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         switch (k.k->type) {
 273         case KEY_TYPE_inode: {
 274                 struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 275
 276                 unpacked->bi_inum       = inode.k->p.offset;
 277                 unpacked->bi_journal_seq= 0;
 278                 unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 279                 unpacked->bi_flags      = le32_to_cpu(inode.v->bi_flags);
 280                 unpacked->bi_mode       = le16_to_cpu(inode.v->bi_mode);
 281
 282                 if (INODE_NEW_VARINT(inode.v)) {
 283                         return bch2_inode_unpack_v2(unpacked, inode.v->fields,
 284                                                     bkey_val_end(inode),
 285                                                     INODE_NR_FIELDS(inode.v));
 286                 } else {
 287                         return bch2_inode_unpack_v1(inode, unpacked);
 288                 }
 289                 break;
 290         }
 291         case KEY_TYPE_inode_v2: {
 292                 struct bkey_s_c_inode_v2 inode = bkey_s_c_to_inode_v2(k);
 293
 294                 unpacked->bi_inum       = inode.k->p.offset;
 295                 unpacked->bi_journal_seq= le64_to_cpu(inode.v->bi_journal_seq);
 296                 unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 297                 unpacked->bi_flags      = le64_to_cpu(inode.v->bi_flags);
 298                 unpacked->bi_mode       = le16_to_cpu(inode.v->bi_mode);
 299
 300                 return bch2_inode_unpack_v2(unpacked, inode.v->fields,
 301                                             bkey_val_end(inode),
 302                                             INODEv2_NR_FIELDS(inode.v));
 303         }
 304         default:
 305                 BUG();
 306         }
 307 }
 308
 309 int bch2_inode_unpack(struct bkey_s_c k,
 310                       struct bch_inode_unpacked *unpacked)
 311 {
 312         if (likely(k.k->type == KEY_TYPE_inode_v3))
 313                 return bch2_inode_unpack_v3(k, unpacked);
 314         return bch2_inode_unpack_slowpath(k, unpacked);
 315 }
 316
 317 int bch2_inode_peek(struct btree_trans *trans,
 318                     struct btree_iter *iter,
 319                     struct bch_inode_unpacked *inode,
 320                     subvol_inum inum, unsigned flags)
 321 {
 322         struct bkey_s_c k;
 323         u32 snapshot;
 324         int ret;
 325
 326         ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 327         if (ret)
 328                 return ret;
 329
 330         bch2_trans_iter_init(trans, iter, BTREE_ID_inodes,
 331                              SPOS(0, inum.inum, snapshot),
 332                              flags|BTREE_ITER_CACHED);
 333         k = bch2_btree_iter_peek_slot(iter);
 334         ret = bkey_err(k);
 335         if (ret)
 336                 goto err;
 337
 338         ret = bkey_is_inode(k.k) ? 0 : -ENOENT;
 339         if (ret)
 340                 goto err;
 341
 342         ret = bch2_inode_unpack(k, inode);
 343         if (ret)
 344                 goto err;
 345
 346         return 0;
 347 err:
 348         bch2_trans_iter_exit(trans, iter);
 349         return ret;
 350 }
 351
 352 int bch2_inode_write(struct btree_trans *trans,
 353                      struct btree_iter *iter,
 354                      struct bch_inode_unpacked *inode)
 355 {
 356         struct bkey_inode_buf *inode_p;
 357
 358         inode_p = bch2_trans_kmalloc(trans, sizeof(*inode_p));
 359         if (IS_ERR(inode_p))
 360                 return PTR_ERR(inode_p);
 361
 362         bch2_inode_pack_inlined(inode_p, inode);
 363         inode_p->inode.k.p.snapshot = iter->snapshot;
 364         return bch2_trans_update(trans, iter, &inode_p->inode.k_i, 0);
 365 }
 366
 367 struct bkey_s_c bch2_inode_to_v3(struct btree_trans *trans, struct bkey_s_c k)
 368 {
 369         struct bch_inode_unpacked u;
 370         struct bkey_inode_buf *inode_p;
