE_ID_alloc, POS_MIN,
 601                                          BTREE_ITER_PREFETCH, k, ({
 602                         /*
 603                          * Not a fsck error because this is checked/repaired by
 604                          * bch2_check_alloc_key() which runs later:
 605                          */
 606                         if (!bch2_dev_bucket_exists(c, k.k->p))
 607                                 continue;
 608
 609                         struct bch_dev *ca = bch_dev_bkey_exists(c, k.k->p.inode);
 610
 611                         struct bch_alloc_v4 a;
 612                         *bucket_gen(ca, k.k->p.offset) = bch2_alloc_to_v4(k, &a)->gen;
 613                         0;
 614                 }));
 615         }
 616
 617         bch2_trans_put(trans);
 618         up_read(&c->gc_lock);
 619
 620         bch_err_fn(c, ret);
 621         return ret;
 622 }
 623
 624 /* Free space/discard btree: */
 625
 626 static int bch2_bucket_do_index(struct btree_trans *trans,
 627                                 struct bkey_s_c alloc_k,
 628                                 const struct bch_alloc_v4 *a,
 629                                 bool set)
 630 {
 631         struct bch_fs *c = trans->c;
 632         struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 633         struct btree_iter iter;
 634         struct bkey_s_c old;
 635         struct bkey_i *k;
 636         enum btree_id btree;
 637         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 638         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 639         struct printbuf buf = PRINTBUF;
 640         int ret;
 641
 642         if (a->data_type != BCH_DATA_free &&
 643             a->data_type != BCH_DATA_need_discard)
 644                 return 0;
 645
 646         k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
 647         if (IS_ERR(k))
 648                 return PTR_ERR(k);
 649
 650         bkey_init(&k->k);
 651         k->k.type = new_type;
 652
 653         switch (a->data_type) {
 654         case BCH_DATA_free:
 655                 btree = BTREE_ID_freespace;
 656                 k->k.p = alloc_freespace_pos(alloc_k.k->p, *a);
 657                 bch2_key_resize(&k->k, 1);
 658                 break;
 659         case BCH_DATA_need_discard:
 660                 btree = BTREE_ID_need_discard;
 661                 k->k.p = alloc_k.k->p;
 662                 break;
 663         default:
 664                 return 0;
 665         }
 666
 667         old = bch2_bkey_get_iter(trans, &iter, btree,
 668                              bkey_start_pos(&k->k),
 669                              BTREE_ITER_INTENT);
 670         ret = bkey_err(old);
 671         if (ret)
 672                 return ret;
 673
 674         if (ca->mi.freespace_initialized &&
 675             c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info &&
 676             bch2_trans_inconsistent_on(old.k->type != old_type, trans,
 677                         "incorrect key when %s %s:%llu:%llu:0 (got %s should be %s)\n"
 678                         "  for %s",
 679                         set ? "setting" : "clearing",
 680                         bch2_btree_id_str(btree),
 681                         iter.pos.inode,
 682                         iter.pos.offset,
 683                         bch2_bkey_types[old.k->type],
 684                         bch2_bkey_types[old_type],
 685                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 686                 ret = -EIO;
 687                 goto err;
 688         }
 689
 690         ret = bch2_trans_update(trans, &iter, k, 0);
 691 err:
 692         bch2_trans_iter_exit(trans, &iter);
 693         printbuf_exit(&buf);
 694         return ret;
 695 }
 696
 697 static noinline int bch2_bucket_gen_update(struct btree_trans *trans,
 698                                            struct bpos bucket, u8 gen)
 699 {
 700         struct btree_iter iter;
 701         unsigned offset;
 702         struct bpos pos = alloc_gens_pos(bucket, &offset);
 703         struct bkey_i_bucket_gens *g;
 704         struct bkey_s_c k;
 705         int ret;
 706
 707         g = bch2_trans_kmalloc(trans, sizeof(*g));
 708         ret = PTR_ERR_OR_ZERO(g);
 709         if (ret)
 710                 return ret;
 711
 712         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_bucket_gens, pos,
 713                                BTREE_ITER_INTENT|
 714                                BTREE_ITER_WITH_UPDATES);
 715         ret = bkey_err(k);
 716         if (ret)
 717                 return ret;
 718
 719         if (k.k->type != KEY_TYPE_bucket_gens) {
 720                 bkey_bucket_gens_init(&g->k_i);
 721                 g->k.p = iter.pos;
 722         } else {
 723                 bkey_reassemble(&g->k_i, k);
 724         }
 725
 726         g->v.gens[offset] = gen;
 727
 728         ret = bch2_trans_update(trans, &iter, &g->k_i, 0);
 729         bch2_trans_iter_exit(trans, &iter);
 730         return ret;
 731 }
 732
 733 int bch2_trigger_alloc(struct btree_trans *trans,
 734                        enum btree_id btree, unsigned level,
 735                        struct bkey_s_c old, struct bkey_s new,
 736                        unsigned flags)
 737 {
 738         struct bch_fs *c = trans->c;
 739         int ret = 0;
 740
 741         if (bch2_trans_inconsistent_on(!bch2_dev_bucket_exists(c, new.k->p), trans,
 742                                        "alloc key for invalid device or bucket"))
 743                 return -EIO;
 744
 745         struct bch_dev *ca = bch_dev_bkey_exists(c, new.k->p.inode);
 746
 747         struct bch_alloc_v4 old_a_convert;
 748         const struct bch_alloc_v4 *old_a = bch2_alloc_to_v4(old, &old_a_convert);
 749
 750         if (flags & BTREE_TRIGGER_TRANSACTIONAL) {
 751                 struct bch_alloc_v4 *new_a = bkey_s_to_alloc_v4(new).v;
 752
 753                 new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
 754
 755                 if (bch2_bucket_sectors(*new_a) > bch2_bucket_sectors(*old_a)) {
 756                         new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 757                         new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 758                         SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
 759                         SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
 760                 }
 761
 762                 if (data_type_is_empty(new_a->data_type) &&
 763                     BCH_ALLOC_V4_NEED_INC_GEN(new_a) &&
 764                     !bch2_bucket_is_open_safe(c, new.k->p.inode, new.k->p.offset)) {
 765                         new_a->gen++;
 766                         SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
 767                 }
 768
 769                 if (old_a->data_type != new_a->data_type ||
 770                     (new_a->data_type == BCH_DATA_free &&
 771                      alloc_freespace_genbits(*old_a) != alloc_freespace_genbits(*new_a))) {
 772                         ret =   bch2_bucket_do_index(trans, old, old_a, false) ?:
 773                                 bch2_bucket_do_index(trans, new.s_c, new_a, true);
 774                         if (ret)
 775                                 return ret;
 776                 }
 777
 778                 if (new_a->data_type == BCH_DATA_cached &&
 779                     !new_a->io_time[READ])
 780                         new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 781
 782                 u64 old_lru = alloc_lru_idx_read(*old_a);
 783                 u64 new_lru = alloc_lru_idx_read(*new_a);
 784                 if (old_lru != new_lru) {
 785                         ret = bch2_lru_change(trans, new.k->p.inode,
 786                                               bucket_to_u64(new.k->p),
 787                                               old_lru, new_lru);
 788                         if (ret)
 789                                 return ret;
 790                 }
 791
 792                 new_a->fragmentation_lru = alloc_lru_idx_fragmentation(*new_a,
 793                                                 bch_dev_bkey_exists(c, new.k->p.inode));
 794                 if (old_a->fragmentation_lru != new_a->fragmentation_lru) {
 795                         ret = bch2_lru_change(trans,
 796                                         BCH_LRU_FRAGMENTATION_START,
 797                                         bucket_to_u64(new.k->p),
 798                                         old_a->fragmentation_lru, new_a->fragmentation_lru);
 799                         if (ret)
 800                                 return ret;
 801                 }
 802
 803                 if (old_a->gen != new_a->gen) {
 804                         ret = bch2_bucket_gen_update(trans, new.k->p, new_a->gen);
 805                         if (ret)
 806                                 return ret;
