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 "buckets.h"
   8 #include "error.h"
   9 #include "extents.h"
  10 #include "extent_update.h"
  11 #include "inode.h"
  12 #include "str_hash.h"
  13 #include "subvolume.h"
  14 #include "varint.h"
  15
  16 #include <linux/random.h>
  17
  18 #include <asm/unaligned.h>
  19
  20 const char * const bch2_inode_opts[] = {
  21 #define x(name, ...)    #name,
  22         BCH_INODE_OPTS()
  23 #undef  x
  24         NULL,
  25 };
  26
  27 static const u8 byte_table[8] = { 1, 2, 3, 4, 6, 8, 10, 13 };
  28
  29 static int inode_decode_field(const u8 *in, const u8 *end,
  30                               u64 out[2], unsigned *out_bits)
  31 {
  32         __be64 be[2] = { 0, 0 };
  33         unsigned bytes, shift;
  34         u8 *p;
  35
  36         if (in >= end)
  37                 return -1;
  38
  39         if (!*in)
  40                 return -1;
  41
  42         /*
  43          * position of highest set bit indicates number of bytes:
  44          * shift = number of bits to remove in high byte:
  45          */
  46         shift   = 8 - __fls(*in); /* 1 <= shift <= 8 */
  47         bytes   = byte_table[shift - 1];
  48
  49         if (in + bytes > end)
  50                 return -1;
  51
  52         p = (u8 *) be + 16 - bytes;
  53         memcpy(p, in, bytes);
  54         *p ^= (1 << 8) >> shift;
  55
  56         out[0] = be64_to_cpu(be[0]);
  57         out[1] = be64_to_cpu(be[1]);
  58         *out_bits = out[0] ? 64 + fls64(out[0]) : fls64(out[1]);
  59
  60         return bytes;
  61 }
  62
  63 static inline void bch2_inode_pack_inlined(struct bkey_inode_buf *packed,
  64                                            const struct bch_inode_unpacked *inode)
  65 {
  66         struct bkey_i_inode_v3 *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_v3_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_sectors      = cpu_to_le64(inode->bi_sectors);
  80         packed->inode.v.bi_size         = cpu_to_le64(inode->bi_size);
  81         packed->inode.v.bi_version      = cpu_to_le64(inode->bi_version);
  82         SET_INODEv3_MODE(&packed->inode.v, inode->bi_mode);
  83         SET_INODEv3_FIELDS_START(&packed->inode.v, INODEv3_FIELDS_START_CUR);
  84
  85
  86 #define x(_name, _bits)                                                 \
  87         nr_fields++;                                                    \
  88                                                                         \
  89         if (inode->_name) {                                             \
  90                 ret = bch2_varint_encode_fast(out, inode->_name);       \
  91                 out += ret;                                             \
  92                                                                         \
  93                 if (_bits > 64)                                         \
  94                         *out++ = 0;                                     \
  95                                                                         \
  96                 last_nonzero_field = out;                               \
  97                 last_nonzero_fieldnr = nr_fields;                       \
  98         } else {                                                        \
  99                 *out++ = 0;                                             \
 100                                                                         \
 101                 if (_bits > 64)                                         \
 102                         *out++ = 0;                                     \
 103         }
 104
 105         BCH_INODE_FIELDS_v3()
 106 #undef  x
 107         BUG_ON(out > end);
 108
 109         out = last_nonzero_field;
 110         nr_fields = last_nonzero_fieldnr;
 111
 112         bytes = out - (u8 *) &packed->inode.v;
 113         set_bkey_val_bytes(&packed->inode.k, bytes);
 114         memset_u64s_tail(&packed->inode.v, 0, bytes);
 115
 116         SET_INODEv3_NR_FIELDS(&k->v, nr_fields);
 117
 118         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
 119                 struct bch_inode_unpacked unpacked;
 120
 121                 int ret = bch2_inode_unpack(bkey_i_to_s_c(&packed->inode.k_i),
 122                                            &unpacked);
 123                 BUG_ON(ret);
 124                 BUG_ON(unpacked.bi_inum         != inode->bi_inum);
 125                 BUG_ON(unpacked.bi_hash_seed    != inode->bi_hash_seed);
 126                 BUG_ON(unpacked.bi_sectors      != inode->bi_sectors);
 127                 BUG_ON(unpacked.bi_size         != inode->bi_size);