 371         int ret;
 372
 373         inode_p = bch2_trans_kmalloc(trans, sizeof(*inode_p));
 374         if (IS_ERR(inode_p))
 375                 return bkey_s_c_err(PTR_ERR(inode_p));
 376
 377         ret = bch2_inode_unpack(k, &u);
 378         if (ret)
 379                 return bkey_s_c_err(ret);
 380
 381         bch2_inode_pack(inode_p, &u);
 382         return bkey_i_to_s_c(&inode_p->inode.k_i);
 383 }
 384
 385 static int __bch2_inode_invalid(struct bkey_s_c k, struct printbuf *err)
 386 {
 387         struct bch_inode_unpacked unpacked;
 388
 389         if (k.k->p.inode) {
 390                 prt_printf(err, "nonzero k.p.inode");
 391                 return -BCH_ERR_invalid_bkey;
 392         }
 393
 394         if (k.k->p.offset < BLOCKDEV_INODE_MAX) {
 395                 prt_printf(err, "fs inode in blockdev range");
 396                 return -BCH_ERR_invalid_bkey;
 397         }
 398
 399         if (bch2_inode_unpack(k, &unpacked)) {
 400                 prt_printf(err, "invalid variable length fields");
 401                 return -BCH_ERR_invalid_bkey;
 402         }
 403
 404         if (unpacked.bi_data_checksum >= BCH_CSUM_OPT_NR + 1) {
 405                 prt_printf(err, "invalid data checksum type (%u >= %u",
 406                         unpacked.bi_data_checksum, BCH_CSUM_OPT_NR + 1);
 407                 return -BCH_ERR_invalid_bkey;
 408         }
 409
 410         if (unpacked.bi_compression >= BCH_COMPRESSION_OPT_NR + 1) {
 411                 prt_printf(err, "invalid data checksum type (%u >= %u)",
 412                        unpacked.bi_compression, BCH_COMPRESSION_OPT_NR + 1);
 413                 return -BCH_ERR_invalid_bkey;
 414         }
 415
 416         if ((unpacked.bi_flags & BCH_INODE_UNLINKED) &&
 417             unpacked.bi_nlink != 0) {
 418                 prt_printf(err, "flagged as unlinked but bi_nlink != 0");
 419                 return -BCH_ERR_invalid_bkey;
 420         }
 421
 422         if (unpacked.bi_subvol && !S_ISDIR(unpacked.bi_mode)) {
 423                 prt_printf(err, "subvolume root but not a directory");
 424                 return -BCH_ERR_invalid_bkey;
 425         }
 426
 427         return 0;
 428 }
 429
 430 int bch2_inode_invalid(const struct bch_fs *c, struct bkey_s_c k,
 431                        int rw, struct printbuf *err)
 432 {
 433         struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 434
 435         if (bkey_val_bytes(k.k) < sizeof(*inode.v)) {
 436                 prt_printf(err, "incorrect value size (%zu < %zu)",
 437                        bkey_val_bytes(k.k), sizeof(*inode.v));
 438                 return -BCH_ERR_invalid_bkey;
 439         }
 440
 441         if (INODE_STR_HASH(inode.v) >= BCH_STR_HASH_NR) {
 442                 prt_printf(err, "invalid str hash type (%llu >= %u)",
 443                        INODE_STR_HASH(inode.v), BCH_STR_HASH_NR);
 444                 return -BCH_ERR_invalid_bkey;
 445         }
 446
 447         return __bch2_inode_invalid(k, err);
 448 }
 449
 450 int bch2_inode_v2_invalid(const struct bch_fs *c, struct bkey_s_c k,
 451                           int rw, struct printbuf *err)
 452 {
 453         struct bkey_s_c_inode_v2 inode = bkey_s_c_to_inode_v2(k);
 454
 455         if (bkey_val_bytes(k.k) < sizeof(*inode.v)) {
 456                 prt_printf(err, "incorrect value size (%zu < %zu)",
 457                        bkey_val_bytes(k.k), sizeof(*inode.v));
 458                 return -BCH_ERR_invalid_bkey;
 459         }
 460
 461         if (INODEv2_STR_HASH(inode.v) >= BCH_STR_HASH_NR) {
 462                 prt_printf(err, "invalid str hash type (%llu >= %u)",
 463                        INODEv2_STR_HASH(inode.v), BCH_STR_HASH_NR);
 464                 return -BCH_ERR_invalid_bkey;
 465         }
 466
 467         return __bch2_inode_invalid(k, err);
 468 }
 469
 470 int bch2_inode_v3_invalid(const struct bch_fs *c, struct bkey_s_c k,
 471                           int rw, struct printbuf *err)
 472 {
