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

   1
   2 #include "bcachefs.h"
   3 #include "bkey_methods.h"
   4 #include "btree_update.h"
   5 #include "error.h"
   6 #include "extents.h"
   7 #include "inode.h"
   8 #include "io.h"
   9 #include "keylist.h"
  10
  11 #include <linux/random.h>
  12
  13 #include <asm/unaligned.h>
  14
  15 #define FIELD_BYTES()                                           \
  16
  17 static const u8 byte_table[8] = { 1, 2, 3, 4, 6, 8, 10, 13 };
  18 static const u8 bits_table[8] = {
  19         1  * 8 - 1,
  20         2  * 8 - 2,
  21         3  * 8 - 3,
  22         4  * 8 - 4,
  23         6  * 8 - 5,
  24         8  * 8 - 6,
  25         10 * 8 - 7,
  26         13 * 8 - 8,
  27 };
  28
  29 static int inode_encode_field(u8 *out, u8 *end, u64 hi, u64 lo)
  30 {
  31         __be64 in[2] = { cpu_to_be64(hi), cpu_to_be64(lo), };
  32         unsigned shift, bytes, bits = likely(!hi)
  33                 ? fls64(lo)
  34                 : fls64(hi) + 64;
  35
  36         for (shift = 1; shift <= 8; shift++)
  37                 if (bits < bits_table[shift - 1])
  38                         goto got_shift;
  39
  40         BUG();
  41 got_shift:
  42         bytes = byte_table[shift - 1];
  43
  44         BUG_ON(out + bytes > end);
  45
  46         memcpy(out, (u8 *) in + 16 - bytes, bytes);
  47         *out |= (1 << 8) >> shift;
  48
  49         return bytes;
  50 }
  51
  52 static int inode_decode_field(const u8 *in, const u8 *end,
  53                               u64 out[2], unsigned *out_bits)
  54 {
  55         __be64 be[2] = { 0, 0 };
  56         unsigned bytes, shift;
  57         u8 *p;
  58
  59         if (in >= end)
  60                 return -1;
  61
  62         if (!*in)
  63                 return -1;
  64
  65         /*
  66          * position of highest set bit indicates number of bytes:
  67          * shift = number of bits to remove in high byte:
  68          */
  69         shift   = 8 - __fls(*in); /* 1 <= shift <= 8 */
  70         bytes   = byte_table[shift - 1];
  71
  72         if (in + bytes > end)
  73                 return -1;
  74
  75         p = (u8 *) be + 16 - bytes;
  76         memcpy(p, in, bytes);
  77         *p ^= (1 << 8) >> shift;
  78
  79         out[0] = be64_to_cpu(be[0]);
  80         out[1] = be64_to_cpu(be[1]);
  81         *out_bits = out[0] ? 64 + fls64(out[0]) : fls64(out[1]);
  82
  83         return bytes;
  84 }
  85
  86 void bch2_inode_pack(struct bkey_inode_buf *packed,
  87                      const struct bch_inode_unpacked *inode)
  88 {
  89         u8 *out = packed->inode.v.fields;
  90         u8 *end = (void *) &packed[1];
  91         u8 *last_nonzero_field = out;
  92         unsigned nr_fields = 0, last_nonzero_fieldnr = 0;
  93
  94         bkey_inode_init(&packed->inode.k_i);
  95         packed->inode.k.p.inode         = inode->bi_inum;
  96         packed->inode.v.bi_hash_seed    = inode->bi_hash_seed;
  97         packed->inode.v.bi_flags        = cpu_to_le32(inode->bi_flags);
  98         packed->inode.v.bi_mode         = cpu_to_le16(inode->bi_mode);
  99
 100 #define BCH_INODE_FIELD(_name, _bits)                                   \
 101         out += inode_encode_field(out, end, 0, inode->_name);           \
 102         nr_fields++;                                                    \
 103                                                                         \
 104         if (inode->_name) {                                             \
 105                 last_nonzero_field = out;                               \
 106                 last_nonzero_fieldnr = nr_fields;                       \
 107         }
 108
 109         BCH_INODE_FIELDS()
 110 #undef  BCH_INODE_FIELD
 111
 112         out = last_nonzero_field;
 113         nr_fields = last_nonzero_fieldnr;
 114
 115         set_bkey_val_bytes(&packed->inode.k, out - (u8 *) &packed->inode.v);
 116         memset(out, 0,
 117                (u8 *) &packed->inode.v +
 118                bkey_val_bytes(&packed->inode.k) - out);
 119
 120         SET_INODE_NR_FIELDS(&packed->inode.v, nr_fields);
 121
 122         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
 123                 struct bch_inode_unpacked unpacked;
 124
 125                 int ret = bch2_inode_unpack(inode_i_to_s_c(&packed->inode),
 126                                            &unpacked);
 127                 BUG_ON(ret);
 128                 BUG_ON(unpacked.bi_inum         != inode->bi_inum);
 129                 BUG_ON(unpacked.bi_hash_seed    != inode->bi_hash_seed);
 130                 BUG_ON(unpacked.bi_mode         != inode->bi_mode);
 131
 132 #define BCH_INODE_FIELD(_name, _bits)   BUG_ON(unpacked._name != inode->_name);