 807                 }
 808
 809                 /*
 810                  * need to know if we're getting called from the invalidate path or
 811                  * not:
 812                  */
 813
 814                 if ((flags & BTREE_TRIGGER_BUCKET_INVALIDATE) &&
 815                     old_a->cached_sectors) {
 816                         ret = bch2_update_cached_sectors_list(trans, new.k->p.inode,
 817                                                               -((s64) old_a->cached_sectors));
 818                         if (ret)
 819                                 return ret;
 820                 }
 821         }
 822
 823         if ((flags & BTREE_TRIGGER_ATOMIC) && (flags & BTREE_TRIGGER_INSERT)) {
 824                 struct bch_alloc_v4 *new_a = bkey_s_to_alloc_v4(new).v;
 825                 u64 journal_seq = trans->journal_res.seq;
 826                 u64 bucket_journal_seq = new_a->journal_seq;
 827
 828                 if ((flags & BTREE_TRIGGER_INSERT) &&
 829                     data_type_is_empty(old_a->data_type) !=
 830                     data_type_is_empty(new_a->data_type) &&
 831                     new.k->type == KEY_TYPE_alloc_v4) {
 832                         struct bch_alloc_v4 *v = bkey_s_to_alloc_v4(new).v;
 833
 834                         /*
 835                          * If the btree updates referring to a bucket weren't flushed
 836                          * before the bucket became empty again, then the we don't have
 837                          * to wait on a journal flush before we can reuse the bucket:
 838                          */
 839                         v->journal_seq = bucket_journal_seq =
 840                                 data_type_is_empty(new_a->data_type) &&
 841                                 (journal_seq == v->journal_seq ||
 842                                  bch2_journal_noflush_seq(&c->journal, v->journal_seq))
 843                                 ? 0 : journal_seq;
 844                 }
 845
 846                 if (!data_type_is_empty(old_a->data_type) &&
 847                     data_type_is_empty(new_a->data_type) &&
 848                     bucket_journal_seq) {
 849                         ret = bch2_set_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
 850                                         c->journal.flushed_seq_ondisk,
 851                                         new.k->p.inode, new.k->p.offset,
 852                                         bucket_journal_seq);
 853                         if (ret) {
 854                                 bch2_fs_fatal_error(c,
 855                                         "error setting bucket_needs_journal_commit: %i", ret);
 856                                 return ret;
 857                         }
 858                 }
 859
 860                 percpu_down_read(&c->mark_lock);
 861                 if (new_a->gen != old_a->gen)
 862                         *bucket_gen(ca, new.k->p.offset) = new_a->gen;
 863
 864                 bch2_dev_usage_update(c, ca, old_a, new_a, journal_seq, false);
 865                 percpu_up_read(&c->mark_lock);
 866
 867 #define eval_state(_a, expr)            ({ const struct bch_alloc_v4 *a = _a; expr; })
 868 #define statechange(expr)               !eval_state(old_a, expr) && eval_state(new_a, expr)
 869 #define bucket_flushed(a)               (!a->journal_seq || a->journal_seq <= c->journal.flushed_seq_ondisk)
 870
 871                 if (statechange(a->data_type == BCH_DATA_free) &&
 872                     bucket_flushed(new_a))
 873                         closure_wake_up(&c->freelist_wait);
 874
 875                 if (statechange(a->data_type == BCH_DATA_need_discard) &&
 876                     !bch2_bucket_is_open(c, new.k->p.inode, new.k->p.offset) &&
 877                     bucket_flushed(new_a))
 878                         bch2_discard_one_bucket_fast(c, new.k->p);
 879
 880                 if (statechange(a->data_type == BCH_DATA_cached) &&
 881                     !bch2_bucket_is_open(c, new.k->p.inode, new.k->p.offset) &&
 882                     should_invalidate_buckets(ca, bch2_dev_usage_read(ca)))
 883                         bch2_do_invalidates(c);
 884
 885                 if (statechange(a->data_type == BCH_DATA_need_gc_gens))
 886                         bch2_do_gc_gens(c);
 887         }
 888
 889         if ((flags & BTREE_TRIGGER_GC) &&
 890             (flags & BTREE_TRIGGER_BUCKET_INVALIDATE)) {
 891                 struct bch_alloc_v4 new_a_convert;
 892                 const struct bch_alloc_v4 *new_a = bch2_alloc_to_v4(new.s_c, &new_a_convert);
 893
 894                 percpu_down_read(&c->mark_lock);
 895                 struct bucket *g = gc_bucket(ca, new.k->p.offset);
 896
 897                 bucket_lock(g);
 898
 899                 g->gen_valid            = 1;
 900                 g->gen                  = new_a->gen;
 901                 g->data_type            = new_a->data_type;
 902                 g->stripe               = new_a->stripe;
 903                 g->stripe_redundancy    = new_a->stripe_redundancy;
 904                 g->dirty_sectors        = new_a->dirty_sectors;
 905                 g->cached_sectors       = new_a->cached_sectors;
 906
 907                 bucket_unlock(g);
 908                 percpu_up_read(&c->mark_lock);
 909         }
 910
 911         return 0;
 912 }
 913
 914 /*
 915  * This synthesizes deleted extents for holes, similar to BTREE_ITER_SLOTS for
 916  * extents style btrees, but works on non-extents btrees:
 917  */
 918 static struct bkey_s_c bch2_get_key_or_hole(struct btree_iter *iter, struct bpos end, struct bkey *hole)
 919 {
 920         struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
 921
 922         if (bkey_err(k))
 923                 return k;
 924
 925         if (k.k->type) {
 926                 return k;
 927         } else {
 928                 struct btree_iter iter2;
 929                 struct bpos next;
 930
 931                 bch2_trans_copy_iter(&iter2, iter);
 932
 933                 struct btree_path *path = btree_iter_path(iter->trans, iter);
 934                 if (!bpos_eq(path->l[0].b->key.k.p, SPOS_MAX))
 935                         end = bkey_min(end, bpos_nosnap_successor(path->l[0].b->key.k.p));
 936
 937                 end = bkey_min(end, POS(iter->pos.inode, iter->pos.offset + U32_MAX - 1));
 938
 939                 /*
 940                  * btree node min/max is a closed interval, upto takes a half
 941                  * open interval:
 942                  */
 943                 k = bch2_btree_iter_peek_upto(&iter2, end);
 944                 next = iter2.pos;
 945                 bch2_trans_iter_exit(iter->trans, &iter2);
 946
 947                 BUG_ON(next.offset >= iter->pos.offset + U32_MAX);
 948
 949                 if (bkey_err(k))
 950                         return k;
 951
 952                 bkey_init(hole);
 953                 hole->p = iter->pos;
 954
 955                 bch2_key_resize(hole, next.offset - iter->pos.offset);
 956                 return (struct bkey_s_c) { hole, NULL };
 957         }
 958 }
 959
 960 static bool next_bucket(struct bch_fs *c, struct bpos *bucket)
 961 {
 962         struct bch_dev *ca;
 963
 964         if (bch2_dev_bucket_exists(c, *bucket))
 965                 return true;
 966
 967         if (bch2_dev_exists2(c, bucket->inode)) {
 968                 ca = bch_dev_bkey_exists(c, bucket->inode);
 969
 970                 if (bucket->offset < ca->mi.first_bucket) {
 971                         bucket->offset = ca->mi.first_bucket;
 972                         return true;
 973                 }
 974
 975                 bucket->inode++;
 976                 bucket->offset = 0;
 977         }
 978
 979         rcu_read_lock();
 980         ca = __bch2_next_dev_idx(c, bucket->inode, NULL);
 981         if (ca)
 982                 *bucket = POS(ca->dev_idx, ca->mi.first_bucket);
 983         rcu_read_unlock();
 984
 985         return ca != NULL;
 986 }
 987
 988 static struct bkey_s_c bch2_get_key_or_real_bucket_hole(struct btree_iter *iter, struct bkey *hole)
 989 {
 990         struct bch_fs *c = iter->trans->c;
 991         struct bkey_s_c k;
 992 again:
 993         k = bch2_get_key_or_hole(iter, POS_MAX, hole);
 994         if (bkey_err(k))
 995                 return k;
 996
 997         if (!k.k->type) {
 998                 struct bpos bucket = bkey_start_pos(k.k);
 999
1000                 if (!bch2_dev_bucket_exists(c, bucket)) {
1001                         if (!next_bucket(c, &bucket))
1002                                 return bkey_s_c_null;
1003
1004                         bch2_btree_iter_set_pos(iter, bucket);
1005                         goto again;
1006                 }
1007
1008                 if (!bch2_dev_bucket_exists(c, k.k->p)) {
1009                         struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