 473         struct bkey_s_c_inode_v3 inode = bkey_s_c_to_inode_v3(k);
 474
 475         if (bkey_val_bytes(k.k) < sizeof(*inode.v)) {
 476                 prt_printf(err, "incorrect value size (%zu < %zu)",
 477                        bkey_val_bytes(k.k), sizeof(*inode.v));
 478                 return -BCH_ERR_invalid_bkey;
 479         }
 480
 481         if (INODEv3_FIELDS_START(inode.v) < INODEv3_FIELDS_START_INITIAL ||
 482             INODEv3_FIELDS_START(inode.v) > bkey_val_u64s(inode.k)) {
 483                 prt_printf(err, "invalid fields_start (got %llu, min %u max %zu)",
 484                        INODEv3_FIELDS_START(inode.v),
 485                        INODEv3_FIELDS_START_INITIAL,
 486                        bkey_val_u64s(inode.k));
 487                 return -BCH_ERR_invalid_bkey;
 488         }
 489
 490         if (INODEv3_STR_HASH(inode.v) >= BCH_STR_HASH_NR) {
 491                 prt_printf(err, "invalid str hash type (%llu >= %u)",
 492                        INODEv3_STR_HASH(inode.v), BCH_STR_HASH_NR);
 493                 return -BCH_ERR_invalid_bkey;
 494         }
 495
 496         return __bch2_inode_invalid(k, err);
 497 }
 498
 499 static void __bch2_inode_unpacked_to_text(struct printbuf *out,
 500                                           struct bch_inode_unpacked *inode)
 501 {
 502         prt_printf(out, "mode %o flags %x journal_seq %llu bi_size %llu bi_sectors %llu bi_version %llu",
 503                inode->bi_mode, inode->bi_flags,
 504                inode->bi_journal_seq,
 505                inode->bi_size,
 506                inode->bi_sectors,
 507                inode->bi_version);
 508
 509 #define x(_name, _bits)                                         \
 510         prt_printf(out, " "#_name " %llu", (u64) inode->_name);
 511         BCH_INODE_FIELDS_v3()
 512 #undef  x
 513 }
 514
 515 void bch2_inode_unpacked_to_text(struct printbuf *out, struct bch_inode_unpacked *inode)
 516 {
 517         prt_printf(out, "inum: %llu ", inode->bi_inum);
 518         __bch2_inode_unpacked_to_text(out, inode);
 519 }
 520
 521 void bch2_inode_to_text(struct printbuf *out, struct bch_fs *c, struct bkey_s_c k)
 522 {
 523         struct bch_inode_unpacked inode;
 524
 525         if (bch2_inode_unpack(k, &inode)) {
 526                 prt_printf(out, "(unpack error)");
 527                 return;
 528         }
 529
 530         __bch2_inode_unpacked_to_text(out, &inode);
 531 }
 532
 533 int bch2_inode_generation_invalid(const struct bch_fs *c, struct bkey_s_c k,
 534                                   int rw, struct printbuf *err)
 535 {
 536         if (k.k->p.inode) {
 537                 prt_printf(err, "nonzero k.p.inode");
 538                 return -BCH_ERR_invalid_bkey;
 539         }
 540
 541         if (bkey_val_bytes(k.k) != sizeof(struct bch_inode_generation)) {
 542                 prt_printf(err, "incorrect value size (%zu != %zu)",
 543                        bkey_val_bytes(k.k), sizeof(struct bch_inode_generation));
 544                 return -BCH_ERR_invalid_bkey;
 545         }
 546
 547         return 0;
 548 }
 549
 550 void bch2_inode_generation_to_text(struct printbuf *out, struct bch_fs *c,
 551                                    struct bkey_s_c k)
 552 {
 553         struct bkey_s_c_inode_generation gen = bkey_s_c_to_inode_generation(k);
 554
 555         prt_printf(out, "generation: %u", le32_to_cpu(gen.v->bi_generation));
 556 }
 557
 558 void bch2_inode_init_early(struct bch_fs *c,
 559                            struct bch_inode_unpacked *inode_u)
 560 {
 561         enum bch_str_hash_type str_hash =
 562                 bch2_str_hash_opt_to_type(c, c->opts.str_hash);
 563
 564         memset(inode_u, 0, sizeof(*inode_u));
 565
 566         /* ick */
 567         inode_u->bi_flags |= str_hash << INODE_STR_HASH_OFFSET;
 568         get_random_bytes(&inode_u->bi_hash_seed,
 569                          sizeof(inode_u->bi_hash_seed));
 570 }
 571
 572 void bch2_inode_init_late(struct bch_inode_unpacked *inode_u, u64 now,
