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 "btree_key_cache.h"
   5 #include "bkey_methods.h"
   6 #include "btree_update.h"
   7 #include "error.h"
   8 #include "extents.h"
   9 #include "extent_update.h"
  10 #include "inode.h"
  11 #include "str_hash.h"
  12 #include "subvolume.h"
  13 #include "varint.h"
  14
  15 #include <linux/random.h>
  16
  17 #include <asm/unaligned.h>
  18
  19 const char * const bch2_inode_opts[] = {
  20 #define x(name, ...)    #name,
  21         BCH_INODE_OPTS()
  22 #undef  x
  23         NULL,
  24 };
  25
  26 static const u8 byte_table[8] = { 1, 2, 3, 4, 6, 8, 10, 13 };
  27 static const u8 bits_table[8] = {
  28         1  * 8 - 1,
  29         2  * 8 - 2,
  30         3  * 8 - 3,
  31         4  * 8 - 4,
  32         6  * 8 - 5,
  33         8  * 8 - 6,
  34         10 * 8 - 7,
  35         13 * 8 - 8,
  36 };
  37
  38 static int inode_decode_field(const u8 *in, const u8 *end,
  39                               u64 out[2], unsigned *out_bits)
  40 {
  41         __be64 be[2] = { 0, 0 };
  42         unsigned bytes, shift;
  43         u8 *p;
  44
  45         if (in >= end)
  46                 return -1;
  47
  48         if (!*in)
  49                 return -1;
  50
  51         /*
  52          * position of highest set bit indicates number of bytes:
  53          * shift = number of bits to remove in high byte:
  54          */
  55         shift   = 8 - __fls(*in); /* 1 <= shift <= 8 */
  56         bytes   = byte_table[shift - 1];
  57
  58         if (in + bytes > end)
  59                 return -1;
  60
  61         p = (u8 *) be + 16 - bytes;
  62         memcpy(p, in, bytes);
  63         *p ^= (1 << 8) >> shift;
  64
  65         out[0] = be64_to_cpu(be[0]);
  66         out[1] = be64_to_cpu(be[1]);
  67         *out_bits = out[0] ? 64 + fls64(out[0]) : fls64(out[1]);
  68
  69         return bytes;
  70 }
  71
  72 void bch2_inode_pack(struct bch_fs *c,
  73                      struct bkey_inode_buf *packed,
  74                      const struct bch_inode_unpacked *inode)
  75 {
  76         struct bkey_i_inode_v2 *k = &packed->inode;
  77         u8 *out = k->v.fields;
  78         u8 *end = (void *) &packed[1];
  79         u8 *last_nonzero_field = out;
  80         unsigned nr_fields = 0, last_nonzero_fieldnr = 0;
  81         unsigned bytes;
  82         int ret;
  83
  84         bkey_inode_v2_init(&packed->inode.k_i);
  85         packed->inode.k.p.offset        = inode->bi_inum;
  86         packed->inode.v.bi_journal_seq  = cpu_to_le64(inode->bi_journal_seq);
  87         packed->inode.v.bi_hash_seed    = inode->bi_hash_seed;
  88         packed->inode.v.bi_flags        = cpu_to_le64(inode->bi_flags);
  89         packed->inode.v.bi_flags        = cpu_to_le64(inode->bi_flags);
  90         packed->inode.v.bi_mode         = cpu_to_le16(inode->bi_mode);
  91
  92 #define x(_name, _bits)                                                 \
  93         nr_fields++;                                                    \
  94                                                                         \
  95         if (inode->_name) {                                             \
  96                 ret = bch2_varint_encode_fast(out, inode->_name);       \
  97                 out += ret;                                             \
  98                                                                         \
  99                 if (_bits > 64)                                         \
 100                         *out++ = 0;                                     \
 101                                                                         \
 102                 last_nonzero_field = out;                               \
 103                 last_nonzero_fieldnr = nr_fields;                       \
 104         } else {                                                        \
 105                 *out++ = 0;                                             \
 106                                                                         \
 107                 if (_bits > 64)                                         \
 108                         *out++ = 0;                                     \
