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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_key_cache.h"
   5 #include "bkey_methods.h"
   6 #include "btree_update.h"
   7 #include "error.h"
   8 #include "extents.h"
   9 #include "extent_update.h"
  10 #include "inode.h"
  11 #include "str_hash.h"
  12 #include "subvolume.h"
  13 #include "varint.h"
  14
  15 #include <linux/random.h>
  16
  17 #include <asm/unaligned.h>
  18
  19 const char * const bch2_inode_opts[] = {
  20 #define x(name, ...)    #name,
  21         BCH_INODE_OPTS()
  22 #undef  x
  23         NULL,
  24 };
  25
  26 static const u8 byte_table[8] = { 1, 2, 3, 4, 6, 8, 10, 13 };
  27 static const u8 bits_table[8] = {
  28         1  * 8 - 1,
  29         2  * 8 - 2,
  30         3  * 8 - 3,
  31         4  * 8 - 4,
  32         6  * 8 - 5,
  33         8  * 8 - 6,
  34         10 * 8 - 7,
  35         13 * 8 - 8,
  36 };
  37
  38 static int inode_encode_field(u8 *out, u8 *end, u64 hi, u64 lo)
  39 {
  40         __be64 in[2] = { cpu_to_be64(hi), cpu_to_be64(lo), };
  41         unsigned shift, bytes, bits = likely(!hi)
  42                 ? fls64(lo)
  43                 : fls64(hi) + 64;
  44
  45         for (shift = 1; shift <= 8; shift++)
  46                 if (bits < bits_table[shift - 1])
  47                         goto got_shift;
  48
  49         BUG();
  50 got_shift:
  51         bytes = byte_table[shift - 1];
  52
  53         BUG_ON(out + bytes > end);
  54
  55         memcpy(out, (u8 *) in + 16 - bytes, bytes);
  56         *out |= (1 << 8) >> shift;
  57
  58         return bytes;
  59 }
  60
  61 static int inode_decode_field(const u8 *in, const u8 *end,
  62                               u64 out[2], unsigned *out_bits)
  63 {
  64         __be64 be[2] = { 0, 0 };
  65         unsigned bytes, shift;
  66         u8 *p;
  67
  68         if (in >= end)
  69                 return -1;
  70
  71         if (!*in)
  72                 return -1;
  73
  74         /*
  75          * position of highest set bit indicates number of bytes:
  76          * shift = number of bits to remove in high byte:
  77          */
  78         shift   = 8 - __fls(*in); /* 1 <= shift <= 8 */
  79         bytes   = byte_table[shift - 1];
  80
  81         if (in + bytes > end)
  82                 return -1;
  83
  84         p = (u8 *) be + 16 - bytes;
  85         memcpy(p, in, bytes);
  86         *p ^= (1 << 8) >> shift;
  87
  88         out[0] = be64_to_cpu(be[0]);
  89         out[1] = be64_to_cpu(be[1]);
  90         *out_bits = out[0] ? 64 + fls64(out[0]) : fls64(out[1]);
  91
  92         return bytes;
  93 }
  94
  95 static noinline void bch2_inode_pack_v1(struct bkey_inode_buf *packed,
  96                                         const struct bch_inode_unpacked *inode)
  97 {
  98         struct bkey_i_inode *k = &packed->inode;
  99         u8 *out = k->v.fields;
 100         u8 *end = (void *) &packed[1];
 101         u8 *last_nonzero_field = out;
 102         unsigned nr_fields = 0, last_nonzero_fieldnr = 0;
 103         unsigned bytes;
 104
 105 #define x(_name, _bits)                                                 \
 106         out += inode_encode_field(out, end, 0, inode->_name);           \
 107         nr_fields++;                                                    \
 108                                                                         \
 109         if (inode->_name) {                                             \
 110                 last_nonzero_field = out;                               \
 111                 last_nonzero_fieldnr = nr_fields;                       \
 112         }
 113
 114         BCH_INODE_FIELDS()
 115 #undef  x
 116
 117         out = last_nonzero_field;
 118         nr_fields = last_nonzero_fieldnr;
 119
 120         bytes = out - (u8 *) &packed->inode.v;
 121         set_bkey_val_bytes(&packed->inode.k, bytes);
 122         memset_u64s_tail(&packed->inode.v, 0, bytes);
 123
 124         SET_INODE_NR_FIELDS(&k->v, nr_fields);
 125 }
 126
 127 static void bch2_inode_pack_v2(struct bkey_inode_buf *packed,
 128                                const struct bch_inode_unpacked *inode)
 129 {
 130         struct bkey_i_inode *k = &packed->inode;