 128                 BUG_ON(unpacked.bi_version      != inode->bi_version);
 129                 BUG_ON(unpacked.bi_mode         != inode->bi_mode);
 130
 131 #define x(_name, _bits) if (unpacked._name != inode->_name)             \
 132                         panic("unpacked %llu should be %llu",           \
 133                               (u64) unpacked._name, (u64) inode->_name);
 134                 BCH_INODE_FIELDS_v3()
 135 #undef  x
 136         }
 137 }
 138
 139 void bch2_inode_pack(struct bkey_inode_buf *packed,
 140                      const struct bch_inode_unpacked *inode)
 141 {
 142         bch2_inode_pack_inlined(packed, inode);
 143 }
 144
 145 static noinline int bch2_inode_unpack_v1(struct bkey_s_c_inode inode,
 146                                 struct bch_inode_unpacked *unpacked)
 147 {
 148         const u8 *in = inode.v->fields;
 149         const u8 *end = bkey_val_end(inode);
 150         u64 field[2];
 151         unsigned fieldnr = 0, field_bits;
 152         int ret;
 153
 154 #define x(_name, _bits)                                 \
 155         if (fieldnr++ == INODE_NR_FIELDS(inode.v)) {                    \
 156                 unsigned offset = offsetof(struct bch_inode_unpacked, _name);\
 157                 memset((void *) unpacked + offset, 0,                   \
 158                        sizeof(*unpacked) - offset);                     \
 159                 return 0;                                               \
 160         }                                                               \
 161                                                                         \
 162         ret = inode_decode_field(in, end, field, &field_bits);          \
 163         if (ret < 0)                                                    \
 164                 return ret;                                             \
 165                                                                         \
 166         if (field_bits > sizeof(unpacked->_name) * 8)                   \
 167                 return -1;                                              \
 168                                                                         \
 169         unpacked->_name = field[1];                                     \
 170         in += ret;
 171
 172         BCH_INODE_FIELDS_v2()
 173 #undef  x
 174
 175         /* XXX: signal if there were more fields than expected? */
 176         return 0;
 177 }
 178
 179 static int bch2_inode_unpack_v2(struct bch_inode_unpacked *unpacked,
 180                                 const u8 *in, const u8 *end,
 181                                 unsigned nr_fields)
 182 {
 183         unsigned fieldnr = 0;
 184         int ret;
 185         u64 v[2];
 186
 187 #define x(_name, _bits)                                                 \
 188         if (fieldnr < nr_fields) {                                      \
 189                 ret = bch2_varint_decode_fast(in, end, &v[0]);          \
 190                 if (ret < 0)                                            \
 191                         return ret;                                     \
 192                 in += ret;                                              \
 193                                                                         \
 194                 if (_bits > 64) {                                       \
 195                         ret = bch2_varint_decode_fast(in, end, &v[1]);  \
 196                         if (ret < 0)                                    \
 197                                 return ret;                             \
 198                         in += ret;                                      \
 199                 } else {                                                \
 200                         v[1] = 0;                                       \
 201                 }                                                       \
 202         } else {                                                        \
 203                 v[0] = v[1] = 0;                                        \
 204         }                                                               \
 205                                                                         \
 206         unpacked->_name = v[0];                                         \
 207         if (v[1] || v[0] != unpacked->_name)                            \
 208                 return -1;                                              \
 209         fieldnr++;
 210
 211         BCH_INODE_FIELDS_v2()
 212 #undef  x
 213
 214         /* XXX: signal if there were more fields than expected? */
 215         return 0;
 216 }
 217
 218 static int bch2_inode_unpack_v3(struct bkey_s_c k,
