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 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
  99         bkey_inode_init(&packed->inode.k_i);
 100         packed->inode.k.p.inode         = inode->bi_inum;
 101         packed->inode.v.bi_hash_seed    = inode->bi_hash_seed;
 102         packed->inode.v.bi_flags        = cpu_to_le32(inode->bi_flags);
 103         packed->inode.v.bi_mode         = cpu_to_le16(inode->bi_mode);
 104
 105 #define x(_name, _bits)                                 \
 106         out += inode_encode_field(out, end, 0, inode->_name);           \
 107         nr_fields++;                                                    \
 108                                                                         \
 109         if (inode->_name) {                                             \
 110                 last_nonzero_field = out;                               \
 111                 last_nonzero_fieldnr = nr_fields;                       \
 112         }
 113
 114         BCH_INODE_FIELDS()
 115 #undef  x
 116
 117         out = last_nonzero_field;
 118         nr_fields = last_nonzero_fieldnr;
 119
 120         set_bkey_val_bytes(&packed->inode.k, out - (u8 *) &packed->inode.v);
 121         memset(out, 0,
 122                (u8 *) &packed->inode.v +
 123                bkey_val_bytes(&packed->inode.k) - out);
 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 const char *bch2_inode_invalid(const struct bch_fs *c, struct bkey_s_c k)
 184 {
 185                 struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 186                 struct bch_inode_unpacked unpacked;
 187
 188         if (k.k->p.offset)
 189                 return "nonzero offset";
 190
 191         if (bkey_val_bytes(k.k) < sizeof(struct bch_inode))
 192                 return "incorrect value size";
 193
 194         if (k.k->p.inode < BLOCKDEV_INODE_MAX)
 195                 return "fs inode in blockdev range";
 196
 197         if (INODE_STR_HASH(inode.v) >= BCH_STR_HASH_NR)
 198                 return "invalid str hash type";
 199
 200         if (bch2_inode_unpack(inode, &unpacked))
 201                 return "invalid variable length fields";
 202
 203         if (unpacked.bi_data_checksum >= BCH_CSUM_OPT_NR + 1)
 204                 return "invalid data checksum type";
 205
 206         if (unpacked.bi_compression >= BCH_COMPRESSION_OPT_NR + 1)
 207                 return "invalid data checksum type";
 208
 209         if ((unpacked.bi_flags & BCH_INODE_UNLINKED) &&
 210             unpacked.bi_nlink != 0)
 211                 return "flagged as unlinked but bi_nlink != 0";
 212
 213         return NULL;
 214 }
 215
 216 void bch2_inode_to_text(struct printbuf *out, struct bch_fs *c,
 217                        struct bkey_s_c k)
 218 {
 219         struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 220         struct bch_inode_unpacked unpacked;
 221
 222         if (bch2_inode_unpack(inode, &unpacked)) {
 223                 pr_buf(out, "(unpack error)");
 224                 return;
 225         }
 226
 227 #define x(_name, _bits)                                         \
 228         pr_buf(out, #_name ": %llu ", (u64) unpacked._name);
 229         BCH_INODE_FIELDS()
 230 #undef  x
 231 }
 232
 233 const char *bch2_inode_generation_invalid(const struct bch_fs *c,
 234                                           struct bkey_s_c k)
 235 {
 236         if (k.k->p.offset)
 237                 return "nonzero offset";
 238
 239         if (bkey_val_bytes(k.k) != sizeof(struct bch_inode_generation))
 240                 return "incorrect value size";
 241
 242         return NULL;
 243 }
 244
 245 void bch2_inode_generation_to_text(struct printbuf *out, struct bch_fs *c,
 246                                    struct bkey_s_c k)
 247 {
 248 }
 249
 250 void bch2_inode_init(struct bch_fs *c, struct bch_inode_unpacked *inode_u,
 251                      uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 252                      struct bch_inode_unpacked *parent)
 253 {
 254         s64 now = timespec_to_bch2_time(c,
 255                 timespec64_trunc(current_kernel_time64(),
 256                                  c->sb.time_precision));
 257
 258         memset(inode_u, 0, sizeof(*inode_u));
 259
 260         /* ick */
 261         inode_u->bi_flags |= c->opts.str_hash << INODE_STR_HASH_OFFSET;
 262         get_random_bytes(&inode_u->bi_hash_seed,
 263                          sizeof(inode_u->bi_hash_seed));
 264
