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
  28 static int inode_decode_field(const u8 *in, const u8 *end,
  29                               u64 out[2], unsigned *out_bits)
  30 {
  31         __be64 be[2] = { 0, 0 };
  32         unsigned bytes, shift;
  33         u8 *p;
  34
  35         if (in >= end)
  36                 return -1;
  37
  38         if (!*in)
  39                 return -1;
  40
  41         /*
  42          * position of highest set bit indicates number of bytes:
  43          * shift = number of bits to remove in high byte:
  44          */
  45         shift   = 8 - __fls(*in); /* 1 <= shift <= 8 */
  46         bytes   = byte_table[shift - 1];
  47
  48         if (in + bytes > end)
  49                 return -1;
  50
  51         p = (u8 *) be + 16 - bytes;
  52         memcpy(p, in, bytes);
  53         *p ^= (1 << 8) >> shift;
  54
  55         out[0] = be64_to_cpu(be[0]);
  56         out[1] = be64_to_cpu(be[1]);
  57         *out_bits = out[0] ? 64 + fls64(out[0]) : fls64(out[1]);
  58
  59         return bytes;
  60 }
  61
  62 void bch2_inode_pack(struct bch_fs *c,
  63                      struct bkey_inode_buf *packed,
  64                      const struct bch_inode_unpacked *inode)
  65 {
  66         struct bkey_i_inode_v2 *k = &packed->inode;
  67         u8 *out = k->v.fields;
  68         u8 *end = (void *) &packed[1];
  69         u8 *last_nonzero_field = out;
  70         unsigned nr_fields = 0, last_nonzero_fieldnr = 0;
  71         unsigned bytes;
  72         int ret;
  73
  74         bkey_inode_v2_init(&packed->inode.k_i);
  75         packed->inode.k.p.offset        = inode->bi_inum;
  76         packed->inode.v.bi_journal_seq  = cpu_to_le64(inode->bi_journal_seq);
  77         packed->inode.v.bi_hash_seed    = inode->bi_hash_seed;
  78         packed->inode.v.bi_flags        = cpu_to_le64(inode->bi_flags);
  79         packed->inode.v.bi_flags        = cpu_to_le64(inode->bi_flags);
  80         packed->inode.v.bi_mode         = cpu_to_le16(inode->bi_mode);
  81
  82 #define x(_name, _bits)                                                 \
  83         nr_fields++;                                                    \
  84                                                                         \
  85         if (inode->_name) {                                             \
  86                 ret = bch2_varint_encode_fast(out, inode->_name);       \
  87                 out += ret;                                             \
  88                                                                         \
  89                 if (_bits > 64)                                         \
  90                         *out++ = 0;                                     \
  91                                                                         \
  92                 last_nonzero_field = out;                               \
  93                 last_nonzero_fieldnr = nr_fields;                       \
  94         } else {                                                        \
  95                 *out++ = 0;                                             \
  96                                                                         \
  97                 if (_bits > 64)                                         \
  98                         *out++ = 0;                                     \
  99         }
 100
 101         BCH_INODE_FIELDS()
 102 #undef  x
 103         BUG_ON(out > end);
 104
 105         out = last_nonzero_field;
 106         nr_fields = last_nonzero_fieldnr;
 107
 108         bytes = out - (u8 *) &packed->inode.v;
 109         set_bkey_val_bytes(&packed->inode.k, bytes);
 110         memset_u64s_tail(&packed->inode.v, 0, bytes);
 111
 112         SET_INODEv2_NR_FIELDS(&k->v, nr_fields);
 113
 114         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
 115                 struct bch_inode_unpacked unpacked;
 116
 117                 int ret = bch2_inode_unpack(bkey_i_to_s_c(&packed->inode.k_i),
 118                                            &unpacked);