 131         u8 *out = k->v.fields;
 132         u8 *end = (void *) &packed[1];
 133         u8 *last_nonzero_field = out;
 134         unsigned nr_fields = 0, last_nonzero_fieldnr = 0;
 135         unsigned bytes;
 136         int ret;
 137
 138 #define x(_name, _bits)                                                 \
 139         nr_fields++;                                                    \
 140                                                                         \
 141         if (inode->_name) {                                             \
 142                 ret = bch2_varint_encode_fast(out, inode->_name);       \
 143                 out += ret;                                             \
 144                                                                         \
 145                 if (_bits > 64)                                         \
 146                         *out++ = 0;                                     \
 147                                                                         \
 148                 last_nonzero_field = out;                               \
 149                 last_nonzero_fieldnr = nr_fields;                       \
 150         } else {                                                        \
 151                 *out++ = 0;                                             \
 152                                                                         \
 153                 if (_bits > 64)                                         \
 154                         *out++ = 0;                                     \
 155         }
 156
 157         BCH_INODE_FIELDS()
 158 #undef  x
 159         BUG_ON(out > end);
 160
 161         out = last_nonzero_field;
 162         nr_fields = last_nonzero_fieldnr;
 163
 164         bytes = out - (u8 *) &packed->inode.v;
 165         set_bkey_val_bytes(&packed->inode.k, bytes);
 166         memset_u64s_tail(&packed->inode.v, 0, bytes);
 167
 168         SET_INODE_NR_FIELDS(&k->v, nr_fields);
 169 }
 170
 171 void bch2_inode_pack(struct bch_fs *c,
 172                      struct bkey_inode_buf *packed,
 173                      const struct bch_inode_unpacked *inode)
 174 {
 175         bkey_inode_init(&packed->inode.k_i);
 176         packed->inode.k.p.offset        = inode->bi_inum;
 177         packed->inode.v.bi_hash_seed    = inode->bi_hash_seed;
 178         packed->inode.v.bi_flags        = cpu_to_le32(inode->bi_flags);
 179         packed->inode.v.bi_mode         = cpu_to_le16(inode->bi_mode);
 180
 181         if (c->sb.features & (1ULL << BCH_FEATURE_new_varint)) {
 182                 SET_INODE_NEW_VARINT(&packed->inode.v, true);
 183                 bch2_inode_pack_v2(packed, inode);
 184         } else {
 185                 bch2_inode_pack_v1(packed, inode);
 186         }
 187
 188         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
 189                 struct bch_inode_unpacked unpacked;
 190
 191                 int ret = bch2_inode_unpack(inode_i_to_s_c(&packed->inode),
 192                                            &unpacked);
 193                 BUG_ON(ret);
 194                 BUG_ON(unpacked.bi_inum         != inode->bi_inum);
 195                 BUG_ON(unpacked.bi_hash_seed    != inode->bi_hash_seed);
 196                 BUG_ON(unpacked.bi_mode         != inode->bi_mode);
 197
 198 #define x(_name, _bits) if (unpacked._name != inode->_name)             \
 199                         panic("unpacked %llu should be %llu",           \
 200                               (u64) unpacked._name, (u64) inode->_name);
 201                 BCH_INODE_FIELDS()
 202 #undef  x
 203         }
 204 }
 205
 206 static noinline int bch2_inode_unpack_v1(struct bkey_s_c_inode inode,
 207                                 struct bch_inode_unpacked *unpacked)
 208 {
 209         const u8 *in = inode.v->fields;
 210         const u8 *end = bkey_val_end(inode);
 211         u64 field[2];
 212         unsigned fieldnr = 0, field_bits;
 213         int ret;
 214
 215 #define x(_name, _bits)                                 \
 216         if (fieldnr++ == INODE_NR_FIELDS(inode.v)) {                    \
