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 static const char *bch2_inode_invalid(const struct bch_fs *c,
 179                                       struct bkey_s_c k)
 180 {
 181         if (k.k->p.offset)
 182                 return "nonzero offset";
 183
 184         switch (k.k->type) {
 185         case BCH_INODE_FS: {
 186                 struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 187                 struct bch_inode_unpacked unpacked;
 188
 189                 if (bkey_val_bytes(k.k) < sizeof(struct bch_inode))
 190                         return "incorrect value size";
 191
 192                 if (k.k->p.inode < BLOCKDEV_INODE_MAX)
 193                         return "fs inode in blockdev range";
 194
 195                 if (INODE_STR_HASH(inode.v) >= BCH_STR_HASH_NR)
 196                         return "invalid str hash type";
 197
 198                 if (bch2_inode_unpack(inode, &unpacked))
 199                         return "invalid variable length fields";
 200
 201                 return NULL;
 202         }
 203         case BCH_INODE_BLOCKDEV:
 204                 if (bkey_val_bytes(k.k) != sizeof(struct bch_inode_blockdev))
 205                         return "incorrect value size";
 206
 207                 if (k.k->p.inode >= BLOCKDEV_INODE_MAX)
 208                         return "blockdev inode in fs range";
 209
 210                 return NULL;
 211         case BCH_INODE_GENERATION:
 212                 if (bkey_val_bytes(k.k) != sizeof(struct bch_inode_generation))
 213                         return "incorrect value size";
 214
 215                 return NULL;
 216         default:
 217                 return "invalid type";
 218         }
 219 }
 220
 221 static void bch2_inode_to_text(struct bch_fs *c, char *buf,
 222                                size_t size, struct bkey_s_c k)
 223 {
 224         struct bkey_s_c_inode inode;
 225         struct bch_inode_unpacked unpacked;
 226
 227         switch (k.k->type) {
 228         case BCH_INODE_FS:
 229                 inode = bkey_s_c_to_inode(k);
 230                 if (bch2_inode_unpack(inode, &unpacked)) {
 231                         scnprintf(buf, size, "(unpack error)");
 232                         break;
 233                 }
 234
 235                 scnprintf(buf, size, "i_size %llu", unpacked.bi_size);
 236                 break;
 237         }
 238 }
 239
 240 const struct bkey_ops bch2_bkey_inode_ops = {
 241         .key_invalid    = bch2_inode_invalid,
 242         .val_to_text    = bch2_inode_to_text,
 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 {
 248         s64 now = timespec_to_bch2_time(c, CURRENT_TIME);
 249
 250         memset(inode_u, 0, sizeof(*inode_u));
 251
 252         /* ick */
 253         inode_u->bi_flags |= c->opts.str_hash << INODE_STR_HASH_OFFSET;
 254         get_random_bytes(&inode_u->bi_hash_seed, sizeof(inode_u->bi_hash_seed));
 255
 256         inode_u->bi_mode        = mode;
 257         inode_u->bi_uid         = uid;
 258         inode_u->bi_gid         = gid;
 259         inode_u->bi_dev         = rdev;
 260         inode_u->bi_atime       = now;
 261         inode_u->bi_mtime       = now;
 262         inode_u->bi_ctime       = now;
 263         inode_u->bi_otime       = now;
 264 }
 265
 266 int bch2_inode_create(struct bch_fs *c, struct bch_inode_unpacked *inode_u,
 267                       u64 min, u64 max, u64 *hint)
 268 {
 269         struct bkey_inode_buf inode_p;
 270         struct btree_iter iter;
 271         bool searched_from_start = false;
 272         int ret;
 273
 274         if (!max)
 275                 max = ULLONG_MAX;
 276
 277         if (c->opts.inodes_32bit)
 278                 max = min_t(u64, max, U32_MAX);
 279
 280         if (*hint >= max || *hint < min)
 281                 *hint = min;
 282
 283         if (*hint == min)
 284                 searched_from_start = true;
 285 again:
 286         bch2_btree_iter_init(&iter, c, BTREE_ID_INODES, POS(*hint, 0),
 287                              BTREE_ITER_INTENT);
 288
 289         while (1) {
 290                 struct bkey_s_c k = bch2_btree_iter_peek_with_holes(&iter);
 291                 u32 bi_generation = 0;
 292
 293                 ret = btree_iter_err(k);
 294                 if (ret) {
 295                         bch2_btree_iter_unlock(&iter);
 296                         return ret;
 297                 }
 298
 299                 switch (k.k->type) {
 300                 case BCH_INODE_BLOCKDEV:
 301                 case BCH_INODE_FS:
 302                         /* slot used */
 303                         if (iter.pos.inode == max)
