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 "bkey_methods.h"
   5 #include "btree_update.h"
   6 #include "error.h"
   7 #include "extents.h"
   8 #include "inode.h"
   9 #include "str_hash.h"
  10
  11 #include <linux/random.h>
  12
  13 #include <asm/unaligned.h>
  14
  15 const char * const bch2_inode_opts[] = {
  16 #define x(name, ...)    #name,
  17         BCH_INODE_OPTS()
  18 #undef  x
  19         NULL,
  20 };
  21
  22 static const u8 byte_table[8] = { 1, 2, 3, 4, 6, 8, 10, 13 };
  23 static const u8 bits_table[8] = {
  24         1  * 8 - 1,
  25         2  * 8 - 2,
  26         3  * 8 - 3,
  27         4  * 8 - 4,
  28         6  * 8 - 5,
  29         8  * 8 - 6,
  30         10 * 8 - 7,
  31         13 * 8 - 8,
  32 };
  33
  34 static int inode_encode_field(u8 *out, u8 *end, u64 hi, u64 lo)
  35 {
  36         __be64 in[2] = { cpu_to_be64(hi), cpu_to_be64(lo), };
  37         unsigned shift, bytes, bits = likely(!hi)
  38                 ? fls64(lo)
  39                 : fls64(hi) + 64;
  40
  41         for (shift = 1; shift <= 8; shift++)
  42                 if (bits < bits_table[shift - 1])
  43                         goto got_shift;
  44
  45         BUG();
  46 got_shift:
  47         bytes = byte_table[shift - 1];
  48
  49         BUG_ON(out + bytes > end);
  50
  51         memcpy(out, (u8 *) in + 16 - bytes, bytes);
  52         *out |= (1 << 8) >> shift;
  53
  54         return bytes;
  55 }
  56
  57 static int inode_decode_field(const u8 *in, const u8 *end,
  58                               u64 out[2], unsigned *out_bits)
  59 {
  60         __be64 be[2] = { 0, 0 };
  61         unsigned bytes, shift;
  62         u8 *p;
  63
  64         if (in >= end)
  65                 return -1;
  66
  67         if (!*in)
  68                 return -1;
  69
  70         /*
  71          * position of highest set bit indicates number of bytes:
  72          * shift = number of bits to remove in high byte:
  73          */
  74         shift   = 8 - __fls(*in); /* 1 <= shift <= 8 */
  75         bytes   = byte_table[shift - 1];
  76
  77         if (in + bytes > end)
  78                 return -1;
  79
  80         p = (u8 *) be + 16 - bytes;
  81         memcpy(p, in, bytes);
  82         *p ^= (1 << 8) >> shift;
  83
  84         out[0] = be64_to_cpu(be[0]);
  85         out[1] = be64_to_cpu(be[1]);
  86         *out_bits = out[0] ? 64 + fls64(out[0]) : fls64(out[1]);
  87
  88         return bytes;
  89 }
  90
  91 void bch2_inode_pack(struct bkey_inode_buf *packed,
  92                      const struct bch_inode_unpacked *inode)
  93 {
  94         u8 *out = packed->inode.v.fields;
  95         u8 *end = (void *) &packed[1];
  96         u8 *last_nonzero_field = out;
  97         unsigned nr_fields = 0, last_nonzero_fieldnr = 0;
  98         unsigned bytes;
  99
 100         bkey_inode_init(&packed->inode.k_i);
 101         packed->inode.k.p.inode         = inode->bi_inum;
 102         packed->inode.v.bi_hash_seed    = inode->bi_hash_seed;
 103         packed->inode.v.bi_flags        = cpu_to_le32(inode->bi_flags);
 104         packed->inode.v.bi_mode         = cpu_to_le16(inode->bi_mode);
 105
 106 #define x(_name, _bits)                                 \
 107         out += inode_encode_field(out, end, 0, inode->_name);           \
 108         nr_fields++;                                                    \
 109                                                                         \
 110         if (inode->_name) {                                             \
 111                 last_nonzero_field = out;                               \
 112                 last_nonzero_fieldnr = nr_fields;                       \
 113         }
 114
 115         BCH_INODE_FIELDS()
 116 #undef  x
 117
 118         out = last_nonzero_field;
 119         nr_fields = last_nonzero_fieldnr;
 120
 121         bytes = out - (u8 *) &packed->inode.v;
 122         set_bkey_val_bytes(&packed->inode.k, bytes);
 123         memset_u64s_tail(&packed->inode.v, 0, bytes);
 124
 125         SET_INODE_NR_FIELDS(&packed->inode.v, nr_fields);
 126