 217                 memset(&unpacked->_name, 0,                             \
 218                        sizeof(*unpacked) -                              \
 219                        offsetof(struct bch_inode_unpacked, _name));     \
 220                 return 0;                                               \
 221         }                                                               \
 222                                                                         \
 223         ret = inode_decode_field(in, end, field, &field_bits);          \
 224         if (ret < 0)                                                    \
 225                 return ret;                                             \
 226                                                                         \
 227         if (field_bits > sizeof(unpacked->_name) * 8)                   \
 228                 return -1;                                              \
 229                                                                         \
 230         unpacked->_name = field[1];                                     \
 231         in += ret;
 232
 233         BCH_INODE_FIELDS()
 234 #undef  x
 235
 236         /* XXX: signal if there were more fields than expected? */
 237         return 0;
 238 }
 239
 240 static int bch2_inode_unpack_v2(struct bkey_s_c_inode inode,
 241                                 struct bch_inode_unpacked *unpacked)
 242 {
 243         const u8 *in = inode.v->fields;
 244         const u8 *end = bkey_val_end(inode);
 245         unsigned fieldnr = 0;
 246         int ret;
 247         u64 v[2];
 248
 249 #define x(_name, _bits)                                                 \
 250         if (fieldnr < INODE_NR_FIELDS(inode.v)) {                       \
 251                 ret = bch2_varint_decode_fast(in, end, &v[0]);          \
 252                 if (ret < 0)                                            \
 253                         return ret;                                     \
 254                 in += ret;                                              \
 255                                                                         \
 256                 if (_bits > 64) {                                       \
 257                         ret = bch2_varint_decode_fast(in, end, &v[1]);  \
 258                         if (ret < 0)                                    \
 259                                 return ret;                             \
 260                         in += ret;                                      \
 261                 } else {                                                \
 262                         v[1] = 0;                                       \
 263                 }                                                       \
 264         } else {                                                        \
 265                 v[0] = v[1] = 0;                                        \
 266         }                                                               \
 267                                                                         \
 268         unpacked->_name = v[0];                                         \
 269         if (v[1] || v[0] != unpacked->_name)                            \
 270                 return -1;                                              \
 271         fieldnr++;
 272
 273         BCH_INODE_FIELDS()
 274 #undef  x
 275
 276         /* XXX: signal if there were more fields than expected? */
 277         return 0;
 278 }
 279
 280 int bch2_inode_unpack(struct bkey_s_c_inode inode,
 281                       struct bch_inode_unpacked *unpacked)
 282 {
 283         unpacked->bi_inum       = inode.k->p.offset;
 284         unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 285         unpacked->bi_flags      = le32_to_cpu(inode.v->bi_flags);
 286         unpacked->bi_mode       = le16_to_cpu(inode.v->bi_mode);
 287
 288         if (INODE_NEW_VARINT(inode.v)) {
 289                 return bch2_inode_unpack_v2(inode, unpacked);
 290         } else {
 291                 return bch2_inode_unpack_v1(inode, unpacked);
 292         }
 293
 294         return 0;
 295 }
 296
 297 int bch2_inode_peek(struct btree_trans *trans,