 133                 BCH_INODE_FIELDS()
 134 #undef  BCH_INODE_FIELD
 135         }
 136 }
 137
 138 int bch2_inode_unpack(struct bkey_s_c_inode inode,
 139                       struct bch_inode_unpacked *unpacked)
 140 {
 141         const u8 *in = inode.v->fields;
 142         const u8 *end = (void *) inode.v + bkey_val_bytes(inode.k);
 143         u64 field[2];
 144         unsigned fieldnr = 0, field_bits;
 145         int ret;
 146
 147         unpacked->bi_inum       = inode.k->p.inode;
 148         unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 149         unpacked->bi_flags      = le32_to_cpu(inode.v->bi_flags);
 150         unpacked->bi_mode       = le16_to_cpu(inode.v->bi_mode);
 151
 152 #define BCH_INODE_FIELD(_name, _bits)                                   \
 153         if (fieldnr++ == INODE_NR_FIELDS(inode.v)) {                    \
 154                 memset(&unpacked->_name, 0,                             \
 155                        sizeof(*unpacked) -                              \
 156                        offsetof(struct bch_inode_unpacked, _name));     \
 157                 return 0;                                               \
 158         }                                                               \
 159                                                                         \
 160         ret = inode_decode_field(in, end, field, &field_bits);          \
 161         if (ret < 0)                                                    \
 162                 return ret;                                             \
 163                                                                         \
 164         if (field_bits > sizeof(unpacked->_name) * 8)                   \
 165                 return -1;                                              \
 166                                                                         \
 167         unpacked->_name = field[1];                                     \
 168         in += ret;
 169
 170         BCH_INODE_FIELDS()
 171 #undef  BCH_INODE_FIELD
 172
 173         /* XXX: signal if there were more fields than expected? */
 174
 175         return 0;
 176 }
 177
 178 const char *bch2_inode_invalid(const struct bch_fs *c, struct bkey_s_c k)
 179 {
 180                 struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 181                 struct bch_inode_unpacked unpacked;
 182
 183         if (k.k->p.offset)
 184                 return "nonzero offset";
 185
 186         if (bkey_val_bytes(k.k) < sizeof(struct bch_inode))
 187                 return "incorrect value size";
 188
 189         if (k.k->p.inode < BLOCKDEV_INODE_MAX)
 190                 return "fs inode in blockdev range";
 191
 192         if (INODE_STR_HASH(inode.v) >= BCH_STR_HASH_NR)
 193                 return "invalid str hash type";
 194
 195         if (bch2_inode_unpack(inode, &unpacked))
 196                 return "invalid variable length fields";
 197
 198         if (unpacked.bi_data_checksum >= BCH_CSUM_OPT_NR + 1)
 199                 return "invalid data checksum type";
 200
 201         if (unpacked.bi_compression >= BCH_COMPRESSION_OPT_NR + 1)
 202                 return "invalid data checksum type";
 203
 204         if ((unpacked.bi_flags & BCH_INODE_UNLINKED) &&
 205             unpacked.bi_nlink != 0)
 206                 return "flagged as unlinked but bi_nlink != 0";
 207
 208         return NULL;
 209 }
 210
 211 void bch2_inode_to_text(struct printbuf *out, struct bch_fs *c,
 212                        struct bkey_s_c k)
 213 {
 214         struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 215         struct bch_inode_unpacked unpacked;
 216
 217         if (bch2_inode_unpack(inode, &unpacked)) {
 218                 pr_buf(out, "(unpack error)");
 219                 return;
 220         }
 221
 222 #define BCH_INODE_FIELD(_name, _bits)                                           \
 223         pr_buf(out, #_name ": %llu ", (u64) unpacked._name);
 224         BCH_INODE_FIELDS()
 225 #undef  BCH_INODE_FIELD
 226 }
 227
 228 const char *bch2_inode_generation_invalid(const struct bch_fs *c,
 229                                           struct bkey_s_c k)
 230 {
 231         if (k.k->p.offset)
 232                 return "nonzero offset";
 233
 234         if (bkey_val_bytes(k.k) != sizeof(struct bch_inode_generation))
 235                 return "incorrect value size";
 236
 237         return NULL;
 238 }
 239
 240 void bch2_inode_generation_to_text(struct printbuf *out, struct bch_fs *c,
 241                                    struct bkey_s_c k)
 242 {
 243 }
 244
 245 void bch2_inode_init(struct bch_fs *c, struct bch_inode_unpacked *inode_u,
