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 "io.h"
  10 #include "keylist.h"
  11
  12 #include <linux/random.h>
  13
  14 #include <asm/unaligned.h>
  15
  16 const char * const bch2_inode_opts[] = {
  17 #define x(name, ...)    #name,
  18         BCH_INODE_OPTS()
  19 #undef  x
  20         NULL,
  21 };
  22
  23 static const u8 byte_table[8] = { 1, 2, 3, 4, 6, 8, 10, 13 };
  24 static const u8 bits_table[8] = {
  25         1  * 8 - 1,
  26         2  * 8 - 2,
  27         3  * 8 - 3,
  28         4  * 8 - 4,
  29         6  * 8 - 5,
  30         8  * 8 - 6,
  31         10 * 8 - 7,
  32         13 * 8 - 8,
  33 };
  34
  35 static int inode_encode_field(u8 *out, u8 *end, u64 hi, u64 lo)
  36 {
  37         __be64 in[2] = { cpu_to_be64(hi), cpu_to_be64(lo), };
  38         unsigned shift, bytes, bits = likely(!hi)
  39                 ? fls64(lo)
  40                 : fls64(hi) + 64;
  41
  42         for (shift = 1; shift <= 8; shift++)
  43                 if (bits < bits_table[shift - 1])
  44                         goto got_shift;
  45
  46         BUG();
  47 got_shift:
  48         bytes = byte_table[shift - 1];
  49
  50         BUG_ON(out + bytes > end);
  51
  52         memcpy(out, (u8 *) in + 16 - bytes, bytes);
  53         *out |= (1 << 8) >> shift;
  54
  55         return bytes;
  56 }
  57
  58 static int inode_decode_field(const u8 *in, const u8 *end,
  59                               u64 out[2], unsigned *out_bits)
  60 {
  61         __be64 be[2] = { 0, 0 };
  62         unsigned bytes, shift;
  63         u8 *p;
  64
  65         if (in >= end)
  66                 return -1;
  67
  68         if (!*in)
  69                 return -1;
  70
  71         /*
  72          * position of highest set bit indicates number of bytes:
  73          * shift = number of bits to remove in high byte:
  74          */
  75         shift   = 8 - __fls(*in); /* 1 <= shift <= 8 */
  76         bytes   = byte_table[shift - 1];
  77
  78         if (in + bytes > end)
  79                 return -1;
  80
  81         p = (u8 *) be + 16 - bytes;
  82         memcpy(p, in, bytes);
  83         *p ^= (1 << 8) >> shift;
  84
  85         out[0] = be64_to_cpu(be[0]);
  86         out[1] = be64_to_cpu(be[1]);
  87         *out_bits = out[0] ? 64 + fls64(out[0]) : fls64(out[1]);
  88
  89         return bytes;
  90 }
  91
  92 void bch2_inode_pack(struct bkey_inode_buf *packed,
  93                      const struct bch_inode_unpacked *inode)
  94 {
  95         u8 *out = packed->inode.v.fields;
  96         u8 *end = (void *) &packed[1];
  97         u8 *last_nonzero_field = out;
  98         unsigned nr_fields = 0, last_nonzero_fieldnr = 0;
  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         set_bkey_val_bytes(&packed->inode.k, out - (u8 *) &packed->inode.v);
 122         memset(out, 0,
 123                (u8 *) &packed->inode.v +
 124                bkey_val_bytes(&packed->inode.k) - out);
 125
 126         SET_INODE_NR_FIELDS(&packed->inode.v, nr_fields);
 127
 128         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
 129                 struct bch_inode_unpacked unpacked;
 130
 131                 int ret = bch2_inode_unpack(inode_i_to_s_c(&packed->inode),
 132                                            &unpacked);
 133                 BUG_ON(ret);
 134                 BUG_ON(unpacked.bi_inum         != inode->bi_inum);
 135                 BUG_ON(unpacked.bi_hash_seed    != inode->bi_hash_seed);
 136                 BUG_ON(unpacked.bi_mode         != inode->bi_mode);
 137
 138 #define x(_name, _bits) BUG_ON(unpacked._name != inode->_name);
 139                 BCH_INODE_FIELDS()
 140 #undef  x
 141         }
 142 }
 143
 144 int bch2_inode_unpack(struct bkey_s_c_inode inode,
 145                       struct bch_inode_unpacked *unpacked)
 146 {
 147         const u8 *in = inode.v->fields;
 148         const u8 *end = (void *) inode.v + bkey_val_bytes(inode.k);
 149         u64 field[2];
 150         unsigned fieldnr = 0, field_bits;
 151         int ret;
 152
 153         unpacked->bi_inum       = inode.k->p.inode;
 154         unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 155         unpacked->bi_flags      = le32_to_cpu(inode.v->bi_flags);
 156         unpacked->bi_mode       = le16_to_cpu(inode.v->bi_mode);
 157
 158 #define x(_name, _bits)                                 \