 298                     struct btree_iter *iter,
 299                     struct bch_inode_unpacked *inode,
 300                     subvol_inum inum, unsigned flags)
 301 {
 302         struct bkey_s_c k;
 303         u32 snapshot;
 304         int ret;
 305
 306         if (0 && trans->c->opts.inodes_use_key_cache)
 307                 flags |= BTREE_ITER_CACHED;
 308
 309         ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 310         if (ret)
 311                 return ret;
 312
 313         bch2_trans_iter_init(trans, iter, BTREE_ID_inodes,
 314                              SPOS(0, inum.inum, snapshot), flags);
 315         k = bch2_btree_iter_peek_slot(iter);
 316         ret = bkey_err(k);
 317         if (ret)
 318                 goto err;
 319
 320         ret = k.k->type == KEY_TYPE_inode ? 0 : -ENOENT;
 321         if (ret)
 322                 goto err;
 323
 324         ret = bch2_inode_unpack(bkey_s_c_to_inode(k), inode);
 325         if (ret)
 326                 goto err;
 327
 328         return 0;
 329 err:
 330         bch2_trans_iter_exit(trans, iter);
 331         return ret;
 332 }
 333
 334 int bch2_inode_write(struct btree_trans *trans,
 335                      struct btree_iter *iter,
 336                      struct bch_inode_unpacked *inode)
 337 {
 338         struct bkey_inode_buf *inode_p;
 339
 340         inode_p = bch2_trans_kmalloc(trans, sizeof(*inode_p));
 341         if (IS_ERR(inode_p))
 342                 return PTR_ERR(inode_p);
 343
 344         bch2_inode_pack(trans->c, inode_p, inode);
 345         inode_p->inode.k.p.snapshot = iter->snapshot;
 346         return bch2_trans_update(trans, iter, &inode_p->inode.k_i, 0);
 347 }
 348
 349 const char *bch2_inode_invalid(const struct bch_fs *c, struct bkey_s_c k)
 350 {
 351         struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 352         struct bch_inode_unpacked unpacked;
 353
 354         if (k.k->p.inode)
 355                 return "nonzero k.p.inode";
 356
 357         if (bkey_val_bytes(k.k) < sizeof(struct bch_inode))
 358                 return "incorrect value size";
 359
 360         if (k.k->p.offset < BLOCKDEV_INODE_MAX)
 361                 return "fs inode in blockdev range";
 362
 363         if (INODE_STR_HASH(inode.v) >= BCH_STR_HASH_NR)
 364                 return "invalid str hash type";
 365
 366         if (bch2_inode_unpack(inode, &unpacked))
 367                 return "invalid variable length fields";
 368
 369         if (unpacked.bi_data_checksum >= BCH_CSUM_OPT_NR + 1)
 370                 return "invalid data checksum type";
 371
 372         if (unpacked.bi_compression >= BCH_COMPRESSION_OPT_NR + 1)
 373                 return "invalid data checksum type";
 374
 375         if ((unpacked.bi_flags & BCH_INODE_UNLINKED) &&
 376             unpacked.bi_nlink != 0)
 377                 return "flagged as unlinked but bi_nlink != 0";
 378
 379         if (unpacked.bi_subvol && !S_ISDIR(unpacked.bi_mode))
 380                 return "subvolume root but not a directory";
 381
 382         return NULL;
 383 }
 384
 385 static void __bch2_inode_unpacked_to_text(struct printbuf *out, struct bch_inode_unpacked *inode)
 386 {
 387         pr_buf(out, "mode %o flags %x ", inode->bi_mode, inode->bi_flags);
 388
 389 #define x(_name, _bits)                                         \
 390         pr_buf(out, #_name " %llu ", (u64) inode->_name);
 391         BCH_INODE_FIELDS()
 392 #undef  x
 393 }
 394
 395 void bch2_inode_unpacked_to_text(struct printbuf *out, struct bch_inode_unpacked *inode)
 396 {
 397         pr_buf(out, "inum: %llu ", inode->bi_inum);
 398         __bch2_inode_unpacked_to_text(out, inode);
 399 }
 400
 401 void bch2_inode_to_text(struct printbuf *out, struct bch_fs *c,
 402                        struct bkey_s_c k)
 403 {
 404         struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 405         struct bch_inode_unpacked unpacked;
 406
 407         if (bch2_inode_unpack(inode, &unpacked)) {
 408                 pr_buf(out, "(unpack error)");
 409                 return;
 410         }
 411
 412         __bch2_inode_unpacked_to_text(out, &unpacked);
 413 }
 414
 415 const char *bch2_inode_generation_invalid(const struct bch_fs *c,