 119                 BUG_ON(ret);
 120                 BUG_ON(unpacked.bi_inum         != inode->bi_inum);
 121                 BUG_ON(unpacked.bi_hash_seed    != inode->bi_hash_seed);
 122                 BUG_ON(unpacked.bi_mode         != inode->bi_mode);
 123
 124 #define x(_name, _bits) if (unpacked._name != inode->_name)             \
 125                         panic("unpacked %llu should be %llu",           \
 126                               (u64) unpacked._name, (u64) inode->_name);
 127                 BCH_INODE_FIELDS()
 128 #undef  x
 129         }
 130 }
 131
 132 static noinline int bch2_inode_unpack_v1(struct bkey_s_c_inode inode,
 133                                 struct bch_inode_unpacked *unpacked)
 134 {
 135         const u8 *in = inode.v->fields;
 136         const u8 *end = bkey_val_end(inode);
 137         u64 field[2];
 138         unsigned fieldnr = 0, field_bits;
 139         int ret;
 140
 141 #define x(_name, _bits)                                 \
 142         if (fieldnr++ == INODE_NR_FIELDS(inode.v)) {                    \
 143                 memset(&unpacked->_name, 0,                             \
 144                        sizeof(*unpacked) -                              \
 145                        offsetof(struct bch_inode_unpacked, _name));     \
 146                 return 0;                                               \
 147         }                                                               \
 148                                                                         \
 149         ret = inode_decode_field(in, end, field, &field_bits);          \
 150         if (ret < 0)                                                    \
 151                 return ret;                                             \
 152                                                                         \
 153         if (field_bits > sizeof(unpacked->_name) * 8)                   \
 154                 return -1;                                              \
 155                                                                         \
 156         unpacked->_name = field[1];                                     \
 157         in += ret;
 158
 159         BCH_INODE_FIELDS()
 160 #undef  x
 161
 162         /* XXX: signal if there were more fields than expected? */
 163         return 0;
 164 }
 165
 166 static int bch2_inode_unpack_v2(struct bch_inode_unpacked *unpacked,
 167                                 const u8 *in, const u8 *end,
 168                                 unsigned nr_fields)
 169 {
 170         unsigned fieldnr = 0;
 171         int ret;
 172         u64 v[2];
 173
 174 #define x(_name, _bits)                                                 \
 175         if (fieldnr < nr_fields) {                                      \
 176                 ret = bch2_varint_decode_fast(in, end, &v[0]);          \
 177                 if (ret < 0)                                            \
 178                         return ret;                                     \
 179                 in += ret;                                              \
 180                                                                         \
 181                 if (_bits > 64) {                                       \
 182                         ret = bch2_varint_decode_fast(in, end, &v[1]);  \
 183                         if (ret < 0)                                    \
 184                                 return ret;                             \
 185                         in += ret;                                      \
 186                 } else {                                                \
 187                         v[1] = 0;                                       \
 188                 }                                                       \
 189         } else {                                                        \
 190                 v[0] = v[1] = 0;                                        \
 191         }                                                               \
