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

   1 // SPDX-License-Identifier: GPL-2.0
   2 #include "bcachefs.h"
   3 #include "alloc_background.h"
   4 #include "alloc_foreground.h"
   5 #include "backpointers.h"
   6 #include "btree_cache.h"
   7 #include "btree_io.h"
   8 #include "btree_key_cache.h"
   9 #include "btree_update.h"
  10 #include "btree_update_interior.h"
  11 #include "btree_gc.h"
  12 #include "buckets.h"
  13 #include "buckets_waiting_for_journal.h"
  14 #include "clock.h"
  15 #include "debug.h"
  16 #include "ec.h"
  17 #include "error.h"
  18 #include "lru.h"
  19 #include "recovery.h"
  20 #include "varint.h"
  21
  22 #include <linux/kthread.h>
  23 #include <linux/math64.h>
  24 #include <linux/random.h>
  25 #include <linux/rculist.h>
  26 #include <linux/rcupdate.h>
  27 #include <linux/sched/task.h>
  28 #include <linux/sort.h>
  29 #include <trace/events/bcachefs.h>
  30
  31 /* Persistent alloc info: */
  32
  33 static const unsigned BCH_ALLOC_V1_FIELD_BYTES[] = {
  34 #define x(name, bits) [BCH_ALLOC_FIELD_V1_##name] = bits / 8,
  35         BCH_ALLOC_FIELDS_V1()
  36 #undef x
  37 };
  38
  39 struct bkey_alloc_unpacked {
  40         u64             journal_seq;
  41         u8              gen;
  42         u8              oldest_gen;
  43         u8              data_type;
  44         bool            need_discard:1;
  45         bool            need_inc_gen:1;
  46 #define x(_name, _bits) u##_bits _name;
  47         BCH_ALLOC_FIELDS_V2()
  48 #undef  x
  49 };
  50
  51 static inline u64 alloc_field_v1_get(const struct bch_alloc *a,
  52                                      const void **p, unsigned field)
  53 {
  54         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  55         u64 v;
  56
  57         if (!(a->fields & (1 << field)))
  58                 return 0;
  59
  60         switch (bytes) {
  61         case 1:
  62                 v = *((const u8 *) *p);
  63                 break;
  64         case 2:
  65                 v = le16_to_cpup(*p);
  66                 break;
  67         case 4:
  68                 v = le32_to_cpup(*p);
  69                 break;
  70         case 8:
  71                 v = le64_to_cpup(*p);
  72                 break;
  73         default:
  74                 BUG();
  75         }
  76
  77         *p += bytes;
  78         return v;
  79 }
  80
  81 static inline void alloc_field_v1_put(struct bkey_i_alloc *a, void **p,
  82                                       unsigned field, u64 v)
  83 {
  84         unsigned bytes = BCH_ALLOC_V1_FIELD_BYTES[field];
  85
  86         if (!v)
  87                 return;
  88
  89         a->v.fields |= 1 << field;
  90
  91         switch (bytes) {
  92         case 1:
  93                 *((u8 *) *p) = v;
  94                 break;
  95         case 2:
  96                 *((__le16 *) *p) = cpu_to_le16(v);
  97                 break;
  98         case 4:
  99                 *((__le32 *) *p) = cpu_to_le32(v);
 100                 break;
 101         case 8:
 102                 *((__le64 *) *p) = cpu_to_le64(v);
 103                 break;
 104         default:
 105                 BUG();
 106         }
 107
 108         *p += bytes;
 109 }
 110
 111 static void bch2_alloc_unpack_v1(struct bkey_alloc_unpacked *out,
 112                                  struct bkey_s_c k)
 113 {
 114         const struct bch_alloc *in = bkey_s_c_to_alloc(k).v;
 115         const void *d = in->data;
 116         unsigned idx = 0;
 117
 118         out->gen = in->gen;
 119
 120 #define x(_name, _bits) out->_name = alloc_field_v1_get(in, &d, idx++);
 121         BCH_ALLOC_FIELDS_V1()
 122 #undef  x
 123 }
 124
 125 static int bch2_alloc_unpack_v2(struct bkey_alloc_unpacked *out,
 126                                 struct bkey_s_c k)
 127 {
 128         struct bkey_s_c_alloc_v2 a = bkey_s_c_to_alloc_v2(k);
 129         const u8 *in = a.v->data;
 130         const u8 *end = bkey_val_end(a);
 131         unsigned fieldnr = 0;
 132         int ret;
 133         u64 v;
 134
 135         out->gen        = a.v->gen;
 136         out->oldest_gen = a.v->oldest_gen;
 137         out->data_type  = a.v->data_type;
 138
 139 #define x(_name, _bits)                                                 \
 140         if (fieldnr < a.v->nr_fields) {                                 \
 141                 ret = bch2_varint_decode_fast(in, end, &v);             \
 142                 if (ret < 0)                                            \
 143                         return ret;                                     \
 144                 in += ret;                                              \
 145         } else {                                                        \
 146                 v = 0;                                                  \
 147         }                                                               \
 148         out->_name = v;                                                 \
 149         if (v != out->_name)                                            \
 150                 return -1;                                              \
 151         fieldnr++;
 152
 153         BCH_ALLOC_FIELDS_V2()
 154 #undef  x
 155         return 0;
 156 }
 157
 158 static int bch2_alloc_unpack_v3(struct bkey_alloc_unpacked *out,
 159                                 struct bkey_s_c k)
 160 {
 161         struct bkey_s_c_alloc_v3 a = bkey_s_c_to_alloc_v3(k);
 162         const u8 *in = a.v->data;
 163         const u8 *end = bkey_val_end(a);
 164         unsigned fieldnr = 0;
 165         int ret;
 166         u64 v;
 167
 168         out->gen        = a.v->gen;
 169         out->oldest_gen = a.v->oldest_gen;
 170         out->data_type  = a.v->data_type;
 171         out->need_discard = BCH_ALLOC_V3_NEED_DISCARD(a.v);
 172         out->need_inc_gen = BCH_ALLOC_V3_NEED_INC_GEN(a.v);
 173         out->journal_seq = le64_to_cpu(a.v->journal_seq);
 174
 175 #define x(_name, _bits)                                                 \
 176         if (fieldnr < a.v->nr_fields) {                                 \
 177                 ret = bch2_varint_decode_fast(in, end, &v);             \
 178                 if (ret < 0)                                            \
 179                         return ret;                                     \
 180                 in += ret;                                              \
 181         } else {                                                        \
 182                 v = 0;                                                  \
 183         }                                                               \
 184         out->_name = v;                                                 \
 185         if (v != out->_name)                                            \
 186                 return -1;                                              \
 187         fieldnr++;
 188
 189         BCH_ALLOC_FIELDS_V2()
 190 #undef  x
 191         return 0;
 192 }
 193
 194 static struct bkey_alloc_unpacked bch2_alloc_unpack(struct bkey_s_c k)
 195 {
 196         struct bkey_alloc_unpacked ret = { .gen = 0 };
 197
 198         switch (k.k->type) {
 199         case KEY_TYPE_alloc:
 200                 bch2_alloc_unpack_v1(&ret, k);
 201                 break;
 202         case KEY_TYPE_alloc_v2:
 203                 bch2_alloc_unpack_v2(&ret, k);
 204                 break;
 205         case KEY_TYPE_alloc_v3:
 206                 bch2_alloc_unpack_v3(&ret, k);
 207                 break;
 208         }
 209
 210         return ret;
 211 }
 212
 213 struct bkey_i_alloc_v4 *
 214 bch2_trans_start_alloc_update(struct btree_trans *trans, struct btree_iter *iter,
 215                               struct bpos pos)
 216 {
 217         struct bkey_s_c k;
 218         struct bkey_i_alloc_v4 *a;
 219         int ret;
 220
 221         bch2_trans_iter_init(trans, iter, BTREE_ID_alloc, pos,
 222                              BTREE_ITER_WITH_UPDATES|
 223                              BTREE_ITER_CACHED|
 224                              BTREE_ITER_INTENT);