 219                                 struct bch_inode_unpacked *unpacked)
 220 {
 221         struct bkey_s_c_inode_v3 inode = bkey_s_c_to_inode_v3(k);
 222         const u8 *in = inode.v->fields;
 223         const u8 *end = bkey_val_end(inode);
 224         unsigned nr_fields = INODEv3_NR_FIELDS(inode.v);
 225         unsigned fieldnr = 0;
 226         int ret;
 227         u64 v[2];
 228
 229         unpacked->bi_inum       = inode.k->p.offset;
 230         unpacked->bi_journal_seq= le64_to_cpu(inode.v->bi_journal_seq);
 231         unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 232         unpacked->bi_flags      = le64_to_cpu(inode.v->bi_flags);
 233         unpacked->bi_sectors    = le64_to_cpu(inode.v->bi_sectors);
 234         unpacked->bi_size       = le64_to_cpu(inode.v->bi_size);
 235         unpacked->bi_version    = le64_to_cpu(inode.v->bi_version);
 236         unpacked->bi_mode       = INODEv3_MODE(inode.v);
 237
 238 #define x(_name, _bits)                                                 \
 239         if (fieldnr < nr_fields) {                                      \
 240                 ret = bch2_varint_decode_fast(in, end, &v[0]);          \
 241                 if (ret < 0)                                            \
 242                         return ret;                                     \
 243                 in += ret;                                              \
 244                                                                         \
 245                 if (_bits > 64) {                                       \
 246                         ret = bch2_varint_decode_fast(in, end, &v[1]);  \
 247                         if (ret < 0)                                    \
 248                                 return ret;                             \
 249                         in += ret;                                      \
 250                 } else {                                                \
 251                         v[1] = 0;                                       \
 252                 }                                                       \
 253         } else {                                                        \
 254                 v[0] = v[1] = 0;                                        \
 255         }                                                               \
 256                                                                         \
 257         unpacked->_name = v[0];                                         \
 258         if (v[1] || v[0] != unpacked->_name)                            \
 259                 return -1;                                              \
 260         fieldnr++;
 261
 262         BCH_INODE_FIELDS_v3()
 263 #undef  x
 264
 265         /* XXX: signal if there were more fields than expected? */
 266         return 0;
 267 }
 268
 269 static noinline int bch2_inode_unpack_slowpath(struct bkey_s_c k,
 270                                                struct bch_inode_unpacked *unpacked)
 271 {
 272         memset(unpacked, 0, sizeof(*unpacked));
 273
 274         switch (k.k->type) {
 275         case KEY_TYPE_inode: {
 276                 struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 277
 278                 unpacked->bi_inum       = inode.k->p.offset;
 279                 unpacked->bi_journal_seq= 0;
 280                 unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 281                 unpacked->bi_flags      = le32_to_cpu(inode.v->bi_flags);
 282                 unpacked->bi_mode       = le16_to_cpu(inode.v->bi_mode);
 283
 284                 if (INODE_NEW_VARINT(inode.v)) {
 285                         return bch2_inode_unpack_v2(unpacked, inode.v->fields,
 286                                                     bkey_val_end(inode),
 287                                                     INODE_NR_FIELDS(inode.v));
 288                 } else {
 289                         return bch2_inode_unpack_v1(inode, unpacked);
 290                 }
 291                 break;
 292         }
 293         case KEY_TYPE_inode_v2: {
 294                 struct bkey_s_c_inode_v2 inode = bkey_s_c_to_inode_v2(k);
 295
 296                 unpacked->bi_inum       = inode.k->p.offset;
 297                 unpacked->bi_journal_seq= le64_to_cpu(inode.v->bi_journal_seq);
 298                 unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 299                 unpacked->bi_flags      = le64_to_cpu(inode.v->bi_flags);
 300                 unpacked->bi_mode       = le16_to_cpu(inode.v->bi_mode);
 301
 302                 return bch2_inode_unpack_v2(unpacked, inode.v->fields,