 573                           uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 574                           struct bch_inode_unpacked *parent)
 575 {
 576         inode_u->bi_mode        = mode;
 577         inode_u->bi_uid         = uid;
 578         inode_u->bi_gid         = gid;
 579         inode_u->bi_dev         = rdev;
 580         inode_u->bi_atime       = now;
 581         inode_u->bi_mtime       = now;
 582         inode_u->bi_ctime       = now;
 583         inode_u->bi_otime       = now;
 584
 585         if (parent && parent->bi_mode & S_ISGID) {
 586                 inode_u->bi_gid = parent->bi_gid;
 587                 if (S_ISDIR(mode))
 588                         inode_u->bi_mode |= S_ISGID;
 589         }
 590
 591         if (parent) {
 592 #define x(_name, ...)   inode_u->bi_##_name = parent->bi_##_name;
 593                 BCH_INODE_OPTS()
 594 #undef x
 595         }
 596 }
 597
 598 void bch2_inode_init(struct bch_fs *c, struct bch_inode_unpacked *inode_u,
 599                      uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 600                      struct bch_inode_unpacked *parent)
 601 {
 602         bch2_inode_init_early(c, inode_u);
 603         bch2_inode_init_late(inode_u, bch2_current_time(c),
 604                              uid, gid, mode, rdev, parent);
 605 }
 606
 607 static inline u32 bkey_generation(struct bkey_s_c k)
 608 {
 609         switch (k.k->type) {
 610         case KEY_TYPE_inode:
 611         case KEY_TYPE_inode_v2:
 612                 BUG();
 613         case KEY_TYPE_inode_generation:
 614                 return le32_to_cpu(bkey_s_c_to_inode_generation(k).v->bi_generation);
 615         default:
 616                 return 0;
 617         }
 618 }
 619
 620 /*
 621  * This just finds an empty slot:
 622  */
 623 int bch2_inode_create(struct btree_trans *trans,
 624                       struct btree_iter *iter,
 625                       struct bch_inode_unpacked *inode_u,
 626                       u32 snapshot, u64 cpu)
 627 {
 628         struct bch_fs *c = trans->c;
 629         struct bkey_s_c k;
 630         u64 min, max, start, pos, *hint;
 631         int ret = 0;
 632         unsigned bits = (c->opts.inodes_32bit ? 31 : 63);
 633
 634         if (c->opts.shard_inode_numbers) {
 635                 bits -= c->inode_shard_bits;
 636
 637                 min = (cpu << bits);
 638                 max = (cpu << bits) | ~(ULLONG_MAX << bits);
 639
 640                 min = max_t(u64, min, BLOCKDEV_INODE_MAX);
 641                 hint = c->unused_inode_hints + cpu;
 642         } else {
 643                 min = BLOCKDEV_INODE_MAX;
 644                 max = ~(ULLONG_MAX << bits);
 645                 hint = c->unused_inode_hints;
 646         }
 647
 648         start = READ_ONCE(*hint);
 649
 650         if (start >= max || start < min)
 651                 start = min;
 652
 653         pos = start;
 654         bch2_trans_iter_init(trans, iter, BTREE_ID_inodes, POS(0, pos),
 655                              BTREE_ITER_ALL_SNAPSHOTS|
 656                              BTREE_ITER_INTENT);
 657 again:
 658         while ((k = bch2_btree_iter_peek(iter)).k &&
 659                !(ret = bkey_err(k)) &&
 660                bkey_cmp(k.k->p, POS(0, max)) < 0) {
 661                 while (pos < iter->pos.offset) {
 662                         if (!bch2_btree_key_cache_find(c, BTREE_ID_inodes, POS(0, pos)))
 663                                 goto found_slot;
 664
 665                         pos++;
 666                 }
 667
 668                 if (k.k->p.snapshot == snapshot &&
 669                     !bkey_is_inode(k.k) &&
 670                     !bch2_btree_key_cache_find(c, BTREE_ID_inodes, SPOS(0, pos, snapshot))) {
 671                         bch2_btree_iter_advance(iter);
 672                         continue;
 673                 }
 674
 675                 /*
 676                  * We don't need to iterate over keys in every snapshot once