1010
1011                         bch2_key_resize(hole, ca->mi.nbuckets - bucket.offset);
1012                 }
1013         }
1014
1015         return k;
1016 }
1017
1018 static noinline_for_stack
1019 int bch2_check_alloc_key(struct btree_trans *trans,
1020                          struct bkey_s_c alloc_k,
1021                          struct btree_iter *alloc_iter,
1022                          struct btree_iter *discard_iter,
1023                          struct btree_iter *freespace_iter,
1024                          struct btree_iter *bucket_gens_iter)
1025 {
1026         struct bch_fs *c = trans->c;
1027         struct bch_dev *ca;
1028         struct bch_alloc_v4 a_convert;
1029         const struct bch_alloc_v4 *a;
1030         unsigned discard_key_type, freespace_key_type;
1031         unsigned gens_offset;
1032         struct bkey_s_c k;
1033         struct printbuf buf = PRINTBUF;
1034         int ret;
1035
1036         if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
1037                         alloc_key_to_missing_dev_bucket,
1038                         "alloc key for invalid device:bucket %llu:%llu",
1039                         alloc_k.k->p.inode, alloc_k.k->p.offset))
1040                 return bch2_btree_delete_at(trans, alloc_iter, 0);
1041
1042         ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
1043         if (!ca->mi.freespace_initialized)
1044                 return 0;
1045
1046         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1047
1048         discard_key_type = a->data_type == BCH_DATA_need_discard ? KEY_TYPE_set : 0;
1049         bch2_btree_iter_set_pos(discard_iter, alloc_k.k->p);
1050         k = bch2_btree_iter_peek_slot(discard_iter);
1051         ret = bkey_err(k);
1052         if (ret)
1053                 goto err;
1054
1055         if (k.k->type != discard_key_type &&
1056             (c->opts.reconstruct_alloc ||
1057              fsck_err(c, need_discard_key_wrong,
1058                       "incorrect key in need_discard btree (got %s should be %s)\n"
1059                       "  %s",
1060                       bch2_bkey_types[k.k->type],
1061                       bch2_bkey_types[discard_key_type],
1062                       (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1063                 struct bkey_i *update =
1064                         bch2_trans_kmalloc(trans, sizeof(*update));
1065
1066                 ret = PTR_ERR_OR_ZERO(update);
1067                 if (ret)
1068                         goto err;
1069
1070                 bkey_init(&update->k);
1071                 update->k.type  = discard_key_type;
1072                 update->k.p     = discard_iter->pos;
1073
1074                 ret = bch2_trans_update(trans, discard_iter, update, 0);
1075                 if (ret)
1076                         goto err;
1077         }
1078
1079         freespace_key_type = a->data_type == BCH_DATA_free ? KEY_TYPE_set : 0;
1080         bch2_btree_iter_set_pos(freespace_iter, alloc_freespace_pos(alloc_k.k->p, *a));
1081         k = bch2_btree_iter_peek_slot(freespace_iter);
1082         ret = bkey_err(k);
1083         if (ret)
1084                 goto err;
1085
1086         if (k.k->type != freespace_key_type &&
1087             (c->opts.reconstruct_alloc ||
1088              fsck_err(c, freespace_key_wrong,
1089                       "incorrect key in freespace btree (got %s should be %s)\n"
1090                       "  %s",
1091                       bch2_bkey_types[k.k->type],
1092                       bch2_bkey_types[freespace_key_type],
1093                       (printbuf_reset(&buf),
1094                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1095                 struct bkey_i *update =
1096                         bch2_trans_kmalloc(trans, sizeof(*update));
1097
1098                 ret = PTR_ERR_OR_ZERO(update);
1099                 if (ret)
1100                         goto err;
1101
1102                 bkey_init(&update->k);
1103                 update->k.type  = freespace_key_type;
1104                 update->k.p     = freespace_iter->pos;
1105                 bch2_key_resize(&update->k, 1);
1106
1107                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1108                 if (ret)
1109                         goto err;
1110         }
1111
1112         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(alloc_k.k->p, &gens_offset));
1113         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1114         ret = bkey_err(k);
1115         if (ret)
1116                 goto err;
1117
1118         if (a->gen != alloc_gen(k, gens_offset) &&
1119             (c->opts.reconstruct_alloc ||
1120              fsck_err(c, bucket_gens_key_wrong,
1121                       "incorrect gen in bucket_gens btree (got %u should be %u)\n"
1122                       "  %s",
1123                       alloc_gen(k, gens_offset), a->gen,
1124                       (printbuf_reset(&buf),
1125                        bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)))) {
1126                 struct bkey_i_bucket_gens *g =
1127                         bch2_trans_kmalloc(trans, sizeof(*g));
1128
1129                 ret = PTR_ERR_OR_ZERO(g);
1130                 if (ret)
1131                         goto err;
1132
1133                 if (k.k->type == KEY_TYPE_bucket_gens) {
1134                         bkey_reassemble(&g->k_i, k);
1135                 } else {
1136                         bkey_bucket_gens_init(&g->k_i);
1137                         g->k.p = alloc_gens_pos(alloc_k.k->p, &gens_offset);
1138                 }
1139
1140                 g->v.gens[gens_offset] = a->gen;
1141
1142                 ret = bch2_trans_update(trans, bucket_gens_iter, &g->k_i, 0);
1143                 if (ret)
1144                         goto err;
1145         }
1146 err:
1147 fsck_err:
1148         printbuf_exit(&buf);
1149         return ret;
1150 }
1151
1152 static noinline_for_stack
1153 int bch2_check_alloc_hole_freespace(struct btree_trans *trans,
1154                                     struct bpos start,
1155                                     struct bpos *end,
1156                                     struct btree_iter *freespace_iter)
1157 {
1158         struct bch_fs *c = trans->c;
1159         struct bch_dev *ca;
1160         struct bkey_s_c k;
1161         struct printbuf buf = PRINTBUF;
1162         int ret;
1163
1164         ca = bch_dev_bkey_exists(c, start.inode);
1165         if (!ca->mi.freespace_initialized)
1166                 return 0;
1167
1168         bch2_btree_iter_set_pos(freespace_iter, start);
1169
1170         k = bch2_btree_iter_peek_slot(freespace_iter);
1171         ret = bkey_err(k);
1172         if (ret)
1173                 goto err;
1174
1175         *end = bkey_min(k.k->p, *end);
1176
1177         if (k.k->type != KEY_TYPE_set &&
1178             (c->opts.reconstruct_alloc ||
1179              fsck_err(c, freespace_hole_missing,
1180                       "hole in alloc btree missing in freespace btree\n"
1181                       "  device %llu buckets %llu-%llu",
1182                       freespace_iter->pos.inode,
1183                       freespace_iter->pos.offset,
1184                       end->offset))) {
1185                 struct bkey_i *update =
1186                         bch2_trans_kmalloc(trans, sizeof(*update));
1187
1188                 ret = PTR_ERR_OR_ZERO(update);
1189                 if (ret)
1190                         goto err;
1191
1192                 bkey_init(&update->k);
1193                 update->k.type  = KEY_TYPE_set;
1194                 update->k.p     = freespace_iter->pos;
1195                 bch2_key_resize(&update->k,
1196                                 min_t(u64, U32_MAX, end->offset -
1197                                       freespace_iter->pos.offset));
1198
1199                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
1200                 if (ret)
1201                         goto err;
1202         }
1203 err:
1204 fsck_err:
1205         printbuf_exit(&buf);
1206         return ret;
1207 }
1208
1209 static noinline_for_stack
1210 int bch2_check_alloc_hole_bucket_gens(struct btree_trans *trans,
1211                                       struct bpos start,
1212                                       struct bpos *end,
1213                                       struct btree_iter *bucket_gens_iter)
1214 {
1215         struct bch_fs *c = trans->c;
1216         struct bkey_s_c k;
1217         struct printbuf buf = PRINTBUF;
1218         unsigned i, gens_offset, gens_end_offset;
1219         int ret;
1220
1221         bch2_btree_iter_set_pos(bucket_gens_iter, alloc_gens_pos(start, &gens_offset));
1222
1223         k = bch2_btree_iter_peek_slot(bucket_gens_iter);
1224         ret = bkey_err(k);
1225         if (ret)
1226                 goto err;
1227