 109         }
 110
 111         BCH_INODE_FIELDS()
 112 #undef  x
 113         BUG_ON(out > end);
 114
 115         out = last_nonzero_field;
 116         nr_fields = last_nonzero_fieldnr;
 117
 118         bytes = out - (u8 *) &packed->inode.v;
 119         set_bkey_val_bytes(&packed->inode.k, bytes);
 120         memset_u64s_tail(&packed->inode.v, 0, bytes);
 121
 122         SET_INODEv2_NR_FIELDS(&k->v, nr_fields);
 123
 124         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG)) {
 125                 struct bch_inode_unpacked unpacked;
 126
 127                 int ret = bch2_inode_unpack(bkey_i_to_s_c(&packed->inode.k_i),
 128                                            &unpacked);
 129                 BUG_ON(ret);
 130                 BUG_ON(unpacked.bi_inum         != inode->bi_inum);
 131                 BUG_ON(unpacked.bi_hash_seed    != inode->bi_hash_seed);
 132                 BUG_ON(unpacked.bi_mode         != inode->bi_mode);
 133
 134 #define x(_name, _bits) if (unpacked._name != inode->_name)             \
 135                         panic("unpacked %llu should be %llu",           \
 136                               (u64) unpacked._name, (u64) inode->_name);
 137                 BCH_INODE_FIELDS()
 138 #undef  x
 139         }
 140 }
 141
 142 static noinline int bch2_inode_unpack_v1(struct bkey_s_c_inode inode,
 143                                 struct bch_inode_unpacked *unpacked)
 144 {
 145         const u8 *in = inode.v->fields;
 146         const u8 *end = bkey_val_end(inode);
 147         u64 field[2];
 148         unsigned fieldnr = 0, field_bits;
 149         int ret;
 150
 151 #define x(_name, _bits)                                 \
 152         if (fieldnr++ == INODE_NR_FIELDS(inode.v)) {                    \
 153                 memset(&unpacked->_name, 0,                             \
 154                        sizeof(*unpacked) -                              \
 155                        offsetof(struct bch_inode_unpacked, _name));     \
 156                 return 0;                                               \
 157         }                                                               \
 158                                                                         \
 159         ret = inode_decode_field(in, end, field, &field_bits);          \
 160         if (ret < 0)                                                    \
 161                 return ret;                                             \
 162                                                                         \
 163         if (field_bits > sizeof(unpacked->_name) * 8)                   \
 164                 return -1;                                              \
 165                                                                         \
 166         unpacked->_name = field[1];                                     \
 167         in += ret;
 168
 169         BCH_INODE_FIELDS()
 170 #undef  x
 171
 172         /* XXX: signal if there were more fields than expected? */
 173         return 0;
 174 }
 175
 176 static int bch2_inode_unpack_v2(struct bch_inode_unpacked *unpacked,
 177                                 const u8 *in, const u8 *end,
 178                                 unsigned nr_fields)
 179 {
 180         unsigned fieldnr = 0;
 181         int ret;
 182         u64 v[2];
 183
 184 #define x(_name, _bits)                                                 \
 185         if (fieldnr < nr_fields) {                                      \
 186                 ret = bch2_varint_decode_fast(in, end, &v[0]);          \
 187                 if (ret < 0)                                            \
 188                         return ret;                                     \
 189                 in += ret;                                              \
 190                                                                         \
 191                 if (_bits > 64) {                                       \
 192                         ret = bch2_varint_decode_fast(in, end, &v[1]);  \
 193                         if (ret < 0)                                    \
 194                                 return ret;                             \