 225         k = bch2_btree_iter_peek_slot(iter);
 226         ret = bkey_err(k);
 227         if (ret) {
 228                 bch2_trans_iter_exit(trans, iter);
 229                 return ERR_PTR(ret);
 230         }
 231
 232         a = bch2_alloc_to_v4_mut(trans, k);
 233         if (IS_ERR(a))
 234                 bch2_trans_iter_exit(trans, iter);
 235         return a;
 236 }
 237
 238 static unsigned bch_alloc_v1_val_u64s(const struct bch_alloc *a)
 239 {
 240         unsigned i, bytes = offsetof(struct bch_alloc, data);
 241
 242         for (i = 0; i < ARRAY_SIZE(BCH_ALLOC_V1_FIELD_BYTES); i++)
 243                 if (a->fields & (1 << i))
 244                         bytes += BCH_ALLOC_V1_FIELD_BYTES[i];
 245
 246         return DIV_ROUND_UP(bytes, sizeof(u64));
 247 }
 248
 249 int bch2_alloc_v1_invalid(const struct bch_fs *c, struct bkey_s_c k,
 250                           int rw, struct printbuf *err)
 251 {
 252         struct bkey_s_c_alloc a = bkey_s_c_to_alloc(k);
 253
 254         /* allow for unknown fields */
 255         if (bkey_val_u64s(a.k) < bch_alloc_v1_val_u64s(a.v)) {
 256                 prt_printf(err, "incorrect value size (%zu < %u)",
 257                        bkey_val_u64s(a.k), bch_alloc_v1_val_u64s(a.v));
 258                 return -EINVAL;
 259         }
 260
 261         return 0;
 262 }
 263
 264 int bch2_alloc_v2_invalid(const struct bch_fs *c, struct bkey_s_c k,
 265                           int rw, struct printbuf *err)
 266 {
 267         struct bkey_alloc_unpacked u;
 268
 269         if (bch2_alloc_unpack_v2(&u, k)) {
 270                 prt_printf(err, "unpack error");
 271                 return -EINVAL;
 272         }
 273
 274         return 0;
 275 }
 276
 277 int bch2_alloc_v3_invalid(const struct bch_fs *c, struct bkey_s_c k,
 278                           int rw, struct printbuf *err)
 279 {
 280         struct bkey_alloc_unpacked u;
 281
 282         if (bch2_alloc_unpack_v3(&u, k)) {
 283                 prt_printf(err, "unpack error");
 284                 return -EINVAL;
 285         }
 286
 287         return 0;
 288 }
 289
 290 int bch2_alloc_v4_invalid(const struct bch_fs *c, struct bkey_s_c k,
 291                           int rw, struct printbuf *err)
 292 {
 293         struct bkey_s_c_alloc_v4 a = bkey_s_c_to_alloc_v4(k);
 294
 295         if (alloc_v4_u64s(a.v) != bkey_val_u64s(k.k)) {
 296                 prt_printf(err, "bad val size (%lu != %u)",
 297                        bkey_val_u64s(k.k), alloc_v4_u64s(a.v));
 298                 return -EINVAL;
 299         }
 300
 301         if (!BCH_ALLOC_V4_BACKPOINTERS_START(a.v) &&
 302             BCH_ALLOC_V4_NR_BACKPOINTERS(a.v)) {
 303                 prt_printf(err, "invalid backpointers_start");
 304                 return -EINVAL;
 305         }
 306
 307         if (rw == WRITE) {
 308                 if (alloc_data_type(*a.v, a.v->data_type) != a.v->data_type) {
 309                         prt_printf(err, "invalid data type (got %u should be %u)",
 310                                a.v->data_type, alloc_data_type(*a.v, a.v->data_type));
 311                         return -EINVAL;
 312                 }
 313
 314                 switch (a.v->data_type) {
 315                 case BCH_DATA_free:
 316                 case BCH_DATA_need_gc_gens:
 317                 case BCH_DATA_need_discard:
 318                         if (a.v->dirty_sectors ||
 319                             a.v->cached_sectors ||
 320                             a.v->stripe) {
 321                                 prt_printf(err, "empty data type free but have data");
 322                                 return -EINVAL;
 323                         }
 324                         break;
 325                 case BCH_DATA_sb:
 326                 case BCH_DATA_journal:
 327                 case BCH_DATA_btree:
 328                 case BCH_DATA_user:
 329                 case BCH_DATA_parity:
 330                         if (!a.v->dirty_sectors) {
 331                                 prt_printf(err, "data_type %s but dirty_sectors==0",
 332                                        bch2_data_types[a.v->data_type]);
 333                                 return -EINVAL;
 334                         }
 335                         break;
 336                 case BCH_DATA_cached:
 337                         if (!a.v->cached_sectors ||
 338                             a.v->dirty_sectors ||
 339                             a.v->stripe) {
 340                                 prt_printf(err, "data type inconsistency");
 341                                 return -EINVAL;
 342                         }
 343
 344                         if (!a.v->io_time[READ] &&
 345                             test_bit(BCH_FS_CHECK_ALLOC_TO_LRU_REFS_DONE, &c->flags)) {
 346                                 prt_printf(err, "cached bucket with read_time == 0");
 347                                 return -EINVAL;
 348                         }
 349                         break;
 350                 case BCH_DATA_stripe:
 351                         if (!a.v->stripe) {
 352                                 prt_printf(err, "data_type %s but stripe==0",
 353                                        bch2_data_types[a.v->data_type]);
 354                                 return -EINVAL;
 355                         }
 356                         break;
 357                 }
 358         }
 359
 360         return 0;
 361 }
 362
 363 static inline u64 swab40(u64 x)
 364 {
 365         return (((x & 0x00000000ffULL) << 32)|
 366                 ((x & 0x000000ff00ULL) << 16)|
 367                 ((x & 0x0000ff0000ULL) >>  0)|
 368                 ((x & 0x00ff000000ULL) >> 16)|
 369                 ((x & 0xff00000000ULL) >> 32));
 370 }
 371
 372 void bch2_alloc_v4_swab(struct bkey_s k)
 373 {
 374         struct bch_alloc_v4 *a = bkey_s_to_alloc_v4(k).v;
 375         struct bch_backpointer *bp, *bps;
 376
 377         a->journal_seq          = swab64(a->journal_seq);
 378         a->flags                = swab32(a->flags);
 379         a->dirty_sectors        = swab32(a->dirty_sectors);
 380         a->cached_sectors       = swab32(a->cached_sectors);
 381         a->io_time[0]           = swab64(a->io_time[0]);
 382         a->io_time[1]           = swab64(a->io_time[1]);
 383         a->stripe               = swab32(a->stripe);
 384         a->nr_external_backpointers = swab32(a->nr_external_backpointers);
 385
 386         bps = alloc_v4_backpointers(a);
 387         for (bp = bps; bp < bps + BCH_ALLOC_V4_NR_BACKPOINTERS(a); bp++) {
 388                 bp->bucket_offset       = swab40(bp->bucket_offset);
 389                 bp->bucket_len          = swab32(bp->bucket_len);
 390                 bch2_bpos_swab(&bp->pos);
 391         }
 392 }
 393
 394 void bch2_alloc_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 395 {
 396         struct bch_alloc_v4 _a;
 397         const struct bch_alloc_v4 *a = &_a;
 398         const struct bch_backpointer *bps;
 399         unsigned i;
 400
 401         if (k.k->type == KEY_TYPE_alloc_v4)
 402                 a = bkey_s_c_to_alloc_v4(k).v;
 403         else
 404                 bch2_alloc_to_v4(k, &_a);
 405
 406         prt_newline(out);
 407         printbuf_indent_add(out, 2);
 408
 409         prt_printf(out, "gen %u oldest_gen %u data_type %s",
 410                a->gen, a->oldest_gen, bch2_data_types[a->data_type]);
 411         prt_newline(out);
 412         prt_printf(out, "journal_seq       %llu",       a->journal_seq);
 413         prt_newline(out);
 414         prt_printf(out, "need_discard      %llu",       BCH_ALLOC_V4_NEED_DISCARD(a));
 415         prt_newline(out);
 416         prt_printf(out, "need_inc_gen      %llu",       BCH_ALLOC_V4_NEED_INC_GEN(a));
 417         prt_newline(out);
 418         prt_printf(out, "dirty_sectors     %u", a->dirty_sectors);
 419         prt_newline(out);
 420         prt_printf(out, "cached_sectors    %u", a->cached_sectors);
 421         prt_newline(out);
 422         prt_printf(out, "stripe            %u", a->stripe);
 423         prt_newline(out);