 265         inode_u->bi_mode        = mode;
 266         inode_u->bi_uid         = uid;
 267         inode_u->bi_gid         = gid;
 268         inode_u->bi_dev         = rdev;
 269         inode_u->bi_atime       = now;
 270         inode_u->bi_mtime       = now;
 271         inode_u->bi_ctime       = now;
 272         inode_u->bi_otime       = now;
 273
 274         if (parent) {
 275 #define x(_name, ...)   inode_u->bi_##_name = parent->bi_##_name;
 276                 BCH_INODE_OPTS()
 277 #undef x
 278         }
 279 }
 280
 281 static inline u32 bkey_generation(struct bkey_s_c k)
 282 {
 283         switch (k.k->type) {
 284         case KEY_TYPE_inode:
 285                 BUG();
 286         case KEY_TYPE_inode_generation:
 287                 return le32_to_cpu(bkey_s_c_to_inode_generation(k).v->bi_generation);
 288         default:
 289                 return 0;
 290         }
 291 }
 292
 293 int __bch2_inode_create(struct btree_trans *trans,
 294                         struct bch_inode_unpacked *inode_u,
 295                         u64 min, u64 max, u64 *hint)
 296 {
 297         struct bch_fs *c = trans->c;
 298         struct bkey_inode_buf *inode_p;
 299         struct btree_iter *iter;
 300         u64 start;
 301         int ret;
 302
 303         if (!max)
 304                 max = ULLONG_MAX;
 305
 306         if (c->opts.inodes_32bit)
 307                 max = min_t(u64, max, U32_MAX);
 308
 309         start = READ_ONCE(*hint);
 310
 311         if (start >= max || start < min)
 312                 start = min;
 313
 314         inode_p = bch2_trans_kmalloc(trans, sizeof(*inode_p));
 315         if (IS_ERR(inode_p))
 316                 return PTR_ERR(inode_p);
 317
 318         iter = bch2_trans_get_iter(trans,
 319                         BTREE_ID_INODES, POS(start, 0),
 320                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 321         if (IS_ERR(iter))
 322                 return PTR_ERR(iter);
 323 again:
 324         while (1) {
 325                 struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
 326
 327                 ret = btree_iter_err(k);
 328                 if (ret)
 329                         return ret;
 330
 331                 switch (k.k->type) {
 332                 case KEY_TYPE_inode:
 333                         /* slot used */
 334                         if (iter->pos.inode >= max)
 335                                 goto out;
 336
 337                         bch2_btree_iter_next_slot(iter);
 338                         break;
 339
 340                 default:
 341                         *hint                   = k.k->p.inode;
 342                         inode_u->bi_inum        = k.k->p.inode;
 343                         inode_u->bi_generation  = bkey_generation(k);
 344
 345                         bch2_inode_pack(inode_p, inode_u);
 346                         bch2_trans_update(trans,
 347                                 BTREE_INSERT_ENTRY(iter, &inode_p->inode.k_i));
 348                         return 0;
 349                 }
 350         }
 351 out:
 352         if (start != min) {
 353                 /* Retry from start */
 354                 start = min;
 355                 bch2_btree_iter_set_pos(iter, POS(start, 0));
 356                 goto again;
 357         }
 358
 359         return -ENOSPC;
 360 }
 361
 362 int bch2_inode_create(struct bch_fs *c, struct bch_inode_unpacked *inode_u,
 363                       u64 min, u64 max, u64 *hint)
 364 {
 365         return bch2_trans_do(c, NULL, BTREE_INSERT_ATOMIC,
 366                         __bch2_inode_create(&trans, inode_u, min, max, hint));
 367 }
 368
 369 int bch2_inode_rm(struct bch_fs *c, u64 inode_nr)
 370 {
 371         struct btree_trans trans;
 372         struct btree_iter *iter;
 373         struct bkey_i_inode_generation delete;
 374         struct bpos start = POS(inode_nr, 0);
 375         struct bpos end = POS(inode_nr + 1, 0);
 376         int ret;
 377
 378         /*
 379          * If this was a directory, there shouldn't be any real dirents left -
 380          * but there could be whiteouts (from hash collisions) that we should
 381          * delete:
 382          *
 383          * XXX: the dirent could ideally would delete whiteouts when they're no
 384          * longer needed
 385          */
 386         ret   = bch2_btree_delete_range(c, BTREE_ID_EXTENTS,
 387                                         start, end, NULL) ?:
 388                 bch2_btree_delete_range(c, BTREE_ID_XATTRS,