 195                         in += ret;                                      \
 196                 } else {                                                \
 197                         v[1] = 0;                                       \
 198                 }                                                       \
 199         } else {                                                        \
 200                 v[0] = v[1] = 0;                                        \
 201         }                                                               \
 202                                                                         \
 203         unpacked->_name = v[0];                                         \
 204         if (v[1] || v[0] != unpacked->_name)                            \
 205                 return -1;                                              \
 206         fieldnr++;
 207
 208         BCH_INODE_FIELDS()
 209 #undef  x
 210
 211         /* XXX: signal if there were more fields than expected? */
 212         return 0;
 213 }
 214
 215 int bch2_inode_unpack(struct bkey_s_c k,
 216                       struct bch_inode_unpacked *unpacked)
 217 {
 218         switch (k.k->type) {
 219         case KEY_TYPE_inode: {
 220                 struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 221
 222                 unpacked->bi_inum       = inode.k->p.offset;
 223                 unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 224                 unpacked->bi_flags      = le32_to_cpu(inode.v->bi_flags);
 225                 unpacked->bi_mode       = le16_to_cpu(inode.v->bi_mode);
 226
 227                 if (INODE_NEW_VARINT(inode.v)) {
 228                         return bch2_inode_unpack_v2(unpacked, inode.v->fields,
 229                                                     bkey_val_end(inode),
 230                                                     INODE_NR_FIELDS(inode.v));
 231                 } else {
 232                         return bch2_inode_unpack_v1(inode, unpacked);
 233                 }
 234                 break;
 235         }
 236         case KEY_TYPE_inode_v2: {
 237                 struct bkey_s_c_inode_v2 inode = bkey_s_c_to_inode_v2(k);
 238
 239                 unpacked->bi_inum       = inode.k->p.offset;
 240                 unpacked->bi_journal_seq= le64_to_cpu(inode.v->bi_journal_seq);
 241                 unpacked->bi_hash_seed  = inode.v->bi_hash_seed;
 242                 unpacked->bi_flags      = le64_to_cpu(inode.v->bi_flags);
 243                 unpacked->bi_mode       = le16_to_cpu(inode.v->bi_mode);
 244
 245                 return bch2_inode_unpack_v2(unpacked, inode.v->fields,
 246                                             bkey_val_end(inode),
 247                                             INODEv2_NR_FIELDS(inode.v));
 248         }
 249         default:
 250                 BUG();
 251         }
 252 }
 253
 254 int bch2_inode_peek(struct btree_trans *trans,
 255                     struct btree_iter *iter,
 256                     struct bch_inode_unpacked *inode,
 257                     subvol_inum inum, unsigned flags)
 258 {
 259         struct bkey_s_c k;
 260         u32 snapshot;
 261         int ret;
 262
 263         if (0 && trans->c->opts.inodes_use_key_cache)
 264                 flags |= BTREE_ITER_CACHED;
 265
 266         ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 267         if (ret)
 268                 return ret;
 269
 270         bch2_trans_iter_init(trans, iter, BTREE_ID_inodes,
 271                              SPOS(0, inum.inum, snapshot), flags);
 272         k = bch2_btree_iter_peek_slot(iter);
 273         ret = bkey_err(k);
 274         if (ret)
 275                 goto err;
 276
 277         ret = bkey_is_inode(k.k) ? 0 : -ENOENT;
 278         if (ret)
 279                 goto err;
 280
 281         ret = bch2_inode_unpack(k, inode);
 282         if (ret)
 283                 goto err;
 284
 285         return 0;
 286 err:
 287         bch2_trans_iter_exit(trans, iter);
 288         return ret;
 289 }
 290
 291 int bch2_inode_write(struct btree_trans *trans,
 292                      struct btree_iter *iter,