 127         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
 128                 struct bch_inode_unpacked unpacked;
 129
 130                 int ret = bch2_inode_unpack(inode_i_to_s_c(&packed->inode),
 131                                            &unpacked);
 132                 BUG_ON(ret);
 133                 BUG_ON(unpacked.bi_inum         != inode->bi_inum);
 134                 BUG_ON(unpacked.bi_hash_seed    != inode->bi_hash_seed);
 135                 BUG_ON(unpacked.bi_mode         != inode->bi_mode);
 136
 137 #define x(_name, _bits) BUG_ON(unpacked._name != inode->_name);
 138                 BCH_INODE_FIELDS()
 139 #undef  x
 140         }
 141 }
 142
 143 int bch2_inode_unpack(struct bkey_s_c_inode inode,
 144                       struct bch_inode_unpacked *unpacked)
 145 {
 146         const u8 *in = inode.v->fields;
 147         const u8 *end = (void *) inode.v + bkey_val_bytes(inode.k);
 148         u64 field[2];
 149         unsigned fieldnr = 0, field_bits;
 150         int ret;
 151
 152         unpacked->bi_inum       = inode.k->p.inode;
 153         unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 154         unpacked->bi_flags      = le32_to_cpu(inode.v->bi_flags);
 155         unpacked->bi_mode       = le16_to_cpu(inode.v->bi_mode);
 156
 157 #define x(_name, _bits)                                 \
 158         if (fieldnr++ == INODE_NR_FIELDS(inode.v)) {                    \
 159                 memset(&unpacked->_name, 0,                             \
 160                        sizeof(*unpacked) -                              \
 161                        offsetof(struct bch_inode_unpacked, _name));     \
 162                 return 0;                                               \
 163         }                                                               \
 164                                                                         \
 165         ret = inode_decode_field(in, end, field, &field_bits);          \
 166         if (ret < 0)                                                    \
 167                 return ret;                                             \
 168                                                                         \
 169         if (field_bits > sizeof(unpacked->_name) * 8)                   \
 170                 return -1;                                              \
 171                                                                         \
 172         unpacked->_name = field[1];                                     \
 173         in += ret;
 174
 175         BCH_INODE_FIELDS()
 176 #undef  x
 177
 178         /* XXX: signal if there were more fields than expected? */
 179
 180         return 0;
 181 }
 182
 183 struct btree_iter *bch2_inode_peek(struct btree_trans *trans,
 184                                    struct bch_inode_unpacked *inode,
 185                                    u64 inum, unsigned flags)
 186 {
 187         struct btree_iter *iter;
 188         struct bkey_s_c k;
 189         int ret;
 190
 191         iter = bch2_trans_get_iter(trans, BTREE_ID_INODES, POS(inum, 0),
 192                                    BTREE_ITER_SLOTS|flags);
 193         if (IS_ERR(iter))
 194                 return iter;
 195
 196         k = bch2_btree_iter_peek_slot(iter);
 197         ret = bkey_err(k);
 198         if (ret)
 199                 goto err;
 200
 201         ret = k.k->type == KEY_TYPE_inode ? 0 : -EIO;
 202         if (ret)
 203                 goto err;
 204
 205         ret = bch2_inode_unpack(bkey_s_c_to_inode(k), inode);
 206         if (ret)
 207                 goto err;
 208
 209         return iter;
 210 err:
 211         bch2_trans_iter_put(trans, iter);
 212         return ERR_PTR(ret);
 213 }
 214
 215 int bch2_inode_write(struct btree_trans *trans,
 216                      struct btree_iter *iter,
 217                      struct bch_inode_unpacked *inode)
 218 {
 219         struct bkey_inode_buf *inode_p;
 220
 221         inode_p = bch2_trans_kmalloc(trans, sizeof(*inode_p));
 222         if (IS_ERR(inode_p))
 223                 return PTR_ERR(inode_p);
 224