 389                                         start, end, NULL) ?:
 390                 bch2_btree_delete_range(c, BTREE_ID_DIRENTS,
 391                                         start, end, NULL);
 392         if (ret)
 393                 return ret;
 394
 395         bch2_trans_init(&trans, c);
 396
 397         iter = bch2_trans_get_iter(&trans, BTREE_ID_INODES, POS(inode_nr, 0),
 398                                    BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 399         do {
 400                 struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
 401                 u32 bi_generation = 0;
 402
 403                 ret = btree_iter_err(k);
 404                 if (ret)
 405                         break;
 406
 407                 bch2_fs_inconsistent_on(k.k->type != KEY_TYPE_inode, c,
 408                                         "inode %llu not found when deleting",
 409                                         inode_nr);
 410
 411                 switch (k.k->type) {
 412                 case KEY_TYPE_inode: {
 413                         struct bch_inode_unpacked inode_u;
 414
 415                         if (!bch2_inode_unpack(bkey_s_c_to_inode(k), &inode_u))
 416                                 bi_generation = inode_u.bi_generation + 1;
 417                         break;
 418                 }
 419                 case KEY_TYPE_inode_generation: {
 420                         struct bkey_s_c_inode_generation g =
 421                                 bkey_s_c_to_inode_generation(k);
 422                         bi_generation = le32_to_cpu(g.v->bi_generation);
 423                         break;
 424                 }
 425                 }
 426
 427                 if (!bi_generation) {
 428                         bkey_init(&delete.k);
 429                         delete.k.p.inode = inode_nr;
 430                 } else {
 431                         bkey_inode_generation_init(&delete.k_i);
 432                         delete.k.p.inode = inode_nr;
 433                         delete.v.bi_generation = cpu_to_le32(bi_generation);
 434                 }
 435
 436                 bch2_trans_update(&trans,
 437                                   BTREE_INSERT_ENTRY(iter, &delete.k_i));
 438
 439                 ret = bch2_trans_commit(&trans, NULL, NULL,
 440                                         BTREE_INSERT_ATOMIC|
 441                                         BTREE_INSERT_NOFAIL);
 442         } while (ret == -EINTR);
 443
 444         bch2_trans_exit(&trans);
 445         return ret;
 446 }
 447
 448 int bch2_inode_find_by_inum(struct bch_fs *c, u64 inode_nr,
 449                             struct bch_inode_unpacked *inode)
 450 {
 451         struct btree_iter iter;
 452         struct bkey_s_c k;
 453         int ret = -ENOENT;
 454
 455         for_each_btree_key(&iter, c, BTREE_ID_INODES,
 456                            POS(inode_nr, 0),
 457                            BTREE_ITER_SLOTS, k) {
 458                 switch (k.k->type) {
 459                 case KEY_TYPE_inode:
 460                         ret = bch2_inode_unpack(bkey_s_c_to_inode(k), inode);
 461                         break;
 462                 default:
 463                         /* hole, not found */
 464                         break;
 465                 }
 466
 467                 break;
 468         }
 469
 470         return bch2_btree_iter_unlock(&iter) ?: ret;
 471 }
 472
 473 #ifdef CONFIG_BCACHEFS_DEBUG
 474 void bch2_inode_pack_test(void)
 475 {
 476         struct bch_inode_unpacked *u, test_inodes[] = {
 477                 {
 478                         .bi_atime       = U64_MAX,
 479                         .bi_ctime       = U64_MAX,
 480                         .bi_mtime       = U64_MAX,
 481                         .bi_otime       = U64_MAX,
 482                         .bi_size        = U64_MAX,
 483                         .bi_sectors     = U64_MAX,
 484                         .bi_uid         = U32_MAX,
 485                         .bi_gid         = U32_MAX,
 486                         .bi_nlink       = U32_MAX,
 487                         .bi_generation  = U32_MAX,
 488                         .bi_dev         = U32_MAX,
 489                 },
 490         };
 491
 492         for (u = test_inodes;
 493              u < test_inodes + ARRAY_SIZE(test_inodes);
 494              u++) {
 495                 struct bkey_inode_buf p;
 496
 497                 bch2_inode_pack(&p, u);
 498         }
 499 }
 500 #endif