 159         if (fieldnr++ == INODE_NR_FIELDS(inode.v)) {                    \
 160                 memset(&unpacked->_name, 0,                             \
 161                        sizeof(*unpacked) -                              \
 162                        offsetof(struct bch_inode_unpacked, _name));     \
 163                 return 0;                                               \
 164         }                                                               \
 165                                                                         \
 166         ret = inode_decode_field(in, end, field, &field_bits);          \
 167         if (ret < 0)                                                    \
 168                 return ret;                                             \
 169                                                                         \
 170         if (field_bits > sizeof(unpacked->_name) * 8)                   \
 171                 return -1;                                              \
 172                                                                         \
 173         unpacked->_name = field[1];                                     \
 174         in += ret;
 175
 176         BCH_INODE_FIELDS()
 177 #undef  x
 178
 179         /* XXX: signal if there were more fields than expected? */
 180
 181         return 0;
 182 }
 183
 184 const char *bch2_inode_invalid(const struct bch_fs *c, struct bkey_s_c k)
 185 {
 186                 struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 187                 struct bch_inode_unpacked unpacked;
 188
 189         if (k.k->p.offset)
 190                 return "nonzero offset";
 191
 192         if (bkey_val_bytes(k.k) < sizeof(struct bch_inode))
 193                 return "incorrect value size";
 194
 195         if (k.k->p.inode < BLOCKDEV_INODE_MAX)
 196                 return "fs inode in blockdev range";
 197
 198         if (INODE_STR_HASH(inode.v) >= BCH_STR_HASH_NR)
 199                 return "invalid str hash type";
 200
 201         if (bch2_inode_unpack(inode, &unpacked))
 202                 return "invalid variable length fields";
 203
 204         if (unpacked.bi_data_checksum >= BCH_CSUM_OPT_NR + 1)
 205                 return "invalid data checksum type";
 206
 207         if (unpacked.bi_compression >= BCH_COMPRESSION_OPT_NR + 1)
 208                 return "invalid data checksum type";
 209
 210         if ((unpacked.bi_flags & BCH_INODE_UNLINKED) &&
 211             unpacked.bi_nlink != 0)
 212                 return "flagged as unlinked but bi_nlink != 0";
 213
 214         return NULL;
 215 }
 216
 217 void bch2_inode_to_text(struct printbuf *out, struct bch_fs *c,
 218                        struct bkey_s_c k)
 219 {
 220         struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 221         struct bch_inode_unpacked unpacked;
 222
 223         if (bch2_inode_unpack(inode, &unpacked)) {
 224                 pr_buf(out, "(unpack error)");
 225                 return;
 226         }
 227
 228 #define x(_name, _bits)                                         \
 229         pr_buf(out, #_name ": %llu ", (u64) unpacked._name);
 230         BCH_INODE_FIELDS()
 231 #undef  x
 232 }
 233
 234 const char *bch2_inode_generation_invalid(const struct bch_fs *c,
 235                                           struct bkey_s_c k)
 236 {
 237         if (k.k->p.offset)
 238                 return "nonzero offset";
 239
 240         if (bkey_val_bytes(k.k) != sizeof(struct bch_inode_generation))
 241                 return "incorrect value size";
 242
 243         return NULL;
 244 }
 245
 246 void bch2_inode_generation_to_text(struct printbuf *out, struct bch_fs *c,
 247                                    struct bkey_s_c k)
 248 {
 249         struct bkey_s_c_inode_generation gen = bkey_s_c_to_inode_generation(k);
 250
 251         pr_buf(out, "generation: %u", le32_to_cpu(gen.v->bi_generation));
 252 }
 253
 254 void bch2_inode_init(struct bch_fs *c, struct bch_inode_unpacked *inode_u,
 255                      uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 256                      struct bch_inode_unpacked *parent)
 257 {
 258         s64 now = bch2_current_time(c);
 259
 260         memset(inode_u, 0, sizeof(*inode_u));
 261
 262         /* ick */