 303                                             bkey_val_end(inode),
 304                                             INODEv2_NR_FIELDS(inode.v));
 305         }
 306         default:
 307                 BUG();
 308         }
 309 }
 310
 311 int bch2_inode_unpack(struct bkey_s_c k,
 312                       struct bch_inode_unpacked *unpacked)
 313 {
 314         if (likely(k.k->type == KEY_TYPE_inode_v3))
 315                 return bch2_inode_unpack_v3(k, unpacked);
 316         return bch2_inode_unpack_slowpath(k, unpacked);
 317 }
 318
 319 int bch2_inode_peek(struct btree_trans *trans,
 320                     struct btree_iter *iter,
 321                     struct bch_inode_unpacked *inode,
 322                     subvol_inum inum, unsigned flags)
 323 {
 324         struct bkey_s_c k;
 325         u32 snapshot;
 326         int ret;
 327
 328         ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 329         if (ret)
 330                 return ret;
 331
 332         bch2_trans_iter_init(trans, iter, BTREE_ID_inodes,
 333                              SPOS(0, inum.inum, snapshot),
 334                              flags|BTREE_ITER_CACHED);
 335         k = bch2_btree_iter_peek_slot(iter);
 336         ret = bkey_err(k);
 337         if (ret)
 338                 goto err;
 339
 340         ret = bkey_is_inode(k.k) ? 0 : -ENOENT;
 341         if (ret)
 342                 goto err;
 343
 344         ret = bch2_inode_unpack(k, inode);
 345         if (ret)
 346                 goto err;
 347
 348         return 0;
 349 err:
 350         bch2_trans_iter_exit(trans, iter);
 351         return ret;
 352 }
 353
 354 int bch2_inode_write(struct btree_trans *trans,
 355                      struct btree_iter *iter,
 356                      struct bch_inode_unpacked *inode)
 357 {
 358         struct bkey_inode_buf *inode_p;
 359
 360         inode_p = bch2_trans_kmalloc(trans, sizeof(*inode_p));
 361         if (IS_ERR(inode_p))
 362                 return PTR_ERR(inode_p);
 363
 364         bch2_inode_pack_inlined(inode_p, inode);
 365         inode_p->inode.k.p.snapshot = iter->snapshot;
 366         return bch2_trans_update(trans, iter, &inode_p->inode.k_i, 0);
 367 }
 368
 369 struct bkey_i *bch2_inode_to_v3(struct btree_trans *trans, struct bkey_i *k)
 370 {
 371         struct bch_inode_unpacked u;
 372         struct bkey_inode_buf *inode_p;
 373         int ret;
 374
 375         if (!bkey_is_inode(&k->k))
 376                 return ERR_PTR(-ENOENT);
 377
 378         inode_p = bch2_trans_kmalloc(trans, sizeof(*inode_p));
 379         if (IS_ERR(inode_p))
 380                 return ERR_CAST(inode_p);
 381
 382         ret = bch2_inode_unpack(bkey_i_to_s_c(k), &u);
 383         if (ret)
 384                 return ERR_PTR(ret);
 385
 386         bch2_inode_pack(inode_p, &u);
 387         return &inode_p->inode.k_i;
 388 }
 389
 390 static int __bch2_inode_invalid(struct bkey_s_c k, struct printbuf *err)
 391 {
 392         struct bch_inode_unpacked unpacked;
 393
 394         if (k.k->p.inode) {
 395                 prt_printf(err, "nonzero k.p.inode");
 396                 return -BCH_ERR_invalid_bkey;
 397         }
 398
 399         if (k.k->p.offset < BLOCKDEV_INODE_MAX) {
 400                 prt_printf(err, "fs inode in blockdev range");
 401                 return -BCH_ERR_invalid_bkey;
 402         }
 403
 404         if (bch2_inode_unpack(k, &unpacked)) {
 405                 prt_printf(err, "invalid variable length fields");
 406                 return -BCH_ERR_invalid_bkey;
 407         }
 408
 409         if (unpacked.bi_data_checksum >= BCH_CSUM_OPT_NR + 1) {
 410                 prt_printf(err, "invalid data checksum type (%u >= %u",
 411                         unpacked.bi_data_checksum, BCH_CSUM_OPT_NR + 1);
 412                 return -BCH_ERR_invalid_bkey;
 413         }
 414
 415         if (unpacked.bi_compression >= BCH_COMPRESSION_OPT_NR + 1) {
 416                 prt_printf(err, "invalid data checksum type (%u >= %u)",
 417                        unpacked.bi_compression, BCH_COMPRESSION_OPT_NR + 1);
 418                 return -BCH_ERR_invalid_bkey;
 419         }
 420
 421         if ((unpacked.bi_flags & BCH_INODE_UNLINKED) &&
 422             unpacked.bi_nlink != 0) {
 423                 prt_printf(err, "flagged as unlinked but bi_nlink != 0");
 424                 return -BCH_ERR_invalid_bkey;
 425         }
 426
 427         if (unpacked.bi_subvol && !S_ISDIR(unpacked.bi_mode)) {
 428                 prt_printf(err, "subvolume root but not a directory");
 429                 return -BCH_ERR_invalid_bkey;
 430         }
 431
 432         return 0;