1228         if (bkey_cmp(alloc_gens_pos(start, &gens_offset),
1229                      alloc_gens_pos(*end,  &gens_end_offset)))
1230                 gens_end_offset = KEY_TYPE_BUCKET_GENS_NR;
1231
1232         if (k.k->type == KEY_TYPE_bucket_gens) {
1233                 struct bkey_i_bucket_gens g;
1234                 bool need_update = false;
1235
1236                 bkey_reassemble(&g.k_i, k);
1237
1238                 for (i = gens_offset; i < gens_end_offset; i++) {
1239                         if (fsck_err_on(g.v.gens[i], c,
1240                                         bucket_gens_hole_wrong,
1241                                         "hole in alloc btree at %llu:%llu with nonzero gen in bucket_gens btree (%u)",
1242                                         bucket_gens_pos_to_alloc(k.k->p, i).inode,
1243                                         bucket_gens_pos_to_alloc(k.k->p, i).offset,
1244                                         g.v.gens[i])) {
1245                                 g.v.gens[i] = 0;
1246                                 need_update = true;
1247                         }
1248                 }
1249
1250                 if (need_update) {
1251                         struct bkey_i *u = bch2_trans_kmalloc(trans, sizeof(g));
1252
1253                         ret = PTR_ERR_OR_ZERO(u);
1254                         if (ret)
1255                                 goto err;
1256
1257                         memcpy(u, &g, sizeof(g));
1258
1259                         ret = bch2_trans_update(trans, bucket_gens_iter, u, 0);
1260                         if (ret)
1261                                 goto err;
1262                 }
1263         }
1264
1265         *end = bkey_min(*end, bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0));
1266 err:
1267 fsck_err:
1268         printbuf_exit(&buf);
1269         return ret;
1270 }
1271
1272 static noinline_for_stack int bch2_check_discard_freespace_key(struct btree_trans *trans,
1273                                               struct btree_iter *iter)
1274 {
1275         struct bch_fs *c = trans->c;
1276         struct btree_iter alloc_iter;
1277         struct bkey_s_c alloc_k;
1278         struct bch_alloc_v4 a_convert;
1279         const struct bch_alloc_v4 *a;
1280         u64 genbits;
1281         struct bpos pos;
1282         enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
1283                 ? BCH_DATA_need_discard
1284                 : BCH_DATA_free;
1285         struct printbuf buf = PRINTBUF;
1286         int ret;
1287
1288         pos = iter->pos;
1289         pos.offset &= ~(~0ULL << 56);
1290         genbits = iter->pos.offset & (~0ULL << 56);
1291
1292         alloc_k = bch2_bkey_get_iter(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
1293         ret = bkey_err(alloc_k);
1294         if (ret)
1295                 return ret;
1296
1297         if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
1298                         need_discard_freespace_key_to_invalid_dev_bucket,
1299                         "entry in %s btree for nonexistant dev:bucket %llu:%llu",
1300                         bch2_btree_id_str(iter->btree_id), pos.inode, pos.offset))
1301                 goto delete;
1302
1303         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1304
1305         if (fsck_err_on(a->data_type != state ||
1306                         (state == BCH_DATA_free &&
1307                          genbits != alloc_freespace_genbits(*a)), c,
1308                         need_discard_freespace_key_bad,
1309                         "%s\n  incorrectly set at %s:%llu:%llu:0 (free %u, genbits %llu should be %llu)",
1310                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
1311                         bch2_btree_id_str(iter->btree_id),
1312                         iter->pos.inode,
1313                         iter->pos.offset,
1314                         a->data_type == state,
1315                         genbits >> 56, alloc_freespace_genbits(*a) >> 56))
1316                 goto delete;
1317 out:
1318 fsck_err:
1319         set_btree_iter_dontneed(&alloc_iter);
1320         bch2_trans_iter_exit(trans, &alloc_iter);
1321         printbuf_exit(&buf);
1322         return ret;
1323 delete:
1324         ret =   bch2_btree_delete_extent_at(trans, iter,
1325                         iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0) ?:
1326                 bch2_trans_commit(trans, NULL, NULL,
1327                         BCH_TRANS_COMMIT_no_enospc);
1328         goto out;
1329 }
1330
1331 /*
1332  * We've already checked that generation numbers in the bucket_gens btree are
1333  * valid for buckets that exist; this just checks for keys for nonexistent
1334  * buckets.
1335  */
1336 static noinline_for_stack
1337 int bch2_check_bucket_gens_key(struct btree_trans *trans,
1338                                struct btree_iter *iter,
1339                                struct bkey_s_c k)
1340 {
1341         struct bch_fs *c = trans->c;
1342         struct bkey_i_bucket_gens g;
1343         struct bch_dev *ca;
1344         u64 start = bucket_gens_pos_to_alloc(k.k->p, 0).offset;
1345         u64 end = bucket_gens_pos_to_alloc(bpos_nosnap_successor(k.k->p), 0).offset;
1346         u64 b;
1347         bool need_update = false, dev_exists;
1348         struct printbuf buf = PRINTBUF;
1349         int ret = 0;
1350
1351         BUG_ON(k.k->type != KEY_TYPE_bucket_gens);
1352         bkey_reassemble(&g.k_i, k);
1353
1354         /* if no bch_dev, skip out whether we repair or not */
1355         dev_exists = bch2_dev_exists2(c, k.k->p.inode);
1356         if (!dev_exists) {
1357                 if (fsck_err_on(!dev_exists, c,
1358                                 bucket_gens_to_invalid_dev,
1359                                 "bucket_gens key for invalid device:\n  %s",
1360                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1361                         ret = bch2_btree_delete_at(trans, iter, 0);
1362                 }
1363                 goto out;
1364         }
1365
1366         ca = bch_dev_bkey_exists(c, k.k->p.inode);
1367         if (fsck_err_on(end <= ca->mi.first_bucket ||
1368                         start >= ca->mi.nbuckets, c,
1369                         bucket_gens_to_invalid_buckets,
1370                         "bucket_gens key for invalid buckets:\n  %s",
1371                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1372                 ret = bch2_btree_delete_at(trans, iter, 0);
1373                 goto out;
1374         }
1375
1376         for (b = start; b < ca->mi.first_bucket; b++)
1377                 if (fsck_err_on(g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK], c,
1378                                 bucket_gens_nonzero_for_invalid_buckets,
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_nonzero_for_invalid_buckets,
1387                                 "bucket_gens key has nonzero gen for invalid bucket")) {
1388                         g.v.gens[b & KEY_TYPE_BUCKET_GENS_MASK] = 0;
1389                         need_update = true;
1390                 }
1391
1392         if (need_update) {
1393                 struct bkey_i *u = bch2_trans_kmalloc(trans, sizeof(g));
1394
1395                 ret = PTR_ERR_OR_ZERO(u);
1396                 if (ret)
1397                         goto out;
1398
1399                 memcpy(u, &g, sizeof(g));
1400                 ret = bch2_trans_update(trans, iter, u, 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 = bch2_trans_get(c);
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_iter_init(trans, &iter, BTREE_ID_alloc, POS_MIN,
1417                              BTREE_ITER_PREFETCH);
1418         bch2_trans_iter_init(trans, &discard_iter, BTREE_ID_need_discard, POS_MIN,
1419                              BTREE_ITER_PREFETCH);
1420         bch2_trans_iter_init(trans, &freespace_iter, BTREE_ID_freespace, POS_MIN,
1421                              BTREE_ITER_PREFETCH);
1422         bch2_trans_iter_init(trans, &bucket_gens_iter, BTREE_ID_bucket_gens, POS_MIN,
1423                              BTREE_ITER_PREFETCH);
1424
1425         while (1) {
1426                 struct bpos next;
1427
1428                 bch2_trans_begin(trans);
1429
1430                 k = bch2_get_key_or_real_bucket_hole(&iter, &hole);
1431                 ret = bkey_err(k);
1432                 if (ret)
1433                         goto bkey_err;
1434
1435                 if (!k.k)
1436                         break;
1437
1438                 if (k.k->type) {
1439                         next = bpos_nosnap_successor(k.k->p);
1440
1441                         ret = bch2_check_alloc_key(trans,
1442                                                    k, &iter,
1443                                                    &discard_iter,
1444                                                    &freespace_iter,