 263         inode_u->bi_flags |= c->opts.str_hash << INODE_STR_HASH_OFFSET;
 264         get_random_bytes(&inode_u->bi_hash_seed,
 265                          sizeof(inode_u->bi_hash_seed));
 266
 267         inode_u->bi_mode        = mode;
 268         inode_u->bi_uid         = uid;
 269         inode_u->bi_gid         = gid;
 270         inode_u->bi_dev         = rdev;
 271         inode_u->bi_atime       = now;
 272         inode_u->bi_mtime       = now;
 273         inode_u->bi_ctime       = now;
 274         inode_u->bi_otime       = now;
 275
 276         if (parent) {
 277 #define x(_name, ...)   inode_u->bi_##_name = parent->bi_##_name;
 278                 BCH_INODE_OPTS()
 279 #undef x
 280         }
 281 }
 282
 283 static inline u32 bkey_generation(struct bkey_s_c k)
 284 {
 285         switch (k.k->type) {
 286         case KEY_TYPE_inode:
 287                 BUG();
 288         case KEY_TYPE_inode_generation:
 289                 return le32_to_cpu(bkey_s_c_to_inode_generation(k).v->bi_generation);
 290         default:
 291                 return 0;
 292         }
 293 }
 294
 295 int __bch2_inode_create(struct btree_trans *trans,
 296                         struct bch_inode_unpacked *inode_u,
 297                         u64 min, u64 max, u64 *hint)
 298 {
 299         struct bch_fs *c = trans->c;
 300         struct bkey_inode_buf *inode_p;
 301         struct btree_iter *iter;
 302         u64 start;
 303         int ret;
 304
 305         if (!max)
 306                 max = ULLONG_MAX;
 307
 308         if (c->opts.inodes_32bit)
 309                 max = min_t(u64, max, U32_MAX);
 310
 311         start = READ_ONCE(*hint);
 312
 313         if (start >= max || start < min)
 314                 start = min;
 315
 316         inode_p = bch2_trans_kmalloc(trans, sizeof(*inode_p));
 317         if (IS_ERR(inode_p))
 318                 return PTR_ERR(inode_p);
 319
 320         iter = bch2_trans_get_iter(trans,
 321                         BTREE_ID_INODES, POS(start, 0),
 322                         BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 323         if (IS_ERR(iter))
 324                 return PTR_ERR(iter);
 325 again:
 326         while (1) {
 327                 struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
 328
 329                 ret = bkey_err(k);
 330                 if (ret)
 331                         return ret;
 332
 333                 switch (k.k->type) {
 334                 case KEY_TYPE_inode:
 335                         /* slot used */
 336                         if (iter->pos.inode >= max)
 337                                 goto out;
 338
 339                         bch2_btree_iter_next_slot(iter);
 340                         break;
 341
 342                 default:
 343                         *hint                   = k.k->p.inode;
 344                         inode_u->bi_inum        = k.k->p.inode;
 345                         inode_u->bi_generation  = bkey_generation(k);
 346
 347                         bch2_inode_pack(inode_p, inode_u);
 348                         bch2_trans_update(trans, iter, &inode_p->inode.k_i);
 349                         return 0;
 350                 }
 351         }
 352 out:
 353         if (start != min) {
 354                 /* Retry from start */
 355                 start = min;
 356                 bch2_btree_iter_set_pos(iter, POS(start, 0));
 357                 goto again;
 358         }
 359
 360         return -ENOSPC;
 361 }
 362
 363 int bch2_inode_create(struct bch_fs *c, struct bch_inode_unpacked *inode_u,
 364                       u64 min, u64 max, u64 *hint)
 365 {
 366         return bch2_trans_do(c, NULL, BTREE_INSERT_ATOMIC,
 367                         __bch2_inode_create(&trans, inode_u, min, max, hint));
 368 }
 369
 370 int bch2_inode_rm(struct bch_fs *c, u64 inode_nr)
 371 {
 372         struct btree_trans trans;
 373         struct btree_iter *iter;
 374         struct bkey_i_inode_generation delete;
 375         struct bpos start = POS(inode_nr, 0);
 376         struct bpos end = POS(inode_nr + 1, 0);
 377         int ret;
 378
 379         /*
 380          * If this was a directory, there shouldn't be any real dirents left -
 381          * but there could be whiteouts (from hash collisions) that we should
 382          * delete:
 383          *
 384          * XXX: the dirent could ideally would delete whiteouts when they're no