 246                      uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 247                      struct bch_inode_unpacked *parent)
 248 {
 249         s64 now = timespec_to_bch2_time(c,
 250                 timespec64_trunc(current_kernel_time64(),
 251                                  c->sb.time_precision));
 252
 253         memset(inode_u, 0, sizeof(*inode_u));
 254
 255         /* ick */
 256         inode_u->bi_flags |= c->opts.str_hash << INODE_STR_HASH_OFFSET;
 257         get_random_bytes(&inode_u->bi_hash_seed, sizeof(inode_u->bi_hash_seed));
 258
 259         inode_u->bi_mode        = mode;
 260         inode_u->bi_uid         = uid;
 261         inode_u->bi_gid         = gid;
 262         inode_u->bi_dev         = rdev;
 263         inode_u->bi_atime       = now;
 264         inode_u->bi_mtime       = now;
 265         inode_u->bi_ctime       = now;
 266         inode_u->bi_otime       = now;
 267
 268         if (parent) {
 269 #define BCH_INODE_FIELD(_name)  inode_u->_name = parent->_name;
 270                 BCH_INODE_FIELDS_INHERIT()
 271 #undef BCH_INODE_FIELD
 272         }
 273 }
 274
 275 static inline u32 bkey_generation(struct bkey_s_c k)
 276 {
 277         switch (k.k->type) {
 278         case KEY_TYPE_inode:
 279                 BUG();
 280         case KEY_TYPE_inode_generation:
 281                 return le32_to_cpu(bkey_s_c_to_inode_generation(k).v->bi_generation);
 282         default:
 283                 return 0;
 284         }
 285 }
 286
 287 int __bch2_inode_create(struct btree_trans *trans,
 288                         struct bch_inode_unpacked *inode_u,
 289                         u64 min, u64 max, u64 *hint)
 290 {
 291         struct bch_fs *c = trans->c;
 292         struct bkey_inode_buf *inode_p;
 293         struct btree_iter *iter;
 294         u64 start;
 295         int ret;
 296
 297         if (!max)
 298                 max = ULLONG_MAX;
 299
 300         if (c->opts.inodes_32bit)
 301                 max = min_t(u64, max, U32_MAX);
 302
 303         start = READ_ONCE(*hint);
 304
 305         if (start >= max || start < min)
 306                 start = min;
 307
 308         inode_p = bch2_trans_kmalloc(trans, sizeof(*inode_p));
 309         if (IS_ERR(inode_p))
 310                 return PTR_ERR(inode_p);
 311
 312         iter = bch2_trans_get_iter(trans,
 313                         BTREE_ID_INODES, POS(start, 0),
 314                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 315         if (IS_ERR(iter))
 316                 return PTR_ERR(iter);
 317 again:
 318         while (1) {
 319                 struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
 320
 321                 ret = btree_iter_err(k);
 322                 if (ret)
 323                         return ret;
 324
 325                 switch (k.k->type) {
 326                 case KEY_TYPE_inode:
 327                         /* slot used */
 328                         if (iter->pos.inode >= max)
 329                                 goto out;
 330
 331                         bch2_btree_iter_next_slot(iter);
 332                         break;
 333
 334                 default:
 335                         *hint                   = k.k->p.inode;
 336                         inode_u->bi_inum        = k.k->p.inode;
 337                         inode_u->bi_generation  = bkey_generation(k);
 338
 339                         bch2_inode_pack(inode_p, inode_u);
 340                         bch2_trans_update(trans,
 341                                 BTREE_INSERT_ENTRY(iter, &inode_p->inode.k_i));
 342                         return 0;
 343                 }
 344         }
 345 out:
 346         if (start != min) {
 347                 /* Retry from start */
 348                 start = min;
 349                 bch2_btree_iter_set_pos(iter, POS(start, 0));
 350                 goto again;
 351         }
 352
 353         return -ENOSPC;
 354 }
 355
 356 int bch2_inode_create(struct bch_fs *c, struct bch_inode_unpacked *inode_u,
 357                       u64 min, u64 max, u64 *hint)
 358 {
 359         return bch2_trans_do(c, NULL, BTREE_INSERT_ATOMIC,
 360                         __bch2_inode_create(&trans, inode_u, min, max, hint));
 361 }
 362
 363 int bch2_inode_rm(struct bch_fs *c, u64 inode_nr)
 364 {
 365         struct btree_iter iter;
 366         struct bkey_i_inode_generation delete;
 367         struct bpos start = POS(inode_nr, 0);
 368         struct bpos end = POS(inode_nr + 1, 0);
 369         int ret;
 370
 371         /*
 372          * If this was a directory, there shouldn't be any real dirents left -
 373          * but there could be whiteouts (from hash collisions) that we should
 374          * delete:
 375          *