1445                                                    &bucket_gens_iter);
1446                         if (ret)
1447                                 goto bkey_err;
1448                 } else {
1449                         next = k.k->p;
1450
1451                         ret = bch2_check_alloc_hole_freespace(trans,
1452                                                     bkey_start_pos(k.k),
1453                                                     &next,
1454                                                     &freespace_iter) ?:
1455                                 bch2_check_alloc_hole_bucket_gens(trans,
1456                                                     bkey_start_pos(k.k),
1457                                                     &next,
1458                                                     &bucket_gens_iter);
1459                         if (ret)
1460                                 goto bkey_err;
1461                 }
1462
1463                 ret = bch2_trans_commit(trans, NULL, NULL,
1464                                         BCH_TRANS_COMMIT_no_enospc);
1465                 if (ret)
1466                         goto bkey_err;
1467
1468                 bch2_btree_iter_set_pos(&iter, next);
1469 bkey_err:
1470                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1471                         continue;
1472                 if (ret)
1473                         break;
1474         }
1475         bch2_trans_iter_exit(trans, &bucket_gens_iter);
1476         bch2_trans_iter_exit(trans, &freespace_iter);
1477         bch2_trans_iter_exit(trans, &discard_iter);
1478         bch2_trans_iter_exit(trans, &iter);
1479
1480         if (ret < 0)
1481                 goto err;
1482
1483         ret = for_each_btree_key(trans, iter,
1484                         BTREE_ID_need_discard, POS_MIN,
1485                         BTREE_ITER_PREFETCH, k,
1486                 bch2_check_discard_freespace_key(trans, &iter));
1487         if (ret)
1488                 goto err;
1489
1490         bch2_trans_iter_init(trans, &iter, BTREE_ID_freespace, POS_MIN,
1491                              BTREE_ITER_PREFETCH);
1492         while (1) {
1493                 bch2_trans_begin(trans);
1494                 k = bch2_btree_iter_peek(&iter);
1495                 if (!k.k)
1496                         break;
1497
1498                 ret = bkey_err(k) ?:
1499                         bch2_check_discard_freespace_key(trans, &iter);
1500                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
1501                         ret = 0;
1502                         continue;
1503                 }
1504                 if (ret) {
1505                         struct printbuf buf = PRINTBUF;
1506                         bch2_bkey_val_to_text(&buf, c, k);
1507
1508                         bch_err(c, "while checking %s", buf.buf);
1509                         printbuf_exit(&buf);
1510                         break;
1511                 }
1512
1513                 bch2_btree_iter_set_pos(&iter, bpos_nosnap_successor(iter.pos));
1514         }
1515         bch2_trans_iter_exit(trans, &iter);
1516         if (ret)
1517                 goto err;
1518
1519         ret = for_each_btree_key_commit(trans, iter,
1520                         BTREE_ID_bucket_gens, POS_MIN,
1521                         BTREE_ITER_PREFETCH, k,
1522                         NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1523                 bch2_check_bucket_gens_key(trans, &iter, k));
1524 err:
1525         bch2_trans_put(trans);
1526         bch_err_fn(c, ret);
1527         return ret;
1528 }
1529
1530 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
1531                                        struct btree_iter *alloc_iter)
1532 {
1533         struct bch_fs *c = trans->c;
1534         struct btree_iter lru_iter;
1535         struct bch_alloc_v4 a_convert;
1536         const struct bch_alloc_v4 *a;
1537         struct bkey_s_c alloc_k, lru_k;
1538         struct printbuf buf = PRINTBUF;
1539         int ret;
1540
1541         alloc_k = bch2_btree_iter_peek(alloc_iter);
1542         if (!alloc_k.k)
1543                 return 0;
1544
1545         ret = bkey_err(alloc_k);
1546         if (ret)
1547                 return ret;
1548
1549         a = bch2_alloc_to_v4(alloc_k, &a_convert);
1550
1551         if (a->data_type != BCH_DATA_cached)
1552                 return 0;
1553
1554         if (fsck_err_on(!a->io_time[READ], c,
1555                         alloc_key_cached_but_read_time_zero,
1556                         "cached bucket with read_time 0\n"
1557                         "  %s",
1558                 (printbuf_reset(&buf),
1559                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
1560                 struct bkey_i_alloc_v4 *a_mut =
1561                         bch2_alloc_to_v4_mut(trans, alloc_k);
1562                 ret = PTR_ERR_OR_ZERO(a_mut);
1563                 if (ret)
1564                         goto err;
1565
1566                 a_mut->v.io_time[READ] = atomic64_read(&c->io_clock[READ].now);
1567                 ret = bch2_trans_update(trans, alloc_iter,
1568                                         &a_mut->k_i, BTREE_TRIGGER_NORUN);
1569                 if (ret)
1570                         goto err;
1571
1572                 a = &a_mut->v;
1573         }
1574
1575         lru_k = bch2_bkey_get_iter(trans, &lru_iter, BTREE_ID_lru,
1576                              lru_pos(alloc_k.k->p.inode,
1577                                      bucket_to_u64(alloc_k.k->p),
1578                                      a->io_time[READ]), 0);
1579         ret = bkey_err(lru_k);
1580         if (ret)
1581                 return ret;
1582
1583         if (fsck_err_on(lru_k.k->type != KEY_TYPE_set, c,
1584                         alloc_key_to_missing_lru_entry,
1585                         "missing lru entry\n"
1586                         "  %s",
1587                         (printbuf_reset(&buf),
1588                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
1589                 ret = bch2_lru_set(trans,
1590                                    alloc_k.k->p.inode,
1591                                    bucket_to_u64(alloc_k.k->p),
1592                                    a->io_time[READ]);
1593                 if (ret)
1594                         goto err;
1595         }
1596 err:
1597 fsck_err:
1598         bch2_trans_iter_exit(trans, &lru_iter);
1599         printbuf_exit(&buf);
1600         return ret;
1601 }
1602
1603 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
1604 {
1605         int ret = bch2_trans_run(c,
1606                 for_each_btree_key_commit(trans, iter, BTREE_ID_alloc,
1607                                 POS_MIN, BTREE_ITER_PREFETCH, k,
1608                                 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
1609                         bch2_check_alloc_to_lru_ref(trans, &iter)));
1610         bch_err_fn(c, ret);
1611         return ret;
1612 }
1613
1614 static int discard_in_flight_add(struct bch_fs *c, struct bpos bucket)
1615 {
1616         int ret;
1617
1618         mutex_lock(&c->discard_buckets_in_flight_lock);
1619         darray_for_each(c->discard_buckets_in_flight, i)
1620                 if (bkey_eq(*i, bucket)) {
1621                         ret = -EEXIST;
1622                         goto out;
1623                 }
1624
1625         ret = darray_push(&c->discard_buckets_in_flight, bucket);
1626 out:
1627         mutex_unlock(&c->discard_buckets_in_flight_lock);
1628         return ret;
1629 }
1630
1631 static void discard_in_flight_remove(struct bch_fs *c, struct bpos bucket)
1632 {
1633         mutex_lock(&c->discard_buckets_in_flight_lock);
1634         darray_for_each(c->discard_buckets_in_flight, i)
1635                 if (bkey_eq(*i, bucket)) {
1636                         darray_remove_item(&c->discard_buckets_in_flight, i);
1637                         goto found;
1638                 }
1639         BUG();
1640 found:
1641         mutex_unlock(&c->discard_buckets_in_flight_lock);
1642 }
1643
1644 struct discard_buckets_state {
1645         u64             seen;
1646         u64             open;
1647         u64             need_journal_commit;
1648         u64             discarded;
1649         struct bch_dev  *ca;
1650         u64             need_journal_commit_this_dev;
1651 };
1652
1653 static void discard_buckets_next_dev(struct bch_fs *c, struct discard_buckets_state *s, struct bch_dev *ca)
1654 {
1655         if (s->ca == ca)
1656                 return;
1657
1658         if (s->ca && s->need_journal_commit_this_dev >
1659             bch2_dev_usage_read(s->ca).d[BCH_DATA_free].buckets)
1660                 bch2_journal_flush_async(&c->journal, NULL);
1661
1662         if (s->ca)
1663                 percpu_ref_put(&s->ca->ref);
1664         if (ca)
1665                 percpu_ref_get(&ca->ref);