 304                                 goto out;
 305
 306                         bch2_btree_iter_advance_pos(&iter);
 307                         break;
 308
 309                 case BCH_INODE_GENERATION: {
 310                         struct bkey_s_c_inode_generation g =
 311                                 bkey_s_c_to_inode_generation(k);
 312                         bi_generation = le32_to_cpu(g.v->bi_generation);
 313                         /* fallthrough: */
 314                 }
 315                 default:
 316                         inode_u->bi_generation = bi_generation;
 317
 318                         bch2_inode_pack(&inode_p, inode_u);
 319                         inode_p.inode.k.p = k.k->p;
 320
 321                         ret = bch2_btree_insert_at(c, NULL, NULL, NULL,
 322                                         BTREE_INSERT_ATOMIC,
 323                                         BTREE_INSERT_ENTRY(&iter,
 324                                                            &inode_p.inode.k_i));
 325
 326                         if (ret != -EINTR) {
 327                                 bch2_btree_iter_unlock(&iter);
 328
 329                                 if (!ret) {
 330                                         inode_u->bi_inum =
 331                                                 inode_p.inode.k.p.inode;
 332                                         *hint = inode_p.inode.k.p.inode + 1;
 333                                 }
 334
 335                                 return ret;
 336                         }
 337
 338                         if (ret == -EINTR)
 339                                 continue;
 340
 341                 }
 342         }
 343 out:
 344         bch2_btree_iter_unlock(&iter);
 345
 346         if (!searched_from_start) {
 347                 /* Retry from start */
 348                 *hint = min;
 349                 searched_from_start = true;
 350                 goto again;
 351         }
 352
 353         return -ENOSPC;
 354 }
 355
 356 int bch2_inode_truncate(struct bch_fs *c, u64 inode_nr, u64 new_size,
 357                         struct extent_insert_hook *hook, u64 *journal_seq)
 358 {
 359         return bch2_btree_delete_range(c, BTREE_ID_EXTENTS,
 360                                        POS(inode_nr, new_size),
 361                                        POS(inode_nr + 1, 0),
 362                                        ZERO_VERSION, NULL, hook,
 363                                        journal_seq);
 364 }
 365
 366 int bch2_inode_rm(struct bch_fs *c, u64 inode_nr)
 367 {
 368         struct btree_iter iter;
 369         struct bkey_i_inode_generation delete;
 370         int ret;
 371
 372         ret = bch2_inode_truncate(c, inode_nr, 0, NULL, NULL);
 373         if (ret < 0)
 374                 return ret;
 375
 376         ret = bch2_btree_delete_range(c, BTREE_ID_XATTRS,
 377                                      POS(inode_nr, 0),
 378                                      POS(inode_nr + 1, 0),
 379                                      ZERO_VERSION, NULL, NULL, NULL);
 380         if (ret < 0)
 381                 return ret;
 382
 383         /*
 384          * If this was a directory, there shouldn't be any real dirents left -
 385          * but there could be whiteouts (from hash collisions) that we should
 386          * delete:
 387          *
 388          * XXX: the dirent could ideally would delete whiteouts when they're no
 389          * longer needed
 390          */
 391         ret = bch2_btree_delete_range(c, BTREE_ID_DIRENTS,
 392                                      POS(inode_nr, 0),
 393                                      POS(inode_nr + 1, 0),
 394                                      ZERO_VERSION, NULL, NULL, NULL);
 395         if (ret < 0)
 396                 return ret;
 397
 398         bch2_btree_iter_init(&iter, c, BTREE_ID_INODES, POS(inode_nr, 0),
 399                              BTREE_ITER_INTENT);
 400         do {
 401                 struct bkey_s_c k = bch2_btree_iter_peek_with_holes(&iter);
 402                 u32 bi_generation = 0;
 403
 404                 ret = btree_iter_err(k);
 405                 if (ret) {
 406                         bch2_btree_iter_unlock(&iter);
 407                         return ret;
 408                 }
 409
 410                 bch2_fs_inconsistent_on(k.k->type != BCH_INODE_FS, c,
 411                                         "inode %llu not found when deleting",
 412                                         inode_nr);
 413
 414                 switch (k.k->type) {
 415                 case BCH_INODE_FS: {
 416                         struct bch_inode_unpacked inode_u;
 417