 416                                           struct bkey_s_c k)
 417 {
 418         if (k.k->p.inode)
 419                 return "nonzero k.p.inode";
 420
 421         if (bkey_val_bytes(k.k) != sizeof(struct bch_inode_generation))
 422                 return "incorrect value size";
 423
 424         return NULL;
 425 }
 426
 427 void bch2_inode_generation_to_text(struct printbuf *out, struct bch_fs *c,
 428                                    struct bkey_s_c k)
 429 {
 430         struct bkey_s_c_inode_generation gen = bkey_s_c_to_inode_generation(k);
 431
 432         pr_buf(out, "generation: %u", le32_to_cpu(gen.v->bi_generation));
 433 }
 434
 435 void bch2_inode_init_early(struct bch_fs *c,
 436                            struct bch_inode_unpacked *inode_u)
 437 {
 438         enum bch_str_hash_type str_hash =
 439                 bch2_str_hash_opt_to_type(c, c->opts.str_hash);
 440
 441         memset(inode_u, 0, sizeof(*inode_u));
 442
 443         /* ick */
 444         inode_u->bi_flags |= str_hash << INODE_STR_HASH_OFFSET;
 445         get_random_bytes(&inode_u->bi_hash_seed,
 446                          sizeof(inode_u->bi_hash_seed));
 447 }
 448
 449 void bch2_inode_init_late(struct bch_inode_unpacked *inode_u, u64 now,
 450                           uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 451                           struct bch_inode_unpacked *parent)
 452 {
 453         inode_u->bi_mode        = mode;
 454         inode_u->bi_uid         = uid;
 455         inode_u->bi_gid         = gid;
 456         inode_u->bi_dev         = rdev;
 457         inode_u->bi_atime       = now;
 458         inode_u->bi_mtime       = now;
 459         inode_u->bi_ctime       = now;
 460         inode_u->bi_otime       = now;
 461
 462         if (parent && parent->bi_mode & S_ISGID) {
 463                 inode_u->bi_gid = parent->bi_gid;
 464                 if (S_ISDIR(mode))
 465                         inode_u->bi_mode |= S_ISGID;
 466         }
 467
 468         if (parent) {
 469 #define x(_name, ...)   inode_u->bi_##_name = parent->bi_##_name;
 470                 BCH_INODE_OPTS()
 471 #undef x
 472         }
 473 }
 474
 475 void bch2_inode_init(struct bch_fs *c, struct bch_inode_unpacked *inode_u,
 476                      uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 477                      struct bch_inode_unpacked *parent)
 478 {
 479         bch2_inode_init_early(c, inode_u);
 480         bch2_inode_init_late(inode_u, bch2_current_time(c),
 481                              uid, gid, mode, rdev, parent);
 482 }
 483
 484 static inline u32 bkey_generation(struct bkey_s_c k)
 485 {
 486         switch (k.k->type) {
 487         case KEY_TYPE_inode:
 488                 BUG();
 489         case KEY_TYPE_inode_generation:
 490                 return le32_to_cpu(bkey_s_c_to_inode_generation(k).v->bi_generation);
 491         default:
 492                 return 0;
 493         }
 494 }
 495
 496 /*
 497  * This just finds an empty slot:
 498  */
 499 int bch2_inode_create(struct btree_trans *trans,
 500                       struct btree_iter *iter,
 501                       struct bch_inode_unpacked *inode_u,
 502                       u32 snapshot, u64 cpu)
 503 {
 504         struct bch_fs *c = trans->c;
 505         struct bkey_s_c k;
 506         u64 min, max, start, pos, *hint;
 507         int ret = 0;
 508         unsigned bits = (c->opts.inodes_32bit ? 31 : 63);
 509
 510         if (c->opts.shard_inode_numbers) {
 511                 bits -= c->inode_shard_bits;
 512
 513                 min = (cpu << bits);
 514                 max = (cpu << bits) | ~(ULLONG_MAX << bits);
 515
 516                 min = max_t(u64, min, BLOCKDEV_INODE_MAX);
 517                 hint = c->unused_inode_hints + cpu;
 518         } else {
 519                 min = BLOCKDEV_INODE_MAX;
 520                 max = ~(ULLONG_MAX << bits);
 521                 hint = c->unused_inode_hints;
 522         }
 523
 524         start = READ_ONCE(*hint);
 525
 526         if (start >= max || start < min)
 527                 start = min;
 528
 529         pos = start;
 530         bch2_trans_iter_init(trans, iter, BTREE_ID_inodes, POS(0, pos),