1666         s->ca = ca;
1667         s->need_journal_commit_this_dev = 0;
1668 }
1669
1670 static int bch2_discard_one_bucket(struct btree_trans *trans,
1671                                    struct btree_iter *need_discard_iter,
1672                                    struct bpos *discard_pos_done,
1673                                    struct discard_buckets_state *s)
1674 {
1675         struct bch_fs *c = trans->c;
1676         struct bpos pos = need_discard_iter->pos;
1677         struct btree_iter iter = { NULL };
1678         struct bkey_s_c k;
1679         struct bch_dev *ca;
1680         struct bkey_i_alloc_v4 *a;
1681         struct printbuf buf = PRINTBUF;
1682         bool discard_locked = false;
1683         int ret = 0;
1684
1685         ca = bch_dev_bkey_exists(c, pos.inode);
1686
1687         if (!percpu_ref_tryget(&ca->io_ref)) {
1688                 bch2_btree_iter_set_pos(need_discard_iter, POS(pos.inode + 1, 0));
1689                 return 0;
1690         }
1691
1692         discard_buckets_next_dev(c, s, ca);
1693
1694         if (bch2_bucket_is_open_safe(c, pos.inode, pos.offset)) {
1695                 s->open++;
1696                 goto out;
1697         }
1698
1699         if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
1700                         c->journal.flushed_seq_ondisk,
1701                         pos.inode, pos.offset)) {
1702                 s->need_journal_commit++;
1703                 s->need_journal_commit_this_dev++;
1704                 goto out;
1705         }
1706
1707         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_alloc,
1708                                need_discard_iter->pos,
1709                                BTREE_ITER_CACHED);
1710         ret = bkey_err(k);
1711         if (ret)
1712                 goto out;
1713
1714         a = bch2_alloc_to_v4_mut(trans, k);
1715         ret = PTR_ERR_OR_ZERO(a);
1716         if (ret)
1717                 goto out;
1718
1719         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
1720                 a->v.gen++;
1721                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1722                 goto write;
1723         }
1724
1725         if (a->v.journal_seq > c->journal.flushed_seq_ondisk) {
1726                 if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
1727                         bch2_trans_inconsistent(trans,
1728                                 "clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
1729                                 "%s",
1730                                 a->v.journal_seq,
1731                                 c->journal.flushed_seq_ondisk,
1732                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1733                         ret = -EIO;
1734                 }
1735                 goto out;
1736         }
1737
1738         if (a->v.data_type != BCH_DATA_need_discard) {
1739                 if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_alloc_info) {
1740                         bch2_trans_inconsistent(trans,
1741                                 "bucket incorrectly set in need_discard btree\n"
1742                                 "%s",
1743                                 (bch2_bkey_val_to_text(&buf, c, k), buf.buf));
1744                         ret = -EIO;
1745                 }
1746
1747                 goto out;
1748         }
1749
1750         if (discard_in_flight_add(c, SPOS(iter.pos.inode, iter.pos.offset, true)))
1751                 goto out;
1752
1753         discard_locked = true;
1754
1755         if (!bkey_eq(*discard_pos_done, iter.pos) &&
1756             ca->mi.discard && !c->opts.nochanges) {
1757                 /*
1758                  * This works without any other locks because this is the only
1759                  * thread that removes items from the need_discard tree
1760                  */
1761                 bch2_trans_unlock_long(trans);
1762                 blkdev_issue_discard(ca->disk_sb.bdev,
1763                                      k.k->p.offset * ca->mi.bucket_size,
1764                                      ca->mi.bucket_size,
1765                                      GFP_KERNEL);
1766                 *discard_pos_done = iter.pos;
1767
1768                 ret = bch2_trans_relock_notrace(trans);
1769                 if (ret)
1770                         goto out;
1771         }
1772
1773         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
1774         a->v.data_type = alloc_data_type(a->v, a->v.data_type);
1775 write:
1776         ret =   bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1777                 bch2_trans_commit(trans, NULL, NULL,
1778                                   BCH_WATERMARK_btree|
1779                                   BCH_TRANS_COMMIT_no_enospc);
1780         if (ret)
1781                 goto out;
1782
1783         count_event(c, bucket_discard);
1784         s->discarded++;
1785 out:
1786         if (discard_locked)
1787                 discard_in_flight_remove(c, iter.pos);
1788         s->seen++;
1789         bch2_trans_iter_exit(trans, &iter);
1790         percpu_ref_put(&ca->io_ref);
1791         printbuf_exit(&buf);
1792         return ret;
1793 }
1794
1795 static void bch2_do_discards_work(struct work_struct *work)
1796 {
1797         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
1798         struct discard_buckets_state s = {};
1799         struct bpos discard_pos_done = POS_MAX;
1800         int ret;
1801
1802         /*
1803          * We're doing the commit in bch2_discard_one_bucket instead of using
1804          * for_each_btree_key_commit() so that we can increment counters after
1805          * successful commit:
1806          */
1807         ret = bch2_trans_run(c,
1808                 for_each_btree_key(trans, iter,
1809                                    BTREE_ID_need_discard, POS_MIN, 0, k,
1810                         bch2_discard_one_bucket(trans, &iter, &discard_pos_done, &s)));
1811
1812         discard_buckets_next_dev(c, &s, NULL);
1813
1814         trace_discard_buckets(c, s.seen, s.open, s.need_journal_commit, s.discarded,
1815                               bch2_err_str(ret));
1816
1817         bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1818 }
1819
1820 void bch2_do_discards(struct bch_fs *c)
1821 {
1822         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_discard) &&
1823             !queue_work(c->write_ref_wq, &c->discard_work))
1824                 bch2_write_ref_put(c, BCH_WRITE_REF_discard);
1825 }
1826
1827 static int bch2_clear_bucket_needs_discard(struct btree_trans *trans, struct bpos bucket)
1828 {
1829         struct btree_iter iter;
1830         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, bucket, BTREE_ITER_INTENT);
1831         struct bkey_s_c k = bch2_btree_iter_peek_slot(&iter);
1832         int ret = bkey_err(k);
1833         if (ret)
1834                 goto err;
1835
1836         struct bkey_i_alloc_v4 *a = bch2_alloc_to_v4_mut(trans, k);
1837         ret = PTR_ERR_OR_ZERO(a);
1838         if (ret)
1839                 goto err;
1840
1841         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
1842         a->v.data_type = alloc_data_type(a->v, a->v.data_type);
1843
1844         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1845 err:
1846         bch2_trans_iter_exit(trans, &iter);
1847         return ret;
1848 }
1849
1850 static void bch2_do_discards_fast_work(struct work_struct *work)
1851 {
1852         struct bch_fs *c = container_of(work, struct bch_fs, discard_fast_work);
1853
1854         while (1) {
1855                 bool got_bucket = false;
1856                 struct bpos bucket;
1857                 struct bch_dev *ca;
1858
1859                 mutex_lock(&c->discard_buckets_in_flight_lock);
1860                 darray_for_each(c->discard_buckets_in_flight, i) {
1861                         if (i->snapshot)
1862                                 continue;
1863
1864                         ca = bch_dev_bkey_exists(c, i->inode);
1865
1866                         if (!percpu_ref_tryget(&ca->io_ref)) {
1867                                 darray_remove_item(&c->discard_buckets_in_flight, i);
1868                                 continue;
1869                         }
1870
1871                         got_bucket = true;
1872                         bucket = *i;
1873                         i->snapshot = true;
1874                         break;
1875                 }
1876                 mutex_unlock(&c->discard_buckets_in_flight_lock);
1877
1878                 if (!got_bucket)
1879                         break;
1880
1881                 blkdev_issue_discard(ca->disk_sb.bdev,
1882                                      bucket.offset * ca->mi.bucket_size,
1883                                      ca->mi.bucket_size,
1884                                      GFP_KERNEL);
1885
1886                 int ret = bch2_trans_do(c, NULL, NULL,