 385          * longer needed
 386          */
 387         ret   = bch2_btree_delete_range(c, BTREE_ID_EXTENTS,
 388                                         start, end, NULL) ?:
 389                 bch2_btree_delete_range(c, BTREE_ID_XATTRS,
 390                                         start, end, NULL) ?:
 391                 bch2_btree_delete_range(c, BTREE_ID_DIRENTS,
 392                                         start, end, NULL);
 393         if (ret)
 394                 return ret;
 395
 396         bch2_trans_init(&trans, c, 0, 0);
 397
 398         iter = bch2_trans_get_iter(&trans, BTREE_ID_INODES, POS(inode_nr, 0),
 399                                    BTREE_ITER_SLOTS|BTREE_ITER_INTENT);
 400         do {
 401                 struct bkey_s_c k = bch2_btree_iter_peek_slot(iter);
 402                 u32 bi_generation = 0;
 403
 404                 ret = bkey_err(k);
 405                 if (ret)
 406                         break;
 407
 408                 bch2_fs_inconsistent_on(k.k->type != KEY_TYPE_inode, c,
 409                                         "inode %llu not found when deleting",
 410                                         inode_nr);
 411
 412                 switch (k.k->type) {
 413                 case KEY_TYPE_inode: {
 414                         struct bch_inode_unpacked inode_u;
 415
 416                         if (!bch2_inode_unpack(bkey_s_c_to_inode(k), &inode_u))
 417                                 bi_generation = inode_u.bi_generation + 1;
 418                         break;
 419                 }
 420                 case KEY_TYPE_inode_generation: {
 421                         struct bkey_s_c_inode_generation g =
 422                                 bkey_s_c_to_inode_generation(k);
 423                         bi_generation = le32_to_cpu(g.v->bi_generation);
 424                         break;
 425                 }
 426                 }
 427
 428                 if (!bi_generation) {
 429                         bkey_init(&delete.k);
 430                         delete.k.p.inode = inode_nr;
 431                 } else {
 432                         bkey_inode_generation_init(&delete.k_i);
 433                         delete.k.p.inode = inode_nr;
 434                         delete.v.bi_generation = cpu_to_le32(bi_generation);
 435                 }
 436
 437                 bch2_trans_update(&trans, 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_trans(struct btree_trans *trans, 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         iter = bch2_trans_get_iter(trans, BTREE_ID_INODES,
 456                         POS(inode_nr, 0), BTREE_ITER_SLOTS);
 457         if (IS_ERR(iter))
 458                 return PTR_ERR(iter);
 459
 460         k = bch2_btree_iter_peek_slot(iter);
 461         if (k.k->type == KEY_TYPE_inode)
 462                 ret = bch2_inode_unpack(bkey_s_c_to_inode(k), inode);
 463
 464         bch2_trans_iter_put(trans, iter);
 465
 466         return ret;
 467 }
 468
 469 int bch2_inode_find_by_inum(struct bch_fs *c, u64 inode_nr,
 470                             struct bch_inode_unpacked *inode)
 471 {
 472         return bch2_trans_do(c, NULL, 0,
 473                 bch2_inode_find_by_inum_trans(&trans, inode_nr, inode));
 474 }
 475
 476 #ifdef CONFIG_BCACHEFS_DEBUG
 477 void bch2_inode_pack_test(void)
 478 {
 479         struct bch_inode_unpacked *u, test_inodes[] = {
 480                 {
 481                         .bi_atime       = U64_MAX,
 482                         .bi_ctime       = U64_MAX,
 483                         .bi_mtime       = U64_MAX,
 484                         .bi_otime       = U64_MAX,
 485                         .bi_size        = U64_MAX,
 486                         .bi_sectors     = U64_MAX,
 487                         .bi_uid         = U32_MAX,
 488                         .bi_gid         = U32_MAX,
 489                         .bi_nlink       = U32_MAX,
 490                         .bi_generation  = U32_MAX,
 491                         .bi_dev         = U32_MAX,
 492                 },
 493         };
 494
 495         for (u = test_inodes;
 496              u < test_inodes + ARRAY_SIZE(test_inodes);
 497              u++) {
 498                 struct bkey_inode_buf p;
 499
 500                 bch2_inode_pack(&p, u);
 501         }
 502 }
 503 #endif