 424         prt_printf(out, "stripe_redundancy %u", a->stripe_redundancy);
 425         prt_newline(out);
 426         prt_printf(out, "io_time[READ]     %llu",       a->io_time[READ]);
 427         prt_newline(out);
 428         prt_printf(out, "io_time[WRITE]    %llu",       a->io_time[WRITE]);
 429         prt_newline(out);
 430         prt_printf(out, "backpointers:     %llu",       BCH_ALLOC_V4_NR_BACKPOINTERS(a));
 431         printbuf_indent_add(out, 2);
 432
 433         bps = alloc_v4_backpointers_c(a);
 434         for (i = 0; i < BCH_ALLOC_V4_NR_BACKPOINTERS(a); i++) {
 435                 prt_newline(out);
 436                 bch2_backpointer_to_text(out, &bps[i]);
 437         }
 438
 439         printbuf_indent_sub(out, 4);
 440 }
 441
 442 void bch2_alloc_to_v4(struct bkey_s_c k, struct bch_alloc_v4 *out)
 443 {
 444         if (k.k->type == KEY_TYPE_alloc_v4) {
 445                 int d;
 446
 447                 *out = *bkey_s_c_to_alloc_v4(k).v;
 448
 449                 d = (int) BCH_ALLOC_V4_U64s -
 450                         (int) (BCH_ALLOC_V4_BACKPOINTERS_START(out) ?: BCH_ALLOC_V4_U64s_V0);
 451                 if (unlikely(d > 0)) {
 452                         memset((u64 *) out + BCH_ALLOC_V4_BACKPOINTERS_START(out),
 453                                0,
 454                                d * sizeof(u64));
 455                         SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 456                 }
 457         } else {
 458                 struct bkey_alloc_unpacked u = bch2_alloc_unpack(k);
 459
 460                 *out = (struct bch_alloc_v4) {
 461                         .journal_seq            = u.journal_seq,
 462                         .flags                  = u.need_discard,
 463                         .gen                    = u.gen,
 464                         .oldest_gen             = u.oldest_gen,
 465                         .data_type              = u.data_type,
 466                         .stripe_redundancy      = u.stripe_redundancy,
 467                         .dirty_sectors          = u.dirty_sectors,
 468                         .cached_sectors         = u.cached_sectors,
 469                         .io_time[READ]          = u.read_time,
 470                         .io_time[WRITE]         = u.write_time,
 471                         .stripe                 = u.stripe,
 472                 };
 473
 474                 SET_BCH_ALLOC_V4_BACKPOINTERS_START(out, BCH_ALLOC_V4_U64s);
 475         }
 476 }
 477
 478 struct bkey_i_alloc_v4 *bch2_alloc_to_v4_mut(struct btree_trans *trans, struct bkey_s_c k)
 479 {
 480         unsigned bytes = k.k->type == KEY_TYPE_alloc_v4
 481                 ? bkey_bytes(k.k)
 482                 : sizeof(struct bkey_i_alloc_v4);
 483         struct bkey_i_alloc_v4 *ret;
 484
 485         /*
 486          * Reserve space for one more backpointer here:
 487          * Not sketchy at doing it this way, nope...
 488          */
 489         ret = bch2_trans_kmalloc(trans, bytes + sizeof(struct bch_backpointer));
 490         if (IS_ERR(ret))
 491                 return ret;
 492
 493         if (k.k->type == KEY_TYPE_alloc_v4) {
 494                 bkey_reassemble(&ret->k_i, k);
 495
 496                 if (BCH_ALLOC_V4_BACKPOINTERS_START(&ret->v) < BCH_ALLOC_V4_U64s) {
 497                         struct bch_backpointer *src, *dst;
 498
 499                         src = alloc_v4_backpointers(&ret->v);
 500                         SET_BCH_ALLOC_V4_BACKPOINTERS_START(&ret->v, BCH_ALLOC_V4_U64s);
 501                         dst = alloc_v4_backpointers(&ret->v);
 502
 503                         memmove(dst, src, BCH_ALLOC_V4_NR_BACKPOINTERS(&ret->v) *
 504                                 sizeof(struct bch_backpointer));
 505                         memset(src, 0, dst - src);
 506                         set_alloc_v4_u64s(ret);
 507                 }
 508         } else {
 509                 bkey_alloc_v4_init(&ret->k_i);
 510                 ret->k.p = k.k->p;
 511                 bch2_alloc_to_v4(k, &ret->v);
 512         }
 513         return ret;
 514 }
 515
 516 int bch2_alloc_read(struct bch_fs *c)
 517 {
 518         struct btree_trans trans;
 519         struct btree_iter iter;
 520         struct bkey_s_c k;
 521         struct bch_alloc_v4 a;
 522         struct bch_dev *ca;
 523         int ret;
 524
 525         bch2_trans_init(&trans, c, 0, 0);
 526
 527         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
 528                            BTREE_ITER_PREFETCH, k, ret) {
 529                 /*
 530                  * Not a fsck error because this is checked/repaired by
 531                  * bch2_check_alloc_key() which runs later:
 532                  */
 533                 if (!bch2_dev_bucket_exists(c, k.k->p))
 534                         continue;
 535
 536                 ca = bch_dev_bkey_exists(c, k.k->p.inode);
 537                 bch2_alloc_to_v4(k, &a);
 538
 539                 *bucket_gen(ca, k.k->p.offset) = a.gen;
 540         }
 541         bch2_trans_iter_exit(&trans, &iter);
 542
 543         bch2_trans_exit(&trans);
 544
 545         if (ret)
 546                 bch_err(c, "error reading alloc info: %i", ret);
 547
 548         return ret;
 549 }
 550
 551 /* Free space/discard btree: */
 552
 553 static int bch2_bucket_do_index(struct btree_trans *trans,
 554                                 struct bkey_s_c alloc_k,
 555                                 const struct bch_alloc_v4 *a,
 556                                 bool set)
 557 {
 558         struct bch_fs *c = trans->c;
 559         struct bch_dev *ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 560         struct btree_iter iter;
 561         struct bkey_s_c old;
 562         struct bkey_i *k;
 563         enum btree_id btree;
 564         enum bch_bkey_type old_type = !set ? KEY_TYPE_set : KEY_TYPE_deleted;
 565         enum bch_bkey_type new_type =  set ? KEY_TYPE_set : KEY_TYPE_deleted;
 566         struct printbuf buf = PRINTBUF;
 567         int ret;
 568
 569         if (a->data_type != BCH_DATA_free &&
 570             a->data_type != BCH_DATA_need_discard)
 571                 return 0;
 572
 573         k = bch2_trans_kmalloc(trans, sizeof(*k));
 574         if (IS_ERR(k))
 575                 return PTR_ERR(k);
 576
 577         bkey_init(&k->k);
 578         k->k.type = new_type;
 579
 580         switch (a->data_type) {
 581         case BCH_DATA_free:
 582                 btree = BTREE_ID_freespace;
 583                 k->k.p = alloc_freespace_pos(alloc_k.k->p, *a);
 584                 bch2_key_resize(&k->k, 1);
 585                 break;
 586         case BCH_DATA_need_discard:
 587                 btree = BTREE_ID_need_discard;
 588                 k->k.p = alloc_k.k->p;
 589                 break;
 590         default:
 591                 return 0;
 592         }
 593
 594         bch2_trans_iter_init(trans, &iter, btree,
 595                              bkey_start_pos(&k->k),
 596                              BTREE_ITER_INTENT);
 597         old = bch2_btree_iter_peek_slot(&iter);
 598         ret = bkey_err(old);
 599         if (ret)
 600                 goto err;
 601
 602         if (ca->mi.freespace_initialized &&
 603             bch2_trans_inconsistent_on(old.k->type != old_type, trans,
 604                         "incorrect key when %s %s btree (got %s should be %s)\n"
 605                         "  for %s",
 606                         set ? "setting" : "clearing",
 607                         bch2_btree_ids[btree],
 608                         bch2_bkey_types[old.k->type],
 609                         bch2_bkey_types[old_type],
 610                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 611                 ret = -EIO;
 612                 goto err;
 613         }
 614
 615         ret = bch2_trans_update(trans, &iter, k, 0);
 616 err:
 617         bch2_trans_iter_exit(trans, &iter);
 618         printbuf_exit(&buf);
 619         return ret;
 620 }
 621
 622 int bch2_trans_mark_alloc(struct btree_trans *trans,