 433 }
 434
 435 int bch2_inode_invalid(const struct bch_fs *c, struct bkey_s_c k,
 436                        int rw, struct printbuf *err)
 437 {
 438         struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 439
 440         if (bkey_val_bytes(k.k) < sizeof(*inode.v)) {
 441                 prt_printf(err, "incorrect value size (%zu < %zu)",
 442                        bkey_val_bytes(k.k), sizeof(*inode.v));
 443                 return -BCH_ERR_invalid_bkey;
 444         }
 445
 446         if (INODE_STR_HASH(inode.v) >= BCH_STR_HASH_NR) {
 447                 prt_printf(err, "invalid str hash type (%llu >= %u)",
 448                        INODE_STR_HASH(inode.v), BCH_STR_HASH_NR);
 449                 return -BCH_ERR_invalid_bkey;
 450         }
 451
 452         return __bch2_inode_invalid(k, err);
 453 }
 454
 455 int bch2_inode_v2_invalid(const struct bch_fs *c, struct bkey_s_c k,
 456                           int rw, struct printbuf *err)
 457 {
 458         struct bkey_s_c_inode_v2 inode = bkey_s_c_to_inode_v2(k);
 459
 460         if (bkey_val_bytes(k.k) < sizeof(*inode.v)) {
 461                 prt_printf(err, "incorrect value size (%zu < %zu)",
 462                        bkey_val_bytes(k.k), sizeof(*inode.v));
 463                 return -BCH_ERR_invalid_bkey;
 464         }
 465
 466         if (INODEv2_STR_HASH(inode.v) >= BCH_STR_HASH_NR) {
 467                 prt_printf(err, "invalid str hash type (%llu >= %u)",
 468                        INODEv2_STR_HASH(inode.v), BCH_STR_HASH_NR);
 469                 return -BCH_ERR_invalid_bkey;
 470         }
 471
 472         return __bch2_inode_invalid(k, err);
 473 }
 474
 475 int bch2_inode_v3_invalid(const struct bch_fs *c, struct bkey_s_c k,
 476                           int rw, struct printbuf *err)
 477 {
 478         struct bkey_s_c_inode_v3 inode = bkey_s_c_to_inode_v3(k);
 479
 480         if (bkey_val_bytes(k.k) < sizeof(*inode.v)) {
 481                 prt_printf(err, "incorrect value size (%zu < %zu)",
 482                        bkey_val_bytes(k.k), sizeof(*inode.v));
 483                 return -BCH_ERR_invalid_bkey;
 484         }
 485
 486         if (INODEv3_FIELDS_START(inode.v) < INODEv3_FIELDS_START_INITIAL ||
 487             INODEv3_FIELDS_START(inode.v) > bkey_val_u64s(inode.k)) {
 488                 prt_printf(err, "invalid fields_start (got %llu, min %u max %zu)",
 489                        INODEv3_FIELDS_START(inode.v),
 490                        INODEv3_FIELDS_START_INITIAL,
 491                        bkey_val_u64s(inode.k));
 492                 return -BCH_ERR_invalid_bkey;
 493         }
 494
 495         if (INODEv3_STR_HASH(inode.v) >= BCH_STR_HASH_NR) {
 496                 prt_printf(err, "invalid str hash type (%llu >= %u)",
 497                        INODEv3_STR_HASH(inode.v), BCH_STR_HASH_NR);
 498                 return -BCH_ERR_invalid_bkey;
 499         }
 500
 501         return __bch2_inode_invalid(k, err);
 502 }
 503
 504 static void __bch2_inode_unpacked_to_text(struct printbuf *out,
 505                                           struct bch_inode_unpacked *inode)
 506 {
 507         prt_printf(out, "mode %o flags %x journal_seq %llu bi_size %llu bi_sectors %llu bi_version %llu",
 508                inode->bi_mode, inode->bi_flags,
 509                inode->bi_journal_seq,
 510                inode->bi_size,
 511                inode->bi_sectors,
 512                inode->bi_version);
 513
 514 #define x(_name, _bits)                                         \
 515         prt_printf(out, " "#_name " %llu", (u64) inode->_name);
 516         BCH_INODE_FIELDS_v3()
 517 #undef  x
 518 }
 519
 520 void bch2_inode_unpacked_to_text(struct printbuf *out, struct bch_inode_unpacked *inode)
 521 {
 522         prt_printf(out, "inum: %llu ", inode->bi_inum);
 523         __bch2_inode_unpacked_to_text(out, inode);
 524 }
 525
 526 void bch2_inode_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 527 {
 528         struct bch_inode_unpacked inode;
 529
 530         if (bch2_inode_unpack(k, &inode)) {
 531                 prt_printf(out, "(unpack error)");
 532                 return;
 533         }
 534
 535         __bch2_inode_unpacked_to_text(out, &inode);
 536 }
 537
 538 int bch2_inode_generation_invalid(const struct bch_fs *c, struct bkey_s_c k,
 539                                   int rw, struct printbuf *err)
 540 {
 541         if (k.k->p.inode) {
 542                 prt_printf(err, "nonzero k.p.inode");
 543                 return -BCH_ERR_invalid_bkey;
 544         }
 545
 546         if (bkey_val_bytes(k.k) != sizeof(struct bch_inode_generation)) {