1887                                         BCH_WATERMARK_btree|
1888                                         BCH_TRANS_COMMIT_no_enospc,
1889                                         bch2_clear_bucket_needs_discard(trans, bucket));
1890                 bch_err_fn(c, ret);
1891
1892                 percpu_ref_put(&ca->io_ref);
1893                 discard_in_flight_remove(c, bucket);
1894
1895                 if (ret)
1896                         break;
1897         }
1898
1899         bch2_write_ref_put(c, BCH_WRITE_REF_discard_fast);
1900 }
1901
1902 static void bch2_discard_one_bucket_fast(struct bch_fs *c, struct bpos bucket)
1903 {
1904         struct bch_dev *ca = bch_dev_bkey_exists(c, bucket.inode);
1905
1906         if (!percpu_ref_is_dying(&ca->io_ref) &&
1907             !discard_in_flight_add(c, bucket) &&
1908             bch2_write_ref_tryget(c, BCH_WRITE_REF_discard_fast) &&
1909             !queue_work(c->write_ref_wq, &c->discard_fast_work))
1910                 bch2_write_ref_put(c, BCH_WRITE_REF_discard_fast);
1911 }
1912
1913 static int invalidate_one_bucket(struct btree_trans *trans,
1914                                  struct btree_iter *lru_iter,
1915                                  struct bkey_s_c lru_k,
1916                                  s64 *nr_to_invalidate)
1917 {
1918         struct bch_fs *c = trans->c;
1919         struct btree_iter alloc_iter = { NULL };
1920         struct bkey_i_alloc_v4 *a = NULL;
1921         struct printbuf buf = PRINTBUF;
1922         struct bpos bucket = u64_to_bucket(lru_k.k->p.offset);
1923         unsigned cached_sectors;
1924         int ret = 0;
1925
1926         if (*nr_to_invalidate <= 0)
1927                 return 1;
1928
1929         if (!bch2_dev_bucket_exists(c, bucket)) {
1930                 prt_str(&buf, "lru entry points to invalid bucket");
1931                 goto err;
1932         }
1933
1934         if (bch2_bucket_is_open_safe(c, bucket.inode, bucket.offset))
1935                 return 0;
1936
1937         a = bch2_trans_start_alloc_update(trans, &alloc_iter, bucket);
1938         ret = PTR_ERR_OR_ZERO(a);
1939         if (ret)
1940                 goto out;
1941
1942         /* We expect harmless races here due to the btree write buffer: */
1943         if (lru_pos_time(lru_iter->pos) != alloc_lru_idx_read(a->v))
1944                 goto out;
1945
1946         BUG_ON(a->v.data_type != BCH_DATA_cached);
1947
1948         if (!a->v.cached_sectors)
1949                 bch_err(c, "invalidating empty bucket, confused");
1950
1951         cached_sectors = a->v.cached_sectors;
1952
1953         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1954         a->v.gen++;
1955         a->v.data_type          = 0;
1956         a->v.dirty_sectors      = 0;
1957         a->v.cached_sectors     = 0;
1958         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1959         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1960
1961         ret =   bch2_trans_update(trans, &alloc_iter, &a->k_i,
1962                                 BTREE_TRIGGER_BUCKET_INVALIDATE) ?:
1963                 bch2_trans_commit(trans, NULL, NULL,
1964                                   BCH_WATERMARK_btree|
1965                                   BCH_TRANS_COMMIT_no_enospc);
1966         if (ret)
1967                 goto out;
1968
1969         trace_and_count(c, bucket_invalidate, c, bucket.inode, bucket.offset, cached_sectors);
1970         --*nr_to_invalidate;
1971 out:
1972         bch2_trans_iter_exit(trans, &alloc_iter);
1973         printbuf_exit(&buf);
1974         return ret;
1975 err:
1976         prt_str(&buf, "\n  lru key: ");
1977         bch2_bkey_val_to_text(&buf, c, lru_k);
1978
1979         prt_str(&buf, "\n  lru entry: ");
1980         bch2_lru_pos_to_text(&buf, lru_iter->pos);
1981
1982         prt_str(&buf, "\n  alloc key: ");
1983         if (!a)
1984                 bch2_bpos_to_text(&buf, bucket);
1985         else
1986                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
1987
1988         bch_err(c, "%s", buf.buf);
1989         if (c->curr_recovery_pass > BCH_RECOVERY_PASS_check_lrus) {
1990                 bch2_inconsistent_error(c);
1991                 ret = -EINVAL;
1992         }
1993
1994         goto out;
1995 }
1996
1997 static void bch2_do_invalidates_work(struct work_struct *work)
1998 {
1999         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
2000         struct btree_trans *trans = bch2_trans_get(c);
2001         int ret = 0;
2002
2003         ret = bch2_btree_write_buffer_tryflush(trans);
2004         if (ret)
2005                 goto err;
2006
2007         for_each_member_device(c, ca) {
2008                 s64 nr_to_invalidate =
2009                         should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
2010
2011                 ret = for_each_btree_key_upto(trans, iter, BTREE_ID_lru,
2012                                 lru_pos(ca->dev_idx, 0, 0),
2013                                 lru_pos(ca->dev_idx, U64_MAX, LRU_TIME_MAX),
2014                                 BTREE_ITER_INTENT, k,
2015                         invalidate_one_bucket(trans, &iter, k, &nr_to_invalidate));
2016
2017                 if (ret < 0) {
2018                         percpu_ref_put(&ca->ref);
2019                         break;
2020                 }
2021         }
2022 err:
2023         bch2_trans_put(trans);
2024         bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
2025 }
2026
2027 void bch2_do_invalidates(struct bch_fs *c)
2028 {
2029         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_invalidate) &&
2030             !queue_work(c->write_ref_wq, &c->invalidate_work))
2031                 bch2_write_ref_put(c, BCH_WRITE_REF_invalidate);
2032 }
2033
2034 int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca,
2035                             u64 bucket_start, u64 bucket_end)
2036 {
2037         struct btree_trans *trans = bch2_trans_get(c);
2038         struct btree_iter iter;
2039         struct bkey_s_c k;
2040         struct bkey hole;
2041         struct bpos end = POS(ca->dev_idx, bucket_end);
2042         struct bch_member *m;
2043         unsigned long last_updated = jiffies;
2044         int ret;
2045
2046         BUG_ON(bucket_start > bucket_end);
2047         BUG_ON(bucket_end > ca->mi.nbuckets);
2048
2049         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc,
2050                 POS(ca->dev_idx, max_t(u64, ca->mi.first_bucket, bucket_start)),
2051                 BTREE_ITER_PREFETCH);
2052         /*
2053          * Scan the alloc btree for every bucket on @ca, and add buckets to the
2054          * freespace/need_discard/need_gc_gens btrees as needed:
2055          */
2056         while (1) {
2057                 if (last_updated + HZ * 10 < jiffies) {
2058                         bch_info(ca, "%s: currently at %llu/%llu",
2059                                  __func__, iter.pos.offset, ca->mi.nbuckets);
2060                         last_updated = jiffies;
2061                 }
2062
2063                 bch2_trans_begin(trans);
2064
2065                 if (bkey_ge(iter.pos, end)) {
2066                         ret = 0;
2067                         break;
2068                 }
2069
2070                 k = bch2_get_key_or_hole(&iter, end, &hole);
2071                 ret = bkey_err(k);
2072                 if (ret)
2073                         goto bkey_err;
2074
2075                 if (k.k->type) {
2076                         /*
2077                          * We process live keys in the alloc btree one at a
2078                          * time:
2079                          */
2080                         struct bch_alloc_v4 a_convert;
2081                         const struct bch_alloc_v4 *a = bch2_alloc_to_v4(k, &a_convert);
2082
2083                         ret =   bch2_bucket_do_index(trans, k, a, true) ?:
2084                                 bch2_trans_commit(trans, NULL, NULL,
2085                                                   BCH_TRANS_COMMIT_no_enospc);
2086                         if (ret)
2087                                 goto bkey_err;
2088
2089                         bch2_btree_iter_advance(&iter);
2090                 } else {
2091                         struct bkey_i *freespace;
2092
2093                         freespace = bch2_trans_kmalloc(trans, sizeof(*freespace));
2094                         ret = PTR_ERR_OR_ZERO(freespace);
2095                         if (ret)
2096                                 goto bkey_err;
2097
2098                         bkey_init(&freespace->k);
2099                         freespace->k.type       = KEY_TYPE_set;
2100                         freespace->k.p          = k.k->p;
2101                         freespace->k.size       = k.k->size;
2102
2103                         ret = bch2_btree_insert_trans(trans, BTREE_ID_freespace, freespace, 0) ?:
2104                                 bch2_trans_commit(trans, NULL, NULL,
2105                                                   BCH_TRANS_COMMIT_no_enospc);
2106                         if (ret)
2107                                 goto bkey_err;
2108
2109                         bch2_btree_iter_set_pos(&iter, k.k->p);
2110                 }
2111 bkey_err:
2112                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
2113                         continue;
2114                 if (ret)
2115                         break;
2116         }
2117
2118         bch2_trans_iter_exit(trans, &iter);
2119         bch2_trans_put(trans);
2120
2121         if (ret < 0) {
2122                 bch_err_msg(ca, ret, "initializing free space");
2123                 return ret;
2124         }
2125
2126         mutex_lock(&c->sb_lock);
2127         m = bch2_members_v2_get_mut(c->disk_sb.sb, ca->dev_idx);
2128         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
2129         mutex_unlock(&c->sb_lock);
2130
2131         return 0;
2132 }
2133
2134 int bch2_fs_freespace_init(struct bch_fs *c)
2135 {
2136         int ret = 0;
2137         bool doing_init = false;
2138
2139         /*
2140          * We can crash during the device add path, so we need to check this on
2141          * every mount:
2142          */
2143
2144         for_each_member_device(c, ca) {
2145                 if (ca->mi.freespace_initialized)
2146                         continue;
2147
2148                 if (!doing_init) {
2149                         bch_info(c, "initializing freespace");
2150                         doing_init = true;
2151                 }
2152
2153                 ret = bch2_dev_freespace_init(c, ca, 0, ca->mi.nbuckets);
2154                 if (ret) {
2155                         percpu_ref_put(&ca->ref);
2156                         bch_err_fn(c, ret);
2157                         return ret;
2158                 }
2159         }
2160
2161         if (doing_init) {
2162                 mutex_lock(&c->sb_lock);
2163                 bch2_write_super(c);
2164                 mutex_unlock(&c->sb_lock);
2165                 bch_verbose(c, "done initializing freespace");
2166         }
2167
2168         return 0;
2169 }
2170
2171 /* Bucket IO clocks: */
2172
2173 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
2174                               size_t bucket_nr, int rw)
2175 {
2176         struct bch_fs *c = trans->c;
2177         struct btree_iter iter;
2178         struct bkey_i_alloc_v4 *a;
2179         u64 now;
2180         int ret = 0;
2181
2182         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
2183         ret = PTR_ERR_OR_ZERO(a);
2184         if (ret)
2185                 return ret;
2186
2187         now = atomic64_read(&c->io_clock[rw].now);
2188         if (a->v.io_time[rw] == now)
2189                 goto out;
2190
2191         a->v.io_time[rw] = now;
2192
2193         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
2194                 bch2_trans_commit(trans, NULL, NULL, 0);
2195 out:
2196         bch2_trans_iter_exit(trans, &iter);
2197         return ret;
2198 }
2199
2200 /* Startup/shutdown (ro/rw): */
2201
2202 void bch2_recalc_capacity(struct bch_fs *c)
2203 {
2204         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
2205         unsigned bucket_size_max = 0;
2206         unsigned long ra_pages = 0;
2207
2208         lockdep_assert_held(&c->state_lock);
2209
2210         for_each_online_member(c, ca) {
2211                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
2212
2213                 ra_pages += bdi->ra_pages;
2214         }
2215
2216         bch2_set_ra_pages(c, ra_pages);
2217
2218         for_each_rw_member(c, ca) {
2219                 u64 dev_reserve = 0;
2220
2221                 /*
2222                  * We need to reserve buckets (from the number
2223                  * of currently available buckets) against
2224                  * foreground writes so that mainly copygc can
2225                  * make forward progress.
2226                  *
2227                  * We need enough to refill the various reserves
2228                  * from scratch - copygc will use its entire
2229                  * reserve all at once, then run against when
2230                  * its reserve is refilled (from the formerly
2231                  * available buckets).
2232                  *
2233                  * This reserve is just used when considering if
2234                  * allocations for foreground writes must wait -
2235                  * not -ENOSPC calculations.
2236                  */
2237
2238                 dev_reserve += ca->nr_btree_reserve * 2;
2239                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
2240
2241                 dev_reserve += 1;       /* btree write point */
2242                 dev_reserve += 1;       /* copygc write point */
2243                 dev_reserve += 1;       /* rebalance write point */
2244
2245                 dev_reserve *= ca->mi.bucket_size;
2246
2247                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
2248                                              ca->mi.first_bucket);
2249
2250                 reserved_sectors += dev_reserve * 2;
2251
2252                 bucket_size_max = max_t(unsigned, bucket_size_max,
2253                                         ca->mi.bucket_size);
2254         }
2255
2256         gc_reserve = c->opts.gc_reserve_bytes
2257                 ? c->opts.gc_reserve_bytes >> 9
2258                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
2259
2260         reserved_sectors = max(gc_reserve, reserved_sectors);
2261
2262         reserved_sectors = min(reserved_sectors, capacity);
2263
2264         c->capacity = capacity - reserved_sectors;
2265
2266         c->bucket_size_max = bucket_size_max;
2267
2268         /* Wake up case someone was waiting for buckets */
2269         closure_wake_up(&c->freelist_wait);
2270 }
2271
2272 u64 bch2_min_rw_member_capacity(struct bch_fs *c)
2273 {
2274         u64 ret = U64_MAX;
2275
2276         for_each_rw_member(c, ca)
2277                 ret = min(ret, ca->mi.nbuckets * ca->mi.bucket_size);
2278         return ret;
2279 }
2280
2281 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
2282 {
2283         struct open_bucket *ob;
2284         bool ret = false;
2285
2286         for (ob = c->open_buckets;
2287              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
2288              ob++) {
2289                 spin_lock(&ob->lock);
2290                 if (ob->valid && !ob->on_partial_list &&
2291                     ob->dev == ca->dev_idx)
2292                         ret = true;
2293                 spin_unlock(&ob->lock);
2294         }
2295
2296         return ret;
2297 }
2298
2299 /* device goes ro: */
2300 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
2301 {
2302         unsigned i;
2303
2304         /* First, remove device from allocation groups: */
2305
2306         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2307                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
2308
2309         /*
2310          * Capacity is calculated based off of devices in allocation groups:
2311          */
2312         bch2_recalc_capacity(c);
2313
2314         bch2_open_buckets_stop(c, ca, false);
2315
2316         /*
2317          * Wake up threads that were blocked on allocation, so they can notice
2318          * the device can no longer be removed and the capacity has changed:
2319          */
2320         closure_wake_up(&c->freelist_wait);
2321
2322         /*
2323          * journal_res_get() can block waiting for free space in the journal -
2324          * it needs to notice there may not be devices to allocate from anymore:
2325          */
2326         wake_up(&c->journal.wait);
2327
2328         /* Now wait for any in flight writes: */
2329
2330         closure_wait_event(&c->open_buckets_wait,
2331                            !bch2_dev_has_open_write_point(c, ca));
2332 }
2333
2334 /* device goes rw: */
2335 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
2336 {
2337         unsigned i;
2338
2339         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
2340                 if (ca->mi.data_allowed & (1 << i))
2341                         set_bit(ca->dev_idx, c->rw_devs[i].d);
2342 }
2343
2344 void bch2_fs_allocator_background_exit(struct bch_fs *c)
2345 {
2346         darray_exit(&c->discard_buckets_in_flight);
2347 }
2348
2349 void bch2_fs_allocator_background_init(struct bch_fs *c)
2350 {
2351         spin_lock_init(&c->freelist_lock);
2352         mutex_init(&c->discard_buckets_in_flight_lock);
2353         INIT_WORK(&c->discard_work, bch2_do_discards_work);
2354         INIT_WORK(&c->discard_fast_work, bch2_do_discards_fast_work);
2355         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
2356 }