 192                                                                         \
 193         unpacked->_name = v[0];                                         \
 194         if (v[1] || v[0] != unpacked->_name)                            \
 195                 return -1;                                              \
 196         fieldnr++;
 197
 198         BCH_INODE_FIELDS()
 199 #undef  x
 200
 201         /* XXX: signal if there were more fields than expected? */
 202         return 0;
 203 }
 204
 205 int bch2_inode_unpack(struct bkey_s_c k,
 206                       struct bch_inode_unpacked *unpacked)
 207 {
 208         switch (k.k->type) {
 209         case KEY_TYPE_inode: {
 210                 struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 211
 212                 unpacked->bi_inum       = inode.k->p.offset;
 213                 unpacked->bi_journal_seq= 0;
 214                 unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 215                 unpacked->bi_flags      = le32_to_cpu(inode.v->bi_flags);
 216                 unpacked->bi_mode       = le16_to_cpu(inode.v->bi_mode);
 217
 218                 if (INODE_NEW_VARINT(inode.v)) {
 219                         return bch2_inode_unpack_v2(unpacked, inode.v->fields,
 220                                                     bkey_val_end(inode),
 221                                                     INODE_NR_FIELDS(inode.v));
 222                 } else {
 223                         return bch2_inode_unpack_v1(inode, unpacked);
 224                 }
 225                 break;
 226         }
 227         case KEY_TYPE_inode_v2: {
 228                 struct bkey_s_c_inode_v2 inode = bkey_s_c_to_inode_v2(k);
 229
 230                 unpacked->bi_inum       = inode.k->p.offset;
 231                 unpacked->bi_journal_seq= le64_to_cpu(inode.v->bi_journal_seq);
 232                 unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 233                 unpacked->bi_flags      = le64_to_cpu(inode.v->bi_flags);
 234                 unpacked->bi_mode       = le16_to_cpu(inode.v->bi_mode);
 235
 236                 return bch2_inode_unpack_v2(unpacked, inode.v->fields,
 237                                             bkey_val_end(inode),
 238                                             INODEv2_NR_FIELDS(inode.v));
 239         }
 240         default:
 241                 BUG();
 242         }
 243 }
 244
 245 int bch2_inode_peek(struct btree_trans *trans,
 246                     struct btree_iter *iter,
 247                     struct bch_inode_unpacked *inode,
 248                     subvol_inum inum, unsigned flags)
 249 {
 250         struct bkey_s_c k;
 251         u32 snapshot;
 252         int ret;
 253
 254         if (0 && trans->c->opts.inodes_use_key_cache)
 255                 flags |= BTREE_ITER_CACHED;
 256
 257         ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 258         if (ret)
 259                 return ret;
 260
 261         bch2_trans_iter_init(trans, iter, BTREE_ID_inodes,
 262                              SPOS(0, inum.inum, snapshot), flags);
 263         k = bch2_btree_iter_peek_slot(iter);
 264         ret = bkey_err(k);
 265         if (ret)
 266                 goto err;
 267
 268         ret = bkey_is_inode(k.k) ? 0 : -ENOENT;
 269         if (ret)
 270                 goto err;
 271
 272         ret = bch2_inode_unpack(k, inode);
 273         if (ret)
 274                 goto err;
 275
 276         return 0;
 277 err:
 278         bch2_trans_iter_exit(trans, iter);
 279         return ret;
 280 }
 281
 282 int bch2_inode_write(struct btree_trans *trans,
 283                      struct btree_iter *iter,
 284                      struct bch_inode_unpacked *inode)
 285 {
 286         struct bkey_inode_buf *inode_p;
 287
 288         inode_p = bch2_trans_kmalloc(trans, sizeof(*inode_p));
 289         if (IS_ERR(inode_p))
 290                 return PTR_ERR(inode_p);
 291
 292         bch2_inode_pack(trans->c, inode_p, inode);
 293         inode_p->inode.k.p.snapshot = iter->snapshot;
 294         return bch2_trans_update(trans, iter, &inode_p->inode.k_i, 0);