 547                 prt_printf(err, "incorrect value size (%zu != %zu)",
 548                        bkey_val_bytes(k.k), sizeof(struct bch_inode_generation));
 549                 return -BCH_ERR_invalid_bkey;
 550         }
 551
 552         return 0;
 553 }
 554
 555 void bch2_inode_generation_to_text(struct printbuf *out, struct bch_fs *c,
 556                                    struct bkey_s_c k)
 557 {
 558         struct bkey_s_c_inode_generation gen = bkey_s_c_to_inode_generation(k);
 559
 560         prt_printf(out, "generation: %u", le32_to_cpu(gen.v->bi_generation));
 561 }
 562
 563 void bch2_inode_init_early(struct bch_fs *c,
 564                            struct bch_inode_unpacked *inode_u)
 565 {
 566         enum bch_str_hash_type str_hash =
 567                 bch2_str_hash_opt_to_type(c, c->opts.str_hash);
 568
 569         memset(inode_u, 0, sizeof(*inode_u));
 570
 571         /* ick */
 572         inode_u->bi_flags |= str_hash << INODE_STR_HASH_OFFSET;
 573         get_random_bytes(&inode_u->bi_hash_seed,
 574                          sizeof(inode_u->bi_hash_seed));
 575 }
 576
 577 void bch2_inode_init_late(struct bch_inode_unpacked *inode_u, u64 now,
 578                           uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 579                           struct bch_inode_unpacked *parent)
 580 {
 581         inode_u->bi_mode        = mode;
 582         inode_u->bi_uid         = uid;
 583         inode_u->bi_gid         = gid;
 584         inode_u->bi_dev         = rdev;
 585         inode_u->bi_atime       = now;
 586         inode_u->bi_mtime       = now;
 587         inode_u->bi_ctime       = now;
 588         inode_u->bi_otime       = now;
 589
 590         if (parent && parent->bi_mode & S_ISGID) {
 591                 inode_u->bi_gid = parent->bi_gid;
 592                 if (S_ISDIR(mode))
 593                         inode_u->bi_mode |= S_ISGID;
 594         }
 595
 596         if (parent) {
 597 #define x(_name, ...)   inode_u->bi_##_name = parent->bi_##_name;
 598                 BCH_INODE_OPTS()
 599 #undef x
 600         }
 601 }
 602
 603 void bch2_inode_init(struct bch_fs *c, struct bch_inode_unpacked *inode_u,
 604                      uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 605                      struct bch_inode_unpacked *parent)
 606 {
 607         bch2_inode_init_early(c, inode_u);
 608         bch2_inode_init_late(inode_u, bch2_current_time(c),
 609                              uid, gid, mode, rdev, parent);
 610 }
 611
 612 static inline u32 bkey_generation(struct bkey_s_c k)
 613 {
 614         switch (k.k->type) {
 615         case KEY_TYPE_inode:
 616         case KEY_TYPE_inode_v2:
 617                 BUG();
 618         case KEY_TYPE_inode_generation:
 619                 return le32_to_cpu(bkey_s_c_to_inode_generation(k).v->bi_generation);
 620         default:
 621                 return 0;
 622         }
 623 }
 624
 625 /*
 626  * This just finds an empty slot:
 627  */
 628 int bch2_inode_create(struct btree_trans *trans,
 629                       struct btree_iter *iter,
 630                       struct bch_inode_unpacked *inode_u,
 631                       u32 snapshot, u64 cpu)
 632 {
 633         struct bch_fs *c = trans->c;
 634         struct bkey_s_c k;
 635         u64 min, max, start, pos, *hint;
 636         int ret = 0;
 637         unsigned bits = (c->opts.inodes_32bit ? 31 : 63);
 638
 639         if (c->opts.shard_inode_numbers) {
 640                 bits -= c->inode_shard_bits;
 641
 642                 min = (cpu << bits);
 643                 max = (cpu << bits) | ~(ULLONG_MAX << bits);
 644
 645                 min = max_t(u64, min, BLOCKDEV_INODE_MAX);
 646                 hint = c->unused_inode_hints + cpu;
 647         } else {
 648                 min = BLOCKDEV_INODE_MAX;
 649                 max = ~(ULLONG_MAX << bits);
 650                 hint = c->unused_inode_hints;
 651         }
 652
 653         start = READ_ONCE(*hint);
 654
 655         if (start >= max || start < min)
 656                 start = min;
 657
 658         pos = start;
 659         bch2_trans_iter_init(trans, iter, BTREE_ID_inodes, POS(0, pos),
 660                              BTREE_ITER_ALL_SNAPSHOTS|
 661                              BTREE_ITER_INTENT);
 662 again:
 663         while ((k = bch2_btree_iter_peek(iter)).k &&
 664                !(ret = bkey_err(k)) &&
 665                bkey_lt(k.k->p, POS(0, max))) {
 666                 while (pos < iter->pos.offset) {
 667                         if (!bch2_btree_key_cache_find(c, BTREE_ID_inodes, POS(0, pos)))
 668                                 goto found_slot;
 669
 670                         pos++;