 623                           enum btree_id btree_id, unsigned level,
 624                           struct bkey_s_c old, struct bkey_i *new,
 625                           unsigned flags)
 626 {
 627         struct bch_fs *c = trans->c;
 628         struct bch_alloc_v4 old_a, *new_a;
 629         u64 old_lru, new_lru;
 630         int ret = 0;
 631
 632         /*
 633          * Deletion only happens in the device removal path, with
 634          * BTREE_TRIGGER_NORUN:
 635          */
 636         BUG_ON(new->k.type != KEY_TYPE_alloc_v4);
 637
 638         bch2_alloc_to_v4(old, &old_a);
 639         new_a = &bkey_i_to_alloc_v4(new)->v;
 640
 641         new_a->data_type = alloc_data_type(*new_a, new_a->data_type);
 642
 643         if (new_a->dirty_sectors > old_a.dirty_sectors ||
 644             new_a->cached_sectors > old_a.cached_sectors) {
 645                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 646                 new_a->io_time[WRITE]= max_t(u64, 1, atomic64_read(&c->io_clock[WRITE].now));
 647                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, true);
 648                 SET_BCH_ALLOC_V4_NEED_DISCARD(new_a, true);
 649         }
 650
 651         if (data_type_is_empty(new_a->data_type) &&
 652             BCH_ALLOC_V4_NEED_INC_GEN(new_a) &&
 653             !bch2_bucket_is_open_safe(c, new->k.p.inode, new->k.p.offset)) {
 654                 new_a->gen++;
 655                 SET_BCH_ALLOC_V4_NEED_INC_GEN(new_a, false);
 656         }
 657
 658         if (old_a.data_type != new_a->data_type ||
 659             (new_a->data_type == BCH_DATA_free &&
 660              alloc_freespace_genbits(old_a) != alloc_freespace_genbits(*new_a))) {
 661                 ret =   bch2_bucket_do_index(trans, old, &old_a, false) ?:
 662                         bch2_bucket_do_index(trans, bkey_i_to_s_c(new), new_a, true);
 663                 if (ret)
 664                         return ret;
 665         }
 666
 667         if (new_a->data_type == BCH_DATA_cached &&
 668             !new_a->io_time[READ])
 669                 new_a->io_time[READ] = max_t(u64, 1, atomic64_read(&c->io_clock[READ].now));
 670
 671         old_lru = alloc_lru_idx(old_a);
 672         new_lru = alloc_lru_idx(*new_a);
 673
 674         if (old_lru != new_lru) {
 675                 ret = bch2_lru_change(trans, new->k.p.inode, new->k.p.offset,
 676                                       old_lru, &new_lru, old);
 677                 if (ret)
 678                         return ret;
 679
 680                 if (new_a->data_type == BCH_DATA_cached)
 681                         new_a->io_time[READ] = new_lru;
 682         }
 683
 684         return 0;
 685 }
 686
 687 static int bch2_check_alloc_key(struct btree_trans *trans,
 688                                 struct btree_iter *alloc_iter,
 689                                 struct btree_iter *discard_iter,
 690                                 struct btree_iter *freespace_iter)
 691 {
 692         struct bch_fs *c = trans->c;
 693         struct bch_dev *ca;
 694         struct bch_alloc_v4 a;
 695         unsigned discard_key_type, freespace_key_type;
 696         struct bkey_s_c alloc_k, k;
 697         struct printbuf buf = PRINTBUF;
 698         int ret;
 699
 700         alloc_k = bch2_dev_bucket_exists(c, alloc_iter->pos)
 701                 ? bch2_btree_iter_peek_slot(alloc_iter)
 702                 : bch2_btree_iter_peek(alloc_iter);
 703         if (!alloc_k.k)
 704                 return 1;
 705
 706         ret = bkey_err(alloc_k);
 707         if (ret)
 708                 return ret;
 709
 710         if (fsck_err_on(!bch2_dev_bucket_exists(c, alloc_k.k->p), c,
 711                         "alloc key for invalid device:bucket %llu:%llu",
 712                         alloc_k.k->p.inode, alloc_k.k->p.offset))
 713                 return bch2_btree_delete_at(trans, alloc_iter, 0);
 714
 715         ca = bch_dev_bkey_exists(c, alloc_k.k->p.inode);
 716         if (!ca->mi.freespace_initialized)
 717                 return 0;
 718
 719         bch2_alloc_to_v4(alloc_k, &a);
 720
 721         discard_key_type = a.data_type == BCH_DATA_need_discard
 722                 ? KEY_TYPE_set : 0;
 723         freespace_key_type = a.data_type == BCH_DATA_free
 724                 ? KEY_TYPE_set : 0;
 725
 726         bch2_btree_iter_set_pos(discard_iter, alloc_k.k->p);
 727         bch2_btree_iter_set_pos(freespace_iter, alloc_freespace_pos(alloc_k.k->p, a));
 728
 729         k = bch2_btree_iter_peek_slot(discard_iter);
 730         ret = bkey_err(k);
 731         if (ret)
 732                 goto err;
 733
 734         if (fsck_err_on(k.k->type != discard_key_type, c,
 735                         "incorrect key in need_discard btree (got %s should be %s)\n"
 736                         "  %s",
 737                         bch2_bkey_types[k.k->type],
 738                         bch2_bkey_types[discard_key_type],
 739                         (bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 740                 struct bkey_i *update =
 741                         bch2_trans_kmalloc(trans, sizeof(*update));
 742
 743                 ret = PTR_ERR_OR_ZERO(update);
 744                 if (ret)
 745                         goto err;
 746
 747                 bkey_init(&update->k);
 748                 update->k.type  = discard_key_type;
 749                 update->k.p     = discard_iter->pos;
 750
 751                 ret = bch2_trans_update(trans, discard_iter, update, 0);
 752                 if (ret)
 753                         goto err;
 754         }
 755
 756         k = bch2_btree_iter_peek_slot(freespace_iter);
 757         ret = bkey_err(k);
 758         if (ret)
 759                 goto err;
 760
 761         if (fsck_err_on(k.k->type != freespace_key_type, c,
 762                         "incorrect key in freespace btree (got %s should be %s)\n"
 763                         "  %s",
 764                         bch2_bkey_types[k.k->type],
 765                         bch2_bkey_types[freespace_key_type],
 766                         (printbuf_reset(&buf),
 767                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf))) {
 768                 struct bkey_i *update =
 769                         bch2_trans_kmalloc(trans, sizeof(*update));
 770
 771                 ret = PTR_ERR_OR_ZERO(update);
 772                 if (ret)
 773                         goto err;
 774
 775                 bkey_init(&update->k);
 776                 update->k.type  = freespace_key_type;
 777                 update->k.p     = freespace_iter->pos;
 778                 bch2_key_resize(&update->k, 1);
 779
 780                 ret = bch2_trans_update(trans, freespace_iter, update, 0);
 781                 if (ret)
 782                         goto err;
 783         }
 784 err:
 785 fsck_err:
 786         printbuf_exit(&buf);
 787         return ret;
 788 }
 789
 790 static int bch2_check_discard_freespace_key(struct btree_trans *trans,
 791                                             struct btree_iter *iter)
 792 {
 793         struct bch_fs *c = trans->c;
 794         struct btree_iter alloc_iter;
 795         struct bkey_s_c k, freespace_k;
 796         struct bch_alloc_v4 a;
 797         u64 genbits;
 798         struct bpos pos;
 799         enum bch_data_type state = iter->btree_id == BTREE_ID_need_discard
 800                 ? BCH_DATA_need_discard
 801                 : BCH_DATA_free;
 802         struct printbuf buf = PRINTBUF;
 803         int ret;
 804
 805         freespace_k = bch2_btree_iter_peek(iter);
 806         if (!freespace_k.k)
 807                 return 1;
 808
 809         ret = bkey_err(freespace_k);
 810         if (ret)
 811                 return ret;
 812
 813         pos = iter->pos;
 814         pos.offset &= ~(~0ULL << 56);
 815         genbits = iter->pos.offset & (~0ULL << 56);
 816
 817         bch2_trans_iter_init(trans, &alloc_iter, BTREE_ID_alloc, pos, 0);
 818
 819         if (fsck_err_on(!bch2_dev_bucket_exists(c, pos), c,
 820                         "entry in %s btree for nonexistant dev:bucket %llu:%llu",