 295 }
 296
 297 const char *bch2_inode_invalid(const struct bch_fs *c, struct bkey_s_c k)
 298 {
 299         struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 300         struct bch_inode_unpacked unpacked;
 301
 302         if (k.k->p.inode)
 303                 return "nonzero k.p.inode";
 304
 305         if (bkey_val_bytes(k.k) < sizeof(struct bch_inode))
 306                 return "incorrect value size";
 307
 308         if (k.k->p.offset < BLOCKDEV_INODE_MAX)
 309                 return "fs inode in blockdev range";
 310
 311         if (INODE_STR_HASH(inode.v) >= BCH_STR_HASH_NR)
 312                 return "invalid str hash type";
 313
 314         if (bch2_inode_unpack(k, &unpacked))
 315                 return "invalid variable length fields";
 316
 317         if (unpacked.bi_data_checksum >= BCH_CSUM_OPT_NR + 1)
 318                 return "invalid data checksum type";
 319
 320         if (unpacked.bi_compression >= BCH_COMPRESSION_OPT_NR + 1)
 321                 return "invalid data checksum type";
 322
 323         if ((unpacked.bi_flags & BCH_INODE_UNLINKED) &&
 324             unpacked.bi_nlink != 0)
 325                 return "flagged as unlinked but bi_nlink != 0";
 326
 327         if (unpacked.bi_subvol && !S_ISDIR(unpacked.bi_mode))
 328                 return "subvolume root but not a directory";
 329
 330         return NULL;
 331 }
 332
 333 const char *bch2_inode_v2_invalid(const struct bch_fs *c, struct bkey_s_c k)
 334 {
 335         struct bkey_s_c_inode_v2 inode = bkey_s_c_to_inode_v2(k);
 336         struct bch_inode_unpacked unpacked;
 337
 338         if (k.k->p.inode)
 339                 return "nonzero k.p.inode";
 340
 341         if (bkey_val_bytes(k.k) < sizeof(struct bch_inode))
 342                 return "incorrect value size";
 343
 344         if (k.k->p.offset < BLOCKDEV_INODE_MAX)
 345                 return "fs inode in blockdev range";
 346
 347         if (INODEv2_STR_HASH(inode.v) >= BCH_STR_HASH_NR)
 348                 return "invalid str hash type";
 349
 350         if (bch2_inode_unpack(k, &unpacked))
 351                 return "invalid variable length fields";
 352
 353         if (unpacked.bi_data_checksum >= BCH_CSUM_OPT_NR + 1)
 354                 return "invalid data checksum type";
 355
 356         if (unpacked.bi_compression >= BCH_COMPRESSION_OPT_NR + 1)
 357                 return "invalid data checksum type";
 358
 359         if ((unpacked.bi_flags & BCH_INODE_UNLINKED) &&
 360             unpacked.bi_nlink != 0)
 361                 return "flagged as unlinked but bi_nlink != 0";
 362
 363         if (unpacked.bi_subvol && !S_ISDIR(unpacked.bi_mode))
 364                 return "subvolume root but not a directory";
 365
 366         return NULL;
 367 }
 368
 369 static void __bch2_inode_unpacked_to_text(struct printbuf *out, struct bch_inode_unpacked *inode)
 370 {
 371         pr_buf(out, "mode %o flags %x journal_seq %llu",
 372                inode->bi_mode, inode->bi_flags,
 373                inode->bi_journal_seq);
 374
 375 #define x(_name, _bits)                                         \
 376         pr_buf(out, " "#_name " %llu", (u64) inode->_name);
 377         BCH_INODE_FIELDS()
 378 #undef  x
 379 }
 380
 381 void bch2_inode_unpacked_to_text(struct printbuf *out, struct bch_inode_unpacked *inode)
 382 {
 383         pr_buf(out, "inum: %llu ", inode->bi_inum);
 384         __bch2_inode_unpacked_to_text(out, inode);
 385 }
 386
 387 void bch2_inode_to_text(struct printbuf *out, struct bch_fs *c,
 388                        struct bkey_s_c k)
 389 {
 390         struct bch_inode_unpacked inode;
 391
 392         if (bch2_inode_unpack(k, &inode)) {
 393                 pr_buf(out, "(unpack error)");
 394                 return;
 395         }
 396
 397         __bch2_inode_unpacked_to_text(out, &inode);
 398 }
 399
 400 const char *bch2_inode_generation_invalid(const struct bch_fs *c,
 401                                           struct bkey_s_c k)
 402 {
 403         if (k.k->p.inode)
 404                 return "nonzero k.p.inode";