 677                  * we've found just one:
 678                  */
 679                 pos = iter->pos.offset + 1;
 680                 bch2_btree_iter_set_pos(iter, POS(0, pos));
 681         }
 682
 683         while (!ret && pos < max) {
 684                 if (!bch2_btree_key_cache_find(c, BTREE_ID_inodes, POS(0, pos)))
 685                         goto found_slot;
 686
 687                 pos++;
 688         }
 689
 690         if (!ret && start == min)
 691                 ret = -BCH_ERR_ENOSPC_inode_create;
 692
 693         if (ret) {
 694                 bch2_trans_iter_exit(trans, iter);
 695                 return ret;
 696         }
 697
 698         /* Retry from start */
 699         pos = start = min;
 700         bch2_btree_iter_set_pos(iter, POS(0, pos));
 701         goto again;
 702 found_slot:
 703         bch2_btree_iter_set_pos(iter, SPOS(0, pos, snapshot));
 704         k = bch2_btree_iter_peek_slot(iter);
 705         ret = bkey_err(k);
 706         if (ret) {
 707                 bch2_trans_iter_exit(trans, iter);
 708                 return ret;
 709         }
 710
 711         /* We may have raced while the iterator wasn't pointing at pos: */
 712         if (bkey_is_inode(k.k) ||
 713             bch2_btree_key_cache_find(c, BTREE_ID_inodes, k.k->p))
 714                 goto again;
 715
 716         *hint                   = k.k->p.offset;
 717         inode_u->bi_inum        = k.k->p.offset;
 718         inode_u->bi_generation  = bkey_generation(k);
 719         return 0;
 720 }
 721
 722 static int bch2_inode_delete_keys(struct btree_trans *trans,
 723                                   subvol_inum inum, enum btree_id id)
 724 {
 725         struct btree_iter iter;
 726         struct bkey_s_c k;
 727         struct bkey_i delete;
 728         u32 snapshot;
 729         int ret = 0;
 730
 731         /*
 732          * We're never going to be deleting extents, no need to use an extent
 733          * iterator:
 734          */
 735         bch2_trans_iter_init(trans, &iter, id, POS(inum.inum, 0),
 736                              BTREE_ITER_INTENT);
 737
 738         while (1) {
 739                 bch2_trans_begin(trans);
 740
 741                 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 742                 if (ret)
 743                         goto err;
 744
 745                 bch2_btree_iter_set_snapshot(&iter, snapshot);
 746
 747                 k = bch2_btree_iter_peek_upto(&iter, POS(inum.inum, U64_MAX));
 748                 ret = bkey_err(k);
 749                 if (ret)
 750                         goto err;
 751
 752                 if (!k.k)
 753                         break;
 754
 755                 bkey_init(&delete.k);
 756                 delete.k.p = iter.pos;
 757
 758                 if (iter.flags & BTREE_ITER_IS_EXTENTS) {
 759                         bch2_key_resize(&delete.k, k.k->p.offset - iter.pos.offset);
 760
 761                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
 762                         if (ret)
 763                                 goto err;
 764                 }
 765
 766                 ret = bch2_trans_update(trans, &iter, &delete, 0) ?:
 767                       bch2_trans_commit(trans, NULL, NULL,
 768                                         BTREE_INSERT_NOFAIL);
 769 err:
 770                 if (ret && !bch2_err_matches(ret, BCH_ERR_transaction_restart))
 771                         break;
 772         }
 773
 774         bch2_trans_iter_exit(trans, &iter);
 775         return ret;
 776 }
 777
 778 int bch2_inode_rm(struct bch_fs *c, subvol_inum inum)
 779 {
 780         struct btree_trans trans;
 781         struct btree_iter iter = { NULL };
 782         struct bkey_i_inode_generation delete;
 783         struct bch_inode_unpacked inode_u;
 784         struct bkey_s_c k;
 785         u32 snapshot;
 786         int ret;
 787
 788         bch2_trans_init(&trans, c, 0, 1024);
 789
 790         /*