 671                 }
 672
 673                 if (k.k->p.snapshot == snapshot &&
 674                     !bkey_is_inode(k.k) &&
 675                     !bch2_btree_key_cache_find(c, BTREE_ID_inodes, SPOS(0, pos, snapshot))) {
 676                         bch2_btree_iter_advance(iter);
 677                         continue;
 678                 }
 679
 680                 /*
 681                  * We don't need to iterate over keys in every snapshot once
 682                  * we've found just one:
 683                  */
 684                 pos = iter->pos.offset + 1;
 685                 bch2_btree_iter_set_pos(iter, POS(0, pos));
 686         }
 687
 688         while (!ret && pos < max) {
 689                 if (!bch2_btree_key_cache_find(c, BTREE_ID_inodes, POS(0, pos)))
 690                         goto found_slot;
 691
 692                 pos++;
 693         }
 694
 695         if (!ret && start == min)
 696                 ret = -BCH_ERR_ENOSPC_inode_create;
 697
 698         if (ret) {
 699                 bch2_trans_iter_exit(trans, iter);
 700                 return ret;
 701         }
 702
 703         /* Retry from start */
 704         pos = start = min;
 705         bch2_btree_iter_set_pos(iter, POS(0, pos));
 706         goto again;
 707 found_slot:
 708         bch2_btree_iter_set_pos(iter, SPOS(0, pos, snapshot));
 709         k = bch2_btree_iter_peek_slot(iter);
 710         ret = bkey_err(k);
 711         if (ret) {
 712                 bch2_trans_iter_exit(trans, iter);
 713                 return ret;
 714         }
 715
 716         /* We may have raced while the iterator wasn't pointing at pos: */
 717         if (bkey_is_inode(k.k) ||
 718             bch2_btree_key_cache_find(c, BTREE_ID_inodes, k.k->p))
 719                 goto again;
 720
 721         *hint                   = k.k->p.offset;
 722         inode_u->bi_inum        = k.k->p.offset;
 723         inode_u->bi_generation  = bkey_generation(k);
 724         return 0;
 725 }
 726
 727 static int bch2_inode_delete_keys(struct btree_trans *trans,
 728                                   subvol_inum inum, enum btree_id id)
 729 {
 730         struct btree_iter iter;
 731         struct bkey_s_c k;
 732         struct bkey_i delete;
 733         u32 snapshot;
 734         int ret = 0;
 735
 736         /*
 737          * We're never going to be deleting extents, no need to use an extent
 738          * iterator:
 739          */
 740         bch2_trans_iter_init(trans, &iter, id, POS(inum.inum, 0),
 741                              BTREE_ITER_INTENT);
 742
 743         while (1) {
 744                 bch2_trans_begin(trans);
 745
 746                 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 747                 if (ret)
 748                         goto err;
 749
 750                 bch2_btree_iter_set_snapshot(&iter, snapshot);
 751
 752                 k = bch2_btree_iter_peek_upto(&iter, POS(inum.inum, U64_MAX));
 753                 ret = bkey_err(k);
 754                 if (ret)
 755                         goto err;
 756
 757                 if (!k.k)
 758                         break;
 759
 760                 bkey_init(&delete.k);
 761                 delete.k.p = iter.pos;
 762
 763                 if (iter.flags & BTREE_ITER_IS_EXTENTS) {
 764                         bch2_key_resize(&delete.k, k.k->p.offset - iter.pos.offset);
 765
 766                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
 767                         if (ret)
 768                                 goto err;
 769                 }
 770
 771                 ret = bch2_trans_update(trans, &iter, &delete, 0) ?:
 772                       bch2_trans_commit(trans, NULL, NULL,
 773                                         BTREE_INSERT_NOFAIL);
 774 err:
 775                 if (ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart))
 776                         break;
 777         }
 778
 779         bch2_trans_iter_exit(trans, &iter);
 780         return ret;
 781 }
 782
 783 int bch2_inode_rm(struct bch_fs *c, subvol_inum inum)
 784 {
 785         struct btree_trans trans;
 786         struct btree_iter iter = { NULL };
 787         struct bkey_i_inode_generation delete;
 788         struct bch_inode_unpacked inode_u;
 789         struct bkey_s_c k;
 790         u32 snapshot;
 791         int ret;
 792
 793         bch2_trans_init(&trans, c, 0, 1024);
 794
 795         /*
 796          * If this was a directory, there shouldn't be any real dirents left -
 797          * but there could be whiteouts (from hash collisions) that we should
 798          * delete:
 799          *