 821                         bch2_btree_ids[iter->btree_id], pos.inode, pos.offset))
 822                 goto delete;
 823
 824         k = bch2_btree_iter_peek_slot(&alloc_iter);
 825         ret = bkey_err(k);
 826         if (ret)
 827                 goto err;
 828
 829         bch2_alloc_to_v4(k, &a);
 830
 831         if (fsck_err_on(a.data_type != state ||
 832                         (state == BCH_DATA_free &&
 833                          genbits != alloc_freespace_genbits(a)), c,
 834                         "%s\n  incorrectly set in %s index (free %u, genbits %llu should be %llu)",
 835                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf),
 836                         bch2_btree_ids[iter->btree_id],
 837                         a.data_type == state,
 838                         genbits >> 56, alloc_freespace_genbits(a) >> 56))
 839                 goto delete;
 840 out:
 841 err:
 842 fsck_err:
 843         bch2_trans_iter_exit(trans, &alloc_iter);
 844         printbuf_exit(&buf);
 845         return ret;
 846 delete:
 847         ret = bch2_btree_delete_extent_at(trans, iter,
 848                         iter->btree_id == BTREE_ID_freespace ? 1 : 0, 0);
 849         goto out;
 850 }
 851
 852 int bch2_check_alloc_info(struct bch_fs *c)
 853 {
 854         struct btree_trans trans;
 855         struct btree_iter iter, discard_iter, freespace_iter;
 856         int ret = 0;
 857
 858         bch2_trans_init(&trans, c, 0, 0);
 859
 860         bch2_trans_iter_init(&trans, &iter, BTREE_ID_alloc, POS_MIN,
 861                              BTREE_ITER_PREFETCH);
 862         bch2_trans_iter_init(&trans, &discard_iter, BTREE_ID_need_discard, POS_MIN,
 863                              BTREE_ITER_PREFETCH);
 864         bch2_trans_iter_init(&trans, &freespace_iter, BTREE_ID_freespace, POS_MIN,
 865                              BTREE_ITER_PREFETCH);
 866         while (1) {
 867                 ret = __bch2_trans_do(&trans, NULL, NULL,
 868                                       BTREE_INSERT_NOFAIL|
 869                                       BTREE_INSERT_LAZY_RW,
 870                         bch2_check_alloc_key(&trans, &iter,
 871                                              &discard_iter,
 872                                              &freespace_iter));
 873                 if (ret)
 874                         break;
 875
 876                 bch2_btree_iter_advance(&iter);
 877         }
 878         bch2_trans_iter_exit(&trans, &freespace_iter);
 879         bch2_trans_iter_exit(&trans, &discard_iter);
 880         bch2_trans_iter_exit(&trans, &iter);
 881
 882         if (ret < 0)
 883                 goto err;
 884
 885         bch2_trans_iter_init(&trans, &iter, BTREE_ID_need_discard, POS_MIN,
 886                              BTREE_ITER_PREFETCH);
 887         while (1) {
 888                 ret = __bch2_trans_do(&trans, NULL, NULL,
 889                                       BTREE_INSERT_NOFAIL|
 890                                       BTREE_INSERT_LAZY_RW,
 891                         bch2_check_discard_freespace_key(&trans, &iter));
 892                 if (ret)
 893                         break;
 894
 895                 bch2_btree_iter_advance(&iter);
 896         }
 897         bch2_trans_iter_exit(&trans, &iter);
 898
 899         if (ret < 0)
 900                 goto err;
 901
 902         bch2_trans_iter_init(&trans, &iter, BTREE_ID_freespace, POS_MIN,
 903                              BTREE_ITER_PREFETCH);
 904         while (1) {
 905                 ret = __bch2_trans_do(&trans, NULL, NULL,
 906                                       BTREE_INSERT_NOFAIL|
 907                                       BTREE_INSERT_LAZY_RW,
 908                         bch2_check_discard_freespace_key(&trans, &iter));
 909                 if (ret)
 910                         break;
 911
 912                 bch2_btree_iter_advance(&iter);
 913         }
 914         bch2_trans_iter_exit(&trans, &iter);
 915 err:
 916         bch2_trans_exit(&trans);
 917         return ret < 0 ? ret : 0;
 918 }
 919
 920 static int bch2_check_alloc_to_lru_ref(struct btree_trans *trans,
 921                                        struct btree_iter *alloc_iter)
 922 {
 923         struct bch_fs *c = trans->c;
 924         struct btree_iter lru_iter;
 925         struct bch_alloc_v4 a;
 926         struct bkey_s_c alloc_k, k;
 927         struct printbuf buf = PRINTBUF;
 928         struct printbuf buf2 = PRINTBUF;
 929         int ret;
 930
 931         alloc_k = bch2_btree_iter_peek(alloc_iter);
 932         if (!alloc_k.k)
 933                 return 0;
 934
 935         ret = bkey_err(alloc_k);
 936         if (ret)
 937                 return ret;
 938
 939         bch2_alloc_to_v4(alloc_k, &a);
 940
 941         if (a.data_type != BCH_DATA_cached)
 942                 return 0;
 943
 944         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
 945                              POS(alloc_k.k->p.inode, a.io_time[READ]), 0);
 946
 947         k = bch2_btree_iter_peek_slot(&lru_iter);
 948         ret = bkey_err(k);
 949         if (ret)
 950                 goto err;
 951
 952         if (fsck_err_on(!a.io_time[READ], c,
 953                         "cached bucket with read_time 0\n"
 954                         "  %s",
 955                 (printbuf_reset(&buf),
 956                  bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf)) ||
 957             fsck_err_on(k.k->type != KEY_TYPE_lru ||
 958                         le64_to_cpu(bkey_s_c_to_lru(k).v->idx) != alloc_k.k->p.offset, c,
 959                         "incorrect/missing lru entry\n"
 960                         "  %s\n"
 961                         "  %s",
 962                         (printbuf_reset(&buf),
 963                          bch2_bkey_val_to_text(&buf, c, alloc_k), buf.buf),
 964                         (bch2_bkey_val_to_text(&buf2, c, k), buf2.buf))) {
 965                 u64 read_time = a.io_time[READ];
 966
 967                 if (!a.io_time[READ])
 968                         a.io_time[READ] = atomic64_read(&c->io_clock[READ].now);
 969
 970                 ret = bch2_lru_set(trans,
 971                                    alloc_k.k->p.inode,
 972                                    alloc_k.k->p.offset,
 973                                    &a.io_time[READ]);
 974                 if (ret)
 975                         goto err;
 976
 977                 if (a.io_time[READ] != read_time) {
 978                         struct bkey_i_alloc_v4 *a_mut =
 979                                 bch2_alloc_to_v4_mut(trans, alloc_k);
 980                         ret = PTR_ERR_OR_ZERO(a_mut);
 981                         if (ret)
 982                                 goto err;
 983
 984                         a_mut->v.io_time[READ] = a.io_time[READ];
 985                         ret = bch2_trans_update(trans, alloc_iter,
 986                                                 &a_mut->k_i, BTREE_TRIGGER_NORUN);
 987                         if (ret)
 988                                 goto err;
 989                 }
 990         }
 991 err:
 992 fsck_err:
 993         bch2_trans_iter_exit(trans, &lru_iter);
 994         printbuf_exit(&buf2);
 995         printbuf_exit(&buf);
 996         return ret;
 997 }
 998
 999 int bch2_check_alloc_to_lru_refs(struct bch_fs *c)
1000 {
1001         struct btree_trans trans;
1002         struct btree_iter iter;
1003         struct bkey_s_c k;
1004         int ret = 0;
1005
1006         bch2_trans_init(&trans, c, 0, 0);
1007
1008         for_each_btree_key(&trans, iter, BTREE_ID_alloc, POS_MIN,
1009                            BTREE_ITER_PREFETCH, k, ret) {
1010                 ret = __bch2_trans_do(&trans, NULL, NULL,
1011                                       BTREE_INSERT_NOFAIL|
1012                                       BTREE_INSERT_LAZY_RW,
1013                         bch2_check_alloc_to_lru_ref(&trans, &iter));
1014                 if (ret)
1015                         break;
1016         }
1017         bch2_trans_iter_exit(&trans, &iter);
1018
1019         bch2_trans_exit(&trans);