 791          * If this was a directory, there shouldn't be any real dirents left -
 792          * but there could be whiteouts (from hash collisions) that we should
 793          * delete:
 794          *
 795          * XXX: the dirent could ideally would delete whiteouts when they're no
 796          * longer needed
 797          */
 798         ret   = bch2_inode_delete_keys(&trans, inum, BTREE_ID_extents) ?:
 799                 bch2_inode_delete_keys(&trans, inum, BTREE_ID_xattrs) ?:
 800                 bch2_inode_delete_keys(&trans, inum, BTREE_ID_dirents);
 801         if (ret)
 802                 goto err;
 803 retry:
 804         bch2_trans_begin(&trans);
 805
 806         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
 807         if (ret)
 808                 goto err;
 809
 810         bch2_trans_iter_init(&trans, &iter, BTREE_ID_inodes,
 811                              SPOS(0, inum.inum, snapshot),
 812                              BTREE_ITER_INTENT|BTREE_ITER_CACHED);
 813         k = bch2_btree_iter_peek_slot(&iter);
 814
 815         ret = bkey_err(k);
 816         if (ret)
 817                 goto err;
 818
 819         if (!bkey_is_inode(k.k)) {
 820                 bch2_fs_inconsistent(trans.c,
 821                                      "inode %llu not found when deleting",
 822                                      inum.inum);
 823                 ret = -EIO;
 824                 goto err;
 825         }
 826
 827         bch2_inode_unpack(k, &inode_u);
 828
 829         /* Subvolume root? */
 830         BUG_ON(inode_u.bi_subvol);
 831
 832         bkey_inode_generation_init(&delete.k_i);
 833         delete.k.p = iter.pos;
 834         delete.v.bi_generation = cpu_to_le32(inode_u.bi_generation + 1);
 835
 836         ret   = bch2_trans_update(&trans, &iter, &delete.k_i, 0) ?:
 837                 bch2_trans_commit(&trans, NULL, NULL,
 838                                 BTREE_INSERT_NOFAIL);
 839 err:
 840         bch2_trans_iter_exit(&trans, &iter);
 841         if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 842                 goto retry;
 843
 844         bch2_trans_exit(&trans);
 845         return ret;
 846 }
 847
 848 int bch2_inode_find_by_inum_trans(struct btree_trans *trans,
 849                                   subvol_inum inum,
 850                                   struct bch_inode_unpacked *inode)
 851 {
 852         struct btree_iter iter;
 853         int ret;
 854
 855         ret = bch2_inode_peek(trans, &iter, inode, inum, 0);
 856         if (!ret)
 857                 bch2_trans_iter_exit(trans, &iter);
 858         return ret;
 859 }
 860
 861 int bch2_inode_find_by_inum(struct bch_fs *c, subvol_inum inum,
 862                             struct bch_inode_unpacked *inode)
 863 {
 864         return bch2_trans_do(c, NULL, NULL, 0,
 865                 bch2_inode_find_by_inum_trans(&trans, inum, inode));
 866 }
 867
 868 int bch2_inode_nlink_inc(struct bch_inode_unpacked *bi)
 869 {
 870         if (bi->bi_flags & BCH_INODE_UNLINKED)
 871                 bi->bi_flags &= ~BCH_INODE_UNLINKED;
 872         else {
 873                 if (bi->bi_nlink == U32_MAX)
 874                         return -EINVAL;
 875
 876                 bi->bi_nlink++;
 877         }
 878
 879         return 0;
 880 }
 881
 882 void bch2_inode_nlink_dec(struct btree_trans *trans, struct bch_inode_unpacked *bi)
 883 {
 884         if (bi->bi_nlink && (bi->bi_flags & BCH_INODE_UNLINKED)) {
 885                 bch2_trans_inconsistent(trans, "inode %llu unlinked but link count nonzero",
 886                                         bi->bi_inum);
 887                 return;
 888         }
 889
 890         if (bi->bi_flags & BCH_INODE_UNLINKED) {
 891                 bch2_trans_inconsistent(trans, "inode %llu link count underflow", bi->bi_inum);
 892                 return;
 893         }
 894
 895         if (bi->bi_nlink)
 896                 bi->bi_nlink--;
 897         else
 898                 bi->bi_flags |= BCH_INODE_UNLINKED;
 899 }