 531                              BTREE_ITER_ALL_SNAPSHOTS|
 532                              BTREE_ITER_INTENT);
 533 again:
 534         while ((k = bch2_btree_iter_peek(iter)).k &&
 535                !(ret = bkey_err(k)) &&
 536                bkey_cmp(k.k->p, POS(0, max)) < 0) {
 537                 while (pos < iter->pos.offset) {
 538                         if (!bch2_btree_key_cache_find(c, BTREE_ID_inodes, POS(0, pos)))
 539                                 goto found_slot;
 540
 541                         pos++;
 542                 }
 543
 544                 if (k.k->p.snapshot == snapshot &&
 545                     k.k->type != KEY_TYPE_inode &&
 546                     !bch2_btree_key_cache_find(c, BTREE_ID_inodes, SPOS(0, pos, snapshot))) {
 547                         bch2_btree_iter_advance(iter);
 548                         continue;
 549                 }
 550
 551                 /*
 552                  * We don't need to iterate over keys in every snapshot once
 553                  * we've found just one:
 554                  */
 555                 pos = iter->pos.offset + 1;
 556                 bch2_btree_iter_set_pos(iter, POS(0, pos));
 557         }
 558
 559         while (!ret && pos < max) {
 560                 if (!bch2_btree_key_cache_find(c, BTREE_ID_inodes, POS(0, pos)))
 561                         goto found_slot;
 562
 563                 pos++;
 564         }
 565
 566         if (!ret && start == min)
 567                 ret = -ENOSPC;
 568
 569         if (ret) {
 570                 bch2_trans_iter_exit(trans, iter);
 571                 return ret;
 572         }
 573
 574         /* Retry from start */
 575         pos = start = min;
 576         bch2_btree_iter_set_pos(iter, POS(0, pos));
 577         goto again;
 578 found_slot:
 579         bch2_btree_iter_set_pos(iter, SPOS(0, pos, snapshot));
 580         k = bch2_btree_iter_peek_slot(iter);
 581         ret = bkey_err(k);
 582         if (ret) {
 583                 bch2_trans_iter_exit(trans, iter);
 584                 return ret;
 585         }
 586
 587         /* We may have raced while the iterator wasn't pointing at pos: */
 588         if (k.k->type == KEY_TYPE_inode ||
 589             bch2_btree_key_cache_find(c, BTREE_ID_inodes, k.k->p))
 590                 goto again;
 591
 592         *hint                   = k.k->p.offset;
 593         inode_u->bi_inum        = k.k->p.offset;
 594         inode_u->bi_generation  = bkey_generation(k);
 595         return 0;
 596 }
 597
 598 static int bch2_inode_delete_keys(struct btree_trans *trans,
 599                                   subvol_inum inum, enum btree_id id)
 600 {
 601         u64 offset = 0;
 602         int ret = 0;
 603
 604         while (!ret || ret == -EINTR) {
 605                 struct btree_iter iter;
 606                 struct bkey_s_c k;
 607                 struct bkey_i delete;
 608                 u32 snapshot;
 609
 610                 bch2_trans_begin(trans);
 611
 612                 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 613                 if (ret)
 614                         continue;
 615
 616                 bch2_trans_iter_init(trans, &iter, id,
 617                                      SPOS(inum.inum, offset, snapshot),
 618                                      BTREE_ITER_INTENT);
 619                 k = bch2_btree_iter_peek(&iter);
 620
 621                 if (!k.k || iter.pos.inode != inum.inum) {
 622                         bch2_trans_iter_exit(trans, &iter);
 623                         break;
 624                 }
 625
 626                 ret = bkey_err(k);
 627                 if (ret)
 628                         goto err;
 629
 630                 bkey_init(&delete.k);
 631                 delete.k.p = iter.pos;
 632
 633                 if (btree_node_type_is_extents(iter.btree_id)) {
 634                         unsigned max_sectors =
 635                                 min_t(u64, U64_MAX - iter.pos.offset,
 636                                       KEY_SIZE_MAX & (~0 << trans->c->block_bits));
 637
 638                         /* create the biggest key we can */
 639                         bch2_key_resize(&delete.k, max_sectors);