 293                      struct bch_inode_unpacked *inode)
 294 {
 295         struct bkey_inode_buf *inode_p;
 296
 297         inode_p = bch2_trans_kmalloc(trans, sizeof(*inode_p));
 298         if (IS_ERR(inode_p))
 299                 return PTR_ERR(inode_p);
 300
 301         bch2_inode_pack(trans->c, inode_p, inode);
 302         inode_p->inode.k.p.snapshot = iter->snapshot;
 303         return bch2_trans_update(trans, iter, &inode_p->inode.k_i, 0);
 304 }
 305
 306 const char *bch2_inode_invalid(const struct bch_fs *c, struct bkey_s_c k)
 307 {
 308         struct bkey_s_c_inode inode = bkey_s_c_to_inode(k);
 309         struct bch_inode_unpacked unpacked;
 310
 311         if (k.k->p.inode)
 312                 return "nonzero k.p.inode";
 313
 314         if (bkey_val_bytes(k.k) < sizeof(struct bch_inode))
 315                 return "incorrect value size";
 316
 317         if (k.k->p.offset < BLOCKDEV_INODE_MAX)
 318                 return "fs inode in blockdev range";
 319
 320         if (INODE_STR_HASH(inode.v) >= BCH_STR_HASH_NR)
 321                 return "invalid str hash type";
 322
 323         if (bch2_inode_unpack(k, &unpacked))
 324                 return "invalid variable length fields";
 325
 326         if (unpacked.bi_data_checksum >= BCH_CSUM_OPT_NR + 1)
 327                 return "invalid data checksum type";
 328
 329         if (unpacked.bi_compression >= BCH_COMPRESSION_OPT_NR + 1)
 330                 return "invalid data checksum type";
 331
 332         if ((unpacked.bi_flags & BCH_INODE_UNLINKED) &&
 333             unpacked.bi_nlink != 0)
 334                 return "flagged as unlinked but bi_nlink != 0";
 335
 336         if (unpacked.bi_subvol && !S_ISDIR(unpacked.bi_mode))
 337                 return "subvolume root but not a directory";
 338
 339         return NULL;
 340 }
 341
 342 const char *bch2_inode_v2_invalid(const struct bch_fs *c, struct bkey_s_c k)
 343 {
 344         struct bkey_s_c_inode_v2 inode = bkey_s_c_to_inode_v2(k);
 345         struct bch_inode_unpacked unpacked;
 346
 347         if (k.k->p.inode)
 348                 return "nonzero k.p.inode";
 349
 350         if (bkey_val_bytes(k.k) < sizeof(struct bch_inode))
 351                 return "incorrect value size";
 352
 353         if (k.k->p.offset < BLOCKDEV_INODE_MAX)
 354                 return "fs inode in blockdev range";
 355
 356         if (INODEv2_STR_HASH(inode.v) >= BCH_STR_HASH_NR)
 357                 return "invalid str hash type";
 358
 359         if (bch2_inode_unpack(k, &unpacked))
 360                 return "invalid variable length fields";
 361
 362         if (unpacked.bi_data_checksum >= BCH_CSUM_OPT_NR + 1)
 363                 return "invalid data checksum type";
 364
 365         if (unpacked.bi_compression >= BCH_COMPRESSION_OPT_NR + 1)
 366                 return "invalid data checksum type";
 367
 368         if ((unpacked.bi_flags & BCH_INODE_UNLINKED) &&
 369             unpacked.bi_nlink != 0)
 370                 return "flagged as unlinked but bi_nlink != 0";
 371
 372         if (unpacked.bi_subvol && !S_ISDIR(unpacked.bi_mode))
 373                 return "subvolume root but not a directory";
 374
 375         return NULL;
 376 }
 377
 378 static void __bch2_inode_unpacked_to_text(struct printbuf *out, struct bch_inode_unpacked *inode)
 379 {
 380         pr_buf(out, "mode %o flags %x journal_seq %llu",
 381                inode->bi_mode, inode->bi_flags,
 382                inode->bi_journal_seq);
 383
 384 #define x(_name, _bits)                                         \
 385         pr_buf(out, " "#_name " %llu", (u64) inode->_name);
 386         BCH_INODE_FIELDS()
 387 #undef  x
 388 }
 389
 390 void bch2_inode_unpacked_to_text(struct printbuf *out, struct bch_inode_unpacked *inode)
 391 {
 392         pr_buf(out, "inum: %llu ", inode->bi_inum);
 393         __bch2_inode_unpacked_to_text(out, inode);
 394 }
 395
 396 void bch2_inode_to_text(struct printbuf *out, struct bch_fs *c,
 397                        struct bkey_s_c k)
 398 {
 399         struct bch_inode_unpacked inode;
 400
 401         if (bch2_inode_unpack(k, &inode)) {
 402                 pr_buf(out, "(unpack error)");
 403                 return;