1020         return ret < 0 ? ret : 0;
1021 }
1022
1023 static int bch2_clear_need_discard(struct btree_trans *trans, struct bpos pos,
1024                                    struct bch_dev *ca, bool *discard_done)
1025 {
1026         struct bch_fs *c = trans->c;
1027         struct btree_iter iter;
1028         struct bkey_s_c k;
1029         struct bkey_i_alloc_v4 *a;
1030         struct printbuf buf = PRINTBUF;
1031         int ret;
1032
1033         bch2_trans_iter_init(trans, &iter, BTREE_ID_alloc, pos,
1034                              BTREE_ITER_CACHED);
1035         k = bch2_btree_iter_peek_slot(&iter);
1036         ret = bkey_err(k);
1037         if (ret)
1038                 goto out;
1039
1040         a = bch2_alloc_to_v4_mut(trans, k);
1041         ret = PTR_ERR_OR_ZERO(a);
1042         if (ret)
1043                 goto out;
1044
1045         if (BCH_ALLOC_V4_NEED_INC_GEN(&a->v)) {
1046                 a->v.gen++;
1047                 SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1048                 goto write;
1049         }
1050
1051         if (bch2_trans_inconsistent_on(a->v.journal_seq > c->journal.flushed_seq_ondisk, trans,
1052                         "clearing need_discard but journal_seq %llu > flushed_seq %llu\n"
1053                         "%s",
1054                         a->v.journal_seq,
1055                         c->journal.flushed_seq_ondisk,
1056                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1057                 ret = -EIO;
1058                 goto out;
1059         }
1060
1061         if (bch2_trans_inconsistent_on(a->v.data_type != BCH_DATA_need_discard, trans,
1062                         "bucket incorrectly set in need_discard btree\n"
1063                         "%s",
1064                         (bch2_bkey_val_to_text(&buf, c, k), buf.buf))) {
1065                 ret = -EIO;
1066                 goto out;
1067         }
1068
1069         if (!*discard_done && ca->mi.discard && !c->opts.nochanges) {
1070                 /*
1071                  * This works without any other locks because this is the only
1072                  * thread that removes items from the need_discard tree
1073                  */
1074                 bch2_trans_unlock(trans);
1075                 blkdev_issue_discard(ca->disk_sb.bdev,
1076                                      k.k->p.offset * ca->mi.bucket_size,
1077                                      ca->mi.bucket_size,
1078                                      GFP_KERNEL, 0);
1079                 *discard_done = true;
1080
1081                 ret = bch2_trans_relock(trans) ? 0 : -EINTR;
1082                 if (ret)
1083                         goto out;
1084         }
1085
1086         SET_BCH_ALLOC_V4_NEED_DISCARD(&a->v, false);
1087         a->v.data_type = alloc_data_type(a->v, a->v.data_type);
1088 write:
1089         ret = bch2_trans_update(trans, &iter, &a->k_i, 0);
1090 out:
1091         bch2_trans_iter_exit(trans, &iter);
1092         printbuf_exit(&buf);
1093         return ret;
1094 }
1095
1096 static void bch2_do_discards_work(struct work_struct *work)
1097 {
1098         struct bch_fs *c = container_of(work, struct bch_fs, discard_work);
1099         struct bch_dev *ca = NULL;
1100         struct btree_trans trans;
1101         struct btree_iter iter;
1102         struct bkey_s_c k;
1103         u64 seen = 0, open = 0, need_journal_commit = 0, discarded = 0;
1104         int ret;
1105
1106         bch2_trans_init(&trans, c, 0, 0);
1107
1108         for_each_btree_key(&trans, iter, BTREE_ID_need_discard,
1109                            POS_MIN, 0, k, ret) {
1110                 bool discard_done = false;
1111
1112                 if (ca && k.k->p.inode != ca->dev_idx) {
1113                         percpu_ref_put(&ca->io_ref);
1114                         ca = NULL;
1115                 }
1116
1117                 if (!ca) {
1118                         ca = bch_dev_bkey_exists(c, k.k->p.inode);
1119                         if (!percpu_ref_tryget(&ca->io_ref)) {
1120                                 ca = NULL;
1121                                 bch2_btree_iter_set_pos(&iter, POS(k.k->p.inode + 1, 0));
1122                                 continue;
1123                         }
1124                 }
1125
1126                 seen++;
1127
1128                 if (bch2_bucket_is_open_safe(c, k.k->p.inode, k.k->p.offset)) {
1129                         open++;
1130                         continue;
1131                 }
1132
1133                 if (bch2_bucket_needs_journal_commit(&c->buckets_waiting_for_journal,
1134                                 c->journal.flushed_seq_ondisk,
1135                                 k.k->p.inode, k.k->p.offset)) {
1136                         need_journal_commit++;
1137                         continue;
1138                 }
1139
1140                 ret = __bch2_trans_do(&trans, NULL, NULL,
1141                                       BTREE_INSERT_USE_RESERVE|
1142                                       BTREE_INSERT_NOFAIL,
1143                                 bch2_clear_need_discard(&trans, k.k->p, ca, &discard_done));
1144                 if (ret)
1145                         break;
1146
1147                 this_cpu_inc(c->counters[BCH_COUNTER_bucket_discard]);
1148                 discarded++;
1149         }
1150         bch2_trans_iter_exit(&trans, &iter);
1151
1152         if (ca)
1153                 percpu_ref_put(&ca->io_ref);
1154
1155         bch2_trans_exit(&trans);
1156
1157         if (need_journal_commit * 2 > seen)
1158                 bch2_journal_flush_async(&c->journal, NULL);
1159
1160         percpu_ref_put(&c->writes);
1161
1162         trace_discard_buckets(c, seen, open, need_journal_commit, discarded, ret);
1163 }
1164
1165 void bch2_do_discards(struct bch_fs *c)
1166 {
1167         if (percpu_ref_tryget_live(&c->writes) &&
1168             !queue_work(system_long_wq, &c->discard_work))
1169                 percpu_ref_put(&c->writes);
1170 }
1171
1172 static int invalidate_one_bucket(struct btree_trans *trans, struct bch_dev *ca,
1173                                  struct bpos *bucket_pos, unsigned *cached_sectors)
1174 {
1175         struct bch_fs *c = trans->c;
1176         struct btree_iter lru_iter, alloc_iter = { NULL };
1177         struct bkey_s_c k;
1178         struct bkey_i_alloc_v4 *a;
1179         u64 bucket, idx;
1180         struct printbuf buf = PRINTBUF;
1181         int ret;
1182
1183         bch2_trans_iter_init(trans, &lru_iter, BTREE_ID_lru,
1184                              POS(ca->dev_idx, 0), 0);
1185 next_lru:
1186         k = bch2_btree_iter_peek(&lru_iter);
1187         ret = bkey_err(k);
1188         if (ret)
1189                 goto out;
1190
1191         if (!k.k || k.k->p.inode != ca->dev_idx) {
1192                 ret = 1;
1193                 goto out;
1194         }
1195
1196         if (k.k->type != KEY_TYPE_lru) {
1197                 prt_printf(&buf, "non lru key in lru btree:\n  ");
1198                 bch2_bkey_val_to_text(&buf, c, k);
1199
1200                 if (!test_bit(BCH_FS_CHECK_LRUS_DONE, &c->flags)) {
1201                         bch_err(c, "%s", buf.buf);
1202                         bch2_btree_iter_advance(&lru_iter);
1203                         goto next_lru;
1204                 } else {
1205                         bch2_trans_inconsistent(trans, "%s", buf.buf);
1206                         ret = -EINVAL;
1207                         goto out;
1208                 }
1209         }
1210
1211         idx     = k.k->p.offset;
1212         bucket  = le64_to_cpu(bkey_s_c_to_lru(k).v->idx);
1213
1214         *bucket_pos = POS(ca->dev_idx, bucket);
1215
1216         a = bch2_trans_start_alloc_update(trans, &alloc_iter, *bucket_pos);
1217         ret = PTR_ERR_OR_ZERO(a);
1218         if (ret)
1219                 goto out;
1220
1221         if (idx != alloc_lru_idx(a->v)) {
1222                 prt_printf(&buf, "alloc key does not point back to lru entry when invalidating bucket:\n  ");
1223                 bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(&a->k_i));