 640
 641                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
 642                         if (ret)
 643                                 goto err;
 644                 }
 645
 646                 ret = bch2_trans_update(trans, &iter, &delete, 0) ?:
 647                       bch2_trans_commit(trans, NULL, NULL,
 648                                         BTREE_INSERT_NOFAIL);
 649 err:
 650                 offset = iter.pos.offset;
 651                 bch2_trans_iter_exit(trans, &iter);
 652         }
 653
 654         return ret;
 655 }
 656
 657 int bch2_inode_rm(struct bch_fs *c, subvol_inum inum, bool cached)
 658 {
 659         struct btree_trans trans;
 660         struct btree_iter iter = { NULL };
 661         struct bkey_i_inode_generation delete;
 662         struct bch_inode_unpacked inode_u;
 663         struct bkey_s_c k;
 664         unsigned iter_flags = BTREE_ITER_INTENT;
 665         u32 snapshot;
 666         int ret;
 667
 668         if (0 && cached && c->opts.inodes_use_key_cache)
 669                 iter_flags |= BTREE_ITER_CACHED;
 670
 671         bch2_trans_init(&trans, c, 0, 1024);
 672
 673         /*
 674          * If this was a directory, there shouldn't be any real dirents left -
 675          * but there could be whiteouts (from hash collisions) that we should
 676          * delete:
 677          *
 678          * XXX: the dirent could ideally would delete whiteouts when they're no
 679          * longer needed
 680          */
 681         ret   = bch2_inode_delete_keys(&trans, inum, BTREE_ID_extents) ?:
 682                 bch2_inode_delete_keys(&trans, inum, BTREE_ID_xattrs) ?:
 683                 bch2_inode_delete_keys(&trans, inum, BTREE_ID_dirents);
 684         if (ret)
 685                 goto err;
 686 retry:
 687         bch2_trans_begin(&trans);
 688
 689         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
 690         if (ret)
 691                 goto err;
 692
 693         bch2_trans_iter_init(&trans, &iter, BTREE_ID_inodes,
 694                              SPOS(0, inum.inum, snapshot), iter_flags);
 695         k = bch2_btree_iter_peek_slot(&iter);
 696
 697         ret = bkey_err(k);
 698         if (ret)
 699                 goto err;
 700
 701         if (k.k->type != KEY_TYPE_inode) {
 702                 bch2_fs_inconsistent(trans.c,
 703                                      "inode %llu not found when deleting",
 704                                      inum.inum);
 705                 ret = -EIO;
 706                 goto err;
 707         }
 708
 709         bch2_inode_unpack(bkey_s_c_to_inode(k), &inode_u);
 710
 711         /* Subvolume root? */
 712         if (inode_u.bi_subvol) {
 713                 ret = bch2_subvolume_delete(&trans, inode_u.bi_subvol, -1);
 714                 if (ret)
 715                         goto err;
 716         }
 717
 718         bkey_inode_generation_init(&delete.k_i);
 719         delete.k.p = iter.pos;
 720         delete.v.bi_generation = cpu_to_le32(inode_u.bi_generation + 1);
 721
 722         ret   = bch2_trans_update(&trans, &iter, &delete.k_i, 0) ?:
 723                 bch2_trans_commit(&trans, NULL, NULL,
 724                                 BTREE_INSERT_NOFAIL);
 725 err:
 726         bch2_trans_iter_exit(&trans, &iter);
 727         if (ret == -EINTR)
 728                 goto retry;
 729
 730         bch2_trans_exit(&trans);
 731         return ret;
 732 }
 733
 734 static int bch2_inode_find_by_inum_trans(struct btree_trans *trans,
 735                                          subvol_inum inum,
 736                                          struct bch_inode_unpacked *inode)
 737 {
 738         struct btree_iter iter;
 739         int ret;
 740
 741         ret = bch2_inode_peek(trans, &iter, inode, inum, 0);
 742         if (!ret)
 743                 bch2_trans_iter_exit(trans, &iter);
 744         return ret;
 745 }
 746
 747 int bch2_inode_find_by_inum(struct bch_fs *c, subvol_inum inum,
 748                             struct bch_inode_unpacked *inode)
 749 {
 750         return bch2_trans_do(c, NULL, NULL, 0,
 751                 bch2_inode_find_by_inum_trans(&trans, inum, inode));
 752 }