 404         }
 405
 406         __bch2_inode_unpacked_to_text(out, &inode);
 407 }
 408
 409 const char *bch2_inode_generation_invalid(const struct bch_fs *c,
 410                                           struct bkey_s_c k)
 411 {
 412         if (k.k->p.inode)
 413                 return "nonzero k.p.inode";
 414
 415         if (bkey_val_bytes(k.k) != sizeof(struct bch_inode_generation))
 416                 return "incorrect value size";
 417
 418         return NULL;
 419 }
 420
 421 void bch2_inode_generation_to_text(struct printbuf *out, struct bch_fs *c,
 422                                    struct bkey_s_c k)
 423 {
 424         struct bkey_s_c_inode_generation gen = bkey_s_c_to_inode_generation(k);
 425
 426         pr_buf(out, "generation: %u", le32_to_cpu(gen.v->bi_generation));
 427 }
 428
 429 void bch2_inode_init_early(struct bch_fs *c,
 430                            struct bch_inode_unpacked *inode_u)
 431 {
 432         enum bch_str_hash_type str_hash =
 433                 bch2_str_hash_opt_to_type(c, c->opts.str_hash);
 434
 435         memset(inode_u, 0, sizeof(*inode_u));
 436
 437         /* ick */
 438         inode_u->bi_flags |= str_hash << INODE_STR_HASH_OFFSET;
 439         get_random_bytes(&inode_u->bi_hash_seed,
 440                          sizeof(inode_u->bi_hash_seed));
 441 }
 442
 443 void bch2_inode_init_late(struct bch_inode_unpacked *inode_u, u64 now,
 444                           uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 445                           struct bch_inode_unpacked *parent)
 446 {
 447         inode_u->bi_mode        = mode;
 448         inode_u->bi_uid         = uid;
 449         inode_u->bi_gid         = gid;
 450         inode_u->bi_dev         = rdev;
 451         inode_u->bi_atime       = now;
 452         inode_u->bi_mtime       = now;
 453         inode_u->bi_ctime       = now;
 454         inode_u->bi_otime       = now;
 455
 456         if (parent && parent->bi_mode & S_ISGID) {
 457                 inode_u->bi_gid = parent->bi_gid;
 458                 if (S_ISDIR(mode))
 459                         inode_u->bi_mode |= S_ISGID;
 460         }
 461
 462         if (parent) {
 463 #define x(_name, ...)   inode_u->bi_##_name = parent->bi_##_name;
 464                 BCH_INODE_OPTS()
 465 #undef x
 466         }
 467 }
 468
 469 void bch2_inode_init(struct bch_fs *c, struct bch_inode_unpacked *inode_u,
 470                      uid_t uid, gid_t gid, umode_t mode, dev_t rdev,
 471                      struct bch_inode_unpacked *parent)
 472 {
 473         bch2_inode_init_early(c, inode_u);
 474         bch2_inode_init_late(inode_u, bch2_current_time(c),
 475                              uid, gid, mode, rdev, parent);
 476 }
 477
 478 static inline u32 bkey_generation(struct bkey_s_c k)
 479 {
 480         switch (k.k->type) {
 481         case KEY_TYPE_inode:
 482         case KEY_TYPE_inode_v2:
 483                 BUG();
 484         case KEY_TYPE_inode_generation:
 485                 return le32_to_cpu(bkey_s_c_to_inode_generation(k).v->bi_generation);
 486         default:
 487                 return 0;
 488         }
 489 }
 490
 491 /*
 492  * This just finds an empty slot:
 493  */
 494 int bch2_inode_create(struct btree_trans *trans,
 495                       struct btree_iter *iter,
 496                       struct bch_inode_unpacked *inode_u,
 497                       u32 snapshot, u64 cpu)
 498 {
 499         struct bch_fs *c = trans->c;
 500         struct bkey_s_c k;
 501         u64 min, max, start, pos, *hint;
 502         int ret = 0;
 503         unsigned bits = (c->opts.inodes_32bit ? 31 : 63);
 504
 505         if (c->opts.shard_inode_numbers) {
 506                 bits -= c->inode_shard_bits;
 507
 508                 min = (cpu << bits);
 509                 max = (cpu << bits) | ~(ULLONG_MAX << bits);
 510
 511                 min = max_t(u64, min, BLOCKDEV_INODE_MAX);
 512                 hint = c->unused_inode_hints + cpu;
 513         } else {
 514                 min = BLOCKDEV_INODE_MAX;
 515                 max = ~(ULLONG_MAX << bits);
 516                 hint = c->unused_inode_hints;
 517         }
 518
 519         start = READ_ONCE(*hint);
 520