 405
 406         if (bkey_val_bytes(k.k) != sizeof(struct bch_inode_generation))
 407                 return "incorrect value size";
 408
 409         return NULL;
 410 }
 411
 412 void bch2_inode_generation_to_text(struct printbuf *out, struct bch_fs *c,
 413                                    struct bkey_s_c k)
 414 {
 415         struct bkey_s_c_inode_generation gen = bkey_s_c_to_inode_generation(k);
 416
 417         pr_buf(out, "generation: %u", le32_to_cpu(gen.v->bi_generation));
 418 }
 419
 420 void bch2_inode_init_early(struct bch_fs *c,
 421                            struct bch_inode_unpacked *inode_u)
 422 {
 423         enum bch_str_hash_type str_hash =
 424                 bch2_str_hash_opt_to_type(c, c->opts.str_hash);
 425
 426         memset(inode_u, 0, sizeof(*inode_u));
 427
 428         /* ick */
 429         inode_u->bi_flags |= str_hash << INODE_STR_HASH_OFFSET;
 430         get_random_bytes(&inode_u->bi_hash_seed,
 431                          sizeof(inode_u->bi_hash_seed));
 432 }
 433
 434 void bch2_inode_init_late(struct bch_inode_unpacked *inode_u, u64 now,
 435                           uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 436                           struct bch_inode_unpacked *parent)
 437 {
 438         inode_u->bi_mode        = mode;
 439         inode_u->bi_uid         = uid;
 440         inode_u->bi_gid         = gid;
 441         inode_u->bi_dev         = rdev;
 442         inode_u->bi_atime       = now;
 443         inode_u->bi_mtime       = now;
 444         inode_u->bi_ctime       = now;
 445         inode_u->bi_otime       = now;
 446
 447         if (parent && parent->bi_mode & S_ISGID) {
 448                 inode_u->bi_gid = parent->bi_gid;
 449                 if (S_ISDIR(mode))
 450                         inode_u->bi_mode |= S_ISGID;
 451         }
 452
 453         if (parent) {
 454 #define x(_name, ...)   inode_u->bi_##_name = parent->bi_##_name;
 455                 BCH_INODE_OPTS()
 456 #undef x
 457         }
 458 }
 459
 460 void bch2_inode_init(struct bch_fs *c, struct bch_inode_unpacked *inode_u,
 461                      uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 462                      struct bch_inode_unpacked *parent)
 463 {
 464         bch2_inode_init_early(c, inode_u);
 465         bch2_inode_init_late(inode_u, bch2_current_time(c),
 466                              uid, gid, mode, rdev, parent);
 467 }
 468
 469 static inline u32 bkey_generation(struct bkey_s_c k)
 470 {
 471         switch (k.k->type) {
 472         case KEY_TYPE_inode:
 473         case KEY_TYPE_inode_v2:
 474                 BUG();
 475         case KEY_TYPE_inode_generation:
 476                 return le32_to_cpu(bkey_s_c_to_inode_generation(k).v->bi_generation);
 477         default:
 478                 return 0;
 479         }
 480 }
 481
 482 /*
 483  * This just finds an empty slot:
 484  */
 485 int bch2_inode_create(struct btree_trans *trans,
 486                       struct btree_iter *iter,
 487                       struct bch_inode_unpacked *inode_u,
 488                       u32 snapshot, u64 cpu)
 489 {
 490         struct bch_fs *c = trans->c;
 491         struct bkey_s_c k;
 492         u64 min, max, start, pos, *hint;
 493         int ret = 0;
 494         unsigned bits = (c->opts.inodes_32bit ? 31 : 63);
 495
 496         if (c->opts.shard_inode_numbers) {
 497                 bits -= c->inode_shard_bits;
 498
 499                 min = (cpu << bits);
 500                 max = (cpu << bits) | ~(ULLONG_MAX << bits);
 501
 502                 min = max_t(u64, min, BLOCKDEV_INODE_MAX);
 503                 hint = c->unused_inode_hints + cpu;
 504         } else {
 505                 min = BLOCKDEV_INODE_MAX;
 506                 max = ~(ULLONG_MAX << bits);
 507                 hint = c->unused_inode_hints;
 508         }
 509
 510         start = READ_ONCE(*hint);
 511
 512         if (start >= max || start < min)
 513                 start = min;
 514
 515         pos = start;
 516         bch2_trans_iter_init(trans, iter, BTREE_ID_inodes, POS(0, pos),