 225         bch2_inode_pack(inode_p, inode);
 226         bch2_trans_update(trans, iter, &inode_p->inode.k_i, 0);
 227         return 0;
 228 }
 229
 230 const char *bch2_inode_invalid(const struct bch_fs *c, struct bkey_s_c k)
 231 {
 232                 struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 233                 struct bch_inode_unpacked unpacked;
 234
 235         if (k.k->p.offset)
 236                 return "nonzero offset";
 237
 238         if (bkey_val_bytes(k.k) < sizeof(struct bch_inode))
 239                 return "incorrect value size";
 240
 241         if (k.k->p.inode < BLOCKDEV_INODE_MAX)
 242                 return "fs inode in blockdev range";
 243
 244         if (INODE_STR_HASH(inode.v) >= BCH_STR_HASH_NR)
 245                 return "invalid str hash type";
 246
 247         if (bch2_inode_unpack(inode, &unpacked))
 248                 return "invalid variable length fields";
 249
 250         if (unpacked.bi_data_checksum >= BCH_CSUM_OPT_NR + 1)
 251                 return "invalid data checksum type";
 252
 253         if (unpacked.bi_compression >= BCH_COMPRESSION_OPT_NR + 1)
 254                 return "invalid data checksum type";
 255
 256         if ((unpacked.bi_flags & BCH_INODE_UNLINKED) &&
 257             unpacked.bi_nlink != 0)
 258                 return "flagged as unlinked but bi_nlink != 0";
 259
 260         return NULL;
 261 }
 262
 263 void bch2_inode_to_text(struct printbuf *out, struct bch_fs *c,
 264                        struct bkey_s_c k)
 265 {
 266         struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 267         struct bch_inode_unpacked unpacked;
 268
 269         if (bch2_inode_unpack(inode, &unpacked)) {
 270                 pr_buf(out, "(unpack error)");
 271                 return;
 272         }
 273
 274 #define x(_name, _bits)                                         \
 275         pr_buf(out, #_name ": %llu ", (u64) unpacked._name);
 276         BCH_INODE_FIELDS()
 277 #undef  x
 278 }
 279
 280 const char *bch2_inode_generation_invalid(const struct bch_fs *c,
 281                                           struct bkey_s_c k)
 282 {
 283         if (k.k->p.offset)
 284                 return "nonzero offset";
 285
 286         if (bkey_val_bytes(k.k) != sizeof(struct bch_inode_generation))
 287                 return "incorrect value size";
 288
 289         return NULL;
 290 }
 291
 292 void bch2_inode_generation_to_text(struct printbuf *out, struct bch_fs *c,
 293                                    struct bkey_s_c k)
 294 {
 295         struct bkey_s_c_inode_generation gen = bkey_s_c_to_inode_generation(k);
 296
 297         pr_buf(out, "generation: %u", le32_to_cpu(gen.v->bi_generation));
 298 }
 299
 300 void bch2_inode_init_early(struct bch_fs *c,
 301                            struct bch_inode_unpacked *inode_u)
 302 {
 303         enum bch_str_hash_type str_hash =
 304                 bch2_str_hash_opt_to_type(c, c->opts.str_hash);
 305
 306         memset(inode_u, 0, sizeof(*inode_u));
 307
 308         /* ick */
 309         inode_u->bi_flags |= str_hash << INODE_STR_HASH_OFFSET;
 310         get_random_bytes(&inode_u->bi_hash_seed,
 311                          sizeof(inode_u->bi_hash_seed));
 312 }
 313
 314 void bch2_inode_init_late(struct bch_inode_unpacked *inode_u, u64 now,
 315                           uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 316                           struct bch_inode_unpacked *parent)
 317 {
 318         inode_u->bi_mode        = mode;
 319         inode_u->bi_uid         = uid;
 320         inode_u->bi_gid         = gid;
 321         inode_u->bi_dev         = rdev;
 322         inode_u->bi_atime       = now;
 323         inode_u->bi_mtime       = now;
 324         inode_u->bi_ctime       = now;
 325         inode_u->bi_otime       = now;
 326
 327         if (parent && parent->bi_mode & S_ISGID) {
 328                 inode_u->bi_gid = parent->bi_gid;
 329                 if (S_ISDIR(mode))
 330                         inode_u->bi_mode |= S_ISGID;
 331         }
 332
 333         if (parent) {
 334 #define x(_name, ...)   inode_u->bi_##_name = parent->bi_##_name;