1224                 prt_printf(&buf, "\n  ");
1225                 bch2_bkey_val_to_text(&buf, c, k);
1226
1227                 if (!test_bit(BCH_FS_CHECK_LRUS_DONE, &c->flags)) {
1228                         bch_err(c, "%s", buf.buf);
1229                         bch2_btree_iter_advance(&lru_iter);
1230                         goto next_lru;
1231                 } else {
1232                         bch2_trans_inconsistent(trans, "%s", buf.buf);
1233                         ret = -EINVAL;
1234                         goto out;
1235                 }
1236         }
1237
1238         if (!a->v.cached_sectors)
1239                 bch_err(c, "invalidating empty bucket, confused");
1240
1241         *cached_sectors = a->v.cached_sectors;
1242
1243         SET_BCH_ALLOC_V4_NEED_INC_GEN(&a->v, false);
1244         a->v.gen++;
1245         a->v.data_type          = 0;
1246         a->v.dirty_sectors      = 0;
1247         a->v.cached_sectors     = 0;
1248         a->v.io_time[READ]      = atomic64_read(&c->io_clock[READ].now);
1249         a->v.io_time[WRITE]     = atomic64_read(&c->io_clock[WRITE].now);
1250
1251         ret = bch2_trans_update(trans, &alloc_iter, &a->k_i,
1252                                 BTREE_TRIGGER_BUCKET_INVALIDATE);
1253         if (ret)
1254                 goto out;
1255 out:
1256         bch2_trans_iter_exit(trans, &alloc_iter);
1257         bch2_trans_iter_exit(trans, &lru_iter);
1258         printbuf_exit(&buf);
1259         return ret;
1260 }
1261
1262 static void bch2_do_invalidates_work(struct work_struct *work)
1263 {
1264         struct bch_fs *c = container_of(work, struct bch_fs, invalidate_work);
1265         struct bch_dev *ca;
1266         struct btree_trans trans;
1267         struct bpos bucket;
1268         unsigned i, sectors;
1269         int ret = 0;
1270
1271         bch2_trans_init(&trans, c, 0, 0);
1272
1273         for_each_member_device(ca, c, i) {
1274                 s64 nr_to_invalidate =
1275                         should_invalidate_buckets(ca, bch2_dev_usage_read(ca));
1276
1277                 while (nr_to_invalidate-- >= 0) {
1278                         ret = __bch2_trans_do(&trans, NULL, NULL,
1279                                               BTREE_INSERT_USE_RESERVE|
1280                                               BTREE_INSERT_NOFAIL,
1281                                         invalidate_one_bucket(&trans, ca, &bucket,
1282                                                               &sectors));
1283                         if (ret)
1284                                 break;
1285
1286                         trace_invalidate_bucket(c, bucket.inode, bucket.offset, sectors);
1287                         this_cpu_inc(c->counters[BCH_COUNTER_bucket_invalidate]);
1288                 }
1289         }
1290
1291         bch2_trans_exit(&trans);
1292         percpu_ref_put(&c->writes);
1293 }
1294
1295 void bch2_do_invalidates(struct bch_fs *c)
1296 {
1297         if (percpu_ref_tryget_live(&c->writes) &&
1298             !queue_work(system_long_wq, &c->invalidate_work))
1299                 percpu_ref_put(&c->writes);
1300 }
1301
1302 static int bucket_freespace_init(struct btree_trans *trans, struct btree_iter *iter)
1303 {
1304         struct bch_alloc_v4 a;
1305         struct bkey_s_c k;
1306         int ret;
1307
1308         k = bch2_btree_iter_peek_slot(iter);
1309         ret = bkey_err(k);
1310         if (ret)
1311                 return ret;
1312
1313         bch2_alloc_to_v4(k, &a);
1314         return bch2_bucket_do_index(trans, k, &a, true);
1315 }
1316
1317 static int bch2_dev_freespace_init(struct bch_fs *c, struct bch_dev *ca)
1318 {
1319         struct btree_trans trans;
1320         struct btree_iter iter;
1321         struct bkey_s_c k;
1322         struct bch_member *m;
1323         int ret;
1324
1325         bch2_trans_init(&trans, c, 0, 0);
1326
1327         for_each_btree_key(&trans, iter, BTREE_ID_alloc,
1328                            POS(ca->dev_idx, ca->mi.first_bucket),
1329                            BTREE_ITER_SLOTS|
1330                            BTREE_ITER_PREFETCH, k, ret) {
1331                 if (iter.pos.offset >= ca->mi.nbuckets)
1332                         break;
1333
1334                 ret = __bch2_trans_do(&trans, NULL, NULL,
1335                                       BTREE_INSERT_LAZY_RW,
1336                                  bucket_freespace_init(&trans, &iter));
1337                 if (ret)
1338                         break;
1339         }
1340         bch2_trans_iter_exit(&trans, &iter);
1341
1342         bch2_trans_exit(&trans);
1343
1344         if (ret) {
1345                 bch_err(ca, "error initializing free space: %i", ret);
1346                 return ret;
1347         }
1348
1349         mutex_lock(&c->sb_lock);
1350         m = bch2_sb_get_members(c->disk_sb.sb)->members + ca->dev_idx;
1351         SET_BCH_MEMBER_FREESPACE_INITIALIZED(m, true);
1352         mutex_unlock(&c->sb_lock);
1353
1354         return ret;
1355 }
1356
1357 int bch2_fs_freespace_init(struct bch_fs *c)
1358 {
1359         struct bch_dev *ca;
1360         unsigned i;
1361         int ret = 0;
1362         bool doing_init = false;
1363
1364         /*
1365          * We can crash during the device add path, so we need to check this on
1366          * every mount:
1367          */
1368
1369         for_each_member_device(ca, c, i) {
1370                 if (ca->mi.freespace_initialized)
1371                         continue;
1372
1373                 if (!doing_init) {
1374                         bch_info(c, "initializing freespace");
1375                         doing_init = true;
1376                 }
1377
1378                 ret = bch2_dev_freespace_init(c, ca);
1379                 if (ret) {
1380                         percpu_ref_put(&ca->ref);
1381                         return ret;
1382                 }
1383         }
1384
1385         if (doing_init) {
1386                 mutex_lock(&c->sb_lock);
1387                 bch2_write_super(c);
1388                 mutex_unlock(&c->sb_lock);
1389
1390                 bch_verbose(c, "done initializing freespace");
1391         }
1392
1393         return ret;
1394 }
1395
1396 /* Bucket IO clocks: */
1397
1398 int bch2_bucket_io_time_reset(struct btree_trans *trans, unsigned dev,
1399                               size_t bucket_nr, int rw)
1400 {
1401         struct bch_fs *c = trans->c;
1402         struct btree_iter iter;
1403         struct bkey_i_alloc_v4 *a;
1404         u64 now;
1405         int ret = 0;
1406
1407         a = bch2_trans_start_alloc_update(trans, &iter,  POS(dev, bucket_nr));
1408         ret = PTR_ERR_OR_ZERO(a);
1409         if (ret)
1410                 return ret;
1411
1412         now = atomic64_read(&c->io_clock[rw].now);
1413         if (a->v.io_time[rw] == now)
1414                 goto out;
1415
1416         a->v.io_time[rw] = now;
1417
1418         ret   = bch2_trans_update(trans, &iter, &a->k_i, 0) ?:
1419                 bch2_trans_commit(trans, NULL, NULL, 0);
1420 out:
1421         bch2_trans_iter_exit(trans, &iter);
1422         return ret;
1423 }
1424
1425 /* Startup/shutdown (ro/rw): */
1426
1427 void bch2_recalc_capacity(struct bch_fs *c)
1428 {
1429         struct bch_dev *ca;
1430         u64 capacity = 0, reserved_sectors = 0, gc_reserve;
1431         unsigned bucket_size_max = 0;
1432         unsigned long ra_pages = 0;
1433         unsigned i;
1434
1435         lockdep_assert_held(&c->state_lock);
1436
1437         for_each_online_member(ca, c, i) {
1438                 struct backing_dev_info *bdi = ca->disk_sb.bdev->bd_disk->bdi;
1439
1440                 ra_pages += bdi->ra_pages;
1441         }
1442
1443         bch2_set_ra_pages(c, ra_pages);
1444
1445         for_each_rw_member(ca, c, i) {
1446                 u64 dev_reserve = 0;
1447
1448                 /*
1449                  * We need to reserve buckets (from the number
1450                  * of currently available buckets) against
1451                  * foreground writes so that mainly copygc can
1452                  * make forward progress.