 517                              BTREE_ITER_ALL_SNAPSHOTS|
 518                              BTREE_ITER_INTENT);
 519 again:
 520         while ((k = bch2_btree_iter_peek(iter)).k &&
 521                !(ret = bkey_err(k)) &&
 522                bkey_cmp(k.k->p, POS(0, max)) < 0) {
 523                 while (pos < iter->pos.offset) {
 524                         if (!bch2_btree_key_cache_find(c, BTREE_ID_inodes, POS(0, pos)))
 525                                 goto found_slot;
 526
 527                         pos++;
 528                 }
 529
 530                 if (k.k->p.snapshot == snapshot &&
 531                     !bkey_is_inode(k.k) &&
 532                     !bch2_btree_key_cache_find(c, BTREE_ID_inodes, SPOS(0, pos, snapshot))) {
 533                         bch2_btree_iter_advance(iter);
 534                         continue;
 535                 }
 536
 537                 /*
 538                  * We don't need to iterate over keys in every snapshot once
 539                  * we've found just one:
 540                  */
 541                 pos = iter->pos.offset + 1;
 542                 bch2_btree_iter_set_pos(iter, POS(0, pos));
 543         }
 544
 545         while (!ret && pos < max) {
 546                 if (!bch2_btree_key_cache_find(c, BTREE_ID_inodes, POS(0, pos)))
 547                         goto found_slot;
 548
 549                 pos++;
 550         }
 551
 552         if (!ret && start == min)
 553                 ret = -ENOSPC;
 554
 555         if (ret) {
 556                 bch2_trans_iter_exit(trans, iter);
 557                 return ret;
 558         }
 559
 560         /* Retry from start */
 561         pos = start = min;
 562         bch2_btree_iter_set_pos(iter, POS(0, pos));
 563         goto again;
 564 found_slot:
 565         bch2_btree_iter_set_pos(iter, SPOS(0, pos, snapshot));
 566         k = bch2_btree_iter_peek_slot(iter);
 567         ret = bkey_err(k);
 568         if (ret) {
 569                 bch2_trans_iter_exit(trans, iter);
 570                 return ret;
 571         }
 572
 573         /* We may have raced while the iterator wasn't pointing at pos: */
 574         if (bkey_is_inode(k.k) ||
 575             bch2_btree_key_cache_find(c, BTREE_ID_inodes, k.k->p))
 576                 goto again;
 577
 578         *hint                   = k.k->p.offset;
 579         inode_u->bi_inum        = k.k->p.offset;
 580         inode_u->bi_generation  = bkey_generation(k);
 581         return 0;
 582 }
 583
 584 static int bch2_inode_delete_keys(struct btree_trans *trans,
 585                                   subvol_inum inum, enum btree_id id)
 586 {
 587         u64 offset = 0;
 588         int ret = 0;
 589
 590         while (!ret || ret == -EINTR) {
 591                 struct btree_iter iter;
 592                 struct bkey_s_c k;
 593                 struct bkey_i delete;
 594                 u32 snapshot;
 595
 596                 bch2_trans_begin(trans);
 597
 598                 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 599                 if (ret)
 600                         continue;
 601
 602                 bch2_trans_iter_init(trans, &iter, id,
 603                                      SPOS(inum.inum, offset, snapshot),
 604                                      BTREE_ITER_INTENT);
 605                 k = bch2_btree_iter_peek(&iter);
 606
 607                 if (!k.k || iter.pos.inode != inum.inum) {
 608                         bch2_trans_iter_exit(trans, &iter);
 609                         break;
 610                 }
 611
 612                 ret = bkey_err(k);
 613                 if (ret)
 614                         goto err;
 615
 616                 bkey_init(&delete.k);
 617                 delete.k.p = iter.pos;
 618
 619                 if (btree_node_type_is_extents(iter.btree_id)) {
 620                         unsigned max_sectors =
 621                                 min_t(u64, U64_MAX - iter.pos.offset,
 622                                       KEY_SIZE_MAX & (~0 << trans->c->block_bits));
 623