 335                 BCH_INODE_OPTS()
 336 #undef x
 337         }
 338 }
 339
 340 void bch2_inode_init(struct bch_fs *c, struct bch_inode_unpacked *inode_u,
 341                      uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 342                      struct bch_inode_unpacked *parent)
 343 {
 344         bch2_inode_init_early(c, inode_u);
 345         bch2_inode_init_late(inode_u, bch2_current_time(c),
 346                              uid, gid, mode, rdev, parent);
 347 }
 348
 349 static inline u32 bkey_generation(struct bkey_s_c k)
 350 {
 351         switch (k.k->type) {
 352         case KEY_TYPE_inode:
 353                 BUG();
 354         case KEY_TYPE_inode_generation:
 355                 return le32_to_cpu(bkey_s_c_to_inode_generation(k).v->bi_generation);
 356         default:
 357                 return 0;
 358         }
 359 }
 360
 361 int bch2_inode_create(struct btree_trans *trans,
 362                       struct bch_inode_unpacked *inode_u,
 363                       u64 min, u64 max, u64 *hint)
 364 {
 365         struct bkey_inode_buf *inode_p;
 366         struct btree_iter *iter = NULL;
 367         struct bkey_s_c k;
 368         u64 start;
 369         int ret;
 370
 371         if (!max)
 372                 max = ULLONG_MAX;
 373
 374         if (trans->c->opts.inodes_32bit)
 375                 max = min_t(u64, max, U32_MAX);
 376
 377         start = READ_ONCE(*hint);
 378
 379         if (start >= max || start < min)
 380                 start = min;
 381
 382         inode_p = bch2_trans_kmalloc(trans, sizeof(*inode_p));
 383         if (IS_ERR(inode_p))
 384                 return PTR_ERR(inode_p);
 385 again:
 386         for_each_btree_key(trans, iter, BTREE_ID_INODES, POS(start, 0),
 387                            BTREE_ITER_SLOTS|BTREE_ITER_INTENT, k, ret) {
 388                 if (iter->pos.inode > max)
 389                         break;
 390
 391                 if (k.k->type != KEY_TYPE_inode)
 392                         goto found_slot;
 393         }
 394
 395         bch2_trans_iter_put(trans, iter);
 396
 397         if (ret)
 398                 return ret;
 399
 400         if (start != min) {
 401                 /* Retry from start */
 402                 start = min;
 403                 goto again;
 404         }
 405
 406         return -ENOSPC;
 407 found_slot:
 408         *hint                   = k.k->p.inode;
 409         inode_u->bi_inum        = k.k->p.inode;
 410         inode_u->bi_generation  = bkey_generation(k);
 411
 412         bch2_inode_pack(inode_p, inode_u);
 413         bch2_trans_update(trans, iter, &inode_p->inode.k_i, 0);
 414         bch2_trans_iter_put(trans, iter);
 415         return 0;
 416 }
 417
 418 int bch2_inode_rm(struct bch_fs *c, u64 inode_nr)
 419 {
 420         struct btree_trans trans;
 421         struct btree_iter *iter;
 422         struct bkey_i_inode_generation delete;
 423         struct bpos start = POS(inode_nr, 0);
 424         struct bpos end = POS(inode_nr + 1, 0);
 425         int ret;
 426
 427         /*
 428          * If this was a directory, there shouldn't be any real dirents left -
 429          * but there could be whiteouts (from hash collisions) that we should
 430          * delete:
 431          *
 432          * XXX: the dirent could ideally would delete whiteouts when they're no
 433          * longer needed
 434          */
 435         ret   = bch2_btree_delete_range(c, BTREE_ID_EXTENTS,
 436                                         start, end, NULL) ?:
 437                 bch2_btree_delete_range(c, BTREE_ID_XATTRS,
 438                                         start, end, NULL) ?:
 439                 bch2_btree_delete_range(c, BTREE_ID_DIRENTS,
 440                                         start, end, NULL);
 441         if (ret)
 442                 return ret;
 443
 444         bch2_trans_init(&trans, c, 0, 0);
 445
 446         iter = bch2_trans_get_iter(&trans, BTREE_ID_INODES, POS(inode_nr, 0),
 447                                    BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 448         do {