 376          * XXX: the dirent could ideally would delete whiteouts when they're no
 377          * longer needed
 378          */
 379         ret   = bch2_btree_delete_range(c, BTREE_ID_EXTENTS,
 380                                         start, end, NULL) ?:
 381                 bch2_btree_delete_range(c, BTREE_ID_XATTRS,
 382                                         start, end, NULL) ?:
 383                 bch2_btree_delete_range(c, BTREE_ID_DIRENTS,
 384                                         start, end, NULL);
 385         if (ret)
 386                 return ret;
 387
 388         bch2_btree_iter_init(&iter, c, BTREE_ID_INODES, POS(inode_nr, 0),
 389                              BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 390         do {
 391                 struct bkey_s_c k = bch2_btree_iter_peek_slot(&iter);
 392                 u32 bi_generation = 0;
 393
 394                 ret = btree_iter_err(k);
 395                 if (ret) {
 396                         bch2_btree_iter_unlock(&iter);
 397                         return ret;
 398                 }
 399
 400                 bch2_fs_inconsistent_on(k.k->type != KEY_TYPE_inode, c,
 401                                         "inode %llu not found when deleting",
 402                                         inode_nr);
 403
 404                 switch (k.k->type) {
 405                 case KEY_TYPE_inode: {
 406                         struct bch_inode_unpacked inode_u;
 407
 408                         if (!bch2_inode_unpack(bkey_s_c_to_inode(k), &inode_u))
 409                                 bi_generation = inode_u.bi_generation + 1;
 410                         break;
 411                 }
 412                 case KEY_TYPE_inode_generation: {
 413                         struct bkey_s_c_inode_generation g =
 414                                 bkey_s_c_to_inode_generation(k);
 415                         bi_generation = le32_to_cpu(g.v->bi_generation);
 416                         break;
 417                 }
 418                 }
 419
 420                 if (!bi_generation) {
 421                         bkey_init(&delete.k);
 422                         delete.k.p.inode = inode_nr;
 423                 } else {
 424                         bkey_inode_generation_init(&delete.k_i);
 425                         delete.k.p.inode = inode_nr;
 426                         delete.v.bi_generation = cpu_to_le32(bi_generation);
 427                 }
 428
 429                 ret = bch2_btree_insert_at(c, NULL, NULL,
 430                                 BTREE_INSERT_ATOMIC|
 431                                 BTREE_INSERT_NOFAIL,
 432                                 BTREE_INSERT_ENTRY(&iter, &delete.k_i));
 433         } while (ret == -EINTR);
 434
 435         bch2_btree_iter_unlock(&iter);
 436         return ret;
 437 }
 438
 439 int bch2_inode_find_by_inum(struct bch_fs *c, u64 inode_nr,
 440                             struct bch_inode_unpacked *inode)
 441 {
 442         struct btree_iter iter;
 443         struct bkey_s_c k;
 444         int ret = -ENOENT;
 445
 446         for_each_btree_key(&iter, c, BTREE_ID_INODES,
 447                            POS(inode_nr, 0),
 448                            BTREE_ITER_SLOTS, k) {
 449                 switch (k.k->type) {
 450                 case KEY_TYPE_inode:
 451                         ret = bch2_inode_unpack(bkey_s_c_to_inode(k), inode);
 452                         break;
 453                 default:
 454                         /* hole, not found */
 455                         break;
 456                 }
 457
 458                 break;
 459         }
 460
 461         return bch2_btree_iter_unlock(&iter) ?: ret;
 462 }
 463
 464 #ifdef CONFIG_BCACHEFS_DEBUG
 465 void bch2_inode_pack_test(void)
 466 {
 467         struct bch_inode_unpacked *u, test_inodes[] = {
 468                 {
 469                         .bi_atime       = U64_MAX,
 470                         .bi_ctime       = U64_MAX,
 471                         .bi_mtime       = U64_MAX,
 472                         .bi_otime       = U64_MAX,
 473                         .bi_size        = U64_MAX,
 474                         .bi_sectors     = U64_MAX,
 475                         .bi_uid         = U32_MAX,
 476                         .bi_gid         = U32_MAX,
 477                         .bi_nlink       = U32_MAX,
 478                         .bi_generation  = U32_MAX,
 479                         .bi_dev         = U32_MAX,
 480                 },
 481         };
 482
 483         for (u = test_inodes;
 484              u < test_inodes + ARRAY_SIZE(test_inodes);
 485              u++) {
 486                 struct bkey_inode_buf p;
 487
 488                 bch2_inode_pack(&p, u);
 489         }
 490 }
 491 #endif