1453                  *
1454                  * We need enough to refill the various reserves
1455                  * from scratch - copygc will use its entire
1456                  * reserve all at once, then run against when
1457                  * its reserve is refilled (from the formerly
1458                  * available buckets).
1459                  *
1460                  * This reserve is just used when considering if
1461                  * allocations for foreground writes must wait -
1462                  * not -ENOSPC calculations.
1463                  */
1464
1465                 dev_reserve += ca->nr_btree_reserve * 2;
1466                 dev_reserve += ca->mi.nbuckets >> 6; /* copygc reserve */
1467
1468                 dev_reserve += 1;       /* btree write point */
1469                 dev_reserve += 1;       /* copygc write point */
1470                 dev_reserve += 1;       /* rebalance write point */
1471
1472                 dev_reserve *= ca->mi.bucket_size;
1473
1474                 capacity += bucket_to_sector(ca, ca->mi.nbuckets -
1475                                              ca->mi.first_bucket);
1476
1477                 reserved_sectors += dev_reserve * 2;
1478
1479                 bucket_size_max = max_t(unsigned, bucket_size_max,
1480                                         ca->mi.bucket_size);
1481         }
1482
1483         gc_reserve = c->opts.gc_reserve_bytes
1484                 ? c->opts.gc_reserve_bytes >> 9
1485                 : div64_u64(capacity * c->opts.gc_reserve_percent, 100);
1486
1487         reserved_sectors = max(gc_reserve, reserved_sectors);
1488
1489         reserved_sectors = min(reserved_sectors, capacity);
1490
1491         c->capacity = capacity - reserved_sectors;
1492
1493         c->bucket_size_max = bucket_size_max;
1494
1495         /* Wake up case someone was waiting for buckets */
1496         closure_wake_up(&c->freelist_wait);
1497 }
1498
1499 static bool bch2_dev_has_open_write_point(struct bch_fs *c, struct bch_dev *ca)
1500 {
1501         struct open_bucket *ob;
1502         bool ret = false;
1503
1504         for (ob = c->open_buckets;
1505              ob < c->open_buckets + ARRAY_SIZE(c->open_buckets);
1506              ob++) {
1507                 spin_lock(&ob->lock);
1508                 if (ob->valid && !ob->on_partial_list &&
1509                     ob->dev == ca->dev_idx)
1510                         ret = true;
1511                 spin_unlock(&ob->lock);
1512         }
1513
1514         return ret;
1515 }
1516
1517 /* device goes ro: */
1518 void bch2_dev_allocator_remove(struct bch_fs *c, struct bch_dev *ca)
1519 {
1520         unsigned i;
1521
1522         /* First, remove device from allocation groups: */
1523
1524         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1525                 clear_bit(ca->dev_idx, c->rw_devs[i].d);
1526
1527         /*
1528          * Capacity is calculated based off of devices in allocation groups:
1529          */
1530         bch2_recalc_capacity(c);
1531
1532         /* Next, close write points that point to this device... */
1533         for (i = 0; i < ARRAY_SIZE(c->write_points); i++)
1534                 bch2_writepoint_stop(c, ca, &c->write_points[i]);
1535
1536         bch2_writepoint_stop(c, ca, &c->copygc_write_point);
1537         bch2_writepoint_stop(c, ca, &c->rebalance_write_point);
1538         bch2_writepoint_stop(c, ca, &c->btree_write_point);
1539
1540         mutex_lock(&c->btree_reserve_cache_lock);
1541         while (c->btree_reserve_cache_nr) {
1542                 struct btree_alloc *a =
1543                         &c->btree_reserve_cache[--c->btree_reserve_cache_nr];
1544
1545                 bch2_open_buckets_put(c, &a->ob);
1546         }
1547         mutex_unlock(&c->btree_reserve_cache_lock);
1548
1549         while (1) {
1550                 struct open_bucket *ob;
1551
1552                 spin_lock(&c->freelist_lock);
1553                 if (!ca->open_buckets_partial_nr) {
1554                         spin_unlock(&c->freelist_lock);
1555                         break;
1556                 }
1557                 ob = c->open_buckets +
1558                         ca->open_buckets_partial[--ca->open_buckets_partial_nr];
1559                 ob->on_partial_list = false;
1560                 spin_unlock(&c->freelist_lock);
1561
1562                 bch2_open_bucket_put(c, ob);
1563         }
1564
1565         bch2_ec_stop_dev(c, ca);
1566
1567         /*
1568          * Wake up threads that were blocked on allocation, so they can notice
1569          * the device can no longer be removed and the capacity has changed:
1570          */
1571         closure_wake_up(&c->freelist_wait);
1572
1573         /*
1574          * journal_res_get() can block waiting for free space in the journal -
1575          * it needs to notice there may not be devices to allocate from anymore:
1576          */
1577         wake_up(&c->journal.wait);
1578
1579         /* Now wait for any in flight writes: */
1580
1581         closure_wait_event(&c->open_buckets_wait,
1582                            !bch2_dev_has_open_write_point(c, ca));
1583 }
1584
1585 /* device goes rw: */
1586 void bch2_dev_allocator_add(struct bch_fs *c, struct bch_dev *ca)
1587 {
1588         unsigned i;
1589
1590         for (i = 0; i < ARRAY_SIZE(c->rw_devs); i++)
1591                 if (ca->mi.data_allowed & (1 << i))
1592                         set_bit(ca->dev_idx, c->rw_devs[i].d);
1593 }
1594
1595 void bch2_fs_allocator_background_init(struct bch_fs *c)
1596 {
1597         spin_lock_init(&c->freelist_lock);
1598         INIT_WORK(&c->discard_work, bch2_do_discards_work);
1599         INIT_WORK(&c->invalidate_work, bch2_do_invalidates_work);
1600 }