 449                 struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
 450                 u32 bi_generation = 0;
 451
 452                 ret = bkey_err(k);
 453                 if (ret)
 454                         break;
 455
 456                 bch2_fs_inconsistent_on(k.k->type != KEY_TYPE_inode, c,
 457                                         "inode %llu not found when deleting",
 458                                         inode_nr);
 459
 460                 switch (k.k->type) {
 461                 case KEY_TYPE_inode: {
 462                         struct bch_inode_unpacked inode_u;
 463
 464                         if (!bch2_inode_unpack(bkey_s_c_to_inode(k), &inode_u))
 465                                 bi_generation = inode_u.bi_generation + 1;
 466                         break;
 467                 }
 468                 case KEY_TYPE_inode_generation: {
 469                         struct bkey_s_c_inode_generation g =
 470                                 bkey_s_c_to_inode_generation(k);
 471                         bi_generation = le32_to_cpu(g.v->bi_generation);
 472                         break;
 473                 }
 474                 }
 475
 476                 if (!bi_generation) {
 477                         bkey_init(&delete.k);
 478                         delete.k.p.inode = inode_nr;
 479                 } else {
 480                         bkey_inode_generation_init(&delete.k_i);
 481                         delete.k.p.inode = inode_nr;
 482                         delete.v.bi_generation = cpu_to_le32(bi_generation);
 483                 }
 484
 485                 bch2_trans_update(&trans, iter, &delete.k_i, 0);
 486
 487                 ret = bch2_trans_commit(&trans, NULL, NULL,
 488                                         BTREE_INSERT_NOFAIL);
 489         } while (ret == -EINTR);
 490
 491         bch2_trans_exit(&trans);
 492         return ret;
 493 }
 494
 495 int bch2_inode_find_by_inum_trans(struct btree_trans *trans, u64 inode_nr,
 496                                   struct bch_inode_unpacked *inode)
 497 {
 498         struct btree_iter *iter;
 499         struct bkey_s_c k;
 500         int ret;
 501
 502         iter = bch2_trans_get_iter(trans, BTREE_ID_INODES,
 503                         POS(inode_nr, 0), BTREE_ITER_SLOTS);
 504         if (IS_ERR(iter))
 505                 return PTR_ERR(iter);
 506
 507         k = bch2_btree_iter_peek_slot(iter);
 508         ret = bkey_err(k);
 509         if (ret)
 510                 goto err;
 511
 512         ret = k.k->type == KEY_TYPE_inode
 513                 ? bch2_inode_unpack(bkey_s_c_to_inode(k), inode)
 514                 : -ENOENT;
 515 err:
 516         bch2_trans_iter_put(trans, iter);
 517         return ret;
 518 }
 519
 520 int bch2_inode_find_by_inum(struct bch_fs *c, u64 inode_nr,
 521                             struct bch_inode_unpacked *inode)
 522 {
 523         return bch2_trans_do(c, NULL, NULL, 0,
 524                 bch2_inode_find_by_inum_trans(&trans, inode_nr, inode));
 525 }
 526
 527 #ifdef CONFIG_BCACHEFS_DEBUG
 528 void bch2_inode_pack_test(void)
 529 {
 530         struct bch_inode_unpacked *u, test_inodes[] = {
 531                 {
 532                         .bi_atime       = U64_MAX,
 533                         .bi_ctime       = U64_MAX,
 534                         .bi_mtime       = U64_MAX,
 535                         .bi_otime       = U64_MAX,
 536                         .bi_size        = U64_MAX,
 537                         .bi_sectors     = U64_MAX,
 538                         .bi_uid         = U32_MAX,
 539                         .bi_gid         = U32_MAX,
 540                         .bi_nlink       = U32_MAX,
 541                         .bi_generation  = U32_MAX,
 542                         .bi_dev         = U32_MAX,
 543                 },
 544         };
 545
 546         for (u = test_inodes;
 547              u < test_inodes + ARRAY_SIZE(test_inodes);
 548              u++) {
 549                 struct bkey_inode_buf p;
 550
 551                 bch2_inode_pack(&p, u);
 552         }
 553 }
 554 #endif