 521         if (start >= max || start < min)
 522                 start = min;
 523
 524         pos = start;
 525         bch2_trans_iter_init(trans, iter, BTREE_ID_inodes, POS(0, pos),
 526                              BTREE_ITER_ALL_SNAPSHOTS|
 527                              BTREE_ITER_INTENT);
 528 again:
 529         while ((k = bch2_btree_iter_peek(iter)).k &&
 530                !(ret = bkey_err(k)) &&
 531                bkey_cmp(k.k->p, POS(0, max)) < 0) {
 532                 while (pos < iter->pos.offset) {
 533                         if (!bch2_btree_key_cache_find(c, BTREE_ID_inodes, POS(0, pos)))
 534                                 goto found_slot;
 535
 536                         pos++;
 537                 }
 538
 539                 if (k.k->p.snapshot == snapshot &&
 540                     !bkey_is_inode(k.k) &&
 541                     !bch2_btree_key_cache_find(c, BTREE_ID_inodes, SPOS(0, pos, snapshot))) {
 542                         bch2_btree_iter_advance(iter);
 543                         continue;
 544                 }
 545
 546                 /*
 547                  * We don't need to iterate over keys in every snapshot once
 548                  * we've found just one:
 549                  */
 550                 pos = iter->pos.offset + 1;
 551                 bch2_btree_iter_set_pos(iter, POS(0, pos));
 552         }
 553
 554         while (!ret && pos < max) {
 555                 if (!bch2_btree_key_cache_find(c, BTREE_ID_inodes, POS(0, pos)))
 556                         goto found_slot;
 557
 558                 pos++;
 559         }
 560
 561         if (!ret && start == min)
 562                 ret = -ENOSPC;
 563
 564         if (ret) {
 565                 bch2_trans_iter_exit(trans, iter);
 566                 return ret;
 567         }
 568
 569         /* Retry from start */
 570         pos = start = min;
 571         bch2_btree_iter_set_pos(iter, POS(0, pos));
 572         goto again;
 573 found_slot:
 574         bch2_btree_iter_set_pos(iter, SPOS(0, pos, snapshot));
 575         k = bch2_btree_iter_peek_slot(iter);
 576         ret = bkey_err(k);
 577         if (ret) {
 578                 bch2_trans_iter_exit(trans, iter);
 579                 return ret;
 580         }
 581
 582         /* We may have raced while the iterator wasn't pointing at pos: */
 583         if (bkey_is_inode(k.k) ||
 584             bch2_btree_key_cache_find(c, BTREE_ID_inodes, k.k->p))
 585                 goto again;
 586
 587         *hint                   = k.k->p.offset;
 588         inode_u->bi_inum        = k.k->p.offset;
 589         inode_u->bi_generation  = bkey_generation(k);
 590         return 0;
 591 }
 592
 593 static int bch2_inode_delete_keys(struct btree_trans *trans,
 594                                   subvol_inum inum, enum btree_id id)
 595 {
 596         u64 offset = 0;
 597         int ret = 0;
 598
 599         while (!ret || ret == -EINTR) {
 600                 struct btree_iter iter;
 601                 struct bkey_s_c k;
 602                 struct bkey_i delete;
 603                 u32 snapshot;
 604
 605                 bch2_trans_begin(trans);
 606
 607                 ret = bch2_subvolume_get_snapshot(trans, inum.subvol, &snapshot);
 608                 if (ret)
 609                         continue;
 610
 611                 bch2_trans_iter_init(trans, &iter, id,
 612                                      SPOS(inum.inum, offset, snapshot),
 613                                      BTREE_ITER_INTENT);
 614                 k = bch2_btree_iter_peek(&iter);
 615
 616                 if (!k.k || iter.pos.inode != inum.inum) {
 617                         bch2_trans_iter_exit(trans, &iter);
 618                         break;
 619                 }
 620
 621                 ret = bkey_err(k);
 622                 if (ret)
 623                         goto err;
 624
 625                 bkey_init(&delete.k);
 626                 delete.k.p = iter.pos;
 627
 628                 if (btree_node_type_is_extents(iter.btree_id)) {
 629                         unsigned max_sectors =
 630                                 min_t(u64, U64_MAX - iter.pos.offset,