 418                         if (!bch2_inode_unpack(bkey_s_c_to_inode(k), &inode_u))
 419                                 bi_generation = cpu_to_le32(inode_u.bi_generation) + 1;
 420                         break;
 421                 }
 422                 case BCH_INODE_GENERATION: {
 423                         struct bkey_s_c_inode_generation g =
 424                                 bkey_s_c_to_inode_generation(k);
 425                         bi_generation = le32_to_cpu(g.v->bi_generation);
 426                         break;
 427                 }
 428                 }
 429
 430                 if (!bi_generation) {
 431                         bkey_init(&delete.k);
 432                         delete.k.p.inode = inode_nr;
 433                 } else {
 434                         bkey_inode_generation_init(&delete.k_i);
 435                         delete.k.p.inode = inode_nr;
 436                         delete.v.bi_generation = cpu_to_le32(bi_generation);
 437                 }
 438
 439                 ret = bch2_btree_insert_at(c, NULL, NULL, NULL,
 440                                 BTREE_INSERT_ATOMIC|
 441                                 BTREE_INSERT_NOFAIL,
 442                                 BTREE_INSERT_ENTRY(&iter, &delete.k_i));
 443         } while (ret == -EINTR);
 444
 445         bch2_btree_iter_unlock(&iter);
 446         return ret;
 447 }
 448
 449 int bch2_inode_find_by_inum(struct bch_fs *c, u64 inode_nr,
 450                             struct bch_inode_unpacked *inode)
 451 {
 452         struct btree_iter iter;
 453         struct bkey_s_c k;
 454         int ret = -ENOENT;
 455
 456         for_each_btree_key(&iter, c, BTREE_ID_INODES,
 457                            POS(inode_nr, 0),
 458                            BTREE_ITER_WITH_HOLES, k) {
 459                 switch (k.k->type) {
 460                 case BCH_INODE_FS:
 461                         ret = bch2_inode_unpack(bkey_s_c_to_inode(k), inode);
 462                         break;
 463                 default:
 464                         /* hole, not found */
 465                         break;
 466                 }
 467
 468                 break;
 469
 470         }
 471
 472         return bch2_btree_iter_unlock(&iter) ?: ret;
 473 }
 474
 475 int bch2_cached_dev_inode_find_by_uuid(struct bch_fs *c, uuid_le *uuid,
 476                                        struct bkey_i_inode_blockdev *ret)
 477 {
 478         struct btree_iter iter;
 479         struct bkey_s_c k;
 480
 481         for_each_btree_key(&iter, c, BTREE_ID_INODES, POS(0, 0), 0, k) {
 482                 if (k.k->p.inode >= BLOCKDEV_INODE_MAX)
 483                         break;
 484
 485                 if (k.k->type == BCH_INODE_BLOCKDEV) {
 486                         struct bkey_s_c_inode_blockdev inode =
 487                                 bkey_s_c_to_inode_blockdev(k);
 488
 489                         pr_debug("found inode %llu: %pU (u64s %u)",
 490                                  inode.k->p.inode, inode.v->i_uuid.b,
 491                                  inode.k->u64s);
 492
 493                         if (CACHED_DEV(inode.v) &&
 494                             !memcmp(uuid, &inode.v->i_uuid, 16)) {
 495                                 bkey_reassemble(&ret->k_i, k);
 496                                 bch2_btree_iter_unlock(&iter);
 497                                 return 0;
 498                         }
 499                 }
 500
 501                 bch2_btree_iter_cond_resched(&iter);
 502         }
 503         bch2_btree_iter_unlock(&iter);
 504         return -ENOENT;
 505 }
 506
 507 #ifdef CONFIG_BCACHEFS_DEBUG
 508 void bch2_inode_pack_test(void)
 509 {
 510         struct bch_inode_unpacked *u, test_inodes[] = {
 511                 {
 512                         .bi_atime       = U64_MAX,
 513                         .bi_ctime       = U64_MAX,
 514                         .bi_mtime       = U64_MAX,
 515                         .bi_otime       = U64_MAX,
 516                         .bi_size        = U64_MAX,
 517                         .bi_sectors     = U64_MAX,
 518                         .bi_uid         = U32_MAX,
 519                         .bi_gid         = U32_MAX,
 520                         .bi_nlink       = U32_MAX,
 521                         .bi_generation  = U32_MAX,
 522                         .bi_dev         = U32_MAX,
 523                 },
 524         };
 525
 526         for (u = test_inodes;
 527              u < test_inodes + ARRAY_SIZE(test_inodes);
 528              u++) {
 529                 struct bkey_inode_buf p;
 530
 531                 bch2_inode_pack(&p, u);
 532         }
 533 }
 534 #endif