 800          * XXX: the dirent could ideally would delete whiteouts when they're no
 801          * longer needed
 802          */
 803         ret   = bch2_inode_delete_keys(&trans, inum, BTREE_ID_extents) ?:
 804                 bch2_inode_delete_keys(&trans, inum, BTREE_ID_xattrs) ?:
 805                 bch2_inode_delete_keys(&trans, inum, BTREE_ID_dirents);
 806         if (ret)
 807                 goto err;
 808 retry:
 809         bch2_trans_begin(&trans);
 810
 811         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
 812         if (ret)
 813                 goto err;
 814
 815         bch2_trans_iter_init(&trans, &iter, BTREE_ID_inodes,
 816                              SPOS(0, inum.inum, snapshot),
 817                              BTREE_ITER_INTENT|BTREE_ITER_CACHED);
 818         k = bch2_btree_iter_peek_slot(&iter);
 819
 820         ret = bkey_err(k);
 821         if (ret)
 822                 goto err;
 823
 824         if (!bkey_is_inode(k.k)) {
 825                 bch2_fs_inconsistent(trans.c,
 826                                      "inode %llu not found when deleting",
 827                                      inum.inum);
 828                 ret = -EIO;
 829                 goto err;
 830         }
 831
 832         bch2_inode_unpack(k, &inode_u);
 833
 834         /* Subvolume root? */
 835         BUG_ON(inode_u.bi_subvol);
 836
 837         bkey_inode_generation_init(&delete.k_i);
 838         delete.k.p = iter.pos;
 839         delete.v.bi_generation = cpu_to_le32(inode_u.bi_generation + 1);
 840
 841         ret   = bch2_trans_update(&trans, &iter, &delete.k_i, 0) ?:
 842                 bch2_trans_commit(&trans, NULL, NULL,
 843                                 BTREE_INSERT_NOFAIL);
 844 err:
 845         bch2_trans_iter_exit(&trans, &iter);
 846         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 847                 goto retry;
 848
 849         bch2_trans_exit(&trans);
 850         return ret;
 851 }
 852
 853 int bch2_inode_find_by_inum_trans(struct btree_trans *trans,
 854                                   subvol_inum inum,
 855                                   struct bch_inode_unpacked *inode)
 856 {
 857         struct btree_iter iter;
 858         int ret;
 859
 860         ret = bch2_inode_peek(trans, &iter, inode, inum, 0);
 861         if (!ret)
 862                 bch2_trans_iter_exit(trans, &iter);
 863         return ret;
 864 }
 865
 866 int bch2_inode_find_by_inum(struct bch_fs *c, subvol_inum inum,
 867                             struct bch_inode_unpacked *inode)
 868 {
 869         return bch2_trans_do(c, NULL, NULL, 0,
 870                 bch2_inode_find_by_inum_trans(&trans, inum, inode));
 871 }
 872
 873 int bch2_inode_nlink_inc(struct bch_inode_unpacked *bi)
 874 {
 875         if (bi->bi_flags & BCH_INODE_UNLINKED)
 876                 bi->bi_flags &= ~BCH_INODE_UNLINKED;
 877         else {
 878                 if (bi->bi_nlink == U32_MAX)
 879                         return -EINVAL;
 880
 881                 bi->bi_nlink++;
 882         }
 883
 884         return 0;
 885 }
 886
 887 void bch2_inode_nlink_dec(struct btree_trans *trans, struct bch_inode_unpacked *bi)
 888 {
 889         if (bi->bi_nlink && (bi->bi_flags & BCH_INODE_UNLINKED)) {
 890                 bch2_trans_inconsistent(trans, "inode %llu unlinked but link count nonzero",
 891                                         bi->bi_inum);
 892                 return;
 893         }
 894
 895         if (bi->bi_flags & BCH_INODE_UNLINKED) {
 896                 bch2_trans_inconsistent(trans, "inode %llu link count underflow", bi->bi_inum);
 897                 return;
 898         }
 899
 900         if (bi->bi_nlink)
 901                 bi->bi_nlink--;
 902         else
 903                 bi->bi_flags |= BCH_INODE_UNLINKED;
 904 }
 905
 906 struct bch_opts bch2_inode_opts_to_opts(struct bch_inode_unpacked *inode)
 907 {
 908         struct bch_opts ret = { 0 };
 909 #define x(_name, _bits)                                                 \
 910         if (inode->bi_##_name)                                          \
 911                 opt_set(ret, _name, inode->bi_##_name - 1);
 912         BCH_INODE_OPTS()
 913 #undef x
 914         return ret;
 915 }
 916
 917 void bch2_inode_opts_get(struct bch_io_opts *opts, struct bch_fs *c,
 918                          struct bch_inode_unpacked *inode)
 919 {
 920 #define x(_name, _bits)         opts->_name = inode_opt_get(c, inode, _name);
 921         BCH_INODE_OPTS()
 922 #undef x
 923
 924         if (opts->nocow)
 925                 opts->compression = opts->background_compression = opts->data_checksum = opts->erasure_code = 0;
 926 }