 631                                       KEY_SIZE_MAX & (~0 << trans->c->block_bits));
 632
 633                         /* create the biggest key we can */
 634                         bch2_key_resize(&delete.k, max_sectors);
 635
 636                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
 637                         if (ret)
 638                                 goto err;
 639                 }
 640
 641                 ret = bch2_trans_update(trans, &iter, &delete, 0) ?:
 642                       bch2_trans_commit(trans, NULL, NULL,
 643                                         BTREE_INSERT_NOFAIL);
 644 err:
 645                 offset = iter.pos.offset;
 646                 bch2_trans_iter_exit(trans, &iter);
 647         }
 648
 649         return ret;
 650 }
 651
 652 int bch2_inode_rm(struct bch_fs *c, subvol_inum inum, bool cached)
 653 {
 654         struct btree_trans trans;
 655         struct btree_iter iter = { NULL };
 656         struct bkey_i_inode_generation delete;
 657         struct bch_inode_unpacked inode_u;
 658         struct bkey_s_c k;
 659         unsigned iter_flags = BTREE_ITER_INTENT;
 660         u32 snapshot;
 661         int ret;
 662
 663         if (0 && cached && c->opts.inodes_use_key_cache)
 664                 iter_flags |= BTREE_ITER_CACHED;
 665
 666         bch2_trans_init(&trans, c, 0, 1024);
 667
 668         /*
 669          * If this was a directory, there shouldn't be any real dirents left -
 670          * but there could be whiteouts (from hash collisions) that we should
 671          * delete:
 672          *
 673          * XXX: the dirent could ideally would delete whiteouts when they're no
 674          * longer needed
 675          */
 676         ret   = bch2_inode_delete_keys(&trans, inum, BTREE_ID_extents) ?:
 677                 bch2_inode_delete_keys(&trans, inum, BTREE_ID_xattrs) ?:
 678                 bch2_inode_delete_keys(&trans, inum, BTREE_ID_dirents);
 679         if (ret)
 680                 goto err;
 681 retry:
 682         bch2_trans_begin(&trans);
 683
 684         ret = bch2_subvolume_get_snapshot(&trans, inum.subvol, &snapshot);
 685         if (ret)
 686                 goto err;
 687
 688         bch2_trans_iter_init(&trans, &iter, BTREE_ID_inodes,
 689                              SPOS(0, inum.inum, snapshot), iter_flags);
 690         k = bch2_btree_iter_peek_slot(&iter);
 691
 692         ret = bkey_err(k);
 693         if (ret)
 694                 goto err;
 695
 696         if (!bkey_is_inode(k.k)) {
 697                 bch2_fs_inconsistent(trans.c,
 698                                      "inode %llu not found when deleting",
 699                                      inum.inum);
 700                 ret = -EIO;
 701                 goto err;
 702         }
 703
 704         bch2_inode_unpack(k, &inode_u);
 705
 706         /* Subvolume root? */
 707         BUG_ON(inode_u.bi_subvol);
 708
 709         bkey_inode_generation_init(&delete.k_i);
 710         delete.k.p = iter.pos;
 711         delete.v.bi_generation = cpu_to_le32(inode_u.bi_generation + 1);
 712
 713         ret   = bch2_trans_update(&trans, &iter, &delete.k_i, 0) ?:
 714                 bch2_trans_commit(&trans, NULL, NULL,
 715                                 BTREE_INSERT_NOFAIL);
 716 err:
 717         bch2_trans_iter_exit(&trans, &iter);
 718         if (ret == -EINTR)
 719                 goto retry;
 720
 721         bch2_trans_exit(&trans);
 722         return ret;
 723 }
 724
 725 static int bch2_inode_find_by_inum_trans(struct btree_trans *trans,
 726                                          subvol_inum inum,
 727                                          struct bch_inode_unpacked *inode)
 728 {
 729         struct btree_iter iter;
 730         int ret;
 731
 732         ret = bch2_inode_peek(trans, &iter, inode, inum, 0);
 733         if (!ret)
 734                 bch2_trans_iter_exit(trans, &iter);
 735         return ret;
 736 }
 737
 738 int bch2_inode_find_by_inum(struct bch_fs *c, subvol_inum inum,
 739                             struct bch_inode_unpacked *inode)
 740 {
 741         return bch2_trans_do(c, NULL, NULL, 0,
 742                 bch2_inode_find_by_inum_trans(&trans, inum, inode));
 743 }