 624                         /* create the biggest key we can */
 625                         bch2_key_resize(&delete.k, max_sectors);
 626
 627                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
 628                         if (ret)
 629                                 goto err;
 630                 }
 631
 632                 ret = bch2_trans_update(trans, &iter, &delete, 0) ?:
 633                       bch2_trans_commit(trans, NULL, NULL,
 634                                         BTREE_INSERT_NOFAIL);
 635 err:
 636                 offset = iter.pos.offset;
 637                 bch2_trans_iter_exit(trans, &iter);
 638         }
 639
 640         return ret;
 641 }
 642
 643 int bch2_inode_rm(struct bch_fs *c, subvol_inum inum, bool cached)
 644 {
 645         struct btree_trans trans;
 646         struct btree_iter iter = { NULL };
 647         struct bkey_i_inode_generation delete;
 648         struct bch_inode_unpacked inode_u;
 649         struct bkey_s_c k;
 650         unsigned iter_flags = BTREE_ITER_INTENT;
 651         u32 snapshot;
 652         int ret;
 653
 654         if (0 && cached && c->opts.inodes_use_key_cache)
 655                 iter_flags |= BTREE_ITER_CACHED;
 656
 657         bch2_trans_init(&trans, c, 0, 1024);
 658
 659         /*
 660          * If this was a directory, there shouldn't be any real dirents left -
 661          * but there could be whiteouts (from hash collisions) that we should
 662          * delete:
 663          *
 664          * XXX: the dirent could ideally would delete whiteouts when they're no
 665          * longer needed
 666          */
 667         ret   = bch2_inode_delete_keys(&trans, inum, BTREE_ID_extents) ?:
 668                 bch2_inode_delete_keys(&trans, inum, BTREE_ID_xattrs) ?:
 669                 bch2_inode_delete_keys(&trans, inum, BTREE_ID_dirents);
 670         if (ret)
 671                 goto err;
 672 retry:
 673         bch2_trans_begin(&trans);
 674
 675         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
 676         if (ret)
 677                 goto err;
 678
 679         bch2_trans_iter_init(&trans, &iter, BTREE_ID_inodes,
 680                              SPOS(0, inum.inum, snapshot), iter_flags);
 681         k = bch2_btree_iter_peek_slot(&iter);
 682
 683         ret = bkey_err(k);
 684         if (ret)
 685                 goto err;
 686
 687         if (!bkey_is_inode(k.k)) {
 688                 bch2_fs_inconsistent(trans.c,
 689                                      "inode %llu not found when deleting",
 690                                      inum.inum);
 691                 ret = -EIO;
 692                 goto err;
 693         }
 694
 695         bch2_inode_unpack(k, &inode_u);
 696
 697         /* Subvolume root? */
 698         BUG_ON(inode_u.bi_subvol);
 699
 700         bkey_inode_generation_init(&delete.k_i);
 701         delete.k.p = iter.pos;
 702         delete.v.bi_generation = cpu_to_le32(inode_u.bi_generation + 1);
 703
 704         ret   = bch2_trans_update(&trans, &iter, &delete.k_i, 0) ?:
 705                 bch2_trans_commit(&trans, NULL, NULL,
 706                                 BTREE_INSERT_NOFAIL);
 707 err:
 708         bch2_trans_iter_exit(&trans, &iter);
 709         if (ret == -EINTR)
 710                 goto retry;
 711
 712         bch2_trans_exit(&trans);
 713         return ret;
 714 }
 715
 716 int bch2_inode_find_by_inum_trans(struct btree_trans *trans,
 717                                   subvol_inum inum,
 718                                   struct bch_inode_unpacked *inode)
 719 {
 720         struct btree_iter iter;
 721         int ret;
 722
 723         ret = bch2_inode_peek(trans, &iter, inode, inum, 0);
 724         if (!ret)
 725                 bch2_trans_iter_exit(trans, &iter);
 726         return ret;
 727 }
 728
 729 int bch2_inode_find_by_inum(struct bch_fs *c, subvol_inum inum,
 730                             struct bch_inode_unpacked *inode)
 731 {
 732         return bch2_trans_do(c, NULL, NULL, 0,
 733                 bch2_inode_find_by_inum_trans(&trans, inum, inode));
 734 }